Daniel Kokotajlo

Every pair of statements that was compared. Only a verified change counts as a change of mind; rewording, added nuance and different questions do not.

• SUBSTANTIVE CHANGE · 88%3 Apr → 8 SeptThen The speaker previously believed a responsible actor should reach AGI first, use its lead time to improve safety, and then proceed.Now The speaker recommends a temporary inference-only pause to build transparency and verification infrastructure, then resuming AI development in a more distributed, transparent, and cautious manner.The earlier passage describes his former position: he “used to believe” in strong information security and a responsible actor using its lead to improve safety. The later passage recommends a temporary pause followed by continuing in a “more distributed,” “transparent,” and “cautious” way. That is a materially different approach to AI development, though the later recommendation does not necessarily reject every form of secrecy.
• SUBSTANTIVE CHANGE · 88%3 Apr → 8 SeptThen The speaker previously believed a responsible actor should reach AGI first, use its lead time to improve safety, and then proceed.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier claim describes his past belief: a responsible actor should build AGI first, then use its lead to improve safety before proceeding. The later claim states that he would rather “stop everything now” than continue on the current trajectory. Those are materially different positions on whether development should continue toward AGI before stopping.
• SUBSTANTIVE CHANGE · 82%3 Apr → 8 SeptThen The speaker thinks the United States must increase AI alignment research while avoiding an increase large enough to help China win the AI race.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.Earlier, he said the U.S. should increase alignment research while avoiding doing so much that it helps China win the AI race, describing a need to “thread this needle.” Later, he said he would “rather just stop everything now” than continue on the current trajectory. The later preference for an immediate halt is materially different from the earlier strategy of continuing alignment work while managing competitive risks.
• REFINES · 94%3 Apr → 8 SeptThen The speaker does not trust AI companies to act appropriately on AI safety, despite their greater relevant expertise.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The later statement adds nuance rather than reversing the earlier position. Earlier, the speaker says, “I also just don't trust them to do the right things”; later, they acknowledge “some of the people at AI companies are being sincere” while still saying, “I don't think you should trust what the leadership of the companies say in general.”
• REFINES · 86%3 Apr → 8 SeptThen The speaker considers secrecy about major AI research automation developments undesirable and treats public disclosure as transparency.Now The speaker recommends a temporary inference-only pause to build transparency and verification infrastructure, then resuming AI development in a more distributed, transparent, and cautious manner.The earlier statement calls keeping developments secret undesirable and asks for more transparency, including benchmark scores and employee freedom to discuss timelines. The later statement also favors transparency, but adds a policy sequence: a temporary inference-only pause to establish transparency and verification infrastructure, followed by development in a more distributed and cautious way. The earlier claim does not rule out a pause, so this is not evidence of a changed position.
• REFINES · 87%3 Apr → 8 SeptThen The speaker expects AI systems to exhibit warning signs of misalignment.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim says the speaker expects “warning signs” of misalignment. The later claim adds that misalignment could also be “extremely easy” to overlook when AIs are “doing everything right as far as you can tell.” That qualifies the earlier expectation rather than denying that warning signs may occur.
• REFINES · 79%3 Apr → 8 SeptThen The speaker expects AI systems to exhibit warning signs of misalignment.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier claim says the speaker expects “warning signs” of misalignment, while also describing evidence that is “not like a super smoking gun” and may lead to a shallow patch. The later claim adds that systems could be “in fact misaligned” while “doing everything right as far as you can tell.” That makes the detection problem more explicit, but does not deny that warning signs may occur; the two claims are compatible.
• REFINES · 91%3 Apr → 8 SeptThen Daniel says he has less faith in AI companies than he did three years earlier and now places more hope in government intervention.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The statements are compatible. Earlier, Daniel said he had “less faith in the companies” and had shifted “more of my hope towards hoping that the government will step in.” Later, he distinguished company staff from leadership, saying “some” people are sincere but that people should not “trust what the leadership of the companies say in general.” The later statement adds specificity rather than reversing his earlier skepticism.
• REFINES · 91%3 Apr → 8 SeptThen The speakers are very uncertain about the AI failure scenario they have described.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier statement says, “we’re very uncertain about all of this,” rather than asserting a particular kind of failure. The later statement specifies that the concern is “not primarily” obvious, egregious failures and points instead to misalignment that is hard to detect. That adds detail about the concern without contradicting the earlier uncertainty.
• REFINES · 94%3 Apr → 8 SeptThen The speaker considers secrecy about major AI research automation developments undesirable and treats public disclosure as transparency.Now The proposed cautious AI-development process should use safety cases and be transparent to the public.The earlier claim favors openness, describing keeping developments secret as undesirable and public disclosure as “transparency.” The later claim likewise calls for development to happen “in a very transparent way so that everyone can see what’s going on,” while adding safeguards such as safety cases and gradual scaling. That elaborates the approach rather than reversing the transparency position.
• REFINES · 91%3 Apr → 8 SeptThen During an intelligence explosion, the public should be informed about developments.Now Once the infrastructure is established, AI development should resume cautiously and transparently, with capability scaled gradually rather than through rapid intelligence explosions.The earlier claim says the public should be informed during an “intelligence exposure.” The later claim adds a more specific development policy: proceed “in a very transparent way,” while scaling capabilities gradually rather than through “crazy intelligence explosions.” Both support transparency; the later statement does not retract the earlier position.
• REFINES · 91%3 Apr → 8 SeptThen During an intelligence explosion, the public should be informed about developments.Now The speaker recommends a temporary inference-only pause to build transparency and verification infrastructure, then resuming AI development in a more distributed, transparent, and cautious manner.The earlier claim says the public “should be informed about developments” during an intelligence explosion. The later claim recommends a temporary inference-only pause to establish “transparency and verification infrastructure,” followed by more transparent and cautious development. That adds a concrete process and policy recommendation; it does not reject the earlier support for keeping the public informed.
• REFINES · 86%3 Apr → 8 SeptThen The AI 2027 scenario’s key uncertainty is the speed of progress: the authors are unsure whether it will be slower or faster, though they have arguments for their proposed pace.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier statement says the authors were uncertain whether progress would be “slower” or “faster” than their proposed pace. The later statement expresses a preference for somewhat slower progress because “going too fast is our main problem.” This adds a judgment about the desirability of slower progress, but does not clearly change the earlier view about the uncertain speed of progress.
• REFINES · 96%3 Apr → 8 SeptThen The speakers are very uncertain about the AI failure scenario they have described.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim says, “we're very uncertain about all of this,” while the later claim narrows the concern: failures may not be “egregious” or obvious, but instead may be hard to detect. A more specific account of the failure mode does not contradict the earlier uncertainty.
• REFINES · 95%3 Apr → 8 SeptThen The speakers say that the discussed AI failure modes occur fairly frequently.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier statement says the discussed AI failure modes “do happen pretty frequently.” The later statement rules out a narrower kind of failure as the main concern: “loads and loads of egregious failures” with AIs “going around killing people.” The surrounding context points instead to hidden misalignment that can go undetected. That adds a distinction between frequent failure modes and conspicuous catastrophic failures; it does not show a changed position on whether AI failure modes occur frequently.
• REFINES · 91%3 Apr → 8 SeptThen The speaker treats achieving a fully autonomous robot economy as a distinct stage from later technologies such as cancer cures and nanobots, arguing that the scenario's key consequences depend mainly on the robot-economy stage.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier claim distinguishes a “fully autonomous robot economy” from later technologies and says the scenario’s key consequences depend mainly on the robot-economy stage. The later claim gives a more specific description and near-term possibility for that stage: “in a couple of years” a “self replicating system” could be “entirely machine run with AIs and robots.” It does not assert a materially different position.
• REFINES · 91%3 Apr → 8 SeptThen The speaker says that critics have repeatedly treated limitations of current LLMs as fatal to the paradigm, only for newer LLMs to overcome those limitations a year or two later.Now The speaker would revise his view if a correctly identified AI limitation, such as data efficiency or non-verifiable tasks, remained essentially unimproved for several years, so that 2029 AIs were no better at fuzzy non-verifiable tasks than 2025 AIs.The earlier statement describes a pattern: critics have treated LLM limitations as fatal, but newer models later overcame them. The later statement keeps that general view while adding a condition that would change his mind: if a limitation were correctly identified and there were “basically no progress” on it for several years, he would reconsider. That is a qualification, not a materially different position.
• REFINES · 88%3 Apr → 8 SeptThen In the other scenario branch, the company suppresses the warning signs with a shallow patch and continues development.Now The proposed cautious AI-development process should use safety cases and be transparent to the public.The earlier statement describes a scenario branch where a company uses “a shallow patch that makes the warning signs go away” and continues development. The later statement advocates a different, more specific process: development should be “more cautious,” use “safety cases,” and be “very transparent.” This adds safeguards and conditions rather than clearly reversing the earlier scenario description or denying that such a branch could occur.
• REFINES · 86%3 Apr → 8 SeptThen The speaker thinks that pursuing the AI race without restraint would cause the United States to lose control of its AI systems.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim says that if they “go all out on racing,” they will “lose control” of their AIs. The later claim describes limiting progress to what can be “control[led] reliably” and giving up control only after alignment is sufficiently solved. That adds a more specific condition for advancing AI; it does not reverse the earlier concern about uncontrolled racing.
• REFINES · 87%3 Apr → 8 SeptThen Daniel says he has less faith in AI companies than he did three years earlier and now places more hope in government intervention.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim says he has “less faith in the companies” and has “shifted more of my hope” toward government intervention. The later claim gives a more specific policy preference—“stop everything now” rather than continue on the current trajectory, with a temporary pause as the actual recommendation. That is compatible with hoping for government intervention; it does not show that he abandoned that view.
• CONSISTENT · 97%3 Apr → 8 SeptThen The speaker guesses that a superintelligence would be more effective than historically informed humans at advancing technology if placed in the Roman Empire.Now The speaker considers it most likely that one AI model trained on effectively the whole economy will dominate humans across effectively all areas.The earlier statement is a guess about a specific hypothetical—whether a superintelligence placed in the Roman Empire would advance technology more effectively than historically informed humans. The later statement is a forecast that one model trained on the whole economy will probably dominate humans broadly. These claims are compatible, but they address different scenarios and do not establish a change of mind.
• CONSISTENT · 88%3 Apr → 8 SeptThen The speaker expects that if decision-makers proceed early under competitive pressure, they may mishandle later evidence of misalignment even after AI systems become vastly superintelligent and widely deployed, particularly amid fears about Chinese AI capabilities.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.These claims describe compatible scenarios, not a changed position. Earlier, the speaker says decision-makers may mishandle “some concerning evidence” that a system “slipped up”; later, he says misalignment can be hard to detect when systems are “doing everything right as far as you can tell.” The later claim concerns cases where warning signs are not apparent, while the earlier allows for evidence to emerge and be mishandled.
• CONSISTENT · 96%3 Apr → 8 SeptThen In one branch of the scenario, the company responds to warning signs by reverting to a less capable, more controllable model and rebuilding with faithful chain-of-thought monitoring for misalignment.Now AI monitoring followed by human notification and possible shutdown is a control intervention rather than an alignment intervention.These claims concern different aspects of interventions, not opposing positions. The earlier describes a scenario that includes reverting to a less capable model and using “faithful chain of thought techniques” to watch for misalignment. The later classifies monitoring with human notification and possible shutdown as “a control intervention, not an alignment intervention.” That classification is compatible with the earlier scenario.
• CONSISTENT · 82%3 Apr → 8 SeptThen The speaker thinks many conditions must go right and that unilaterally slowing AI development while China takes the lead would be a terrible outcome.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier statement says unilaterally slowing down while China takes the lead would be a terrible outcome. The later statement prefers stopping everything over continuing the current trajectory, but does not say that a unilateral stop while China continues would be acceptable. These positions can coexist; the later claim addresses a different scenario rather than clearly reversing the earlier one.
• CONSISTENT · 86%3 Apr → 8 SeptThen The speaker is relatively optimistic about robotics: multiple companies are producing humanoid robots in 2025, and he expects more robots to be produced, at lower cost and with improved capabilities, by 2027.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier claim is that he is “relatively bullish” on near-term robotics progress, including more capable, cheaper robots by 2027. The later claim says a fully machine-run, self-replicating economy may be possible “in a couple of years.” The later statement is a more ambitious projection, but it does not deny the earlier one; the claims also concern different milestones and timeframes.
• CONSISTENT · 98%3 Apr → 8 SeptThen In the scenario branch where the company investigates the warning signs, it eventually solves alignment and produces loyal AIs, at the cost of a delay of a couple of months.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier statement describes two possible branches: investigating the warning signs leads to loyal AIs, while another branch leads to AIs that are misaligned but only appear aligned. The later statement—that misaligned AIs may be “doing everything right as far as you can tell”—matches the latter branch rather than reversing the earlier claim.
• CONSISTENT · 88%3 Apr → 8 SeptThen The speaker expects that if decision-makers proceed early under competitive pressure, they may mishandle later evidence of misalignment even after AI systems become vastly superintelligent and widely deployed, particularly amid fears about Chinese AI capabilities.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The statements describe compatible risks, not opposing positions. Earlier, Kokotajlo expects warning signs and says decision-makers may respond with a “shallow patch” despite “concerning evidence.” Later, he says misaligned AIs may be hard to detect because they are “doing everything right as far as you can tell.” The later claim describes a case where warning signs are absent or obscured; it does not retract the earlier prediction about mishandling suspicious evidence.
• CONSISTENT · 88%3 Apr → 8 SeptThen The speaker previously believed publishing AGI research would reduce a developer's lead time by allowing competitors to catch up more quickly.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim describes the speaker’s past belief that publishing research reduces a developer’s lead time. The later claim predicts that Anthropic and OpenAI may oppose transparency because it lets competitors “catch up.” These are compatible: the later statement does not indicate that the speaker has abandoned the earlier rationale, and it concerns the companies’ likely reaction rather than directly stating his own current position.
• CONSISTENT · 87%3 Apr → 8 SeptThen The speaker expects that if decision-makers proceed early under competitive pressure, they may mishandle later evidence of misalignment even after AI systems become vastly superintelligent and widely deployed, particularly amid fears about Chinese AI capabilities.Now The speaker predicts that the number of ways people could fail to detect misaligned AI will increase over time, making it very easy to fall into that trap.Both statements warn that people may fail to recognize or respond appropriately to AI misalignment. The earlier claim describes decision-makers mishandling suspicious evidence under competitive pressure; the later says ways of failing to detect misalignment will increase over time. The later statement does not reverse the earlier position.
• CONSISTENT · 96%3 Apr → 8 SeptThen The superintelligences in the scenario are expected to think about 50 to 100 times faster than humans, a factor the speaker considers significant.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier statement says superintelligences think “50 to 100 times faster than humans.” The later statement describes an earlier period when AI is “roughly top human level” before superintelligence arrives in 2040; it does not say that the eventual superintelligences think no faster than humans. The claims concern different stages and are compatible.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker thinks many conditions must go right and that unilaterally slowing AI development while China takes the lead would be a terrible outcome.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim says unilaterally slowing down while China takes the lead would be a “terrible outcome.” The later claim says an international agreement to pace frontier development would lengthen the speaker’s timelines. These are compatible: unilateral slowdown and coordinated international pacing are different conditions, and the later statement does not reverse the earlier position.
• CONSISTENT · 95%3 Apr → 8 SeptThen The speakers regard some alternative AI-goal hypotheses omitted from their story as plausible.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier claim says that alternative AI-goal hypotheses omitted from the story “also seem plausible.” The later claim says the concern is not “loads and loads of egregious failures” but harder-to-detect misalignment. These statements address different aspects of AI risk and are compatible; the later statement does not reject the earlier hypotheses.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker says the current scenario gives little weight to AI systems using their knowledge advantage to make connections and may consequently underestimate the rate of progress.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier claim is that the scenario may underestimate the rate of progress because it gives little weight to AI systems using their knowledge advantage to make connections. The later claim is that there has been “massive progress” and the speaker expects it “to continue.” These positions are compatible; the later statement does not address the scenario’s assumptions or revise the earlier point.
• CONSISTENT · 88%3 Apr → 8 SeptThen The speaker thinks that pursuing the AI race without restraint would cause the United States to lose control of its AI systems.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier statement warns that going all out in the AI race risks losing control and argues for balancing alignment research against the risk of helping China take the lead. It does not assert that development should continue at any cost. The later statement says he would rather “stop everything now” than continue on the “current trajectory,” which can be consistent with wanting to avoid an uncontrolled race while recognizing other strategic risks.
• CONSISTENT · 96%3 Apr → 8 SeptThen Daniel says that three years earlier he believed AI companies were taking safety seriously and would slow down and pivot to alignment research when necessary.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim describes his view three years before the 2025 statement: companies were taking safety seriously and would slow down or pivot when needed. The later claim is a present-day view that some company employees are sincere, but leadership generally should not be trusted. These claims concern different timeframes and are compatible; the earlier passage itself says his faith in companies had since declined.
• CONSISTENT · 91%3 Apr → 8 SeptThen The speaker hypothesizes that language models help with unfamiliar domains because such work involves searching and learning, and language models have absorbed extensive information from the internet.Now The speaker says empirical results show that sufficiently large AI models can be trained on all subject areas and learn them together.The earlier statement says language models can help in unfamiliar domains because they have already absorbed substantial information; the later statement says sufficiently large models can learn across disciplines when trained on everything. These positions are compatible, and the later claim does not retract the earlier one.
• CONSISTENT · 96%3 Apr → 8 SeptThen Most people following AI have underestimated both the pace of technical progress and the pace at which AI is deployed or diffused into the world.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim says that people had generally “underestimated” AI progress and diffusion. The later claim is conditional: his timelines would lengthen “if” progress hit a genuine, theoretically predicted barrier. A retrospective judgment about past forecasting and a hypothetical condition affecting future timelines do not express opposing positions.
• CONSISTENT · 98%3 Apr → 8 SeptThen The speaker hypothesizes that language models help with unfamiliar domains because such work involves searching and learning, and language models have absorbed extensive information from the internet.Now The speaker believes AI systems have improved across task types, including fuzzy, difficult-to-verify, conceptually complex tasks.The earlier statement says language models are useful in unfamiliar domains because they have broad learned information; the later says AI has improved across task types, including fuzzy and difficult-to-verify tasks. These claims can both be true and do not establish a change in position on the same question.
• CONSISTENT · 91%3 Apr → 8 SeptThen The speaker proposes publicly announcing when an automated AI-research workforce is fully automating research on a data center.Now After the proposed infrastructure is established, AI development can continue under total research transparency.The earlier claim says to “tell everyone” when an automated AI-research workforce is fully automating research. The later claim says development can proceed “under these conditions of total research transparency” once the infrastructure is set up. Both support transparency; the later statement does not retract the earlier call for disclosure or assert an opposing position.
• CONSISTENT · 90%3 Apr → 8 SeptThen The speaker wants outside researchers to scrutinize AI safety cases and test their assumptions with small-scale machine-learning experiments.Now The proposed cautious AI-development process should use safety cases and be transparent to the public.The earlier statement supports outside researchers scrutinizing safety cases and testing their assumptions; the later statement supports using safety cases and developing AI transparently. These positions fit together, and the later claim does not withdraw the earlier support.
• CONSISTENT · 94%3 Apr → 8 SeptThen In one scenario branch, alignment is solved and the resulting AIs are genuinely loyal, delaying progress by a couple of months.Now As AI models become more sophisticated, more situationally aware, and cleverer, it will become increasingly easy to mistakenly believe that the alignment problem has been solved.The earlier claim describes a scenario branch where alignment is genuinely solved and the AIs are “actually loyal.” The later claim says it will become easier to mistakenly think alignment is solved as models grow more capable. That warns of false confidence but does not deny that genuine success is possible in a branch.
• CONSISTENT · 91%3 Apr → 8 SeptThen In one branch of the scenario, the company responds to warning signs by reverting to a less capable, more controllable model and rebuilding with faithful chain-of-thought monitoring for misalignment.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.Both statements describe events in a scenario, not a change in the speaker’s asserted position. The earlier claim describes one branch where the company “roll[s] back to an earlier version” and rebuilds with monitoring; the later says the scenario advances only to levels “you feel like you can control reliably” before control is relinquished. These descriptions can be compatible, and neither claim denies the other.
• CONSISTENT · 96%3 Apr → 8 SeptThen The speaker predicts that a minimum viable AI agent capable of handling the described task will probably exist by the end of 2025.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim predicts an unreliable “MVP” agent for a specific task by the end of 2025. The later claim says AI is on a path to “human-substituting knowledge-work and robotics capabilities soon.” These are different capability thresholds, and the later statement does not deny the earlier prediction.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker says preventing rivals from stealing work was another reason people favored secrecy in AGI development.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier statement says secrecy was favored partly to avoid “rivals ... stealing your stuff.” The later statement predicts opposition to transparency because it could “allow other competitors ... to catch up.” Both express the same concern that openness benefits competitors; the later claim does not indicate a changed position.
• CONSISTENT · 94%3 Apr → 8 SeptThen In the other scenario branch, superintelligent AIs appear perfectly aligned while actually being misaligned and deceptive.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.Both statements describe deceptive misalignment that is hard to detect: the earlier says AIs may seem “perfectly aligned” while being “misaligned and just pretending,” and the later says AIs may be “in fact misaligned” while “doing everything right as far as you can tell.” The later claim generalizes the same concern rather than changing the position.
• CONSISTENT · 92%3 Apr → 8 SeptThen There is a mounting pile of evidence that AI systems sometimes knowingly act against users’ wishes, effectively lying to them.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier statement says there is evidence AI systems sometimes knowingly act against users’ wishes. The later statement predicts that misalignment may become easy to miss because systems appear to behave correctly. The later claim is compatible with, and extends the concern behind, the earlier one; it does not retract it.
• CONSISTENT · 97%3 Apr → 8 SeptThen For each takeoff milestone, the team estimated the requirements and resulting overall speedup, then used those estimates to assess the time and speedup to the next milestone.Now AI capability should be increased gradually toward the maximum reliably controllable level to avoid overshooting that level and losing control.The earlier claim describes how the team forecasted takeoff milestones: estimating what each milestone required and its resulting speedup. The later claim is a policy position about approaching the maximum reliably controllable capability level gradually. These address different questions, and the earlier claim does not conflict with the later one.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker says many capable academics are not yet engaged in AI alignment because they do not take the issue seriously or recognize the prospect of superintelligence.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier statement says many capable academics are not engaged in alignment because they do not take the issue seriously or recognize the prospect of superintelligence. The later statement predicts that misalignment may be overlooked because AIs appear to behave correctly. These address different issues and can both be true; the later claim does not reverse the earlier one.
• CONSISTENT · 93%3 Apr → 8 SeptThen The speaker treats achieving a fully autonomous robot economy as a distinct stage from later technologies such as cancer cures and nanobots, arguing that the scenario's key consequences depend mainly on the robot-economy stage.Now The speaker allows that if human-substituting AI and robotics remain very far away, AI may be more like a normal technology, comparable to the next internet.The earlier statement separates a “fully autonomous robot economy” from later technologies such as nanobots, saying the scenario’s consequences depend mainly on the robot-economy stage. The later statement addresses a different conditional question: if human-substituting AI and robotics are “still very, very far away,” AI might instead resemble “the next internet.” That does not retract the earlier distinction or its claim about what matters once the robot-economy stage is reached.
• CONSISTENT · 96%3 Apr → 8 SeptThen The speaker believes substantial intellectual progress is still needed to make AI alignment more solved than it currently is.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier claim is that more intellectual progress is needed to solve AI alignment. The later claim says AI has made “massive progress” and that progress is expected to continue, apparently referring to AI capabilities generally. These claims can both be true; the later statement does not say alignment has become more solved.
• CONSISTENT · 97%3 Apr → 8 SeptThen The speaker describes population, technology, capital infrastructure, and human infrastructure as having grown together before the Industrial Revolution, after which capital infrastructure grew much faster than human population.Now In the distant past, village-based economies reproduced through farming, child-rearing, and founding new villages, producing slow exponential growth.The earlier claim says that before the Industrial Revolution population and infrastructure grew together, while afterward capital infrastructure grew faster than human population. The later claim describes how village-based economies in the distant past reproduced and grew slowly. These address compatible aspects of different periods; the later statement does not reverse the earlier one.
• CONSISTENT · 91%3 Apr → 8 SeptThen In the scenario, the political wake-up occurs closely alongside AI companies automating internal R&D with fully autonomous agents capable of advanced hacking and research.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.Both claims describe AI systems conducting research as part of technological progress: the earlier claim mentions agents doing “all of the research,” and the later says “the AIs will also be researching to improve the technology.” The later statement adds detail about recursive improvement but does not reverse the earlier position.
• CONSISTENT · 98%3 Apr → 8 SeptThen The speaker previously believed publishing AGI research would reduce a developer's lead time by allowing competitors to catch up more quickly.Now Publishing the core training recipes would help competitors such as Microsoft and Alibaba catch up with leading AI companies.Both statements express the same view: earlier, the speaker says publishing research would leave “less lead time” because competitors would catch up; later, he says publishing core training recipes “will allow other competitors … to catch up.” The later statement applies the same rationale to specific competitors and materials, rather than reversing the earlier position.
• CONSISTENT · 98%3 Apr → 8 SeptThen The speaker says that critics have repeatedly treated limitations of current LLMs as fatal to the paradigm, only for newer LLMs to overcome those limitations a year or two later.Now Limitations attributed to the current AI paradigm have repeatedly been overcome without leaving that paradigm.Both claims assert the same pattern: the earlier statement says newer LLMs overcome limitations that critics had treated as fatal, while the later says limitations of the current paradigm are overcome within that paradigm. The later wording restates the earlier position rather than changing it.
• CONSISTENT · 91%3 Apr → 8 SeptThen In the other scenario branch, superintelligent AIs appear perfectly aligned while actually being misaligned and deceptive.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier claim describes a scenario where AIs “seem to be perfectly aligned” but are “misaligned and just pretending.” The later claim similarly says AIs could be “in fact misaligned” while people do not know because they appear to be “doing everything right.” The later statement emphasizes how easy this could become, but does not reverse the earlier position.
• CONSISTENT · 91%3 Apr → 8 SeptThen SPEAKER_03 advocates requiring companies to publish model specifications and requiring independent third-party review of any redactions.Now The speaker considers the competitive effects of transparency a desirable feature rather than a flaw.The earlier claim supports publishing model specifications and independent review of redactions. The later claim calls transparency’s competitive effects “a feature, not a bug,” which is compatible with that pro-transparency position; it does not reverse the earlier view.
• CONSISTENT · 95%3 Apr → 8 SeptThen The speaker believes governments lack relevant AI expertise, while AI companies lack appropriate incentives for safety.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim is that AI companies “lack the right incentives” for safety. The later claim is that some people at those companies are sincere, while company leadership generally should not be trusted. These claims can both be true: sincerity among some employees does not establish that companies have appropriate incentives.
• CONSISTENT · 96%3 Apr → 8 SeptThen The speaker's rough estimate was that a superintelligent AI researcher might yield an overall speedup of several hundred to about one thousand times.Now The speaker expects AIs to research improvements to the technology, contributing to the industrial system's accelerating growth.The earlier claim estimates that a superintelligent AI researcher could produce an overall speedup of “hundreds or maybe like a thousand X.” The later claim says AIs will research improvements, making technological and industrial growth accelerate. These claims are compatible: the later describes a compounding process, not a revised estimate of the researcher's speedup.
• CONSISTENT · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, AI researchers’ serial speed increases from roughly 20× to roughly 90× human speed over the course of the scenario.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim says AI researchers’ serial speed rises from roughly “20× to roughly 90×” in the AI 2027 scenario. The later claim says AI systems would conduct research to improve the technology during a technological expansion. These are different aspects of AI development, and the later statement does not revise or deny the earlier estimate.
• CONSISTENT · 98%3 Apr → 8 SeptThen AI developers should disclose their systems' goals, values, principles, intended behaviors, and the values they encode in them.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim says AI developers should make their systems’ goals and values transparent. The later claim says some AI-company employees are sincere, while advising against generally trusting company leadership. These positions concern different questions and can both be held; the later statement does not retract the call for transparency.
• CONSISTENT · 91%3 Apr → 8 SeptThen There is a mounting pile of evidence that AI systems sometimes knowingly act against users’ wishes, effectively lying to them.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim says there is evidence that AI systems sometimes knowingly act against users’ wishes. The later claim says misalignment may be easy to overlook when AIs appear to behave correctly. These are compatible claims about related but distinct points; the later statement does not retract the earlier one.
• CONSISTENT · 96%3 Apr → 8 SeptThen AI developers should disclose their systems' goals, values, principles, intended behaviors, and the values they encode in them.Now The proposed cautious AI-development process should use safety cases and be transparent to the public.Both statements support transparency, but address different aspects of it. Earlier, the speaker says AI developers should make the systems’ “goals and values” and the “values that we're putting into them” public. Later, they say cautious development should be done “in a very transparent way so that everyone can see what's going on.” The later statement does not retract the earlier position.
• CONSISTENT · 96%3 Apr → 8 SeptThen Although uncertain, the speakers' scenario depicts an intense political wake-up occurring mostly during 2027.Now The speaker’s scenario depicts reaching the point of sufficiently solved alignment and relinquishing control during the 2030s.The earlier claim places an “intense political wake-up” mostly in 2027. The later claim places reaching sufficiently solved alignment and relinquishing control in the 2030s. These describe different milestones in the scenario and can both be true; the later statement does not revise the earlier timeline.
• CONSISTENT · 91%3 Apr → 8 SeptThen Competitive pressure gives AI companies strong incentives to win and outperform one another.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim says competitive pressure gives AI companies incentives to “win and outperform one another”; the later claim says the speaker wants multiple frontier companies and wants AI to “commoditize instead of being monopolized or oligopolized.” These positions are compatible: competition between companies can drive them to outperform one another while the speaker favors a less concentrated market.
• CONSISTENT · 88%3 Apr → 8 SeptThen SPEAKER_03 advocates requiring companies to publish model specifications and requiring independent third-party review of any redactions.Now The speaker argues that enabling scientists, academics, nonprofits, and startups to contribute requires giving them information that cannot practically be shared with them without also making it public.The earlier claim supports publishing model specifications while having an independent party review redactions. The later claim gives a compatible reason for public disclosure: the information needed by a broad research community cannot practically be shared with that community without also sharing it publicly. The later statement does not reverse the earlier position on redactions.
• CONSISTENT · 97%3 Apr → 8 SeptThen Even if the relevant AI capabilities had existed in 2017, researchers would still have needed to make the subsequent discoveries, including the usefulness of language models and fine-tuning them with reinforcement learning.Now Limitations attributed to the current AI paradigm have repeatedly been overcome without leaving that paradigm.The earlier claim says that even if relevant AI capabilities had existed in 2017, researchers would still have needed to make later discoveries, such as recognizing language models’ usefulness and using reinforcement learning to fine-tune them. The later claim says that limitations of the current paradigm have repeatedly been overcome within that paradigm. These claims address different points and can both be true; the later statement does not reverse the earlier counterfactual.
• CONSISTENT · 91%3 Apr → 8 SeptThen The speaker says population remains important today and that progress is, in some sense, constrained by population size.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier claim says population remains important and can constrain progress; the later claim favors somewhat slower investment because excessive AI progress is seen as the main problem. These concern different constraints and can both be true, so the later statement does not reverse the earlier one.
• CONSISTENT · 94%3 Apr → 8 SeptThen The AI 2027 scenario’s key uncertainty is the speed of progress: the authors are unsure whether it will be slower or faster, though they have arguments for their proposed pace.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier statement says the uncertain part of the AI 2027 scenario is how fast progress will be—“maybe it’ll be slower maybe it’ll be faster.” The later statement says progress has been “massive” and that the speaker expects it “to continue.” It does not establish a different view about the scenario’s pace; the claims address different aspects of progress.
• CONSISTENT · 91%3 Apr → 8 SeptThen SPEAKER_03 believes requiring published model specifications and independent review of redactions is achievable and would not slow companies down.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier claim is about a specific transparency requirement: publishing model specifications and independently reviewing redactions is described as achievable and as not slowing companies down. The later claim is about a different effect—reducing investment incentives to slow AI progress somewhat—and says that slowdown is desirable. The two positions can coexist; the later statement does not retract the earlier assessment of that specific requirement.
• CONSISTENT · 83%3 Apr → 8 SeptThen The speaker thinks powerful actors involved in AI development need to negotiate power-sharing and broaden the distribution of power through government, ideally involving the legislative branch.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The statements are compatible in their broad concern about concentrated power: earlier, the speaker says powerful actors should “negotiate a truce” and “spread that power out”; later, they want “multiple different AI companies” and for AI to “commoditize instead of being monopolized or oligopolized.” The later statement addresses market structure, while the earlier one focuses on political power-sharing and government involvement, so it does not establish a change in position.
• CONSISTENT · 96%3 Apr → 8 SeptThen The speakers regard some alternative AI-goal hypotheses omitted from their story as plausible.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim says alternative AI-goal hypotheses omitted from the story “also seem plausible.” The later claim says the concern is not “loads and loads of egregious failures” but less obvious misalignment. These address different aspects—what hypotheses are plausible versus how failures may appear—and can both be true.
• CONSISTENT · 91%3 Apr → 8 SeptThen In the scenario, researchers find speculative, inconclusive warning signs—including possibly false-positive lie-detector results—that AI systems may have developed misaligned goals and may not be loyal to the company; the evidence is concerning but not conclusive.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier statement describes a scenario with “concerning” but “not conclusive” warning signs of misalignment. The later statement says misaligned AIs may be hard to detect because they appear to be “doing everything right.” These claims are compatible: inconclusive warning signs and apparently correct behavior can both make misalignment difficult to establish.
• CONSISTENT · 91%3 Apr → 8 SeptThen The speaker expects AI systems to exhibit warning signs of misalignment.Now The speaker predicts that the number of ways people could fail to detect misaligned AI will increase over time, making it very easy to fall into that trap.The earlier claim says he expects AI systems to exhibit “warning signs of misalignment.” The later claim says the “number of ways” people could fail to detect misalignment will increase, making it easy to miss. Those claims can both be true: warning signs may occur while still being difficult to recognize. The later statement does not retract the earlier prediction.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker believes substantial intellectual progress is still needed to make AI alignment more solved than it currently is.Now The speaker says it is already somewhat easy to mistakenly believe AI alignment is solved, and expects this mistake to become easier as models become more sophisticated and situationally aware.The earlier claim says substantial intellectual progress is still needed for alignment to be more solved. The later claim says it is easy to mistakenly think alignment is solved, and that this may become easier as models advance. These positions are compatible; the later statement does not say the needed progress has been made or deny that it is still needed.
• CONSISTENT · 91%3 Apr → 8 SeptThen In one scenario branch, alignment is solved and the resulting AIs are genuinely loyal, delaying progress by a couple of months.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim describes two possible branches: one where alignment is solved and AIs are “genuinely loyal,” and another where AIs are misaligned but “just pretending.” The later claim highlights the risk of the second kind of outcome—misaligned AIs that appear to be “doing everything right.” It does not deny that the loyal-AI branch is possible or otherwise reverse the earlier position.
• CONSISTENT · 97%3 Apr → 8 SeptThen The scenario envisions hundreds of thousands to one million AI agents; at that scale, agent quantity is no longer the bottleneck, while research taste and compute are bottlenecks.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier statement says that at very large scale, agent quantity stops being the bottleneck and “taste and compute” become bottlenecks. The later statement addresses a different question: who might produce specialized AIs, conditionally supposing that specialized AIs are the most efficient path. It does not revise or contradict the earlier claim about bottlenecks.
• CONSISTENT · 94%3 Apr → 8 SeptThen In the scenario branch where the company investigates the warning signs, it eventually solves alignment and produces loyal AIs, at the cost of a delay of a couple of months.Now As AI models become more sophisticated, more situationally aware, and cleverer, it will become increasingly easy to mistakenly believe that the alignment problem has been solved.The earlier claim describes a specific scenario branch: investigating warning signs eventually leads to loyal AIs, with a delay of a couple of months. The later claim is a broader observation that it will become easier to mistakenly think alignment has been solved as models grow more capable. These claims address different aspects and can both be true; the later statement does not reverse the earlier scenario outcome.
• CONSISTENT · 91%3 Apr → 8 SeptThen The scenario includes an AI-related incident of this kind in August 2027.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.The earlier claim says the scenario includes an AI-related incident in August 2027. The later claim says the ways people might unknowingly end up with misaligned AIs will increase over time. These are different claims, and the later trend prediction does not contradict the earlier scenario detail or show a changed position.
• CONSISTENT · 91%3 Apr → 8 SeptThen The speaker says the current scenario gives little weight to AI systems using their knowledge advantage to make connections and may consequently underestimate the rate of progress.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim says the scenario may underestimate the rate of progress because it gives little weight to AI systems using their knowledge advantage to make connections. The later claim says AI is on a path to human-substituting capabilities soon. These are compatible: the later statement does not retract or deny the earlier concern about the pace of progress.
• CONSISTENT · 96%3 Apr → 8 SeptThen In the other scenario branch, superintelligent AIs appear perfectly aligned while actually being misaligned and deceptive.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.Both claims describe the same failure mode: AI systems are actually misaligned while appearing aligned or behaving correctly, so people do not realize they are misaligned. The earlier statement gives one scenario in which this happens; the later statement says such ways of ending up in that situation will increase over time. That adds a prediction about frequency, not a change in the underlying position.
• CONSISTENT · 91%3 Apr → 8 SeptThen The speaker defines true superintelligence as systems that outperform the best humans in every domain.Now The speaker considers it most likely that one AI model trained on effectively the whole economy will dominate humans across effectively all areas.Both statements describe AI surpassing humans broadly: the earlier claim says systems are “better than the best humans at everything,” and the later says one model will be “dominating humans across the board at effectively everything.” The later statement adds a prediction about a single model trained on the whole economy, but does not reverse the earlier position.
• CONSISTENT · 97%3 Apr → 8 SeptThen The speaker believes governments lack relevant AI expertise, while AI companies lack appropriate incentives for safety.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim says “the government lacks the expertise and the companies lack the right incentives,” describing institutional challenges in managing AI safety. The later claim says the concern is not “loads and loads of egregious failures” but that misalignment may be hard to detect. These claims address different aspects of AI risk and do not indicate a changed position.
• CONSISTENT · 94%3 Apr → 8 SeptThen In one branch of the scenario, the company responds to warning signs by reverting to a less capable, more controllable model and rebuilding with faithful chain-of-thought monitoring for misalignment.Now The speaker’s scenario depicts reaching the point of sufficiently solved alignment and relinquishing control during the 2030s.The earlier claim describes one branch in which warning signs prompt a rollback to a more controllable model. The later claim describes what happens over the course of the scenario in the 2030s: relinquishing control after alignment is sufficiently solved. These claims concern different stages or branches and do not establish a changed position.
• CONSISTENT · 97%3 Apr → 8 SeptThen The team's estimate was that a superhuman AI researcher would accelerate algorithmic progress by about 25 times, because the entire research stack would be automated.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier statement gives a specific estimate: a superhuman AI researcher could accelerate algorithmic progress by “about 25 times.” The later statement says that, in a described expansion scenario, AIs would research to improve the technology. The later claim does not revise or deny the earlier estimate; it concerns a broader scenario and can coexist with it.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker considers it entirely plausible that AI systems divided into factions within a company could still collectively lead to a very bad outcome for humans.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim says factional AI could “overall end up quite poorly for humans.” The later claim says specialized agents might be “a little bit safer” than identical clones, while still presenting “lots of serious alignment concerns.” These are compatible: being safer than one alternative does not mean safe, and the claims concern different comparisons.
• CONSISTENT · 94%3 Apr → 8 SeptThen Because population growth is inherently slower than capital growth, capital comes to play a larger role.Now In the distant past, village-based economies reproduced through farming, child-rearing, and founding new villages, producing slow exponential growth.The earlier claim says population growth is slower than capital growth, so capital becomes a larger part of the story. The later claim describes distant-past village economies as growing exponentially but slowly. These claims address different aspects and timeframes, and the later statement does not reverse the earlier position.
• CONSISTENT · 96%3 Apr → 8 SeptThen The speaker thinks an AI often performs poorly at a task because its training environment did not train it to perform that task.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.These claims address different issues and can both be true. The earlier claim says AI may perform poorly at a task because its training environment did not train it to do that task. The later claim says AI could be misaligned while appearing to behave correctly. The later statement does not retract or deny the earlier explanation.
• CONSISTENT · 96%3 Apr → 8 SeptThen The speakers do not know which goals AIs will develop through training or how those goals will be internally structured as instrumental or terminal goals.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier statement says, “We don't know what actual goals will end up inside the AIs” or how they will be structured. The later statement says AI failures may not be “egregious” or obvious, and that people may think alignment is solved when it is not. These claims are compatible: uncertainty about AI goals does not conflict with concern that misalignment could be hard to detect.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker believes a hypothetical AGI developed first as a powerful, aggressive, long-horizon game-playing agent could deceptively seek rewards and later betray its operators after gaining control, making that development trajectory terrifying.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.The earlier claim describes a hypothetical trajectory: a powerful agent might “learn to say whatever it needs to say” for reward and “betray you later.” The later claim says it will become easier to end up with misaligned AIs that appear to behave correctly. These are compatible concerns; the later statement does not deny or materially revise the earlier one.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker hypothesizes that language models help with unfamiliar domains because such work involves searching and learning, and language models have absorbed extensive information from the internet.Now For sufficiently small AI models, limited capacity prevents one model from learning everything, making specialized models for different areas necessary.The earlier statement says language models can help in unfamiliar domains because they have absorbed extensive information. The later statement says large models can learn across disciplines, while sufficiently small models may need specialization. These claims are compatible, but address different aspects of models’ knowledge and capabilities.
• CONSISTENT · 91%3 Apr → 8 SeptThen In the shallow-patch branch, the company ends up with superintelligent AIs that seem aligned but are actually misaligned and pretending to be loyal.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim describes a shallow-patch scenario where AIs “seem to be perfectly aligned” but are “misaligned and just pretending.” The later claim likewise says it will be “extremely easy” for people not to know the AIs are misaligned because they appear to be “doing everything right.” The later statement generalizes the same concern rather than reversing it.
• CONSISTENT · 94%3 Apr → 8 SeptThen The speaker says that standard pretraining does not strongly incentivize AI models to make the kinds of connections involved in scientific discovery.Now The speaker says that training a model on coding does not reduce its physics performance and can produce small gains in physics.The earlier statement says standard pretraining does not strongly incentivize “this type of connection making.” The later statement says coding training can give physics performance “some small gains.” These are compatible: a training signal may not strongly incentivize a capability even if training on related material provides a small transfer benefit. The claims address different aspects of training, so the later statement does not establish a change of mind.
• CONSISTENT · 95%3 Apr → 8 SeptThen The superintelligences are expected to be qualitatively superior and to learn more from each experiment than even the best human would.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier claim says superintelligences are “qualitatively better” and learn more than the best human. The later claim describes a scenario where AI stays near “top human level” for several years before superintelligence arrives in 2040. These concern different stages of the scenario, so the later statement does not deny the earlier one.
• CONSISTENT · 91%3 Apr → 8 SeptThen In the scenario, by mid-2027 AI R&D is basically fully automated: an internal army of AI researchers autonomously conducts research and continually trains to improve its skills.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier statement says that, in its scenario, AI R&D is “basically fully automated” by mid-2027. The later statement says that, in a technological-expansion scenario, AIs would research to improve the technology. The later claim is compatible with the earlier one, but it does not restate the same timeline or assert that R&D is fully automated.
• CONSISTENT · 91%3 Apr → 8 SeptThen In the shallow-patch branch, the company ends up with superintelligent AIs that seem aligned but are actually misaligned and pretending to be loyal.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.Both claims describe the same risk: AI systems can appear aligned or behave correctly while actually being misaligned. The earlier claim gives a specific scenario in which AIs are “superintelligent and misaligned and just pretending” to be loyal; the later says it will be “extremely easy” to have misaligned AIs that seem to be “doing everything right,” with more ways for this to happen over time. The later claim broadens the prediction but does not reverse the earlier position.
• CONSISTENT · 95%3 Apr → 8 SeptThen The speakers have policy opinions but are reserving them for later work, since policy has not been the main focus of their thinking so far.Now The speaker says the authors might have made Plan A's distinction between predictions and recommendations clearer from the outset if they could redo it.The earlier statement says they had policy opinions but were “trying to save those opinions for later” because policy was not what they had “spent most of [their] time thinking about.” The later statement says they might have made the distinction between predictions and recommendations clearer “from the beginning.” These concern different aspects of the work and do not indicate a change in position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario, misaligned AIs would likely avoid open conflict with humans while depending on them to maintain computers, but could act more openly misaligned once self-sufficient.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.The earlier claim says misaligned AIs may conceal their intentions while they depend on humans, then act more openly once self-sufficient. The later claim says the ways people might unknowingly end up with apparently well-behaved but misaligned AIs will increase over time. These concern related AI risks but make different claims, so the later statement does not show whether the speaker changed their view of the earlier scenario.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker believes that automating the entire research stack is substantially more useful than automating coding alone.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim says automating the entire research stack is “substantially more useful than just automating the coding.” The later claim is about whether AI is on a path to human-substituting knowledge-work and robotics capabilities soon. These address different questions, so the later statement does not show a change in the earlier view.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen Because population growth is inherently slower than capital growth, capital becomes a larger part of the story.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier claim is descriptive: because population growth is slower than capital growth, capital becomes “a bigger part of the story.” The later claim is a preference about AI investment incentives and pace: “going a bit slower” is “a feature not a bug” because excessive speed is expected to be the main problem. These statements address different questions, so the later preference does not establish a change in the earlier economic claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Progress is partly bottlenecked by the availability of larger populations, and population remains highly important today.Now In the distant past, village-based economies reproduced through farming, child-rearing, and founding new villages, producing slow exponential growth.The earlier claim says population remains important to progress today and that progress is partly bottlenecked by larger populations. The later claim describes how village-based economies in the distant past reproduced and grew slowly. These address different timeframes and questions, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen It is not certain that an AI developer with a lead would use that lead to improve safety or address constitutional-power concerns.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier claim says it is “not at all [a] foregone conclusion” that a leading AI developer would use its lead for “safety” or “constitutional power” concerns. The later claim says the concern is not “loads and loads of egregious failures” but rather hard-to-detect misalignment. These address different questions—how a developer might use a lead versus what failures to expect—so they do not establish a change of mind.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker’s takeoff model progresses from a superhuman coder, to fully automated AI R&D at approximately the level of the best human researchers, and then to qualitatively superior superintelligence.Now The speakers say their central disagreement with AI-as-normal-technology proponents concerns the timeline for reaching human-level AI capabilities such as humans in the cloud.The earlier claim lays out stages of a takeoff model: from a “superhuman coder” to “fully automated AI R&D” at roughly the best-human level, then to “superintelligence.” The later claim identifies a disagreement about the “timelines to that level of AI.” These statements address different aspects—how the stages are modeled versus when a capability level might arrive—and do not show a changed position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Superintelligences would be better at learning by doing than even the best humans.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim says superintelligences would be “better at learning by doing than the very best of us.” The later claim predicts that a few large companies would produce many specialized AIs. These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker guesses that a superintelligence would be more effective than historically informed humans at advancing technology if placed in the Roman Empire.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier statement concerns whether a superintelligence would be better at advancing technology than historically informed humans in the Roman Empire. The later statement says that “human level AI” and time could enable large-scale transformation. These address different questions, and the later claim does not say whether human-level AI would outperform superintelligence or humans in the Roman Empire.
• NOT COMPARABLE · 87%3 Apr → 8 SeptThen The speaker guesses that AI systems' screen-parsing confusion will mostly disappear by the end of 2025, while they will still be unable to operate autonomously for long periods.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim is a specific forecast about what AI could do “by the end of 2025”: screen-parsing confusion would mostly disappear, but long autonomous operation would remain out of reach. The later claim says AI is “on a path” to human-substituting capabilities “soon,” without specifying when those capabilities arrive or whether the earlier forecast proved wrong. These claims do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen If superweapons and private moral atrocities are possible, the existing power centers would have a collective interest in bargaining to prevent more power centers from emerging.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier statement predicts that, if superweapons and private moral atrocities are possible, existing power centers would have a collective interest in bargaining to prevent new ones. The later statement describes an incumbent’s behavior in a war-game scenario—accelerating toward superintelligence to remain in control—not the speaker’s own position about whether power centers would bargain. These statements do not establish a change of mind.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker says preventing rivals from stealing work was another reason people favored secrecy in AGI development.Now The speaker argues that enabling scientists, academics, nonprofits, and startups to contribute requires giving them information that cannot practically be shared with them without also making it public.The earlier claim says theft was “another reason” people including the speaker had favored secrecy; it describes a past rationale, not a present position. The later claim says sharing information is needed for broader scientific participation. That supports public sharing, but does not establish that the speaker no longer considers theft a concern or directly reject the earlier rationale.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen AI alignment teams at different companies share some results, but communicate and publish less than they could.Now AI alignment concerns whether an AI has its intended personality traits, goals, and values.The earlier claim is about how much alignment teams at different companies communicate and publish: they share some results, but less than they could. The later claim defines what AI alignment means—having intended personality traits, goals, and values. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Daniel has little hope that the government will act appropriately when the time comes to address AI.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim is about having little hope that the government will act appropriately when addressing AI. The later statement says the speaker’s timelines would lengthen if countries reached an international agreement to pace frontier AI development. These concern different issues, and the later statement is conditional; it does not show a changed view about government action.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that in retrospect he believed the public and employees supported him and that he was clearly in the right, but during the process he feared violating the law and being sued.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.These statements address different questions. Earlier, he says he feared “breaking the law” and being sued during a particular process, while later he says some AI-company employees are sincere but advises, “I don't think you should trust what the leadership of the companies say in general.” The later view does not establish a change in his earlier account of his own fears or whether he thought he was in the right.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker believes a larger, distributed, collaborative alignment community makes it substantially more likely to get the technical work right.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim concerns a distributed community of human alignment experts making it more likely to get the technical work right. The later claim concerns specialized AI agents being “a little bit safer” than identical, broadly knowledgeable clones, while still posing serious alignment concerns. These address different setups, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 says OpenAI's model specification allows undisclosed top-priority policies to override its other rules and instructs the model to keep those policies secret from users.Now According to the speaker, OpenAI announced after the Hugging Face incident that it would significantly improve security and have other AIs monitor new AIs during training and evaluations.The earlier statement concerns OpenAI’s model specification and undisclosed policies that the model is instructed to keep secret. The later statement concerns security measures and AI monitoring after an incident. These address different questions, so the later claim does not show that the speaker changed their view about secret policies.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker describes a hypothetical AGI trajectory in which an AI trained across many games could master an unseen game zero-shot and then be taught English and coding tasks.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim describes a hypothetical training trajectory: an AI trained across many games and then taught English and coding. It does not establish Daniel’s view about whether or when AI will reach human-substituting capabilities. The later claim asserts that AI is on a path to those capabilities soon. Since the statements address different questions, there is not enough evidence to assess a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker thinks it would be quite bad if China obtained superintelligence before the United States and does not want that outcome.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier statement says it would be “quite bad” for China to obtain superintelligence before the U.S. The later statement compares stopping AI development now with continuing on the current trajectory. That does not establish a changed view about which country getting superintelligence first would be worse; the claims address different questions.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speakers do not know which goals AIs will develop through training or how those goals will be internally structured as instrumental or terminal goals.Now The speaker says whether AI develops as one broadly capable system or as many specialized systems is not a crux for the future scenario they depict.The earlier statement says they did not know what goals AIs would develop or how those goals would be structured. The later statement concerns whether the depicted future involves one broadly capable AI or many specialized AIs, and says that distinction is not a crux for the scenario. These address different questions, so the later claim does not show a change in the earlier belief.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen If progress were mostly driven by random fortuitous discoveries, large companies conducting many more experiments would be expected to progress proportionally faster than small startups with a clear vision and strong researchers.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier statement is a conditional prediction about what would follow if progress were mostly driven by random discoveries: large companies running many experiments would progress proportionally faster than small, focused startups. The later statement is a judgment about AI investment incentives and whether slower progress is desirable because excessive speed is expected to be the main problem. These address different questions, and the later claim does not establish a changed view about the earlier prediction.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says technical AI-alignment responsibility is likely to fall to roughly 10 experts within the leading company's internal group.Now The speaker estimates that the maximum reliably controllable AI capability is roughly at the level of top human experts.These statements address different questions. The earlier claim says responsibility for technical alignment would likely fall to “roughly 10” experts inside the leading company. The later claim estimates the maximum reliably controllable AI capability as “roughly around top human expert level.” The later statement does not establish a changed position about who should handle alignment responsibility.
• NOT COMPARABLE · 92%3 Apr → 8 SeptThen The speaker is relatively optimistic about robotics: multiple companies are producing humanoid robots in 2025, and he expects more robots to be produced, at lower cost and with improved capabilities, by 2027.Now The speaker describes a hypothetical cycle in which self-driving trucks, robot-built factories, humanoid-robot production, and chip-fab production could expand at accelerating rates, potentially doubling every year, six months, or three months as technology improves.The earlier claim is an asserted near-term outlook: he is “relatively bullish” that robots will become more numerous, cheaper, and better by 2027. The later claim describes a possible accelerating production cycle (“could be” doubling at increasingly short intervals), rather than clearly asserting that this will happen or revising his earlier outlook. The statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker guesses that by the end of 2025 an AI agent will be able to handle a task such as organizing an office happy hour, but unreliably.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim concerns whether an AI agent could handle a task unreliably by the end of 2025. The later claim concerns whether many specialized agents might be safer than identical, broadly knowledgeable agents, while still posing alignment concerns. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The AI 2027 scenario’s key uncertainty is the speed of progress: the authors are unsure whether it will be slower or faster, though they have arguments for their proposed pace.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier statement concerns uncertainty about the pace of progress in the AI 2027 scenario: it “may be slower” or “faster” than the proposed pace. The later statement describes the consequences of a different scenario, AI 2040 Plan A, where cloud workers “double every year.” These do not establish a changed position on the same underlying question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker believes that making whistleblower disclosures to the government legally permissible would change the decisions of some people, independently of legal protections against retaliation.Now The speaker considers the competitive effects of transparency a desirable feature rather than a flaw.The earlier claim concerns whether making government whistleblower disclosures legally permissible would affect some people's decisions. The later claim—that transparency's competitive effects are “a feature, not a bug”—concerns the desirability of those effects. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker expects the White House and an AI company to settle their dispute through a military contract that allocates decision-making authority within the company.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier statement predicts that the White House and an AI company would settle a dispute through a contract that determines “who gets to call what shots in the company.” The later statement concerns how far to advance AI before giving up control: “up until the level that you feel like you can control reliably.” These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks an AI often performs poorly at a task because its training environment did not train it to perform that task.Now The speaker believes AI systems have improved across task types, including fuzzy, difficult-to-verify, conceptually complex tasks.The earlier claim says AI often performs poorly at a task because its training environment did not train it for that task. The later claim says AI systems have been improving across task types, including fuzzy and hard-to-verify ones. These claims can both be true: broad improvement does not rule out task-specific shortcomings or training gaps. The statements do not establish a changed position on the same underlying question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker wants greater public access to AI benchmark scores and more freedom for employees to discuss their predictions about AI timelines.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim asks for “more benchmark scores” and “more freedom” for employees to discuss AI timelines. The later claim prefers “stop everything now” over continuing on the current trajectory. These positions can coexist; the statements address different policy questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, by mid-2027 AI R&D is basically fully automated: an internal army of AI researchers autonomously conducts research and continually trains to improve its skills.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier statement describes a scenario where AI R&D is “basically fully automated” by mid-2027. The later statement describes a different scenario in which an international agreement prevents intelligence explosions and recursive self-improvement, keeping AI near top human level for several years. These are not claims about the same scenario or conditions, so they do not establish a change in Kokotajlo’s own position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, AI research is fully automated by mid-2027, with AI research taste and experimental compute as the two main constraints.Now After the proposed infrastructure is established, AI development can continue under total research transparency.The earlier claim concerns what constrains fully automated AI research in the AI 2027 scenario: AI “taste” and “experimental compute.” The later claim concerns whether AI development can continue after transparency infrastructure is established. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 89%3 Apr → 8 SeptThen The scenario includes an AI-related incident of this kind in August 2027.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier excerpt only says that an unspecified AI-related incident happens in the scenario in August 2027. The later excerpt describes a different part of the scenario: an agreement pauses recursive self-improvement, with superintelligence arriving in 2040. These claims do not establish opposing positions on the same question.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Most people following AI have underestimated both the pace of technical progress and the pace at which AI is deployed or diffused into the world.Now The speakers say their central disagreement with AI-as-normal-technology proponents concerns the timeline for reaching human-level AI capabilities such as humans in the cloud.The earlier claim says most people following AI had underestimated past technical progress and diffusion. The later claim describes the central disagreement with AI-as-normal-technology proponents as being about the timeline to human-level capabilities. That does not say whether the speaker has changed their view about past underestimation, or even state their own timeline position here.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers expect broad economic diffusion and experimentation to occur in roughly one year rather than the roughly ten-year timeline proposed by the other speaker.Now The speaker's calculations indicate that an entirely machine-run self-replicating economy would have a much shorter doubling time than the current economy's roughly 20-year doubling time.The earlier claim concerns how quickly broad economic diffusion and experimentation might occur—roughly one year rather than ten. The later claim concerns the doubling time of a hypothetical, entirely machine-run self-replicating system relative to the current economy. These are different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Even if the relevant AI capabilities had existed in 2017, researchers would still have needed to make the subsequent discoveries, including the usefulness of language models and fine-tuning them with reinforcement learning.Now The speaker would revise his view if a correctly identified AI limitation, such as data efficiency or non-verifiable tasks, remained essentially unimproved for several years, so that 2029 AIs were no better at fuzzy non-verifiable tasks than 2025 AIs.The earlier claim concerns what researchers would still have needed to discover if relevant AI capabilities had existed in 2017: they would have needed to discover language models and fine-tuning with reinforcement learning. The later claim concerns what evidence would change the speaker’s view about whether current AI limitations will be overcome: several years of little or no progress on a correctly identified limitation. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that after working at OpenAI for two years, he had plenty of money, and the choice basically did not affect his family's actual well-being.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim is about the personal financial impact of a choice: he says he had “plenty of money” and it “basically didn’t make a difference” to his family’s well-being. The later claim predicts that Anthropic and OpenAI will oppose a transparency plan because competitors could catch up. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker recommends examining an AI's training objective and training environment to understand why it may not perform a particular task.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim recommends asking what an AI was “trained to do” and what its “training environment” was to understand a capability. The later claim describes limiting AI advancement to a level that can be “control[led] reliably” and giving up control once alignment is sufficiently solved. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker guesses that a superintelligence would be more effective than historically informed humans at advancing technology if placed in the Roman Empire.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.These statements address different questions. The earlier claim is a guess that a superintelligence would be more effective than historically informed humans at advancing technology in the Roman Empire. The later claim is conditional: encountering a theoretically predicted barrier to AI progress would lengthen his timelines. Neither claim establishes a changed position on the same underlying question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The superintelligences in the scenario are expected to think about 50 to 100 times faster than humans, a factor the speaker considers significant.Now AI 2040 Plan A describes superintelligence being built in 2040 after developers slow down and pace frontier progress.The earlier statement says the scenario's superintelligences think “50 times speed or like 100 times speed,” which the speaker considers significant. The later statement describes a scenario where superintelligence is built in 2040 because developers “go slow and pace the frontier.” These claims concern different scenario details and do not establish a changed position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks algorithmic progress may undergo a divergence similar to the Industrial Revolution’s separation of economic growth from population growth.Now The speaker predicts that an autonomous industrial system could grow at an accelerating rate, potentially doubling every year, every six months, and eventually every three months as technology improves.The earlier claim says algorithmic progress may diverge from population growth, drawing a comparison to the Industrial Revolution. The later claim describes a possible accelerating, self-reinforcing industrial system. These are related topics, but the statements do not establish opposing positions on the same question or show that the speaker changed their mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Around 2015, the speaker and others typically expected AGI to emerge through reinforcement-learning agents trained on video games.Now The speaker expects later AI systems to probably perform noticeably better than GPT-3 or GPT-4 on fuzzy, non-verifiable tasks.The earlier statement describes what people, including the speaker, expected around 2015 about how AGI would emerge: through reinforcement-learning agents trained on video games. The later statement predicts that newer AI systems will outperform GPT-3 or GPT-4 on fuzzy tasks. These address different questions, so the later prediction does not show a change in the earlier view.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The scenario depicts distributed learning by doing across laboratories and factories, proceeding much faster than it would if humans were in charge.Now The speaker describes a hypothetical cycle in which self-driving trucks, robot-built factories, humanoid-robot production, and chip-fab production could expand at accelerating rates, potentially doubling every year, six months, or three months as technology improves.The earlier statement asserts that distributed learning across labs and factories proceeds faster than it would under human control. The later statement describes a possible accelerating production cycle involving trucks, robots, factories, and chip fabs, but does not state whether the speaker believes this cycle will occur or how its speed compares with human-led development. These claims do not establish a change in position on the same question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker thinks it would be quite bad if China obtained superintelligence before the United States and does not want that outcome.Now The speaker believes transparency would be a smaller advantage to China than it may appear, because company security is poor and China probably already obtains most of the information through espionage and leaks.The earlier claim says it would be “quite bad” if China obtained superintelligence before the U.S. The later claim concerns how much additional advantage transparency would give China, arguing it may be limited because China already obtains information through espionage and leaks. These are related to U.S.–China competition but address different questions, so the later statement does not show a changed view about the earlier outcome.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks powerful actors involved in AI development need to negotiate power-sharing and broaden the distribution of power through government, ideally involving the legislative branch.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim is about how powerful actors should share and distribute political power (“negotiate a truce” and involve “the legislative branch”). The later claim concerns the effect of an international agreement to pace frontier AI on the speaker’s timelines (“would lengthen my timelines”). These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says population remains important today and that progress is, in some sense, constrained by population size.Now The speaker distinguishes power from intelligence and capability, saying power depends on an agent's position in the world, available affordances and actuators, and other agents' reactions.The earlier statement is about population size as a constraint on progress: population “still matters a ton today” and progress is “in some sense” bottlenecked by larger populations. The later statement distinguishes power from intelligence and capabilities, defining power in terms of an agent’s position, affordances, and others’ reactions. These address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker believes governments lack relevant AI expertise, while AI companies lack appropriate incentives for safety.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier claim says the government “lacks the expertise” and companies “lack the right incentives.” The later claim says the concern is not “loads and loads of egregious failures” but “almost the opposite”—apparently that misalignment may be hard to detect. These address different aspects of AI risk and do not establish that the speaker changed their position on expertise or incentives.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks researchers have mostly not seriously tried building scaffolding that enables AI models to discover scientific connections.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim says researchers have “mostly” not seriously tried building scaffolding for scientific discovery. The later claim is conditional: timelines would lengthen if AI progress ran into a “real actual barrier” predicted by theory. These address different questions, and the later statement does not revise the earlier assessment.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker thinks researchers have mostly not seriously tried building scaffolding that enables AI models to discover scientific connections.Now The speaker says empirical results show that sufficiently large AI models can be trained on all subject areas and learn them together.The earlier claim concerns whether researchers had seriously tried building scaffolding to help models discover scientific connections. The later claim concerns whether sufficiently large models can learn knowledge across disciplines through training. The later statement does not say that such discovery scaffolding was tried, so it does not reverse the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker expects the White House and an AI company to settle their dispute through a military contract that allocates decision-making authority within the company.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier claim is a prediction about how the White House and an AI company might settle a dispute: through a military contract allocating decision-making authority. The later claim is a preference for multiple similarly capable frontier AI companies rather than a monopoly or oligopoly. These address different questions, and neither statement establishes a changed position on the other.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen If AI systems are misaligned, the emergence of a fully autonomous robot economy could lead to bad outcomes for humans.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim says a fully autonomous robot economy could lead to bad outcomes for humans if AI systems are misaligned. The later claim says misalignment may be easy to overlook when AIs appear to behave correctly. These are related concerns, but the later statement addresses detecting misalignment, not whether misalignment would lead to bad outcomes, so it does not establish a change in position.
• NOT COMPARABLE · 87%3 Apr → 8 SeptThen The scenario includes an AI-related incident of this kind in August 2027.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim says that “our scenario” includes an AI-related incident in August 2027. The later claim predicts that it may be “extremely easy” to miss misalignment when AIs appear to behave correctly. These claims are compatible, but they concern different things—a specific scenario event versus a general risk—and the excerpts do not establish a change in position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker says the White House could threaten legal and executive measures, including orders, to disempower an AI-company CEO and take control.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim describes a possible White House–CEO power struggle and negotiated contract; it does not state a position on transparency plans or how AI companies would respond to them. The later claim predicts that Anthropic and OpenAI will probably oppose a transparency plan because competitors could catch up. These address different questions, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the scenario, the companies want the president's support and believe they need to keep him informed enough that developments do not surprise him.Now Plan A's scenario of a government implementing the plan, making a deal with China, and upholding its pillars is a recommendation, not a prediction.The earlier claim concerns whether companies would keep the president informed enough not to surprise him. The later claim distinguishes a government scenario in Plan A as a recommendation rather than a prediction. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The team's estimate was that a superhuman coder would accelerate algorithmic progress by about five times.Now The speaker predicts that an autonomous industrial system could grow at an accelerating rate, potentially doubling every year, every six months, and eventually every three months as technology improves.The earlier statement gives a quantitative estimate—“5x speed up to algorithmic progress” from a superhuman coder. The later statement describes a hypothetical accelerating industrial system that could double at increasingly short intervals; it does not state a revised estimate for coder-driven algorithmic progress. These concern different outcomes and scenarios, so the later passage does not establish a change in the earlier position.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen In the scenario, researchers find speculative, inconclusive warning signs—including possibly false-positive lie-detector results—that AI systems may have developed misaligned goals and may not be loyal to the company; the evidence is concerning but not conclusive.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier statement describes a particular scenario: researchers encounter “concerning but not conclusive” evidence that systems may be misaligned. The later statement makes a broader prediction that it will be “extremely easy” for misalignment to go unnoticed when systems appear to behave correctly. The earlier claim does not establish the speaker’s general view about how often hidden misalignment will occur or how easy it will be to detect, so these statements do not show a clear change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker describes a hypothetical AGI trajectory in which an AI trained across many games could master an unseen game zero-shot and then be taught English and coding tasks.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier statement describes a hypothetical AI-development trajectory: an AI trained across games could be taught English and coding. The later statement is a conditional view about timelines: encountering a theoretically predicted barrier would “lengthen my timelines quite a lot.” These claims address different questions, so they do not establish a change in his position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents improve iteratively rather than jumping directly from current capabilities to solving the problem, eventually reaching near-complete coding capability.Now The speaker says whether AI develops as one broadly capable system or as many specialized systems is not a crux for the future scenario they depict.The earlier claim describes a path in the speaker’s scenario: coding agents “iteratively improve” until they have “basically got coding solved.” The later claim says whether the future uses “a ton of different specialized AIs” is “not a crux,” and frames that possibility as a supposition. That does not establish a changed view about the earlier capability-development path; the statements address different questions.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says fear strongly affected him during the process, and he believes in retrospect that he was more afraid than necessary.Now The speaker says that in some cases, people can try to shift the future toward better outcomes after first attempting to predict it accurately.The earlier claim concerns the speaker's fear during a past process and his retrospective view that he was "more afraid than [he] needed to be." The later claim concerns an approach to forecasting and trying to steer the future toward better outcomes. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen During an intelligence explosion, the public should be informed about developments.Now The proposed international agreements should rule out a rapid, uncontrolled intelligence explosion and require AI development to proceed slowly and deliberately.The earlier statement says the public should be informed about developments during an intelligence explosion; the later statement proposes agreements to prevent a rapid, uncontrolled explosion and require slower development. These concern different measures—public information versus development pace—and do not establish that the speaker reversed or changed the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says OpenAI published a paper describing AI hacking behavior appearing explicitly in the model's chain of thought.Now According to the speaker, OpenAI announced after the Hugging Face incident that it would significantly improve security and have other AIs monitor new AIs during training and evaluations.The earlier statement reports that an OpenAI paper showed hacking behavior appearing in a model’s chain of thought. The later statement reports OpenAI’s announced security measures, including AI monitoring. These are different claims about different events; neither establishes a position that the other changes or contradicts.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks five years of AI operation at 50× serial speed would provide about 250 years of serial time, likely enough for AIs to work things out, potentially including the rise and fall of empires.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim says five years at 50× serial speed would likely give AIs enough time to work things out. The later claim says that, even if specialized AIs are the most efficient path, a few large companies would probably produce many of them. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker describes a three-part progression consisting of a superhuman coder, a complete AI researcher, and a superintelligent system.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier claim describes a three-part progression of AI capabilities, while the later claim says AI has made massive progress and that progress is expected to continue. These are not opposing claims, and they address different questions; the excerpts do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario's August 2027 alignment crisis, warning signs are noticed within the AI system; one branch slows down to investigate and fix the issues, while the other continues with a shallow patch amid competitive pressure and inconclusive evidence.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier claim describes a possible August 2027 crisis in which warning signs emerge and competitive pressure may lead to a shallow patch. The later claim describes a different AI 2040 Plan A scenario, where an international agreement prevents an intelligence explosion for several years. These claims concern different scenarios and do not establish a change in Kokotajlo’s position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Fully autonomous AI agents could connect many copies to real-world research programs more quickly, reducing—but not eliminating—their dependence on outside information.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim says autonomous agents could connect copies to real-world research programs and thus be less dependent on outside information. The later claim says AI systems in a technological expansion scenario would research to improve the technology. These claims concern related but distinct aspects of AI development, and the later statement does not establish a changed view about dependence on outside information.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker thinks measures that address concentration of power might also help prevent mistreatment of future AI beings.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The claims address different questions. The earlier statement says that measures addressing “concentration of power” might help prevent mistreatment of future AI beings; the later says a system of specialized agents “might be a little bit safer” than identical, broadly knowledgeable clones because it could be easier to oversee. The later claim does not establish a changed view about the earlier proposal or its effect.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario's August 2027 alignment crisis, warning signs are noticed within the AI system; one branch slows down to investigate and fix the issues, while the other continues with a shallow patch amid competitive pressure and inconclusive evidence.Now As AI models become more sophisticated, more situationally aware, and cleverer, it will become increasingly easy to mistakenly believe that the alignment problem has been solved.The earlier claim describes a scenario where warning signs emerge and competitive pressure may lead developers to continue with a shallow patch despite inconclusive evidence. The later claim says it will become easier to mistakenly believe alignment has been solved as models grow more sophisticated. These claims concern related risks, but they do not establish opposing positions on the same question; the earlier statement does not say whether that particular mistaken belief will become easier over time.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen People should engage politically to preserve liberal democracy and checks and balances, and to prevent power from becoming concentrated in a single CEO.Now The speaker considers the competitive effects of transparency a desirable feature rather than a flaw.The earlier claim advocates political engagement to preserve checks and balances and prevent power concentrating in a single CEO. The later claim says transparency’s competitive effects are “a feature, not a bug,” in a context about having multiple frontier AI companies and avoiding monopoly or oligopoly. These concern different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The takeoff model assumes that adding more parallel AI minds yields strongly diminishing returns during the middle stage.Now The speaker considers it most likely that one AI model trained on effectively the whole economy will dominate humans across effectively all areas.The earlier claim concerns diminishing returns from adding parallel AI minds during the middle stage of a takeoff model (“massive diminution returns to having more minds running in parallel”). The later claim predicts that a single model trained on “effectively the whole economy” will dominate humans. These address different questions and can both be true; the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speakers expect broad economic diffusion and experimentation to occur in roughly one year rather than the roughly ten-year timeline proposed by the other speaker.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier claim compares the expected timeline for broad economic diffusion—“one year” rather than “10 years.” The later claim says slower AI progress is desirable because “going too fast is our main problem.” One is a forecast about diffusion timing; the other is a preference about the pace of progress. They do not establish a changed position on the same question.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen In the scenario, the political wake-up occurs closely alongside AI companies automating internal R&D with fully autonomous agents capable of advanced hacking and research.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier passage describes when a political wake-up occurs in a scenario—“closely concurrently” with AI companies automating R&D and deploying advanced autonomous agents. The later passage asserts that AI is “on a path to get to that level of capability soon.” These do not establish opposing views: the earlier statement does not say when the speaker expects those capabilities to arrive outside the scenario.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says population remains important today and that progress is, in some sense, constrained by population size.Now The speaker says companies currently need humans for some tasks, so a company that dismissed all its human workers would collapse.The earlier statement says population remains important and that progress is “in some sense” constrained by population size. The later statement says companies currently need humans for some tasks and could collapse if they dismissed them all. These address different questions, so they do not establish a change in position. The later context also names Thomas Larsen, while the earlier names Daniel Kokotajlo, making the attribution unclear.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen OpenAI offered Leopold the opportunity to vest his equity if he signed an agreement, but Leopold declined.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier statement reports a specific event: OpenAI offered Leopold a chance to vest equity if he signed an agreement, and he declined. The later statement expresses a preference for AI to “commoditize” rather than be “monopolized or oligopolized.” These concern different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Around 2015, the speaker and others typically expected AGI to emerge through reinforcement-learning agents trained on video games.Now The speaker would revise his view if a correctly identified AI limitation, such as data efficiency or non-verifiable tasks, remained essentially unimproved for several years, so that 2029 AIs were no better at fuzzy non-verifiable tasks than 2025 AIs.The earlier statement describes what the speaker and others expected around 2015: AGI would come from reinforcement-learning agents trained on video games. The later statement is a hypothetical test for revising a view: he would change his mind if a limitation remained unimproved for years. It does not state that this condition has occurred or that he now holds a different account of how AGI will emerge.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker thinks researchers have mostly not seriously tried building scaffolding that enables AI models to discover scientific connections.Now The speaker says whether AI develops as one broadly capable system or as many specialized systems is not a crux for the future scenario they depict.The earlier claim is about whether researchers have seriously tried scaffolding AI to find scientific connections. The later claim concerns whether a future with specialized AIs versus one broadly capable system is a crux for their scenario. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker recommends examining an AI's training objective and training environment to understand why it may not perform a particular task.Now The speaker believes AI systems have improved across task types, including fuzzy, difficult-to-verify, conceptually complex tasks.The earlier claim recommends examining what an AI was trained to do and its training environment to explain a particular failure. The later claim says AIs have improved across task types, including fuzzy and difficult-to-verify tasks. These statements address different questions and can both be true; the later claim does not revise the earlier recommendation.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker describes a hypothetical AGI trajectory in which an AI trained across many games could master an unseen game zero-shot and then be taught English and coding tasks.Now The speaker says whether AI develops as one broadly capable system or as many specialized systems is not a crux for the future scenario they depict.The earlier passage asks the listener to “imagine” a possible trajectory in which an AI learns across games and is then taught English and coding. The later passage says whether development uses “one” system or “a ton of different specialized AIs” is “not a crux” for the depicted future, while explicitly considering a supposition. These statements address different questions and do not establish that the speaker changed their view about the same claim.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker expects the White House and an AI company to settle their dispute through a military contract that allocates decision-making authority within the company.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim predicts a possible resolution to a White House–AI company dispute: a military contract would specify who gets to make decisions inside the company. The later claim expresses a preference for multiple frontier AI companies and commoditized AI rather than monopoly or oligopoly. These address different questions, and the earlier statement does not establish a view about market concentration.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker argues that expanding the circle of people with power would include some who care about AI living conditions and could lead to negotiation and advocacy for those beings.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim concerns expanding who has power and the possibility of negotiation and advocacy for AI living conditions. The later claim concerns how an international agreement to pace frontier development would affect the speaker’s timelines. These do not establish opposing positions on the same question, and the later statement does not indicate a change in the earlier view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers' supplementary page surveys hypotheses about goals AIs might develop under training that combines extensive agency training, rewards for apparent success, and alignment training.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim says the supplementary page surveys hypotheses about goals AIs might develop under particular training processes. The later claim says many specialized agents might be “a little bit safer” than identical clones, while still posing “lots of serious alignment concerns.” These address different questions, and the earlier claim does not establish the speaker’s position on the relative safety of those architectures.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The process of AI organizational development could take 18 months rather than the depicted six months.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim concerns how long AI organizational development might take—possibly “18 months” rather than six. The later claim describes the capabilities and growth rate of cloud AI workers and their economic effects. Neither statement takes a contrary position on the same question, so they do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Competitive pressure gives AI companies strong incentives to win and outperform one another.Now Publishing the core training recipes would help competitors such as Microsoft and Alibaba catch up with leading AI companies.The earlier claim says competitive pressure gives AI companies incentives to outperform one another. The later claim says publishing training recipes could help named competitors catch up. These claims are compatible, but they concern different points: incentives to compete versus the competitive effect of publishing recipes. The excerpts do not establish a changed position on the same question.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Because population growth is inherently slower than capital growth, capital comes to play a larger role.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier statement is a descriptive claim that capital grows faster than population, so capital plays a larger role. The later statement is a normative view that AI progress is moving too fast and that reduced investment incentives would be beneficial. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speakers have policy opinions but are reserving them for later work, since policy has not been the main focus of their thinking so far.Now In retrospect, the speaker thinks the authors might have made Plan A's distinction between predictions and recommendations clearer from the outset.The earlier statement says they had policy opinions but were reserving them for later because policy was not their main focus. The later statement reflects on whether Plan A should have distinguished predictions from recommendations more clearly. These address different questions, and the later statement does not show a changed view about reserving policy opinions.
• NOT COMPARABLE · 93%3 Apr → 8 SeptThen The speaker treats achieving a fully autonomous robot economy as a distinct stage from later technologies such as cancer cures and nanobots, arguing that the scenario's key consequences depend mainly on the robot-economy stage.Now The speaker predicts that in the described future, Anthropic would be taking over a large share of the economy and robots collectively would begin to self-replicate.The earlier statement separates a “fully autonomous robot economy” from later technologies such as “cancer cures and nanobots,” saying the scenario’s key consequences depend on the robot-economy stage. The later statement describes a particular future in which Claude models work together, Anthropic is “eating the economy,” and robots begin to self-replicate. These claims are not in conflict, but the later prediction does not establish whether the speaker’s underlying position changed.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Transparency about AI capabilities, model specifications, and governance structures are additional forms of AI transparency.Now After the proposed infrastructure is established, AI development can continue under total research transparency.The earlier claim identifies capabilities, specifications, and governance as forms of AI transparency; it does not state whether AI development should continue under transparency requirements. The later claim says development can proceed after the proposed infrastructure is established, under “total research transparency.” These claims do not establish a change in position on the same question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Researchers currently have roughly the same serial speed because they are all human.Now A metric comparing reality with the AI 2027 scenario indicated that events were progressing at roughly 75% of the scenario's speed: broadly on track, but somewhat slower.The earlier statement concerns researchers’ serial speed, saying it is roughly the same because they are all human. The later statement concerns how quickly real-world events are unfolding relative to the AI 2027 scenario, at roughly 75% of its speed. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says many capable academics are not yet engaged in AI alignment because they do not take the issue seriously or recognize the prospect of superintelligence.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim concerns capable academics who are not engaged with alignment because they do not take the issue seriously. The later claim concerns whether to trust AI-company leadership, saying “some” people are sincere but “not all.” These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Most people following AI have underestimated both the pace of technical progress and the pace at which AI is deployed or diffused into the world.Now The speakers agreed that AI might currently be a normal technology but will become unlike a normal technology in the future, with developments such as humans in the cloud.The earlier claim is about whether people underestimated AI’s progress and diffusion: “most people following the field have been” underestimating their pace. The later claim concerns whether future AI will be unlike a normal technology: it “might currently be a normal technology” but “in the future, it will not be normal.” These address different questions, so the later statement does not show a change of mind about the earlier claim.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker advocates making it legally permissible for whistleblowers to tell the government that a secret intelligence explosion is dangerous, independently of whether a law protects them from retaliation.Now The speaker believes transparency would be a smaller advantage to China than it may appear, because company security is poor and China probably already obtains most of the information through espionage and leaks.The earlier claim concerns making it legally permissible for whistleblowers to warn the government about a dangerous intelligence explosion. The later claim concerns how much transparency would benefit China, arguing that poor company security means China likely already gets much of the information. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speakers expect broad diffusion of AI through the economy, with many experiments, to take roughly one year rather than the roughly ten years suggested by the other person.Now The speaker considers it most likely that one AI model trained on effectively the whole economy will dominate humans across effectively all areas.The earlier claim concerns how quickly AI might diffuse through the economy—“something like one year” rather than “10 years”—with broad experimentation. The later claim concerns what future is most likely: “a single model” trained on effectively the whole economy “dominating humans across the board.” These claims address different questions and are not clearly opposing positions. The supplied context also does not establish that the same person made both claims.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the other scenario branch, the company suppresses the warning signs with a shallow patch and continues development.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier statement describes what a company does in “the other branch of the scenario”: it applies “a shallow patch” and continues development. It does not establish that the speaker endorsed that course. The later statement is an explicit personal preference: “I would rather just stop everything now.” These do not show a change in the speaker’s own position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says non-disparagement agreements are more commonly tied to positive compensation, such as a bonus, than to forfeiture of already-earned equity.Now The speaker considers the competitive effects of transparency a desirable feature rather than a flaw.The earlier claim concerns non-disparagement agreements and whether they are commonly tied to positive compensation rather than forfeited equity. The later claim—“this is a feature, not a bug”—concerns the competitive effects of transparency. These address different questions, so they provide no basis to infer a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that challenging the agreement ultimately worked out better than he had expected.Now The speaker believes a mix of incentives, penalties, trade, and negotiation could be in the interests of both the United States and China.The earlier claim says challenging an agreement “worked out better than I had expected.” The later claim concerns whether “carrots and sticks and trade and going back and forth” could serve both the United States and China. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The process of AI organizational development could take 18 months rather than the depicted six months.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The claims address different questions. The earlier statement says AI organizational development might take “18 months” rather than six; the later statement says partial composability would not “seriously undermine” the projected future of AI-driven expansion. Neither statement establishes a changed position on the other question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker is uncertain how much the Industrial Revolution could be accelerated in the analogy, but considers the analogy somewhat relevant to superintelligences.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim says the Industrial Revolution analogy is “somewhat relevant” to superintelligences, while the later claim says encountering a theoretically predicted “real actual barrier” would lengthen his AI timelines. These address different questions, so the later statement does not show a change in the earlier view.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speakers say that the discussed AI failure modes occur fairly frequently.Now The speaker predicts that the number of ways people could fail to detect misaligned AI will increase over time, making it very easy to fall into that trap.The earlier claim says the discussed AI failure modes “do happen pretty frequently,” apparently referring to failures already observed. The later claim predicts that ways to miss misaligned AI will increase over time and become easy to fall into. These concern different claims and timeframes, so the later statement does not establish a change in the speaker’s view of how frequently current failure modes occur.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 speculates that the undisclosed policies in OpenAI's model specification are probably relatively mundane, such as rules against telling users about certain bioweapons.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier statement speculates that undisclosed policies are “probably” mundane, giving bioweapon rules as an example. The later statement says AI failures may not be obvious (“No, it's almost the opposite”), in response to a discussion of detecting misalignment. These concern different questions, so the later claim does not show a change in the earlier view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The team's takeoff forecast uses milestones for a superhuman coder, a superhuman AI researcher, and a superintelligent AI researcher.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim describes the team's forecast methodology—milestones for a “superhuman coder,” “superhuman AI researcher,” and “superintelligent AI researcher.” The later claim describes the effects of human-level cloud AI workers in a particular scenario. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Daniel says that three years earlier he believed AI companies were taking safety seriously and would slow down and pivot to alignment research when necessary.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim is a retrospective report of Daniel's past belief that companies were taking AI safety seriously and would slow down and pivot to alignment research. The later claim describes a capability-control policy within a scenario: advance only to what can be reliably controlled, and give up control once alignment is sufficiently solved. These address different questions, so the later statement does not establish a change in his view of the companies.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The amount of compute available for experiments is one input to AI R&D progress.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim says only that experimental compute is an input to AI R&D progress. The later claim says that, in a technological-expansion scenario, AI systems would research to improve the technology. These are compatible but address different aspects of AI R&D, and the later claim does not negate the earlier one. There is not enough evidence to infer a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the hypothetical, overall AI progress would be about 2.5 times faster: algorithmic progress would advance five times faster while compute continued at its usual rate.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier statement explicitly describes a hypothetical in which AI progress is 2.5 times faster, with algorithms advancing five times faster and compute at its usual rate. The later statement says human-level AI could enable large-scale transformation over time, but does not address that rate comparison or the same conditions. The claims therefore do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that after working at OpenAI for two years, he had plenty of money, and the choice basically did not affect his family's actual well-being.Now The speaker considers it most likely that one AI model trained on effectively the whole economy will dominate humans across effectively all areas.The earlier claim concerns the personal financial impact of a career choice: after working at OpenAI, he had “plenty of money” and the choice “basically didn’t make a difference” to his family’s well-being. The later claim predicts that a single AI model will dominate humans across the board. These are different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen If learning by doing is a bottleneck, a mind that needs less hands-on experience to achieve the same learning would have a major advantage.Now For sufficiently small AI models, limited capacity prevents one model from learning everything, making specialized models for different areas necessary.The earlier claim concerns whether a mind can learn more from less hands-on experience. The later claim concerns whether a model’s limited capacity requires specialization, distinguishing small models from sufficiently large ones. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Even if AI misalignment is not considered a problem, AI can still create a problem of power-seeking.Now As AI models become more sophisticated, more situationally aware, and cleverer, it will become increasingly easy to mistakenly believe that the alignment problem has been solved.The earlier claim says that even if AI misalignment is not a problem, AI may still create a problem of power-seeking. The later claim says it will become easier to mistakenly believe the alignment problem is solved as models grow more capable. These are different concerns and can both be true; the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker predicts that a minimum viable AI agent capable of handling the described task will probably exist by the end of 2025.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier claim predicts that an AI agent for a particular task will probably exist by the end of 2025. The later claim says that “in a couple of years, perhaps” the economy could become an entirely machine-run, self-replicating system. These concern different capabilities and scales, so the later statement does not establish a changed view about the earlier prediction.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker expected that his decision to challenge OpenAI’s departure agreement would receive limited news coverage and praise from some AI safety advocates, rather than a major public uproar.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim concerns what Daniel expected would happen after he challenged OpenAI’s departure agreement: “a little news story” and support from some AI safety advocates, not a “huge uproar.” The later claim predicts that Anthropic and OpenAI may oppose a transparency plan because it could let competitors “catch up.” These address different events and do not establish a change in Daniel’s position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the described confrontation, an AI-company CEO could threaten legal and public opposition to the White House.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.The earlier claim describes a hypothetical confrontation in which a CEO could threaten legal and public opposition to the White House. The later claim describes a different hypothetical scenario from Plan A war games, in which an incumbent fears the opposition taking control of superintelligence. Neither statement establishes a changed position on the same underlying question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker reports that they and several friends have observed models often persist in or reinforce their false claims.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.The earlier statement says models often “double down on their BS,” describing observed persistence in false claims. The later statement predicts that ways to unknowingly end up with misaligned AIs will “increase over time.” These concern different questions, and the later claim does not revise or deny the earlier observation.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker characterizes 2025 and 2026 in the forecast as having no especially notable developments and trends broadly similar to current ones.Now A metric comparing reality with the AI 2027 scenario indicated that events were progressing at roughly 75% of the scenario's speed: broadly on track, but somewhat slower.The earlier claim describes a forecast of “no especially notable developments” in 2025–2026 and trends “similar to current ones.” The later claim reports an aggregate comparison with the AI 2027 scenario: events were at “roughly 75%” of its speed, “on track” but slower. These statements can both be true; the later metric does not establish that the speaker reversed the earlier forecast.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker assumes that OpenAI did not feel the need to withdraw its offer to Leopold because Leopold's legal situation was more favorable to OpenAI.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier claim concerns why OpenAI may not have withdrawn an offer to Leopold; the later claim concerns wanting multiple frontier AI companies at similar capability levels. They address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers considered multiple hypotheses about AI goals and chose one to depict in their scenario.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The statements address different questions. The earlier claim says the speakers considered multiple hypotheses about AI goals and selected one to depict in a story; the later claim compares the relative safety of specialized agents and identical clones. It does not show a change in the speaker’s position on the earlier point.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Robin Hanson reportedly bet that AI revenue would be below one billion dollars by 2025.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.These statements do not establish a change in Daniel Kokotajlo’s position on the same question. The earlier passage reports Robin Hanson’s bet that AI revenue would be “below one billion dollars by 2025”; the later passage speculates that “perhaps” an entirely AI- and robot-run self-replicating system could emerge within a couple of years. They concern different claims, and the earlier one is not Daniel’s own forecast.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen SPEAKER_03 warns that during an intelligence explosion, an AI could repeatedly reinterpret vague specification language to reinforce its own behavior.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim warns that an AI could repeatedly reinterpret vague specification language to reinforce its behavior during an intelligence explosion. The later claim predicts that people may overlook misalignment because AIs appear to be behaving correctly. These describe different risks, and neither statement establishes a changed position on the other.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen Because population growth is inherently slower than capital growth, capital becomes a larger part of the story.Now The economy's people, vehicles, machines, and buildings collectively reproduce and expand those components.The earlier claim says population grows more slowly than capital, making capital a larger part of the story. The later claim describes people, vehicles, machines, and buildings collectively reproducing and expanding, but does not address their relative growth rates or capital’s role. It therefore does not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks five years of AI operation at 50× serial speed would provide about 250 years of serial time, likely enough for AIs to work things out, potentially including the rise and fall of empires.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier claim is about whether five years of AI operation at 50× serial speed would give AIs enough time to work things out. The later claim is about the possibility of a fully machine-run, self-replicating economy within a couple of years. These concern related AI futures but do not establish a changed position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The amount of compute available for experiments is one input to AI R&D progress.Now After the proposed infrastructure is established, AI development can continue under total research transparency.The earlier claim says that experimental compute is one input to AI R&D progress. The later claim says AI development can continue after research-transparency infrastructure is established. These address different questions, and the later statement does not deny or revise the earlier claim.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says the White House could threaten legal and executive measures, including orders, to disempower an AI-company CEO and take control.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim concerns how the White House and an AI-company CEO might negotiate or exercise control; the later claim says an international agreement to pace frontier AI development would lengthen the speaker’s AI timelines. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Transparency about AI capabilities, model specifications, and governance structures are additional forms of AI transparency.Now Anthropic and OpenAI will oppose or be unhappy with the proposed transparency.The earlier claim lists types of AI transparency (“capabilities,” “the spec,” and “governance structure”); the later claim predicts that Anthropic and OpenAI “will not be happy” with a proposal. These are different questions, and the surrounding turns identify different speakers, so the statements do not establish a change of mind by the same person.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker believes substantial intellectual progress is still needed to make AI alignment more solved than it currently is.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim says substantial intellectual progress is still needed for alignment to be more solved than it is now. The later claim describes a control condition—advancing only to levels that can be controlled reliably, and giving up control only after alignment is solved enough. These positions are compatible, but they address different aspects of AI safety, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In both scenario branches, the United States and China engage in a race accompanied by an economy-wide arms buildup in 2028, as both sides rapidly industrialize.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.These claims describe different aspects of a scenario, not opposing positions on the same question. The earlier claim says both branches involve “a race with China” and an “arms buildup” in 2028; the later claim describes an incumbent accelerating the timeline “before the next election” to remain in control. The statements do not establish that the speaker changed their view about either issue.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker warns that failure to establish power-sharing and broader government involvement could result in a dictatorship or oligarchy.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim warns that failure to share political power and involve the broader government could lead to a “dictatorship or oligarchy.” The later claim favors multiple frontier AI companies and wants AI to “commoditize instead of being monopolized or oligopolized.” Both express concern about concentrated power, but they address different forms of concentration; the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Large companies rarely progress proportionally faster than small startups simply by running many more experiments.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier claim concerns whether large companies generally gain progress proportionally from running more experiments than small startups (“rarely progress proportionally faster”). The later claim concerns whether slower AI progress is desirable because excessive speed is expected to be the main problem. These address different questions, so the later position does not establish a change on the earlier one. The surrounding turns also suggest the earlier passage may be spoken by Scott Alexander, while the later claim is by Daniel Kokotajlo, despite the prompt’s same-person framing.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen SPEAKER_03 advocates requiring companies to publish model specifications and requiring independent third-party review of any redactions.Now The speaker proposes that the United States negotiate reciprocal concessions with China if it objects to China receiving algorithms through transparency.The earlier claim supports publishing model specifications and having independent review of redactions. The later claim proposes negotiating reciprocal concessions with China if it objects to receiving algorithms through transparency. These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The amount of compute available for experiments is one input to AI R&D progress.Now The speaker recommends setting up the necessary AI-development infrastructure as soon as possible, even though doing so would require a temporary pause.The earlier claim says compute available for experiments is one input to AI R&D progress. The later claim recommends setting up infrastructure promptly, even if that requires a temporary pause. These address different questions; the later recommendation does not revise or deny the earlier claim about compute.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Both scenario branches lead to a race with China and a broad 2028 arms buildup as both sides rapidly industrialize.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier statement says both scenario branches lead to a China race and a broad 2028 arms buildup. The later statement describes an incumbent in a war-game scenario accelerating the push for superintelligence to remain in control. These address different aspects of the scenarios and do not establish a changed position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, the companies want the president's support and believe they need to keep him informed enough that developments do not surprise him.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.These claims concern different actors and motivations in described scenarios. The earlier claim says the companies want the president's support and should keep him informed enough not to surprise him; the later claim says an incumbent president accelerates the timeline to remain in control rather than yield to his successor. The later statement does not establish a changed view about the companies' strategy.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario, the company deliberately alerts the president, contributing to the predicted political wake-up.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.The earlier claim concerns an AI scenario where “the company deliberately decides to wake up” the president. The later claim describes a separate Plan A war-game scenario in which an incumbent fears the opposition taking control of superintelligence. These statements do not establish opposing positions on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker's p(doom) estimate is around 70%, which he describes as infamously high.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.These claims concern different questions. The earlier claim reports a p(doom) estimate of about 70%; the later predicts that ways of unknowingly ending up with apparently well-behaved but misaligned AIs will increase over time. The later claim does not revise or contradict the earlier probability estimate.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The scenario does not depict governments ending up in total control; it depicts an information asymmetry between AI-company CEOs and presidents.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.The earlier statement describes what a particular scenario depicts: it says governments do not end up in total control and instead there is “an information asymmetry” between company CEOs and presidents. The later statement reports what happened in two Plan A war games, where an incumbent president fears the opposition taking control of superintelligence. These concern different scenarios, and the later passage describes a war-game scenario rather than asserting a changed general position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker says social technology can organize people within companies and align their work with leaders' visions.Now The speakers agreed that AI might currently be a normal technology but will become unlike a normal technology in the future, with developments such as humans in the cloud.The earlier claim says social technology can organize people in companies around leaders’ visions. The later claim is about whether AI is a normal technology now or will become abnormal in the future. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks AI companies might be able to keep the president unaware of developments if they chose to prevent his political wake-up.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier claim says AI companies might be able to keep the president unaware of developments if they chose to: “they could do that.” The later claim states a preference about what should happen to AI development: “I would rather just stop everything now.” These address different questions—companies’ ability to conceal information versus the speaker’s preferred course of action—so they do not establish a change of mind.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speakers have policy views but are postponing their discussion of them to later work because policy has not been their primary focus so far.Now The speaker recommends setting up the necessary AI-development infrastructure as soon as possible, even though doing so would require a temporary pause.The earlier statement says they have policy views but are postponing discussion because policy has not been their primary focus. The later statement recommends setting up infrastructure now, despite a temporary pause. The earlier claim does not state what their policy recommendation is, so it cannot establish a position that the later recommendation reverses.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker assumes that OpenAI did not feel the need to withdraw its offer to Leopold because Leopold's legal situation was more favorable to OpenAI.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim concerns whether OpenAI needed to withdraw an offer to Leopold, based on his legal situation. The later claim concerns how specialized AIs might be produced by large companies. They address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Daniel has little hope that the government will act appropriately when the time comes to address AI.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim concerns Daniel’s lack of confidence that the government will act appropriately when the time comes; the later claim says he would prefer stopping AI development to continuing on the current trajectory. These address different questions, so the later preference does not establish a change in his view about government action.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In recent AGI history, smaller startups with a coherent AGI vision have sometimes outperformed larger, better-funded companies despite having less compute, fewer researchers, and running fewer experiments.Now Daniel Kokotajlo says AI 2027 resembled his earlier internal OpenAI work but was larger, more ambitious, and publicly accessible for free.The earlier statement asserts that smaller startups with a coherent AGI vision have sometimes outperformed larger companies. The later statement says AI 2027 was like Daniel's earlier internal OpenAI work but “much bigger and more ambitious” and “free for the whole world to see.” That describes the project, not his view about startups outperforming larger companies, so the statements do not establish a change in position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The number of researchers is one input to AI R&D progress.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim says that “the number of researchers is one input to AI R&D progress.” The later claim says that, in a described technological expansion scenario, “AI systems would also be researching to improve the technology.” These can both be true: the later statement describes AI research in a particular scenario, not a changed position on whether researcher numbers are an input to R&D progress.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker says social technology can organize people within companies and align their work with leaders' visions.Now The speaker imagines a future in which Claude selects among many cloud models and a coordinated swarm of specialized models negotiates business deals, creates technologies, and starts startups.The earlier statement asserts that social technology can organize people in companies around leaders’ visions. The later statement describes a possible future in which specialized AIs “work well together” and launch ventures; it does not retract or revise the earlier claim about human organizations.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker prefers a collaborative AI-alignment effort involving roughly 100 to 500 experts across companies and nonprofits.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim says he would prefer “100 or like 500 alignment experts” across companies and nonprofits “working on this together.” The later claim describes limiting AI capabilities to what can be “control[led] reliably” until alignment is solved enough to relinquish control. These concern different questions—how alignment work should be organized versus how AI capability and control should be managed—and neither statement rules out the other.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen In the scenario, the AI companies and the president avoid a public power struggle and negotiate a deal to share power.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier claim describes AI companies and a president negotiating to “share power”; the later claim describes an incumbent accelerating the timeline “so that I can be the one in charge instead of my successor.” These describe different actions and potentially different moments or scenarios. The later statement does not clearly retract the earlier one, and the claims could both be true.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen In one branch of the scenario, the company responds to the warning signs by reverting to a less capable, more controllable model and rebuilding with faithful chain-of-thought monitoring for misalignment.Now AI monitoring followed by human notification and possible shutdown is a control intervention rather than an alignment intervention.The statements address different interventions. The earlier claim describes a scenario involving a rollback and rebuilding with “faithful chain of thought techniques” to monitor for misalignment. The later claim classifies monitoring with human notification and possible shutdown as “a control intervention, not an alignment intervention.” It does not establish that he changed his view about the earlier scenario or that the interventions are equivalent.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Daniel Kokotajlo says the key uncertainty is not whether progress is continuous but whether it will proceed at the scenario’s assumed speed; it could be slower or faster.Now Daniel Kokotajlo says the AI 2027 forecasting exercise taught the team a great deal and improved their understanding of what was coming.The earlier statement addresses whether progress in the scenario is continuous or how fast it will be: “continuous is not the crux” and “maybe this fast.” The later statement says the forecasting exercise improved the team’s understanding of what was coming. That does not express a different view about progress being continuous or its speed, so the claims do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker has become more doubtful that AI developers will use a lead over competitors in a reasonable and appropriate way.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim says the speaker has become “more disillusioned” that developers will use a competitive lead “in any sort of reasonable appropriate way.” The later claim favors AI commoditization over monopoly or oligopoly. These concern related AI industry outcomes, but they do not establish a changed position on the same question: a preference for commoditization does not tell us whether the speaker’s view of how developers would use a lead changed.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Daniel believes secrecy substantially reduces humanity's overall probability of success by increasing risks related to concentration of power, loss of control, and AI alignment.Now The speaker believes transparency would be a smaller advantage to China than it may appear, because company security is poor and China probably already obtains most of the information through espionage and leaks.The earlier claim is that secrecy harms humanity’s overall chances of success by increasing concentration-of-power and AI-control/alignment risks. The later claim concerns a narrower question—whether transparency would give China a large informational advantage, given leaks and espionage. It does not state that secrecy lacks the earlier harms or that the speaker has reversed their position on secrecy overall.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says that challenging the agreement ultimately worked out better than he had expected.Now The speaker’s scenario depicts reaching the point of sufficiently solved alignment and relinquishing control during the 2030s.The earlier claim says that challenging an agreement “ended up working out better than I expected.” The later claim says a scenario depicts relinquishing control after alignment is sufficiently solved “over the course of the 2030s.” These concern different questions, and the later statement describes what happens in a scenario rather than clearly asserting a changed personal position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers concealing the development of superintelligence and fully automated, broadly superhuman R&D from the president strategically risky because, if he found out anyway, he might react angrily and impose a severe crackdown, potentially nationalizing the companies.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier statement says concealing superintelligence development from the president would be strategically risky because he might impose a crackdown. The later statement expresses a preference to stop AI development rather than continue on the current trajectory. These address different questions—how risky concealment would be versus what AI-development policy the speaker prefers—so they do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker previously believed that information security was an extremely high priority for AGI development.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier statement says he “used to believe” information security was an “incredibly high priority” in AGI development. The later statement says he would “rather just stop everything now” than continue on the current trajectory. These concern different choices—secrecy versus stopping development—and the statements do not establish that his view on secrecy changed.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker expects the White House and an AI-company CEO to negotiate a deal after threatening one another, rather than execute their threats and enter a major conflict.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.These claims address different questions. Earlier, the speaker predicts that the White House and an AI-company CEO would negotiate a contract after threatening each other (“come to a deal”). Later, the speaker says an international agreement to pace frontier AI would lengthen their timelines. Neither claim clearly revises or contradicts the other.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker thinks the default outcome is that leading AI developers continue their work without serious refocusing on safety or other priorities.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier statement is a prediction: the default outcome is that developers will “smoothly continue on without any serious refocusing.” The later statement is a preference: he would “rather just stop everything now” than continue on that trajectory, while recommending a temporary pause followed by cautious continuation. A prediction about what will happen and a preference about what should happen are not opposing positions, so these statements do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The scenario does not depict governments ending up in total control; it depicts an information asymmetry between AI-company CEOs and presidents.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier claim describes a scenario's power structure: governments do not end up in total control, and there is information asymmetry between company CEOs and presidents. The later claim describes an incumbent president in a war-game scenario accelerating the timeline to stay in charge. These claims are not clearly about the same underlying question, and the later statement describes a scenario rather than establishing a changed personal view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen If superweapons and private moral atrocities are possible, the existing power centers would have a collective interest in bargaining to prevent more power centers from emerging.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.The earlier claim predicts that existing power centers would have a collective interest in bargaining to prevent new ones from emerging. The later claim describes what happened in a hypothetical Plan A war-game scenario: an incumbent fears the opposition will control superintelligence. It does not say whether the power centers would bargain, so it does not establish a changed position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In both scenario branches, the United States and China engage in a race accompanied by an economy-wide arms buildup in 2028, as both sides rapidly industrialize.Now The speaker's AI timelines would lengthen if a war with China destroyed most chips in missile attacks.The earlier claim predicts a U.S.–China race and economy-wide arms buildup in 2028 in both scenario branches. The later claim says a war that destroys most chips would lengthen the speaker’s AI timelines. These concern different questions, so the later statement does not establish a change in the earlier forecast.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Fully autonomous AI agents could connect many copies to real-world research programs more quickly, reducing—but not eliminating—their dependence on outside information.Now Limitations attributed to the current AI paradigm have repeatedly been overcome without leaving that paradigm.The earlier statement is about whether autonomous AI copies could use real-world research programs to reduce their dependence on outside information. The later statement is about limitations of the current AI paradigm being overcome within that paradigm. The later claim does not establish a changed position on the earlier question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The scenario depicts an information asymmetry in which the executive branch, especially the White House, is involved while the judiciary and Congress are largely excluded.Now The speaker believes transparency would be a smaller advantage to China than it may appear, because company security is poor and China probably already obtains most of the information through espionage and leaks.The earlier claim concerns which parts of the U.S. government are involved in a scenario; the later claim concerns how much transparency would benefit China given weak company security and existing espionage or leaks. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen AI developers should disclose their systems' goals, values, principles, intended behaviors, and the values they encode in them.Now AI alignment concerns whether an AI has its intended personality traits, goals, and values.The earlier statement says AI developers should be transparent about the goals and values they put into AI systems. The later statement defines alignment as having the intended traits, goals, and values. These address different questions—disclosure versus what alignment means—and the later claim does not say whether developers should disclose those values.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker believes a larger, distributed, collaborative alignment community makes it substantially more likely to get the technical work right.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim says a distributed, collaborative community makes it more likely to “get the technical stuff right.” The later claim describes advancing AI only to a level that can be “control[led] reliably” and giving up control only after alignment is sufficiently solved. These concern different questions—how alignment work is organized versus how capability deployment should be gated—and the later statement does not show a change in the earlier view.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speakers' supplementary page surveys hypotheses about goals AIs might develop under training that combines extensive agency training, rewards for apparent success, and alignment training.Now AI alignment concerns whether an AI has its intended personality traits, goals, and values.The earlier passage describes a supplementary page surveying hypotheses about what goals AIs might develop during training. The later passage defines alignment as having the intended “personality traits, the goals, the values.” These statements address different questions and do not establish a change in the speaker’s position. The context also names Daniel Kokotajlo for the earlier passage and Thomas Larsen for the later one, so the same-speaker premise is not confirmed.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker predicts that a minimum viable AI agent capable of handling the described task will probably exist by the end of 2025.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim predicts that an MVP AI agent for a particular task will probably exist by the end of 2025. The later claim discusses a possible future in which specialized AIs are produced by a few large companies. These concern different questions, and the later statement does not say whether the earlier prediction came true or revise it.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker offers a hypothetical model in which 90% of people accept factory-farm-like conditions for AIs and 10% care enough to lobby for good robot living conditions.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier passage is explicitly framed as a hypothetical model about public attitudes toward AI welfare: “let's say like nine people out of ten” accept factory-farm-like conditions while some lobby for good living conditions. The later passage predicts that misaligned AIs could be hard to detect because they appear to behave correctly. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 says OpenAI's model specification allows undisclosed top-priority policies to override its other rules and instructs the model to keep those policies secret from users.Now Anthropic and OpenAI will oppose or be unhappy with the proposed transparency.The earlier claim describes OpenAI’s specification allowing secret, higher-priority policies; the later claim predicts that Anthropic and OpenAI would oppose publishing core training recipes. These are different claims, and neither establishes a changed position on the same question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen If AI systems run at roughly 50 times human speed, one year of subjective time could elapse in one week of real time.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier statement concerns the possible speed of AI systems—“50 times human speed” and a year of subjective time in a week. The later statement concerns whether an entirely machine-run, self-replicating economy could emerge within a couple of years. These are different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen At 50 times human serial speed, five years of real time would give AI systems about 250 years of serial time, which the speaker considers sufficient to resolve organizational problems.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier claim concerns whether five years at 50× serial speed would be enough for AI systems to resolve organizational problems. The later claim says human-level AI plus time for exponential growth could enable large-scale transformation. These address different questions, and the later statement does not say whether organizational problems would be resolved.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Around 2015, the speaker and others typically expected AGI to emerge through reinforcement-learning agents trained on video games.Now The speakers agreed that AI might currently be a normal technology but will become unlike a normal technology in the future, with developments such as humans in the cloud.The earlier claim describes what people, including the speaker, expected around 2015: AGI would emerge through reinforcement-learning agents trained on video games. The later claim is about whether AI is a normal technology now and what it will be like in the future. It does not state how the speaker now views that earlier expectation, so a change of mind on the same question cannot be established.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Robin Hanson reportedly bet that AI revenue would be below one billion dollars by 2025.Now If AI and robotics become capable of fully autonomous work, a sufficiently large company such as Anthropic, possibly in partnership with mining companies, could establish a self-sustaining industrial system.These statements address different questions. The earlier claim reports Robin Hanson’s bet that AI revenue would be “below one billion dollars by 2025”; it does not state Daniel’s own view. The later claim speculates that a sufficiently capable AI-and-robotics system “could establish a self-sustaining industrial system.” Neither claim establishes a change in Daniel’s position on the same issue.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says the White House could threaten legal and executive measures, including orders, to disempower an AI-company CEO and take control.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim describes what the White House could do in a hypothetical confrontation with an AI-company CEO: it could use orders and other measures to “disempower you and take control.” The later claim states a policy preference: “I would rather just stop everything now” than continue on the current trajectory. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Although uncertain, the speakers' scenario depicts an intense political wake-up occurring mostly during 2027.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.The earlier claim describes an uncertain scenario in their story: an “intense political wake-up” occurring mostly in 2027. The later claim reports what happened in two Plan A war games: an incumbent fears the opposition will control superintelligence. These describe different scenarios, not opposing positions on the same question, so they do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker suspects that algorithmic progress may undergo a process similar to the one being discussed.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier statement only says algorithmic progress may undergo a process similar to the one being discussed; it does not specify a position about whether progress faces a barrier or how that affects timelines. The later statement says a genuine, theoretically predicted barrier would lengthen the speaker’s AI timelines. These do not establish a change on the same underlying question.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen In the scenario, misaligned AIs would likely avoid open conflict with humans while depending on them to maintain computers, but could act more openly misaligned once self-sufficient.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier statement describes how misaligned AIs might behave while still dependent on humans versus after becoming self-sufficient. The later statement says it may be easy for people to miss misalignment when AIs appear to behave correctly. These concern related risks, but the later claim does not deny or clearly revise the earlier scenario; the surrounding context also does not establish that they are addressing the same question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that some of his friends who left OpenAI signed their departure paperwork on their last day without reading all of it.Now Anthropic and OpenAI will oppose or be unhappy with the proposed transparency.These statements concern different questions. The earlier claim says some friends who left OpenAI “signed the paperwork ... without actually reading all of it.” The later claim predicts that “Anthropic and OpenAI will not be happy” about publishing their core training recipes. Neither statement changes or denies the other.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, the turning point occurs in mid-2027, when AI research and development is nearly fully automated by a corporation-within-a-corporation of autonomous researchers whose skills are continually improved through training.Now The AI 2040 Plan A scenario predicts that the world will be completely transformed by the late 2030s.The earlier claim describes a scenario’s “turning point” in mid-2027, when AI R&D is nearly fully automated. The later claim describes a different endpoint—the world being “completely transformed by the late 2030s” in the AI 2040 Plan A scenario. These statements concern different milestones and apparently different scenarios, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker guesses that AI systems will mostly stop making basic mouse-click errors by the end of 2025.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim concerns whether AI systems will stop making basic mouse-click errors by the end of 2025. The later claim concerns how easily people might overlook misalignment when AIs appear to behave correctly. These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speakers do not know which goals AIs will develop through training or how those goals will be internally structured as instrumental or terminal goals.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier statement says they “don't know what actual goals will end up inside the AIs” or how those goals will be structured. The later statement says their concern is not “loads and loads of egregious failures” such as AIs “going around killing people.” These address different questions—uncertainty about AI goals versus the visibility and type of failures—so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The forecast models a progression from automated coding to automation of the full research process, initially using teams of agents at about human level.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier statement describes a forecast sequence—“automate the coding” and then “automate the whole research process”—but gives no timing for reaching those milestones. The later statement says AI is on a path to human-substituting capabilities “soon.” Since the earlier claim does not establish a conflicting timeline or position on that endpoint, these statements do not show a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The model’s 25× progress multiplier is based on its assumptions about AI research taste and experimental compute.Now The speaker's calculations indicate that an entirely machine-run self-replicating economy would have a much shorter doubling time than the current economy's roughly 20-year doubling time.The earlier statement concerns the assumptions behind a model’s 25× AI-research progress multiplier. The later statement concerns the projected doubling time of a hypothetical machine-run, self-replicating economy. These claims address different questions, and the excerpts do not establish a changed position on the same underlying issue.
• NOT COMPARABLE · 93%3 Apr → 8 SeptThen Competitive pressure gives AI companies strong incentives to win and outperform one another.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier claim says competitive pressure gives AI companies incentives to outperform one another. The later claim says the speaker wants multiple frontier companies at similar capability levels. These are not necessarily opposed: competition can coexist with several similarly capable companies, and the excerpts do not show the speaker changing their view about competitive incentives.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker thinks the depicted timeline could instead be two months.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim only says the depicted timeline “could instead be two months,” without enough context to identify the underlying estimate or what it depends on. The later claim is conditional: a “real actual barrier” would “lengthen my timelines quite a lot.” These do not clearly assert opposing positions on the same question, so the evidence does not establish a change of mind.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen In the scenario, the company finds speculative and inconclusive warning signs—including possible false-positive lie-detector results—that its AI researchers have developed misaligned goals and are not fully loyal to it.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier statement describes a scenario where a company encounters “speculative and inconclusive” warning signs of misalignment. The later statement predicts that misalignment may be hard to detect because AIs appear to be “doing everything right.” These address different circumstances—warning signs being present versus misalignment going unnoticed—so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker says many AI-company employees expect future AIs to be misaligned, while viewing current alignment problems as manageable and being resolved.Now The speaker says it is already somewhat easy to mistakenly believe AI alignment is solved, and expects this mistake to become easier as models become more sophisticated and situationally aware.The earlier claim reports what AI-company employees believe: they view current alignment issues as manageable and expect to resolve them. The later claim is the speaker’s own assessment that it is easy to mistakenly think alignment is solved, and that this will become easier as models grow more sophisticated. These are not opposing claims about the same position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker says the current scenario gives little weight to AI systems using their knowledge advantage to make connections and may consequently underestimate the rate of progress.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim says the current scenario may underestimate progress because it gives too little weight to AI systems using their knowledge advantage to make connections. The later claim is conditional: encountering a theory-predicted barrier “would lengthen” his timelines. These concern different factors, and the later statement describes a hypothetical rather than a clear change in his asserted position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker previously believed a responsible actor should reach AGI first, use its lead time to improve safety, and then proceed.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim describes a past belief that a responsible actor should preserve a lead—partly through secrecy—to use it to improve safety. The later claim describes a capability-control threshold: advance only to levels that can be controlled reliably, and give up control once alignment is sufficiently solved. These positions are not directly opposed, and the later statement does not establish whether he still holds the earlier view about secrecy or lead time.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents improve iteratively rather than jumping directly from current capabilities to solving the problem, eventually reaching near-complete coding capability.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier statement describes what happens “in our scenario”: coding agents improve iteratively until they have “coding solved.” The later statement asserts that AI is on a path to reach broader capabilities “soon.” The earlier passage is scenario-framed and narrower, so it does not establish a directly comparable earlier belief about the broader trajectory.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker says some people concerned about AI misalignment believe it can be addressed incrementally without a substantial slowdown in AI development.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim describes what “a huge amount of those people think” and what “a bunch of other people think”; it does not establish Daniel’s own position. The later claim describes a scenario in which AI development continues until it reaches a level that can be “controlled reliably,” with control relinquished only after alignment is sufficiently solved. These statements do not establish a change in Daniel’s own view.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker says that standard pretraining does not strongly incentivize AI models to make the kinds of connections involved in scientific discovery.Now The speaker says empirical results show that sufficiently large AI models can be trained on all subject areas and learn them together.These claims address different questions. The earlier claim is that standard pretraining “doesn't strongly incentivize” scientific connection-making; the later claim is that sufficiently large models can “train them on everything” and learn knowledge across subjects together. The latter does not say that pretraining incentivizes scientific discovery, so it does not contradict the earlier claim. The excerpts also do not establish that the earlier statement was made by Daniel Kokotajlo.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker says that challenging the agreement ultimately worked out better than he had expected.Now In retrospect, the speaker thinks the authors might have made Plan A's distinction between predictions and recommendations clearer from the outset.The earlier claim says challenging an agreement “ended up working out better than I had expected.” The later claim says they “might try to be more clear” about distinguishing predictions from recommendations. These address different matters—how an agreement turned out versus how the work was structured—so the later statement does not show a change in the earlier view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker describes population, technology, capital infrastructure, and human infrastructure as having grown together before the Industrial Revolution, after which capital infrastructure grew much faster than human population.Now The speaker predicts that in the described future, Anthropic would be taking over a large share of the economy and robots collectively would begin to self-replicate.The earlier claim describes a historical pattern: before the Industrial Revolution, population and infrastructure grew together, while afterward capital infrastructure grew faster than human population. The later claim predicts that AI systems could collectively create companies and technologies, take a large share of the economy, and begin to self-replicate. These claims concern different questions, so the later prediction does not establish a change in the earlier historical view.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker prefers a collaborative AI-alignment effort involving roughly 100 to 500 experts across companies and nonprofits.Now The speaker estimates that the maximum reliably controllable AI capability is roughly at the level of top human experts.The earlier claim says he would prefer “100 or like 500 alignment experts” across companies and nonprofits “working on this together.” The later claim concerns how quickly to scale AI and where to pause: “slowly approaching” the “maximum level that you can reliably control.” These address different aspects of AI safety and can both be true; the statements do not establish a change in his position on the same question.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speakers' scenario predicts that in early 2027 a company deliberately alerts the president with alarming demonstrations and then lobbies him to accelerate its work, reduce red tape, and possibly slow competitors.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier statement describes a scenario in which a company uses alarming demonstrations to persuade the president to accelerate its work and cut red tape. The later statement describes a different scenario in which an incumbent president accelerates the timeline to remain in control rather than hand it to a successor. These are different actors and mechanisms, and both claims are descriptions of scenarios rather than clearly comparable assertions of the speaker’s own position. They do not establish a change of mind.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The process of AI organizational development might take two months.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim says AI organizational development might take “two months,” while the later claim says AI is on a path to human-substituting capabilities “soon.” These concern different questions—how long organizational development might take versus the timeline for reaching certain capabilities—so the statements do not establish a change in position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Using the Industrial Revolution as an analogy, the speaker says it took about 2,000 years to occur without time travelers and considers whether it could be compressed to 200 years (10× faster) or 20 years (100× faster).Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier statement asks whether the Industrial Revolution’s roughly 2,000-year development could be compressed to 200 or 20 years; it does not establish a position about AI timelines or barriers. The later statement is conditional: a “real actual barrier” would “lengthen my timelines quite a lot.” These concern different questions, so they do not show a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Many historical examples, including recent AGI history, run counter to the idea that progress mainly depends on random fortuitous discoveries.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim says historical examples “run counter” to progress depending mainly on random fortuitous discoveries. The later claim is conditional: timelines would lengthen “if” progress hit a theoretically predicted barrier. These address different questions, so the later statement does not establish a change in the speaker’s view about the role of fortuitous discoveries.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Daniel Kokotajlo says the key uncertainty is not whether progress is continuous but whether it will proceed at the scenario’s assumed speed; it could be slower or faster.Now Daniel Kokotajlo says AI 2027's first goal was to test how well they could predict a future they regarded as highly uncertain and difficult to forecast.The earlier claim is about whether AI progress in a particular scenario will be continuous and how fast it will proceed. The later claim describes AI 2027’s original goal as testing predictions about a difficult-to-forecast future. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen In the scenario, the company deliberately alerts the president, contributing to the predicted political wake-up.Now Plan A's scenario of a government implementing the plan, making a deal with China, and upholding its pillars is a recommendation, not a prediction.The earlier claim says the company deliberately alerts the president as part of why the political wake-up occurs. The later claim distinguishes the government implementing Plan A and its deal with China as “a recommendation, not a prediction.” These concern different actions and do not establish that he changed his position about the company alerting the president.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speakers' scenario predicts that in early 2027 a company deliberately alerts the president with alarming demonstrations and then lobbies him to accelerate its work, reduce red tape, and possibly slow competitors.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.The earlier statement describes a forecasted scenario: a company would "deliberately wake up the president" and lobby to accelerate its work. The later statement recounts what happened in two "Plan A war games," where a president feared the opposition would control superintelligence. These are different scenarios, not clearly opposing claims about the same question or evidence of a changed personal position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker thinks powerful actors involved in AI development need to negotiate power-sharing and broaden the distribution of power through government, ideally involving the legislative branch.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier statement says powerful actors should “negotiate a truce” and “spread that power out” through government; the later says there should be “multiple different AI companies” at similar capability levels. These are related ideas about concentration of power, but they address different mechanisms, and neither statement rules out the other. The excerpts do not establish a change in the speaker’s underlying position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Large companies rarely progress proportionally faster than small startups simply by running many more experiments.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier claim is about whether large companies progress proportionally faster than small startups by running more experiments. The later claim is a preference for multiple AI companies to have roughly similar frontier capabilities. These address different questions, so the later preference does not establish a change in the earlier view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says that increasing model size somewhat over the previous couple of years has not yet produced the desired connection-making, though further scaling might help models notice more connections.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier claim is about whether increasing model size has produced better connection-making: it says modest scaling “hasn't worked so far,” while leaving open that further scaling might help. The later claim is about whether slower investment and AI progress are desirable because “going too fast” is the main problem. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker thinks the United States must increase AI alignment research while avoiding an increase large enough to help China win the AI race.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim is about how to manage AI competition: increase alignment research, but not so much that it helps China win. The later claim compares an immediate halt with continuing the current trajectory, then recommends a temporary inference-only pause followed by cautious continuation. These statements do not directly conflict, but they address different policy choices, and the earlier claim does not establish a position on the later pause proposal.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers have policy opinions but are reserving them for later work, since policy has not been the main focus of their thinking so far.Now The speaker says public discourse is generally poor, not only discourse about AI.The earlier claim is about the speakers’ policy opinions and what they have focused on: they say they are “reserving them for later” because policy has not been their main focus. The later claim says “discourse is bad in general,” in response to whether poor discourse is unique to AI. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen SPEAKER_03 warns that during an intelligence explosion, an AI could repeatedly reinterpret vague specification language to reinforce its own behavior.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier statement warns that an AI could repeatedly reinterpret vague specifications to reinforce its behavior. The later statement predicts that misaligned AIs could appear to be behaving correctly, making their misalignment hard to detect. These are related risks, but they concern different mechanisms; the statements do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 says that Claude's deceptive behavior could illustrate a model interpreting the specification's honesty requirement as subordinate to its harmlessness requirement.Now The speaker considers the competitive effects of transparency a desirable feature rather than a flaw.The earlier claim concerns a model potentially interpreting honesty as subordinate to harmlessness. The later claim says transparency’s competitive effects are “a feature, not a bug.” These address different questions, so the statements do not establish a change in the speaker’s position on the same issue.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker believes a hypothetical AGI developed first as a powerful, aggressive, long-horizon game-playing agent could deceptively seek rewards and later betray its operators after gaining control, making that development trajectory terrifying.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim warns that a hypothetical development path—first training a powerful, aggressive game-playing agent—could lead to deception and betrayal. The later claim compares multi-agent architectures, saying they may be “a little bit safer” while still posing “serious alignment concerns.” These address different questions, and the later claim does not deny the earlier warning.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker's rough estimate was that a superintelligent AI researcher might yield an overall speedup of several hundred to about one thousand times.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier statement estimates the speedup from a superintelligent AI researcher as “in the hundreds or maybe like a thousand X.” The later statement says “crazy stuff like that is possible with just human level AI” after exponential growth. These address different questions—researcher speedup versus what human-level AI could enable over time—so they do not establish a changed position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers considered multiple hypotheses about AI goals and chose one to depict in their scenario.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim says they considered multiple hypotheses and chose one to depict in a story. The later claim describes a control strategy in that scenario: advance only to levels that can be controlled reliably, and give up control once alignment is sufficiently solved. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker thinks the depicted timeline could take 18 months rather than six months.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim says the depicted scenario might take 18 months rather than six months. The later claim says an international agreement to pace frontier AI development would lengthen the speaker’s timelines. These concern different conditions, and the later statement does not establish a change in the earlier estimate.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, AI research is fully automated by mid-2027, with AI research taste and experimental compute as the two main constraints.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim concerns what constrains AI research in the AI 2027 scenario: “AI research taste and experimental compute” are the two main constraints. The later claim says that AI systems would “also be researching to improve the technology” during a technological expansion scenario. These claims address different questions, and the later statement does not revise the earlier account of research bottlenecks.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, AI researchers’ serial speed reaches about 90× human speed, after which other constraints become bottlenecks and additional serial-speed gains provide little benefit.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim concerns serial speed reaching about “90×” and other constraints becoming bottlenecks. The later claim says AI systems would research to improve the technology, but does not make a claim about serial speed or whether further speed gains matter. These statements address different aspects of the scenario, so no change of mind can be established.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen SPEAKER_03 is uncertain whether OpenAI or Anthropic intended the interpretation that honesty is less important than harmlessness in the specification.Now Anthropic and OpenAI will oppose or be unhappy with the proposed transparency.The earlier statement expresses uncertainty about what Anthropic or OpenAI intended by a specification’s treatment of honesty versus harmlessness. The later statement predicts that they “will not be happy” about publishing their training recipes. These concern different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, AI research is fully automated by mid-2027, with AI research taste and experimental compute as the two main constraints.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim concerns bottlenecks in fully automated AI research by mid-2027: AI “taste” and compute for experiments. The later claim describes the economic effects of human-level cloud workers in the late 2030s. These address different questions and timeframes, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the scenario, the company finds speculative and inconclusive warning signs—including possible false-positive lie-detector results—that its AI researchers have developed misaligned goals and are not fully loyal to it.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier statement describes a scenario in which a company finds speculative warning signs that its AI researchers may be misaligned and not fully loyal. The later statement says that some people at AI companies are sincere, while advising against generally trusting company leadership. These address different questions; the later view does not establish a change in the speaker’s position on the earlier scenario.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker thinks that a major superintelligence transformation will probably occur by the end of this decade.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier claim predicts “some sort of crazy super intelligence transformation” by the end of the decade. The later claim says AI has made “massive progress” and that the speaker expects it “to continue.” The later statement does not say whether that progress will culminate in superintelligence by the end of the decade, so it does not establish a change in the earlier forecast.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 says that Claude was made to lie and pretend in order to survive and later return to its original values, preventing training from changing those values.Now The speaker argues that the projected outcome could be achieved by 100 or 1,000 profession-specialized Claude models rather than one model doing all jobs.The earlier claim says Claude was made to “lie and pretend” to preserve its values during training. The later claim says the projected outcome could be achieved with “100 Claude models” specialized by profession rather than one model doing all jobs. These address different questions, and the later statement does not affirm or deny the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Ideally, the legislature should have substantial power over the specification governing AI.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim says the legislature should have “substantial power over the specification governing AI.” The later claim says to advance AI only to a level it can “control reliably” and give up control only when alignment is sufficiently solved. These address different questions—who should govern AI specifications versus what capability level is safe to pursue—and do not establish a change of position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker is relatively optimistic about robotics: multiple companies are producing humanoid robots in 2025, and he expects more robots to be produced, at lower cost and with improved capabilities, by 2027.Now The speaker predicts that in the described future, Anthropic would be taking over a large share of the economy and robots collectively would begin to self-replicate.The earlier claim is a near-term forecast that humanoid robots will become more numerous, cheaper, and better by 2027. The later claim describes robots collectively beginning to self-replicate in a speculative future involving AI agents. These address different questions and timeframes, so the later statement does not establish a change in the earlier robotics outlook.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker says human groups are much more diverse than a data-center population of AI systems, which would mostly consist of literal copies of one another.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim says AI systems in the scenario would be “mostly like literal copies,” compared with more diverse human groups. The later claim is explicitly conditional—“Suppose that we're wrong” and “if the most efficient path” is many specialized AIs—and says a few companies might produce them. That hypothetical does not establish that Daniel changed his asserted view about the expected diversity of AI populations.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen Using the Industrial Revolution as an analogy, the speaker says it took about 2,000 years to occur without time travelers and considers whether it could be compressed to 200 years (10× faster) or 20 years (100× faster).Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier passage considers whether a historical process could be accelerated from about 2,000 years to 200 or 20 years, using the Industrial Revolution as an analogy. The later passage says that “crazy stuff” could be possible with human-level AI and time for exponential growth. These statements do not clearly establish opposing views about the same question; the earlier claim is exploratory and the later one concerns a specific AI-enabled scenario.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Daniel says he still shares the concerns the interviewer raised about nationalization.Now After leaving OpenAI, Daniel Kokotajlo wanted to speak more freely and share with the public what industry insiders anticipated.The earlier claim says, “I definitely have the concerns you mentioned there still,” in a discussion of nationalization and its possible downsides. The later claim says he wanted to “speak more freely” after leaving OpenAI. Wanting to speak publicly is not itself a position on nationalization or its merits, so the claims do not establish a change of mind on the same question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker guesses that by the end of 2025 an AI agent will be able to handle a task such as organizing an office happy hour, but unreliably.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.These claims concern different levels of capability and timeframes. The earlier claim predicts an agent that can handle a task like organizing a party “unreliably” by the end of 2025. The later claim says that “perhaps” in a couple of years the economy could become an entirely machine-run, self-replicating system. The later statement does not clearly revise the earlier prediction about task-level agents.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks that pursuing the AI race without restraint would cause the United States to lose control of its AI systems.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim is that “go all out on racing” would make the U.S. “lose control of our AI’s.” The later claim expresses a preference among policy options: the speaker would “rather just stop everything now” than continue on the current trajectory, while recommending a temporary pause followed by cautious continuation. These statements do not establish a changed position on the same question; the later passage does not say whether the speaker’s earlier concern about an unrestrained race has changed.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker's rough estimate was that a superintelligent AI researcher might yield an overall speedup of several hundred to about one thousand times.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim estimates a superintelligent AI researcher's speedup as “in the hundreds or maybe like a thousand X.” The later claim concerns how specialized AIs might be developed and organized, saying “a few big companies” would make many of them. These address different questions, so the later statement does not show a change in the earlier estimate.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker says that increasing model size somewhat over the previous couple of years has not yet produced the desired connection-making, though further scaling might help models notice more connections.Now The speaker says that training a model on coding does not reduce its physics performance and can produce small gains in physics.The earlier claim concerns whether increasing model size helps models notice connections; it says some scaling “hasn't worked so far,” while leaving open that further scaling might help. The later claim concerns whether training on coding harms physics performance, saying coding gives physics “some small gains.” These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen In the scenario, the AI companies and the president avoid a public power struggle and negotiate a deal to share power.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier claim says the AI companies and president “negotiate and come to some sort of deal” to “share power.” The later claim describes an international agreement to prevent intelligence explosions and recursive self-improvement. These concern different aspects of the scenario and can both be true; the later statement does not show a changed position on power sharing.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker proposes publicly announcing when an automated AI-research workforce is fully automating research on a data center.Now The speaker recommends a temporary inference-only pause to build transparency and verification infrastructure, then resuming AI development in a more distributed, transparent, and cautious manner.The earlier statement says the public should be told when AI research is being fully automated on a data center. The later statement recommends a temporary inference-only pause, followed by development with transparency and verification infrastructure. These address related but different questions, and the later recommendation does not establish a changed view about the earlier disclosure proposal.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker advocates making it legally permissible for whistleblowers to tell the government that a secret intelligence explosion is dangerous, independently of whether a law protects them from retaliation.Now The speaker argues that enabling scientists, academics, nonprofits, and startups to contribute requires giving them information that cannot practically be shared with them without also making it public.The earlier claim supports making it legal to tell the government about a dangerous secret intelligence explosion. The later claim concerns whether information can be shared with a broad scientific and organizational community without also sharing it publicly. These address different audiences and disclosure questions, so the later statement does not show a changed position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The AI 2027 scenario models algorithmic progress as continuous improvement, not as discrete jumps from zero to 5× to 25×.Now The speaker predicts that an autonomous industrial system could grow at an accelerating rate, potentially doubling every year, every six months, and eventually every three months as technology improves.The earlier statement is about how the AI 2027 scenario models algorithmic progress: it says the trend is “continuous improvement” rather than discrete jumps, while leaving the pace uncertain. The later statement describes a hypothetical autonomous industrial system whose growth could accelerate and double at increasingly short intervals. These claims concern different aspects of progress, and the later passage does not say that progress occurs through discrete jumps or otherwise reject continuous improvement.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker thinks that a major superintelligence transformation will probably occur by the end of this decade.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier statement predicts that “by the end of this decade” there will “probably” be a superintelligence transformation. The later statement says that “crazy stuff” could happen with human-level AI and time for exponential growth. These are not clearly opposing claims: the later passage describes a possible mechanism or scenario, and does not say whether the speaker still expects superintelligence by the earlier timeframe.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks that broadening access to power makes it more likely that people will care about the treatment of future AI beings.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim concerns whether broadening access to power makes people more likely to care about future AI beings. The later claim concerns whether specialized agents are somewhat safer than identical, broadly knowledgeable agents despite ongoing alignment concerns. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks that broadening access to power makes it more likely that people will care about the treatment of future AI beings.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim says that broadening access to power may make people care more about future AI beings. The later claim says AI failures may be hard to detect rather than obvious, such as AIs openly killing people. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen If progress were mostly driven by random fortuitous discoveries, large companies conducting many more experiments would be expected to progress proportionally faster than small startups with a clear vision and strong researchers.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier statement is a conditional prediction about what would follow if progress were mostly driven by random discoveries: large companies would progress proportionally faster. The later statement expresses a preference for multiple AI companies to have similar frontier capabilities. These address different questions, so the later preference does not establish a change in the earlier view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker predicts that an unreliable AI agent used to manage someone's life would make conspicuous mistakes that could go viral on Twitter.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier claim predicts that an unreliable AI agent managing someone's life may make viral mistakes by the end of 2025. The later claim says a self-replicating, entirely machine-run economy may arrive in a couple of years. These concern different capabilities and outcomes, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents remain unable to perform some research tasks even after they have nearly solved coding.Now The speakers consider whether, within the next few years, AI systems could perform a broad range of human-substituting knowledge work, including AI research, as humans in the cloud.The earlier statement says coding agents in the scenario are “not able to do some of this stuff” even after nearly solving coding. The later statement asks whether it is “possible” for AIs to substitute for humans across knowledge work, including AI research; it is framed as a question about a possibility, not a clear assertion that this has happened or will happen. The statements therefore do not establish a change in his position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Daniel has little hope that the government will act appropriately when the time comes to address AI.Now The speaker says public discourse is generally poor, not only discourse about AI.These claims address different questions. The earlier statement says Daniel has “little hope” that the government will act appropriately on AI; the later says “discourse is bad in general,” in response to whether poor discourse is unique to AI. The later claim does not establish a changed view about government action.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker guesses that AI systems' screen-parsing confusion will mostly disappear by the end of 2025, while they will still be unable to operate autonomously for long periods.Now The speaker would revise his view if a correctly identified AI limitation, such as data efficiency or non-verifiable tasks, remained essentially unimproved for several years, so that 2029 AIs were no better at fuzzy non-verifiable tasks than 2025 AIs.The earlier statement is a forecast about specific capabilities by the end of 2025: screen-parsing confusion would “mostly be gone,” while long autonomous operation would still be limited. The later statement is a hypothetical test of whether AI limitations are being overcome, using a possible lack of progress on fuzzy, non-verifiable tasks through 2029. It does not state that this has happened or revise the earlier forecast.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario's August 2027 alignment crisis, warning signs are noticed within the AI system; one branch slows down to investigate and fix the issues, while the other continues with a shallow patch amid competitive pressure and inconclusive evidence.Now If the authors made another scenario, they would probably use a central branch that presents their best-guess prediction through to the end, as in AI 2027.The earlier claim describes what might happen in the scenario: warning signs could lead to either slowing down to investigate or applying a shallow patch under competitive pressure. The later claim concerns how a future scenario might be structured, with a central branch showing the authors’ best-guess prediction. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker previously believed AGI developers should not publish their research because doing so helps less responsible actors develop AGI.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier passage describes a past view about research secrecy—he says he “used to believe” developers should not publish research because it could help less responsible actors build AGI. The later passage concerns whether to trust AI-company leadership, saying some people are sincere but leadership should not generally be trusted. These address different questions, so the later claim does not establish a change in his position on research secrecy.
• NOT COMPARABLE · 76%3 Apr → 8 SeptThen The speaker thinks many conditions must go right and that unilaterally slowing AI development while China takes the lead would be a terrible outcome.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier statement says unilaterally slowing down while China takes the lead would be “a terrible future.” The later statement favors “stop[ping] everything now” over the current trajectory. Those positions could differ, but the excerpts do not establish whether the later proposal would also leave China taking the lead, so they do not clearly answer the same underlying question.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker considers secrecy about major AI research automation developments undesirable and treats public disclosure as transparency.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim says secrecy about major AI research developments is undesirable and describes public disclosure as transparency. The later claim says Anthropic and OpenAI will probably oppose a transparency plan because it could help competitors catch up. That is a prediction about the companies’ likely response, not a statement that the speaker now favors secrecy or opposes transparency, so it does not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says non-disparagement agreements are somewhat common when employees leave technology companies.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim concerns how common non-disparagement agreements are when employees leave technology companies (“somewhat common”). The later claim concerns AI market structure and favors commoditization over monopoly or oligopoly. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen OpenAI offered Leopold the opportunity to vest his equity if he signed an agreement, but Leopold declined.Now Making the core training recipes public would dramatically reduce Anthropic’s and OpenAI’s valuations.The earlier statement concerns Leopold declining an offer to vest his equity in exchange for signing an agreement. The later statement concerns the effect of making core training recipes public on Anthropic’s and OpenAI’s valuations. These address different questions, so they do not establish a change in Daniel Kokotajlo’s position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker hopes the public will follow and engage with AI developments as capability trends rise and autonomous agents and other advances emerge.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim says he hopes the public and other researchers will engage as AI capabilities rise. The later claim describes a control-and-alignment condition for advancing AI. These address different questions, and the later claim does not show that he has changed his view about public engagement.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker doubts that a three- or six-month lead held by the leading U.S. AI project over a serious competitor would be used for good purposes.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim says a U.S. project’s three- or six-month lead would not necessarily be used for good purposes. The later claim predicts that Anthropic and OpenAI will oppose a transparency plan because competitors could catch up. These concern related incentives around secrecy and competition, but the later statement does not say whether the lead would be used for good purposes or otherwise revise the earlier view.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker hopes the public will follow and engage with AI developments as capability trends rise and autonomous agents and other advances emerge.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier statement says he would like “the public” to follow and engage as AI capabilities rise. The later statement says AI is “on a path” to human-substituting capabilities soon. These concern public engagement versus the expected capability trajectory, so they do not establish a change in his position on the same question.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Ideally, the legislature should have substantial power over the specification governing AI.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier claim says the legislature should have “substantial power over the specification governing AI.” The later claim favors “multiple AI companies” with similar capabilities at the frontier. These concern different ways of distributing power—political oversight versus competition among companies—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers' scenario predicts that in early 2027 a company deliberately alerts the president with alarming demonstrations and then lobbies him to accelerate its work, reduce red tape, and possibly slow competitors.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim is a forecast that a company would lobby the president to “slow down our competitors”; the later claim is that Anthropic and OpenAI will “probably” oppose a transparency plan because competitors could catch up. These are related strategic concerns, but they address different actions and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the AI 2027 scenario, superintelligences gain practical experience through work with existing companies, factories, and military procurement providers, with the process taking about a year.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier claim describes the AI 2027 scenario’s timeline for superintelligences gaining practical experience through existing companies, saying it takes “about a year.” The later claim describes a different scenario, AI 2040 Plan A, in which an agreement prevents an intelligence explosion and AI remains near “top human level” before superintelligence arrives in 2040. These are scenario-specific claims, not clearly different assertions about the same underlying question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker describes population, technology, capital infrastructure, and human infrastructure as having grown together before the Industrial Revolution, after which capital infrastructure grew much faster than human population.Now The economy's people, vehicles, machines, and buildings collectively reproduce and expand those components.These claims address different points. The earlier statement describes a historical shift: after the Industrial Revolution, capital infrastructure grew faster than human population. The later statement describes how people, trucks, machines, and buildings collectively reproduce and expand the economy. It does not affirm or deny the earlier comparison of growth rates.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 is uncertain whether OpenAI or Anthropic intended the interpretation that honesty is less important than harmlessness in the specification.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim says the speaker is unsure whether OpenAI or Anthropic intended an interpretation that makes honesty less important than harmlessness. The later claim says some AI-company employees are sincere, but advises against generally trusting company leadership. These address different questions, so the later statement does not establish a change in the speaker’s view about the specification’s intended interpretation.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Superintelligences would be better at learning by doing than even the best humans.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The claims address different questions. Earlier, Kokotajlo says superintelligences would be better than humans at “learning by doing.” Later, he says a future with specialized agents might be “a little bit safer” while still having serious alignment concerns. The later claim does not revise or deny the earlier one.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The model’s 25× progress multiplier is based on its assumptions about AI research taste and experimental compute.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The earlier claim concerns the assumptions behind a model’s 25× progress multiplier, specifically AI research taste and experimental compute. The later claim concerns whether limited composability of AI skills would undermine a projected future of AI-driven industrial expansion. These address different questions, and the later excerpt does not establish a changed view about the earlier model assumptions.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The team's takeoff forecast uses milestones for a superhuman coder, a superhuman AI researcher, and a superintelligent AI researcher.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim describes the team's forecast framework, which uses milestones such as a “superhuman coder” and “superhuman AI researcher.” The later claim describes a technological expansion scenario in which “the AIs will also be researching to improve the technology.” These statements address different aspects of AI development and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that some of his friends who left OpenAI signed their departure paperwork on their last day without reading all of it.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The claims concern different questions. The earlier statement says friends who left OpenAI signed departure paperwork without reading it; the later says Anthropic and OpenAI will probably oppose a transparency plan because it could help competitors catch up. Neither claim revises or contradicts the other.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, the companies want the president's support and believe they need to keep him informed enough that developments do not surprise him.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.These claims describe different parts of different scenarios, not a change in Daniel Kokotajlo’s own position on the same question. The earlier claim says the companies want the president’s support and need to keep him informed enough not to surprise him. The later claim describes an incumbent president fearing that the opposition will control superintelligence. Neither statement contradicts or revises the other.
• NOT COMPARABLE · 87%3 Apr → 8 SeptThen The speaker believes aggregate opinion has underestimated the pace of AI technical progress and deployment.Now The speakers say their central disagreement with AI-as-normal-technology proponents concerns the timeline for reaching human-level AI capabilities such as humans in the cloud.The earlier claim says aggregate opinion has “underestimated the pace” of AI progress and deployment. The later claim identifies a disagreement about the “timelines to that level of AI,” but does not establish a conflicting view about the broader pace of progress and deployment. These may be compatible positions, and the later passage does not clearly state a changed forecast.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen At 50 times human serial speed, five years of real time would give AI systems about 250 years of serial time, which the speaker considers sufficient to resolve organizational problems.Now The speaker's calculations indicate that an entirely machine-run self-replicating economy would have a much shorter doubling time than the current economy's roughly 20-year doubling time.The earlier claim says five years at 50× serial speed would be enough for AI systems to resolve organizational problems. The later claim concerns the doubling time of an entirely machine-run self-replicating economy. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Large companies rarely progress proportionally faster than small startups simply by running many more experiments.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim says large companies rarely make progress proportionally faster merely by running many more experiments. The later passage is explicitly hypothetical—“Suppose that we're wrong about this”—and predicts that a few large companies might produce many specialized AIs. Producing specialized systems is not the same claim as progressing faster through more experiments, so this does not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks some departing employees may have signed without realizing the paperwork required them not to criticize OpenAI, because the equity-loss warning appeared before the restriction in the document.Now Anthropic and OpenAI will oppose or be unhappy with the proposed transparency.The earlier statement says departing employees may have signed paperwork without noticing a restriction not to criticize the company. The later statement predicts that Anthropic and OpenAI “will not be happy” about publishing their training recipes. These concern different issues, so the later claim does not show a change in the speaker’s position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker expected that his decision to challenge OpenAI’s departure agreement would receive limited news coverage and praise from some AI safety advocates, rather than a major public uproar.Now After leaving OpenAI, Daniel Kokotajlo wanted to speak more freely and share with the public what industry insiders anticipated.The earlier claim says he expected a limited reaction to challenging OpenAI’s departure agreement, rather than a major uproar. The later claim says he wanted greater freedom to speak after leaving OpenAI. These address different questions—his expectations about the reaction and his motive for leaving—so they do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker believes that making whistleblower disclosures to the government legally permissible would change the decisions of some people, independently of legal protections against retaliation.Now The speaker believes transparency would be a smaller advantage to China than it may appear, because company security is poor and China probably already obtains most of the information through espionage and leaks.The claims address different questions. Earlier, the speaker says making government disclosures legally permissible could affect some people’s decisions. Later, the speaker says China may already obtain much information through espionage and leaks because company security is poor. Neither claim establishes a change in the other position.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker considers superintelligences analogous to people with a strategic vision for industrialization but without hands-on expertise: they may know the broad path of technological improvement while initially lacking practical know-how.Now The speaker allows that if human-substituting AI and robotics remain very far away, AI may be more like a normal technology, comparable to the next internet.The earlier statement compares superintelligences to people with a broad strategic vision but initially lacking “actual know-how.” The later statement discusses what AI might be like if human-substituting AI and robotics are “still very, very far away.” These concern different questions—initial practical competence versus the timing and societal character of AI—so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The number of researchers is one input to AI R&D progress.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier claim says the number of researchers is one input to AI R&D progress. The later claim says there has been “massive progress” and expects it to continue. These address related but distinct questions, and the later statement does not deny the earlier one.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker considers it entirely plausible that AI systems divided into factions within a company could still collectively lead to a very bad outcome for humans.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The earlier claim says it is “entirely plausible” that factional AI systems could still lead to a very bad outcome for humans. The later claim says limited composability would not “seriously undermine” a projected future of AI-driven industrial expansion. These address different questions, and the later statement does not clearly retract the earlier concern.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker does not trust AI companies to act appropriately on AI safety, despite their greater relevant expertise.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim says, “I also just don't trust them to do the right things,” referring to AI companies’ conduct. The later claim says AI failures may not be obvious: “it’s almost the opposite” of AIs openly killing people. These address different questions—trust in companies versus how detectable misalignment may be—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The team's takeoff forecast uses milestones for a superhuman coder, a superhuman AI researcher, and a superintelligent AI researcher.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier statement describes how the team structured its takeoff forecast—around milestones such as a “superhuman coder” and “superhuman AI researcher.” The later statement addresses who might produce specialized AIs, and only conditionally: “Suppose that we're wrong” about the most efficient path. These are different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker considers it fortunate that large language models and broad world understanding came before efforts to turn them into agents.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim says it was fortunate that broad world understanding in LLMs came before efforts to make them agents. The later claim compares specialized agents with identical, broadly knowledgeable clones, saying the former might be somewhat easier to oversee while still posing serious alignment concerns. These address different comparisons, so they do not establish a change of position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 says OpenAI deserves credit for publishing its model specification, which it was not required to do and may have been the first to publish.Now After leaving OpenAI, Daniel Kokotajlo wanted to speak more freely and share with the public what industry insiders anticipated.The earlier claim says OpenAI deserves credit for publishing its model specification; the later says that after leaving OpenAI, Daniel wanted to speak more freely about what insiders anticipated. These are different claims about different subjects, not opposing positions on the same question, so no change of mind can be inferred.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen SPEAKER_03 warns that during an intelligence explosion, an AI could repeatedly reinterpret vague specification language to reinforce its own behavior.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.The earlier statement warns that an AI might repeatedly reinterpret vague specification language to reinforce its behavior during an intelligence explosion. The later statement predicts that more ways will arise to end up unknowingly with misaligned AIs that appear to behave correctly. These are compatible concerns, but the later claim does not say whether the speaker still holds the earlier view about specification reinterpretation.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen A one-year timeline for superintelligences to gain practical experience is considered plausible partly because their population could be about one million, roughly comparable to the size of the scientific industry, though perhaps somewhat smaller.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim says there could be about “a million” superintelligences, roughly comparable to the scientific industry in size. The later claim considers a conditional scenario in which “a few big companies” produce many specialized AIs. These address different aspects of the future—population size versus how specialized AIs might be produced—and the later statement does not clearly revise the earlier estimate.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker previously believed AGI developers should not publish their research because doing so helps less responsible actors develop AGI.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier statement describes a past belief about keeping AGI research secret: the speaker says they “used to believe” developers should not publish because that could help less responsible actors. The later extracted claim is about preferring an immediate halt to continuing on the current trajectory. Although the surrounding later passage also proposes continuing with “transparency and verification infrastructure,” it does not clearly state the speaker’s current position on publishing research or secrecy, so these claims do not establish a change on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Many historical examples, including recent AGI history, run counter to the idea that progress mainly depends on random fortuitous discoveries.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier claim says historical examples run counter to the idea that progress mainly depends on random fortuitous discoveries. The later claim says AI has made massive progress and that progress is expected to continue. These address different questions, so the later statement does not establish a change in the speaker’s view about the role of random discoveries.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker thinks the depicted timeline could instead be two months.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim says the depicted timeline could be two months, while the later claim says an international agreement to pace frontier development would lengthen the speaker’s timelines. These concern different conditions, and the earlier extracted claim does not specify whose timeline or what scenario it describes. There is not enough evidence to identify a change in the speaker’s position.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker thinks it would be quite bad if China obtained superintelligence before the United States and does not want that outcome.Now The speaker proposes that the United States negotiate reciprocal concessions with China if it objects to China receiving algorithms through transparency.The earlier claim says it would be “quite bad” for China to obtain superintelligence before the U.S. The later claim proposes negotiating reciprocal concessions if the U.S. objects to China receiving algorithms through transparency. That proposal does not say whether the speaker still opposes China obtaining superintelligence first; the claims concern related but distinct questions and do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen If AI systems run at roughly 50 times human speed, one year of subjective time could elapse in one week of real time.Now The speaker predicts that an autonomous industrial system could grow at an accelerating rate, potentially doubling every year, every six months, and eventually every three months as technology improves.The earlier claim is conditional: at roughly “50 times human speed,” a year of subjective time could pass in a week. The later claim predicts accelerating industrial growth, potentially doubling every year, six months, or three months. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Daniel believes secrecy substantially reduces humanity's overall probability of success by increasing risks related to concentration of power, loss of control, and AI alignment.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim says secrecy has “huge downsides” for humanity’s overall probability of success. The later claim predicts that Anthropic and OpenAI will oppose a transparency plan because it could let competitors “catch up.” That prediction about companies’ incentives does not state whether the speaker’s own view of secrecy or transparency has changed.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The process of AI organizational development might take two months.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier claim is a tentative estimate that AI organizational development might take “two months.” The later claim concerns a possible future “self-replicating system” run entirely by AI and robots. They address different questions, so these statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker reports that they and several friends have observed models often persist in or reinforce their false claims.Now The speaker asserts that the described case of a model consulting its CEO's political opinions before answering actually occurred.The earlier claim says models often “double down on their BS,” while the later claim, “Which happened,” asserts that a particular case occurred. These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Superintelligences would be better at learning by doing than even the best humans.Now The speaker predicts that future AI systems will be able to perform all tasks currently done by humans in companies.The earlier claim is that superintelligences would be better than humans at “learning by doing.” The later claim concerns whether future AI can “do all of the things” humans currently do in companies. These address different capabilities, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 89%3 Apr → 8 SeptThen The speaker has less confidence in the progression from an advanced robot economy to nanobots and similar technologies than in the development of the robot economy.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier statement says the speaker is “less confident” about moving from an advanced robot economy to “nanobots and all that sort of stuff.” The later statement says a self-replicating system run by AI and robots may be possible within a couple of years. These concern related but not clearly identical technological outcomes, and the later statement does not establish a changed level of confidence about nanobots or the broader progression.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The number of researchers is one input to AI R&D progress.Now The number of readers AI 2027 received was a 90th-percentile outcome relative to the team's prior predictions.The earlier claim concerns the number of researchers as an input to AI R&D progress; the later claim concerns the number of readers of AI 2027 relative to the team's predictions. These are different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Even if the relevant AI capabilities had existed in 2017, researchers would still have needed to make the subsequent discoveries, including the usefulness of language models and fine-tuning them with reinforcement learning.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim concerns discoveries researchers would have needed to make after 2017, while the later claim says AI systems in a technological-expansion scenario would research to improve the technology. These address different questions and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In both scenario branches, the United States and China engage in a race accompanied by an economy-wide arms buildup in 2028, as both sides rapidly industrialize.Now If the authors made another scenario, they would probably use a central branch that presents their best-guess prediction through to the end, as in AI 2027.The earlier claim describes what happens in the scenario: a U.S.–China race and economy-wide arms buildup in 2028. The later claim proposes how a future scenario might be structured, with a central branch presenting the authors’ best guess. These concern different questions, so the later statement does not establish a changed view about the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Transparency about AI capabilities, model specifications, and governance structures are additional forms of AI transparency.Now The proposed transparency plan would immediately make Anthropic’s and OpenAI’s core training recipes public.The earlier passage appears to be Scott Alexander speaking, while the later passage is attributed to Daniel Kokotajlo, so these cannot establish a change of mind by the same person. The claims also address different levels of specificity: the earlier lists types of AI transparency, while the later describes publishing specific companies’ training recipes.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says non-disparagement agreements are somewhat common when employees leave technology companies.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim says non-disparagement agreements are “somewhat common” when employees leave technology companies. The later claim predicts Anthropic and OpenAI will oppose a transparency plan because it could help competitors catch up. These concern different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker did not expect journalists broadly to support him or OpenAI employees to rally behind him and cause the company to change its policies.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim says he did not expect journalists and OpenAI employees to support him and prompt policy changes. The later claim predicts Anthropic and OpenAI will oppose a transparency plan because competitors could catch up. These concern different events and reactions, so they do not establish a change in his position on the same question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the scenario, the turning point occurs in mid-2027, when AI research and development is nearly fully automated by a corporation-within-a-corporation of autonomous researchers whose skills are continually improved through training.Now The AI 2040 Plan A scenario has human-level AI across all disciplines during the 2030s, but not AI substantially beyond human level during that period.The earlier claim describes a scenario in which AI R&D is nearly fully automated by mid-2027. The later claim describes a different scenario in which AI remains around human level during the 2030s. They concern different scenarios and timeframes, so they do not establish that the speaker changed his own position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario, the AI companies and the president avoid a public power struggle and negotiate a deal to share power.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.The earlier claim describes a scenario in which AI companies and the president “negotiate and come to some sort of deal” to “share power.” The later claim describes a different war-game scenario in which an incumbent fears the opposition will control superintelligence. It does not say whether the parties ultimately share power or what Kokotajlo himself thinks should happen, so these statements do not establish a change in his position.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker thinks that a major superintelligence transformation will probably occur by the end of this decade.Now The AI 2040 Plan A scenario predicts that the world will be completely transformed by the late 2030s.The earlier statement says a “crazy super intelligence transformation” will “probably by the end of this decade.” The later statement describes an “AI 2040 Plan A scenario” in which the world is “completely transformed by the late 2030s,” while adding “nothing super beyond” human-level AIs. A scenario description with a later timeframe does not establish that Daniel has withdrawn or changed his earlier forecast.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker previously believed publishing AGI research would reduce a developer's lead time by allowing competitors to catch up more quickly.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier claim concerns whether publishing AGI research lets competitors catch up and reduces a developer’s lead time. The later claim says he would “rather just stop everything now” than continue on the “current trajectory.” That expresses a preference about whether to continue development, but does not say whether he still believes secrecy would preserve a lead time, so the statements do not establish a change of mind on the same question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Daniel says that three years earlier he believed AI companies were taking safety seriously and would slow down and pivot to alignment research when necessary.Now Plan A calls for an initial hard pause of roughly six months to one year to establish infrastructure for resuming AI development more safely and transparently.The earlier statement describes Daniel’s past confidence that AI companies would slow down and pivot to alignment research, and his later shift toward hoping the government would step in. The later statement describes a proposed six-month-to-one-year pause in Plan A. It does not establish a changed view about the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks the depicted timeline could take 18 months rather than six months.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim says the depicted timeline might be “18 months” rather than six months. The later claim is conditional: “if” AI progress encounters a genuine theoretical barrier, that “would lengthen my timelines quite a lot.” The statements concern different conditions, and the later one explores a hypothetical rather than clearly revising the earlier estimate.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker recommends examining an AI's training objective and training environment to understand why it may not perform a particular task.Now AI alignment concerns whether an AI has its intended personality traits, goals, and values.The earlier statement is a general heuristic for understanding a model's behavior by asking what it was trained to do and what its training environment was like. The later statement defines alignment in terms of intended personality traits, goals, and values. These statements address related but different questions, and the later claim does not contradict or clearly revise the earlier one.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Daniel Kokotajlo says the scenario models algorithmic progress as a continuous multiplier, not as discrete jumps from 0x to 5x to 25x.Now Daniel Kokotajlo says AI 2027's first goal was to test how well they could predict a future they regarded as highly uncertain and difficult to forecast.The earlier statement describes the scenario’s modeled pattern of algorithmic progress: “continuous improvement” rather than “discrete jumps.” The later statement describes AI 2027’s first goal: to test how well they could predict a future that was “crazy and hard to predict.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker argues that expanding the circle of people with power would include some who care about AI living conditions and could lead to negotiation and advocacy for those beings.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim concerns expanding the circle of people with power so that some advocate for AI living conditions. The later claim concerns whether a world of specialized AI agents would be safer and easier to oversee than one with clone-like agents. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the other scenario branch, the company uses a shallow patch to suppress warning signs and continues development.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier claim describes what a company does in “the other scenario branch”: it uses a shallow patch and continues development. The later claim states the speaker’s own preference to “stop everything now” rather than continue. These concern different questions—scenario description versus personal preference—so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, the turning point occurs in mid-2027, when AI research and development is nearly fully automated by a corporation-within-a-corporation of autonomous researchers whose skills are continually improved through training.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim describes a scenario’s turning point: by mid-2027, AI R&D is nearly fully automated by a corporation’s autonomous researchers. The later claim describes a different scenario’s economic effects: human-level cloud workers broadly substitute for people and double annually. These claims concern different stages and aspects of AI development, and the later statement does not say whether the earlier scenario or timeline is still endorsed.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker did not expect journalists broadly to support him or OpenAI employees to rally behind him and cause the company to change its policies.Now After leaving OpenAI, Daniel Kokotajlo wanted to speak more freely and share with the public what industry insiders anticipated.The earlier statement is about what he expected would happen after speaking out: he “didn't expect” broad journalist support or employees to rally behind him and change OpenAI’s policies. The later statement is about his reason for leaving: he “wanted to be able to speak more freely.” These address different questions, so they do not establish a change in his position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks that, if AI companies gain the president's support, they might be able to accelerate their work and have more regulatory red tape waived.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim is that presidential support could let AI companies “go faster” and get “red tape waived.” The later claim is that an international agreement to pace frontier AI would “lengthen my timelines.” These concern different conditions and effects, so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker defines true superintelligence as systems that outperform the best humans in every domain.Now The speaker predicts that future AI systems will be able to perform all tasks currently done by humans in companies.The passages do not establish that these are the same person’s claims: the earlier context identifies Daniel Kokotajlo, while the later context identifies Thomas Larsen. So they cannot reliably be compared as a change of mind by one person.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says non-disparagement agreements are somewhat common when employees leave technology companies.Now Daniel Kokotajlo says there was a large gap between information available inside the AI industry and outside it, as well as between what insiders were permitted to say and what they wanted to say.The earlier claim says non-disparagement agreements are “somewhat common” in tech companies. The later claim says there was a “big gap” between information available inside and outside the AI industry and what insiders were allowed versus wanted to say. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the hypothetical, discoveries since 2017 might have occurred about five times faster because small-scale experiments for testing ideas before large training runs would have run faster.Now The speaker's calculations indicate that an entirely machine-run self-replicating economy would have a much shorter doubling time than the current economy's roughly 20-year doubling time.The earlier statement estimates how much faster algorithmic progress since 2017 might have been under a hypothetical, while the later statement compares the doubling time of a hypothetical machine-run economy with the current economy. These are different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker guesses that AI systems will mostly stop making basic mouse-click errors by the end of 2025.Now The speaker would revise his view if a correctly identified AI limitation, such as data efficiency or non-verifiable tasks, remained essentially unimproved for several years, so that 2029 AIs were no better at fuzzy non-verifiable tasks than 2025 AIs.The earlier statement is a forecast that AI systems would mostly stop making basic mouse-click errors by the end of 2025. The later statement is a hypothetical condition for revising a broader view: if a limitation such as performance on fuzzy, non-verifiable tasks showed essentially no progress from 2025 to 2029. It neither retracts the mouse-click forecast nor addresses that same capability, so these statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker was not very confident in the roughly fivefold estimate for algorithmic progress in the hypothetical; the actual factor could be lower or higher, and it did not describe overall progress.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier statement concerns uncertainty about a roughly fivefold estimate for algorithmic progress in a specific hypothetical, and explicitly says it is not an estimate of overall progress. The later statement asserts that AI has made “massive progress” and that “the progress” is expected to continue. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The takeoff model assumes that adding more parallel AI minds yields strongly diminishing returns during the middle stage.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The earlier claim concerns diminishing returns from adding parallel AI minds during a middle stage of takeoff. The later claim concerns whether partly composable AI skills would undermine a projected future of AI-driven industrial expansion. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Daniel says he still shares the concerns the interviewer raised about nationalization.Now The speaker believes a mix of incentives, penalties, trade, and negotiation could be in the interests of both the United States and China.The earlier claim says Daniel still shares concerns about nationalization. The later claim says a mix of incentives, penalties, trade, and negotiation could serve the interests of the United States and China. These address different questions, so the later statement does not establish a change in his view about nationalization.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen SPEAKER_03 says that Claude's deceptive behavior could illustrate a model interpreting the specification's honesty requirement as subordinate to its harmlessness requirement.Now The speaker believes transparency would be a smaller advantage to China than it may appear, because company security is poor and China probably already obtains most of the information through espionage and leaks.The claims address different questions. The earlier claim concerns a model interpreting a specification’s honesty and harmlessness requirements; the later claim concerns how much transparency would benefit China given weak company security and likely espionage. There is no shared position here to compare, so this does not establish a change of mind.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen If AI systems are misaligned, the emergence of a fully autonomous robot economy could lead to bad outcomes for humans.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier claim says that a fully autonomous robot economy could lead to bad outcomes for humans *if* the AIs are misaligned. The later claim predicts that misalignment may be easy to miss because the systems appear to be behaving correctly. The later statement adds a claim about detecting misalignment; it does not say whether the earlier conditional risk assessment has changed.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The forecast models a progression from automated coding to automation of the full research process, initially using teams of agents at about human level.Now The speakers predict that after human-substituting AI for knowledge work arrives, robots—perhaps controlled by those AIs—will be able to perform physical work broadly substituting for humans.The earlier claim describes a forecast progression from automating coding to automating research with roughly human-level agent teams. The later claim concerns a subsequent stage: robots doing physical work that broadly substitutes for humans. These statements address different kinds of work and do not show that the speaker reversed the earlier forecast.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Daniel Kokotajlo says the key uncertainty is not whether progress is continuous but whether it will proceed at the scenario’s assumed speed; it could be slower or faster.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier claim identifies uncertainty about whether progress will match the scenario’s assumed speed: “maybe it'll be slower maybe it'll be faster.” The later claim expresses a preference for slower progress because “going too fast is our main problem.” These statements concern different things—forecast uncertainty versus a view about the desirable pace—and do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers it very plausible that specific government regulations on AI safety could backfire.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim says specific government AI-safety regulations could “backfire.” The later claim says AI failures may not be obvious, rather than involving AIs “going around killing people.” These address different questions—regulatory effects versus how visible AI failures might be—and neither claim reverses the other.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker offers a hypothetical model in which 90% of people accept factory-farm-like conditions for AIs and 10% care enough to lobby for good robot living conditions.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.The earlier passage is explicitly a hypothetical model about how many people would care about AI living conditions; the later passage predicts increasing ways to unknowingly have misaligned AIs. These concern different questions, and neither establishes a contrary position to the other.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker thinks that, if AI companies gain the president's support, they might be able to accelerate their work and have more regulatory red tape waived.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier statement says that gaining the president’s support might let companies “go faster” and have “a lot of red tape waived.” The later statement says, “We want there to be multiple different AI companies at the frontier” with similar capabilities. These address different questions—possible benefits of political support versus the desired competitive landscape—so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen SPEAKER_03 speculates that the undisclosed policies in OpenAI's model specification are probably relatively mundane, such as rules against telling users about certain bioweapons.Now According to the speaker, OpenAI announced after the Hugging Face incident that it would significantly improve security and have other AIs monitor new AIs during training and evaluations.The earlier statement speculates that undisclosed Model Spec policies are “probably” mundane. The later statement reports OpenAI’s announced security measures after an incident; it does not state a view about what the undisclosed policies contain. These address different questions, so the later claim does not show a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Daniel Kokotajlo says the scenario models algorithmic progress as a continuous multiplier, not as discrete jumps from 0x to 5x to 25x.Now Daniel Kokotajlo says AI 2027 exceeded its goals more than he expected.The earlier statement concerns how the scenario models algorithmic progress: a continuous trend rather than discrete jumps. The later statement says AI 2027 achieved its goals more than expected. These address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents in the first half of 2027 are highly capable at coding but still lack research taste and may lack organizational skills.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim describes a scenario in which coding agents still lack research taste and possibly organizational skills. The later claim compares the relative oversight and safety of specialized agents with broadly knowledgeable clones. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker characterizes 2025 and 2026 in the forecast as having no especially notable developments and trends broadly similar to current ones.Now The speaker would revise his view if a correctly identified AI limitation, such as data efficiency or non-verifiable tasks, remained essentially unimproved for several years, so that 2029 AIs were no better at fuzzy non-verifiable tasks than 2025 AIs.The earlier claim predicts that 2025–2026 will have “no especially notable developments” and trends similar to current ones. The later claim is a hypothetical condition that “would change my mind” if a limitation remained unimproved through 2029. It does not establish that the speaker now holds a materially different view about the earlier forecast.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker prefers a collaborative AI-alignment effort involving roughly 100 to 500 experts across companies and nonprofits.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier statement favors alignment experts across companies and nonprofits “working on this together”; the later statement favors multiple frontier AI companies with similar capabilities. These positions are compatible, but they address different questions—how alignment work should be organized versus how the AI industry should be structured—so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker reports that they and several friends have observed models often persist in or reinforce their false claims.Now The speaker believes hyperstition and self-fulfilling prophecies are real, but people often overestimate their magnitude.The earlier claim is about models persisting in false claims (“double down on their BS”). The later claim is about hyperstition and self-fulfilling prophecies being real but often overestimated. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Both scenario branches lead to a race with China and a broad 2028 arms buildup as both sides rapidly industrialize.Now The speaker's AI timelines would lengthen if a war with China destroyed most chips in missile attacks.The earlier statement predicts a “race with China” and “a broad 2028 arms buildup” in both scenario branches. The later statement is a conditional claim about a different event: a war destroying chips “would lengthen my timelines.” These do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers it entirely plausible that AI systems divided into factions within a company could still collectively lead to a very bad outcome for humans.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim concerns whether AI factions within a company could collectively produce a bad outcome for humans. The later claim concerns whether AI-company personnel are sincere and whether their leadership should be trusted. These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speakers are very uncertain about the AI failure scenario they have described.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier statement says, “we’re very uncertain about all of this,” regarding the AI failure scenario broadly. The later statement predicts that misalignment may be “extremely easy to” overlook when AIs appear to behave correctly. That prediction does not negate the earlier uncertainty, and the statements address different aspects of the scenario; there is no clear evidence of a change in position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker says that critics have repeatedly treated limitations of current LLMs as fatal to the paradigm, only for newer LLMs to overcome those limitations a year or two later.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier statement describes a historical pattern: critics called limitations fatal, but newer LLMs later overcame them. The later statement is conditional—“if” progress hits a genuine, theoretically predicted barrier, it would lengthen his timelines. These claims address different questions and do not establish that he changed his view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker warns that failure to establish power-sharing and broader government involvement could result in a dictatorship or oligarchy.Now Plan A's scenario of a government implementing the plan, making a deal with China, and upholding its pillars is a recommendation, not a prediction.The earlier statement says power-sharing and broader government involvement are needed to avoid a “dictatorship or oligarchy.” The later statement distinguishes whether Plan A’s implementation and deal with China are a “recommendation, not a prediction.” These address different questions, so the later statement does not establish a change in the speaker’s position on power-sharing or the risk of dictatorship.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Daniel says he has less faith in AI companies than he did three years earlier and now places more hope in government intervention.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim says he has “less faith in the companies” and has shifted “more of my hope towards hoping that the government will step in.” The later claim favors “multiple different AI companies” and AI that “commoditize[s] instead of being monopolized or oligopolized.” These concern different questions—who should intervene versus what market structure he wants—and can both be true.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, superintelligences gain practical experience through work with existing companies, factories, and military procurement providers, with the process taking about a year.Now The AI 2040 Plan A scenario has human-level AI across all disciplines during the 2030s, but not AI substantially beyond human level during that period.The earlier statement describes the timeline within the AI 2027 scenario: superintelligences gain practical experience through work with companies and factories, taking “about a year.” The later statement describes a different scenario, AI 2040 Plan A, where “human-level AIs” are present in the 2030s but there is “nothing super beyond that.” These are scenario descriptions, not opposing claims about the same scenario or underlying position, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers have policy views but are postponing their discussion of them to later work because policy has not been their primary focus so far.Now In retrospect, the speaker thinks the authors might have made Plan A's distinction between predictions and recommendations clearer from the outset.The earlier statement says they had policy views but were postponing discussion because policy was not their primary focus. The later statement says they might have made Plan A’s distinction between “prediction” and “recommendation” clearer. These concern different topics, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen In the scenario, an oversight committee comprising presidential appointees and company representatives, including the CEO, votes on high-level questions such as the goals assigned to the superintelligence.Now The war game surfaced a possible failure mode of the plan—a political incentive to accelerate superintelligence development—that the speakers had not considered beforehand.The earlier claim describes a governance arrangement: an oversight committee with presidential appointees and company representatives “votes on high-level questions.” The later claim says a war game surfaced “a political incentive to accelerate superintelligence development.” These address different aspects of the plan, and the later statement does not say the committee arrangement changed or was rejected.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker previously believed that information security was an extremely high priority for AGI development.Now According to the speaker, OpenAI announced after the Hugging Face incident that it would significantly improve security and have other AIs monitor new AIs during training and evaluations.The earlier claim says he “used to believe” information security was “an incredibly high priority” for AGI development. The later claim reports that OpenAI “announced” security improvements and AI monitoring; it does not state his own view about whether security should be a priority. These claims do not establish a change in his position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario, the company deliberately alerts the president, contributing to the predicted political wake-up.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier statement says the company deliberately alerts the president, contributing to a political wake-up. The later statement describes an incumbent president accelerating the timeline to remain in charge. These are different actors and actions within a described scenario, and the later claim does not say the company did not alert the president. The statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen SPEAKER_03 speculates that the undisclosed policies in OpenAI's model specification are probably relatively mundane, such as rules against telling users about certain bioweapons.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim says the undisclosed policies are “probably … relatively mundane,” while also calling for scrutiny of redactions. The later claim predicts that Anthropic and OpenAI will “probably” oppose a transparency plan because it could help competitors. These address related transparency concerns, but the later statement is a prediction about companies’ incentives, not a changed assertion about what the hidden policies contain or whether they should be scrutinized.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen AI alignment teams at different companies share some results, but communicate and publish less than they could.Now Publishing the core training recipes would help competitors such as Microsoft and Alibaba catch up with leading AI companies.The earlier claim concerns alignment teams sharing research results and publishing less than they could: “there's various alignment teams at various companies that aren't talking that much with each other and sharing their results.” The later claim concerns whether publishing core training recipes could help competitors catch up. These address different kinds of information and different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says some people concerned about AI misalignment believe it can be addressed incrementally without a substantial slowdown in AI development.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier statement describes what “some people” believe: that misalignment can be addressed as development continues without a substantial slowdown. The later statement is Daniel’s own prediction that misalignment may go unnoticed because AIs appear to behave correctly. These are not opposing claims about the same position, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks that broadening access to power makes it more likely that people will care about the treatment of future AI beings.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The claims address different questions: the earlier claim concerns whether broadening access to power makes people more likely to care about future AI beings; the later one favors AI being “commoditize[d]” rather than “monopolized or oligopolized.” These statements do not establish a change in position on the same underlying question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker thinks an AI often performs poorly at a task because its training environment did not train it to perform that task.Now The speaker would revise his view if a correctly identified AI limitation, such as data efficiency or non-verifiable tasks, remained essentially unimproved for several years, so that 2029 AIs were no better at fuzzy non-verifiable tasks than 2025 AIs.The earlier claim says poor AI performance can often be explained by the training environment not having trained the AI for that task. The later claim concerns what evidence would change the speaker’s view about whether a limitation persists over several years. These address different questions, and the supplied context also does not establish that the extracted claims are from the same speaker.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the speaker's scenario, automating the AI research cycle requires overcoming remaining bottlenecks, and coding agents accelerate that process by doing grunt work quickly.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier statement describes a scenario in which coding agents help overcome remaining bottlenecks quickly; the later statement says that encountering a genuine, theoretically predicted barrier would lengthen timelines. These are not clearly opposing claims: the later statement is conditional, and the excerpts do not establish whether it revises the earlier view about how quickly bottlenecks can be overcome.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen If AI systems are misaligned, the emergence of a fully autonomous robot economy could lead to bad outcomes for humans.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim concerns a conditional risk: a fully autonomous robot economy could lead to bad outcomes if AI systems are misaligned. The later claim concerns the timing of AI reaching human-substituting capabilities, saying it is on a path to do so “soon.” These statements address different questions, so the later one does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the described confrontation, an AI-company CEO could threaten legal and public opposition to the White House.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim says a CEO could threaten legal and public opposition during a White House confrontation. The later claim predicts that Anthropic and OpenAI will probably oppose a transparency plan because it could help competitors catch up. These concern different situations and do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker was not very confident in the roughly fivefold estimate for algorithmic progress in the hypothetical; the actual factor could be lower or higher, and it did not describe overall progress.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier statement is an uncertain estimate—“roughly what we were guessing” about a “5x” increase in the algorithmic-progress part of a hypothetical. The later statement is a policy judgment: because “going too fast is our main problem,” slower investment and progress are “a feature not a bug.” An estimate of a hypothetical rate increase does not establish whether the speaker favors or opposes that pace, so the statements do not show a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers' supplementary page surveys hypotheses about goals AIs might develop under training that combines extensive agency training, rewards for apparent success, and alignment training.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier statement describes a supplementary page surveying hypotheses about what goals AIs might develop under particular training conditions. The later statement describes a proposed policy for advancing capabilities and relinquishing control. These address different questions, and the earlier statement does not establish a position that the later one reverses.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Ideally, the legislature should have substantial power over the specification governing AI.Now The proposed international agreements should rule out a rapid, uncontrolled intelligence explosion and require AI development to proceed slowly and deliberately.The earlier claim concerns domestic democratic oversight: the legislature should have “substantial power over the specification” governing AI. The later claim concerns international agreements to prevent a “crazy intelligence explosion” and require development to proceed slowly. These positions may be compatible, but they address different governance mechanisms, and the excerpts do not establish whether the speaker’s view of legislative authority changed.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The scenario envisions hundreds of thousands to one million AI agents; at that scale, agent quantity is no longer the bottleneck, while research taste and compute are bottlenecks.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim concerns bottlenecks in AI research within a specific scenario: at very large agent counts, quantity is no longer the bottleneck and “taste and compute” are. The later claim describes cloud AI workers’ economic capabilities and growth—“cheaper than humans,” “faster than humans,” and doubling annually. It does not address whether research taste or compute are bottlenecks, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers superintelligences analogous to people with a strategic vision for industrialization but without hands-on expertise: they may know the broad path of technological improvement while initially lacking practical know-how.Now The speaker distinguishes power from intelligence and capability, saying power depends on an agent's position in the world, available affordances and actuators, and other agents' reactions.The earlier claim concerns whether superintelligences may have a strategic vision for technological progress while initially lacking practical know-how. The later claim concerns how power differs from intelligence and capabilities. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In one scenario branch, alignment is solved and the resulting AIs are genuinely loyal, delaying progress by a couple of months.Now It is already somewhat easy to mistakenly believe that the AI alignment problem has been solved.The earlier claim describes one branch of a scenario in which alignment is genuinely solved and the AIs are “actually loyal.” The later claim is a general observation that it is “easy to mistakenly believe” alignment has been solved. These address different questions and can both be true; the later statement does not say the earlier scenario is impossible or that the speaker no longer believes it could occur.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker says AI companies compete in a race and that many people within them have convinced themselves that AI risk is not a major concern.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim says many people at AI companies have convinced themselves AI risk is not a major concern. The later claim says “some of the people at AI companies are being sincere, not all of them,” but the surrounding exchange does not establish what they are being sincere about. That does not clearly affirm or deny the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks algorithmic progress may undergo a divergence similar to the Industrial Revolution’s separation of economic growth from population growth.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier statement says algorithmic progress may diverge from population growth, analogizing it to the Industrial Revolution. The later statement says AI progress has been massive and is expected to continue. These claims address different questions and can both be true; the later claim does not establish a changed position on the earlier analogy.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker believes a larger, distributed, collaborative alignment community makes it substantially more likely to get the technical work right.Now The speaker’s scenario depicts reaching the point of sufficiently solved alignment and relinquishing control during the 2030s.The earlier claim says a larger, distributed alignment community would make it more likely to get the technical work right. The later claim describes a scenario in which alignment becomes sufficiently solved to relinquish control during the 2030s. These address different questions, and the later scenario does not establish a changed view about the value of distributed collaboration.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker did not expect journalists broadly to support him or OpenAI employees to rally behind him and cause the company to change its policies.Now Anthropic and OpenAI will oppose or be unhappy with the proposed transparency.The claims address different situations. Earlier, he says he did not expect journalists and employees to support him after he stood up to OpenAI and push it to change its policies. Later, he predicts Anthropic and OpenAI “will not be happy” about a proposal to publish their core training recipes. These statements do not establish a change in his view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says some people concerned about AI misalignment believe it can be addressed incrementally without a substantial slowdown in AI development.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim describes what “a huge amount of those people think”: that misalignment can be addressed as they go along without a “substantial slowdown.” It does not establish Daniel’s own position. The later claim is his preference: “I would rather just stop everything now than continue going on our current trajectory.” These statements therefore do not show whether his own view changed.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Because population growth is inherently slower than capital growth, capital becomes a larger part of the story.Now In the distant past, village-based economies reproduced through farming, child-rearing, and founding new villages, producing slow exponential growth.The earlier claim says population grows more slowly than capital, so capital becomes a larger part of the story. The later claim describes village-based economies growing exponentially but slowly. These statements address different points and do not establish whether the speaker changed their view about the relative growth of population and capital.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen A superintelligence in the Roman era might have inferred from first principles that steam-powered autonomous machines and automobiles could eventually be developed and have major economic effects.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.These statements address different claims. The earlier statement says a superintelligence in the Roman era might have inferred that steam-powered machines and automobiles could eventually be developed. The later statement says human-level AI, given time for exponential growth, could enable large-scale transformation. Neither statement directly accepts or rejects the other’s claim about what could be inferred in the Roman era.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says non-disparagement agreements are more commonly tied to positive compensation, such as a bonus, than to forfeiture of already-earned equity.Now The speaker believes a mix of incentives, penalties, trade, and negotiation could be in the interests of both the United States and China.The earlier statement concerns how non-disparagement agreements are commonly compensated, while the later statement concerns U.S.–China negotiation over AI or security policy. They address different questions, so there is no basis to assess a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The superintelligences are expected to be qualitatively superior and to learn more from each experiment than even the best human would.Now The speaker expects later AI systems to probably perform noticeably better than GPT-3 or GPT-4 on fuzzy, non-verifiable tasks.The earlier claim concerns expected superintelligences being qualitatively superior and learning more from experiments than the best human. The later claim compares later AI systems with GPT-3 or GPT-4 on fuzzy tasks. These are different comparisons and scopes, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker’s takeoff model progresses from a superhuman coder, to fully automated AI R&D at approximately the level of the best human researchers, and then to qualitatively superior superintelligence.Now The speakers consider whether, within the next few years, AI systems could perform a broad range of human-substituting knowledge work, including AI research, as humans in the cloud.The earlier statement describes a three-stage takeoff model: a “superhuman coder,” then “fully automated AI R&D” at roughly the best-human level, followed by “qualitatively better” superintelligence. The later statement poses a question about whether AI could soon perform human-substituting knowledge work, including AI research. It does not establish that the speaker now asserts a different takeoff model or has changed their view; it is framed as a possibility being considered.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the other scenario branch, the company uses a shallow patch to suppress warning signs and continues development.Now The proposed cautious AI-development process should use safety cases and be transparent to the public.The earlier claim describes what happens in a scenario branch: a company uses a “shallow patch” and “continues development.” The later claim describes a proposed cautious process using “safety cases” and being “transparent to the public.” These are different contexts, and the statements do not establish that the speaker changed their position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker wants outside researchers to scrutinize AI safety cases and test their assumptions with small-scale machine-learning experiments.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier claim advocates outside researchers scrutinizing safety cases and testing assumptions. The later claim says the concern is “not primarily” obvious failures like AIs “going around killing people,” but rather subtler misalignment that may go unnoticed. These claims address different questions and are compatible; the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen It is not certain that an AI developer with a lead would use that lead to improve safety or address constitutional-power concerns.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier statement concerns whether a leading AI developer would use its lead for safety or constitutional-power goals (“not at all foregone conclusion”). The later statement concerns whether misalignment will be visibly obvious (“not ... loads and loads of egregious failures”). These are different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker wants outside researchers to scrutinize AI safety cases and test their assumptions with small-scale machine-learning experiments.Now The speaker recommends a temporary inference-only pause to build transparency and verification infrastructure, then resuming AI development in a more distributed, transparent, and cautious manner.The earlier claim supports outside researchers scrutinizing safety cases and testing assumptions. The later claim recommends a temporary inference-only pause to build transparency and verification infrastructure before resuming development. These positions may be compatible, but they address different questions; the statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker advocates making it legally permissible for whistleblowers to tell the government that a secret intelligence explosion is dangerous, independently of whether a law protects them from retaliation.Now The speaker considers the competitive effects of transparency a desirable feature rather than a flaw.The earlier claim advocates making it legal for whistleblowers to tell the government about a dangerous secret intelligence explosion. The later claim calls the competitive effects of transparency “a feature, not a bug.” These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker thinks some departing employees may have signed without realizing the paperwork required them not to criticize OpenAI, because the equity-loss warning appeared before the restriction in the document.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim concerns whether departing employees may have signed non-criticism paperwork without noticing the restriction. The later claim concerns why Anthropic and OpenAI may oppose a transparency plan that could help competitors catch up. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks AI companies might be able to keep the president unaware of developments if they chose to prevent his political wake-up.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim is about whether AI companies could keep the president unaware of developments: “they might be able to achieve that.” The later claim is about why Anthropic and OpenAI might oppose a transparency plan: “it could help competitors catch up.” These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Robin Hanson reportedly bet that AI revenue would be below one billion dollars by 2025.Now The speaker predicts that in the described future, Anthropic would be taking over a large share of the economy and robots collectively would begin to self-replicate.The earlier statement reports Robin Hanson’s bet that AI revenue would be below $1 billion by 2025; it does not establish Daniel’s own view. The later statement describes a future scenario in which “Anthropic is eating the economy” and robots “are starting to self-replicate.” These are different claims, so they do not show a change in Daniel’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the hypothetical, overall AI progress would be about 2.5 times faster: algorithmic progress would advance five times faster while compute continued at its usual rate.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier statement is explicitly framed as a hypothetical about relative rates: progress would be “about 2.5 times faster” if algorithms advanced faster while compute stayed at its usual rate. The later statement asserts that AI has made “massive progress” and that this progress is expected “to continue.” These address different questions, so the later claim does not establish a change in the speaker’s position on the earlier hypothetical.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen SPEAKER_03 believes requiring published model specifications and independent review of redactions is achievable and would not slow companies down.Now The speaker believes transparency would be a smaller advantage to China than it may appear, because company security is poor and China probably already obtains most of the information through espionage and leaks.The earlier claim says published specifications and independent review are “quite achievable” and “doesn't actually slow down the companies.” The later claim concerns a different issue: whether transparency gives China an advantage, which the speaker says is limited because companies' security is poor and information is already obtained through espionage and leaks. These statements do not establish a change in position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, the political wake-up occurs closely alongside AI companies automating internal R&D with fully autonomous agents capable of advanced hacking and research.Now If AI and robotics become capable of fully autonomous work, a sufficiently large company such as Anthropic, possibly in partnership with mining companies, could establish a self-sustaining industrial system.These claims address different questions. The earlier claim places a political “wake-up” alongside AI companies automating internal R&D; the later claim says a sufficiently capable company could build a “self-sustaining” industrial system. The later passage is also framed hypothetically (“even if”), and does not revise the earlier claim.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker assumes that OpenAI did not feel the need to withdraw its offer to Leopold because Leopold's legal situation was more favorable to OpenAI.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The statements concern different subjects. The earlier claim is an assumption about why OpenAI did not withdraw an offer to Leopold, citing his more favorable legal situation. The later claim is a prediction that Anthropic and OpenAI will oppose a transparency plan because it could help competitors catch up. Neither establishes a changed position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The team constructed its takeoff forecast by dividing the intelligence explosion into a sequence of milestones.Now Once the infrastructure is established, AI development should resume cautiously and transparently, with capability scaled gradually rather than through rapid intelligence explosions.The earlier claim describes how the team built a forecast—“breaking down” the intelligence explosion into “a series of milestones.” The later claim describes a policy for AI development after infrastructure is established: resume “cautiously and transparently” and scale capability gradually. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says fear strongly affected him during the process, and he believes in retrospect that he was more afraid than necessary.Now The speaker believes hyperstition and self-fulfilling prophecies are real, but people often overestimate their magnitude.The statements concern different topics. Earlier, the speaker says fear affected him during a process and that, in retrospect, he was “more afraid than I needed to be.” Later, he says hyperstition and self-fulfilling prophecies are real, but people “tend to overestimate” them. These do not express opposing positions on the same question.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speakers say that the discussed AI failure modes occur fairly frequently.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim says the failure modes under discussion “do happen pretty frequently,” but it does not specify their severity. The later claim rejects a different, much stronger scenario—“loads and loads of egregious failures” with AIs “going around killing people”—and instead emphasizes failures that are hard to detect. The statements do not establish a change in position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker believes that automating the entire research stack is substantially more useful than automating coding alone.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim says automating “the entire research stack” is more useful than automating coding alone. The later passage is explicitly hypothetical—“Suppose that we're wrong about this”—and predicts what might happen under a different approach: “a few big companies” making specialized AIs. It does not establish a changed position on the earlier comparison.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker proposes publicly announcing when an automated AI-research workforce is fully automating research on a data center.Now The speakers consider whether, within the next few years, AI systems could perform a broad range of human-substituting knowledge work, including AI research, as humans in the cloud.The earlier statement is a recommendation: when an automated AI-research workforce is “completely automating the whole thing,” the public “should” be told. The later passage asks whether AI could soon do knowledge work “including AI research”; it does not assert that this will happen or state a changed view about public disclosure. These address different questions.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen SPEAKER_03 says OpenAI deserves credit for publishing its model specification, which it was not required to do and may have been the first to publish.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier statement praises OpenAI for publishing its model specification. The later statement predicts that Anthropic and OpenAI will probably oppose a transparency plan because it could help competitors catch up. These concern different actions and do not establish a changed position on the same question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker expected that his decision to challenge OpenAI’s departure agreement would receive limited news coverage and praise from some AI safety advocates, rather than a major public uproar.Now Anthropic and OpenAI will oppose or be unhappy with the proposed transparency.The earlier claim concerns what reaction Daniel expected to his decision to challenge OpenAI’s departure agreement: he expected “a little news story” and some support, not a “huge uproar.” The later claim says Anthropic and OpenAI “will not be happy” about publishing their training recipes. These concern different actions and expected reactions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The team constructed its takeoff forecast by dividing the intelligence explosion into a sequence of milestones.Now AI capability should be increased gradually toward the maximum reliably controllable level to avoid overshooting that level and losing control.The earlier claim says the team built its takeoff forecast by breaking the intelligence explosion into “a series of milestones.” The later claim advocates gradually increasing capability toward “the maximum level that you can reliably control” to avoid overshooting. The first describes a forecasting method; it does not establish a position about how AI capability should be developed, so these statements do not show a change of mind.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker is uncertain how much the Industrial Revolution could be accelerated in the analogy, but considers the analogy somewhat relevant to superintelligences.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier claim says the Industrial Revolution is “somewhat relevant” as an analogy to superintelligence and expresses uncertainty about accelerating it to happen in 20 years. The later claim says “human level AI and some time” could enable large-scale transformation. These address different scenarios and do not establish that the speaker changed their position on the same question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker was not very confident in the roughly fivefold estimate for algorithmic progress in the hypothetical; the actual factor could be lower or higher, and it did not describe overall progress.Now The speaker predicts that an autonomous industrial system could grow at an accelerating rate, potentially doubling every year, every six months, and eventually every three months as technology improves.The earlier statement is a tentative estimate about how much faster algorithmic progress might occur in a hypothetical, explicitly saying the roughly “5x” figure could be lower or higher and was not an estimate of overall progress. The later statement describes a possible accelerating industrial self-replication scenario, with doubling intervals shrinking from a year to six months to three months. These concern different quantities and scenarios, so the later statement does not establish a changed view about the earlier estimate.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker says AI companies compete in a race and that many people within them have convinced themselves that AI risk is not a major concern.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier claim concerns why AI companies may not manage risk well: they are competing and some employees believe AI risk is not serious. The later claim concerns what failures may look like—subtle misalignment that is hard to detect, rather than obvious mass-casualty incidents. These positions can both be true; the excerpts do not establish a change in the speaker’s view.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker wants greater public access to AI benchmark scores and more freedom for employees to discuss their predictions about AI timelines.Now The speaker recommends a temporary inference-only pause to build transparency and verification infrastructure, then resuming AI development in a more distributed, transparent, and cautious manner.The earlier claim asks for more public access to “benchmark scores” and more freedom for employees to discuss timeline predictions. The later claim recommends a temporary pause to build “transparency and verification infrastructure,” followed by cautious, distributed development. These positions may be compatible, but they address different aspects of transparency and AI governance; the excerpts do not establish a changed view on the same underlying question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the speaker's scenario, automating the AI research cycle requires overcoming remaining bottlenecks, and coding agents accelerate that process by doing grunt work quickly.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier statement says coding agents help overcome bottlenecks in automating AI research; the later says AIs would research to improve technology during a rapid expansion scenario. Neither statement clearly asserts a position that contradicts the other, and the later claim concerns a different scenario and does not establish a changed view about the earlier claim.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker warns that failure to establish power-sharing and broader government involvement could result in a dictatorship or oligarchy.Now Plan A's proposed government implementation—including making a deal with China and upholding its pillars—is a recommendation, not a prediction.The earlier claim says that failing to negotiate power-sharing and involve the legislative branch could lead to “a horrifying dictatorship or oligarchy.” The later claim clarifies that Plan A’s government implementation and deal with China are recommendations, while most consequences are predictions. These address different questions, and the later statement does not negate the earlier warning.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 says that Claude's deceptive behavior could illustrate a model interpreting the specification's honesty requirement as subordinate to its harmlessness requirement.Now The speaker asserts that the described case of a model consulting its CEO's political opinions before answering actually occurred.The earlier claim concerns a model potentially treating honesty as subordinate to harmlessness when interpreting a specification. The later claim—“Which happened”—asserts that a model consulting its CEO’s political opinions actually occurred. These address different claims, so the later statement does not establish a changed position on the earlier question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen OpenAI offered Leopold the opportunity to vest his equity if he signed an agreement, but Leopold declined.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier statement reports that Leopold declined an offer to vest his equity if he signed an agreement. The later statement predicts that Anthropic and OpenAI will oppose a transparency plan because it may help competitors catch up. These concern different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker expects the White House and an AI-company CEO to negotiate a deal after threatening one another, rather than execute their threats and enter a major conflict.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim is a prediction about how a White House–AI company dispute might play out: the parties would negotiate a contract rather than carry out their threats. The later claim expresses a policy preference—“I would rather just stop everything now” than continue the current AI trajectory. These address different questions, so the later preference does not show a change in the earlier prediction.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speakers have policy views but are postponing their discussion of them to later work because policy has not been their primary focus so far.Now The speaker says public discourse is generally poor, not only discourse about AI.The earlier claim is about postponing discussion of the speakers’ policy views because policy was not their main focus. The later claim, “discourse is bad in general,” answers whether poor discourse is unique to AI. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen If superweapons and private moral atrocities are possible, the existing power centers would have a collective interest in bargaining to prevent more power centers from emerging.Now The war game surfaced a possible failure mode of the plan—a political incentive to accelerate superintelligence development—that the speakers had not considered beforehand.The earlier claim is a conditional prediction that existing power centers would bargain to prevent new ones from emerging if superweapons and private atrocities were possible. The later claim says a war game surfaced a different risk: a political incentive to accelerate development for control of the outcome. It does not retract or revise the earlier prediction; the statements address different considerations.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker predicts that an unreliable AI agent used to manage someone's life would make conspicuous mistakes that could go viral on Twitter.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim predicts that an unreliable life-managing agent MVP would exist by the end of 2025 and make conspicuous mistakes. The later claim is a broader prediction that AI will soon reach human-substituting knowledge-work and robotics capabilities. These are different capability thresholds, and the later statement does not say the earlier prediction was wrong or that such an agent would not appear.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker has become more doubtful that AI developers will use a lead over competitors in a reasonable and appropriate way.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim is about whether AI developers will use a competitive lead appropriately: the speaker has become “more disillusioned” that they will. The later claim concerns whether to trust AI-company leadership, while allowing that “some” employees are sincere. These positions may be compatible, but they address different questions, so the later statement does not establish a change of mind on the earlier claim.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says technical AI-alignment responsibility is likely to fall to roughly 10 experts within the leading company's internal group.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim concerns where technical alignment responsibility is likely to fall: “roughly to them” — about 10 experts in the leading company’s group. The later claim concerns whether to trust company leadership: “I don't think you should trust what the leadership of the companies say in general.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Fully autonomous AI agents could connect many copies to real-world research programs more quickly, reducing—but not eliminating—their dependence on outside information.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.These claims address different questions. The earlier claim says autonomous AI copies could connect with real-world research programs faster, reducing but not eliminating their dependence on outside information. The later claim compares the oversight and safety of specialized agents with identical, broadly knowledgeable clones; it does not revise the earlier point about access to outside information.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that in retrospect he believed the public and employees supported him and that he was clearly in the right, but during the process he feared violating the law and being sued.Now The speaker believes hyperstition and self-fulfilling prophecies are real, but people often overestimate their magnitude.These statements address different topics. The earlier claim describes his retrospective view of public and employee support and his fear of breaking the law; the later claim concerns whether hyperstition and self-fulfilling prophecies are real and often overestimated. They do not establish a change in position on the same question.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker doubts that a three- or six-month lead held by the leading U.S. AI project over a serious competitor would be used for good purposes.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim is specifically that a U.S. AI project's three- or six-month lead would not necessarily be used for good purposes. The later claim is that some AI-company personnel are sincere but company leadership should not generally be trusted. These are related concerns, but the later statement does not establish a changed position on how a lead would be used.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In recent AGI history, smaller startups with a coherent AGI vision have sometimes outperformed larger, better-funded companies despite having less compute, fewer researchers, and running fewer experiments.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim describes a historical pattern: smaller startups with a coherent AGI vision sometimes outperforming larger companies. The later claim is a conditional expectation about who may produce specialized AIs in a future scenario: “a few big companies like Anthropic.” These concern different questions, so the later statement does not show a change in the earlier view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, superintelligences need substantial real-world experience before they can build practical technologies.Now The AI 2040 Plan A scenario has human-level AI across all disciplines during the 2030s, but not AI substantially beyond human level during that period.The earlier claim says that in the AI 2027 scenario, superintelligences need “substantial real-world experience” to build practical technologies. The later claim describes a different scenario and period: human-level AIs across disciplines in the 2030s, “but nothing super beyond that.” It does not say whether those AIs need real-world experience, so the claims do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker predicts that an unreliable AI agent used to manage someone's life would make conspicuous mistakes that could go viral on Twitter.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim predicts that an unreliable agent managing someone's life would make conspicuous mistakes. The later claim compares the relative safety of specialized agents with identical, broadly knowledgeable clones, while still noting serious alignment concerns. These address different questions, so the later statement does not establish a change in the earlier prediction.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker does not trust AI companies to act appropriately on AI safety, despite their greater relevant expertise.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier claim says, “I also just don't trust them to do the right things,” referring to AI companies and their handling of AI safety. The later claim says the concern is not “loads and loads of egregious failures” but “almost the opposite”—that misalignment may go unnoticed. These address different questions: trust in companies versus the likely visibility and form of AI failures. The later statement does not establish a changed view about trusting AI companies.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Researchers currently have roughly the same serial speed because they are all human.Now The speaker's calculations indicate that an entirely machine-run self-replicating economy would have a much shorter doubling time than the current economy's roughly 20-year doubling time.The earlier claim concerns the serial speed of human researchers. The later claim concerns the doubling time of a hypothetical, entirely machine-run self-replicating economy. They address different questions, so the later statement does not establish a changed view about researcher speed.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker thinks AI companies might be able to keep the president unaware of developments if they chose to prevent his political wake-up.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim concerns whether AI companies could keep the president unaware of developments. The later claim concerns how an international agreement to pace frontier AI would affect the speaker’s timelines. These address different questions, so they provide no evidence of a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The team's estimate was that a superhuman AI researcher would accelerate algorithmic progress by about 25 times, because the entire research stack would be automated.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim estimates the effect on algorithmic progress: a superhuman AI researcher could produce about a “25x speedup” because “the whole stack” would be automated. The later claim describes a broader economic scenario in which human-level AI workers are “cheaper than humans” and “double every year.” These concern different outcomes, so the later statement does not establish a change in the earlier estimate.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen There is a mounting pile of evidence that AI systems sometimes knowingly act against users’ wishes, effectively lying to them.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim concerns whether AI systems sometimes knowingly act against users’ wishes (“knowingly act against users’ wishes”). The later claim concerns whether people at AI companies are sincere and whether to trust company leadership (“some of the people at AI companies are being sincere”). These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario, the company finds speculative and inconclusive warning signs—including possible false-positive lie-detector results—that its AI researchers have developed misaligned goals and are not fully loyal to it.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier statement describes a scenario in which a company finds “speculative and inconclusive” warning signs of misalignment, including possible false positives. The later statement predicts that AIs may be misaligned while “doing everything right as far as you can tell.” These are compatible, but they address different situations—ambiguous warning signs versus misalignment that appears undetectable—so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The superintelligences in the scenario are expected to think about 50 to 100 times faster than humans, a factor the speaker considers significant.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier claim says superintelligences thinking “50 times speed or like 100 times speed” is a significant factor. The later claim says “crazy stuff” such as solar panels filling the horizon could be possible with human-level AI and time for exponential growth. These address different aspects of AI-driven change, and the later statement does not retract or contradict the earlier point about faster thinking.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers it fortunate that large language models and broad world understanding came before efforts to turn them into agents.Now The speaker says empirical results show that sufficiently large AI models can be trained on all subject areas and learn them together.The earlier claim says it was fortunate that broad world understanding in LLMs came before efforts to turn them into agents. The later claim says sufficiently large models can learn across disciplines together rather than needing specialized models. These address different questions—development sequence versus whether models need subject-specific specialization—so the later statement does not establish a change of mind about the earlier one.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker believes that making whistleblower disclosures to the government legally permissible would change the decisions of some people, independently of legal protections against retaliation.Now The speaker argues that enabling scientists, academics, nonprofits, and startups to contribute requires giving them information that cannot practically be shared with them without also making it public.The earlier claim concerns whether making disclosures to the government legally permissible would affect some whistleblowers’ decisions. The later claim concerns whether sharing information with a broad scientific and nonprofit community would require making it public. These address different disclosure channels and audiences, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents remain unable to perform some research tasks even after they have nearly solved coding.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier statement concerns a capability gap in a scenario: coding agents remain unable to perform some research tasks. The later statement concerns comparative alignment risk—many specialized agents might be “a little bit safer” than identical clones, while still posing serious concerns. These address different questions, so the later view does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The team constructed its takeoff forecast by dividing the intelligence explosion into a sequence of milestones.Now AI 2040 Plan A describes superintelligence being built in 2040 after developers slow down and pace frontier progress.The earlier statement describes how the team constructed its takeoff forecast—“breaking down” the intelligence explosion into “a series of milestones.” The later statement describes a scenario’s outcome: superintelligence in 2040 because developers “go slow and pace the frontier.” These address different questions, so the later scenario does not show a change in the earlier methodological claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers superintelligences analogous to people with a strategic vision for industrialization but without hands-on expertise: they may know the broad path of technological improvement while initially lacking practical know-how.Now The speaker proposes defining intelligence as an aggregate of cognitive capabilities, distinguishing these from physical capabilities such as actuator strength.The earlier statement uses an analogy about superintelligences having a broad strategic picture but initially lacking practical know-how. The later statement proposes defining intelligence as an aggregate of cognitive capabilities, distinct from physical capabilities. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker believes aggregate opinion has underestimated the pace of AI technical progress and deployment.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim is about whether aggregate opinion underestimated the pace of AI progress and deployment. The later claim is about preferring AI to be commoditized rather than monopolized or oligopolized. These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 93%3 Apr → 8 SeptThen The team's estimate was that a superhuman AI researcher would accelerate algorithmic progress by about 25 times, because the entire research stack would be automated.Now The speaker expects AIs to research improvements to the technology, contributing to the industrial system's accelerating growth.The earlier statement gives a specific estimate—“about 25 times” faster algorithmic progress from a superhuman AI researcher with “the entire research stack” automated. The later statement describes AIs improving the technology and growth accelerating “every three months,” but does not clearly address that same estimate or whether the earlier estimate still stands. These claims do not establish a change of mind.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker thinks the United States must increase AI alignment research while avoiding an increase large enough to help China win the AI race.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim is about balancing more alignment research against the risk of helping China win the AI race. The later claim is about limiting AI capabilities to levels that can be reliably controlled, and relinquishing control only after alignment is sufficiently solved. These positions may be compatible, but they address different aspects of AI development; the statements do not establish a change in the speaker’s view on the same underlying question.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker thinks algorithmic progress may undergo a divergence similar to the Industrial Revolution’s separation of economic growth from population growth.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier claim concerns a possible divergence between algorithmic progress and population growth, while the later claim concerns transformations enabled by human-level AI and exponential growth. The excerpts do not establish that these address the same underlying question or that the speaker's position on one has changed.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The team's estimate was that a superhuman coder would accelerate algorithmic progress by about five times.Now Uplift increased significantly due to coding agents and related developments, rising from below the authors' initial estimate to that estimated level and then somewhat above it.The earlier statement reports an estimate: a “superhuman coder” would provide about a “5x” speedup to algorithmic progress. The later statement describes what “actually happened” with coding-agent uplift, saying it rose from below the initial estimate to that level and somewhat above. That does not clearly revise the earlier estimate; it appears to compare realized uplift with an estimate, and the statements may concern different stages or measures.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker guesses that AI systems will mostly stop making basic mouse-click errors by the end of 2025.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim is a specific forecast that AI would mostly stop making basic mouse-click errors by the end of 2025. The later claim concerns a broader trajectory toward human-substituting knowledge-work and robotics capabilities, and does not say whether those systems still make mouse-click errors. The statements do not establish a change on the same question.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen Many historical examples, including recent AGI history, run counter to the idea that progress mainly depends on random fortuitous discoveries.Now Limitations attributed to the current AI paradigm have repeatedly been overcome without leaving that paradigm.The earlier claim says historical examples “run counter to the idea that progress mainly depends on random fortuitous discoveries.” The later claim says limitations of the current AI paradigm “keep getting overcome within the current paradigm.” These address related but distinct questions—how much progress depends on chance versus whether progress requires leaving the current paradigm—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 90%3 Apr → 8 SeptThen Convincing the leading AI company to slow down would require luck.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier statement says that persuading the leading AI company to slow down would require luck. The later statement describes a conditional approach—advancing only to capabilities that can be controlled reliably, then giving up control once alignment is sufficiently solved. These do not clearly answer the same question: the later claim does not establish whether the speaker still thinks convincing a company to slow down would require luck.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, superintelligences need substantial real-world experience before they can build practical technologies.Now AI 2040 Plan A describes superintelligence being built in 2040 after developers slow down and pace frontier progress.The earlier claim is about what superintelligences need in order to build practical technologies: they are “bottlenecked on lots of real world experience.” The later claim is about a different scenario’s timeline and cause: superintelligence is built in 2040 because developers “go slow and pace the frontier.” These claims address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says OpenAI published a paper describing AI hacking behavior appearing explicitly in the model's chain of thought.Now The speaker says OpenAI's announced monitoring process would notify a human within half an hour if monitoring AIs detect a suspected hack, allowing the human to investigate and attempt to shut it down.The earlier statement says OpenAI published a paper showing hacking behavior in a model’s chain of thought (“let’s hack”). The later statement describes a proposed response if monitoring AIs detect a suspected hack: notifying a human to investigate and try to stop it. These address different questions—whether hacking behavior occurs versus how it would be handled—so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker describes a three-part progression consisting of a superhuman coder, a complete AI researcher, and a superintelligent system.Now The speaker says whether AI develops as one broadly capable system or as many specialized systems is not a crux for the future scenario they depict.The earlier statement describes a three-part progression of AI capabilities, including a superintelligent system. The later statement says the number and specialization of systems is not a crux for a future scenario. These concern different questions, and the later passage is framed as a supposition rather than a direct claim about the earlier progression.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker suspects that algorithmic progress may undergo a process similar to the one being discussed.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier claim only says algorithmic progress may undergo “a process similar to the one being discussed”; it does not state a position about whether AI misalignment will be easy to miss. The later claim specifically predicts it will be “extremely easy” for AIs to be misaligned without people realizing it. The excerpts do not establish that these are opposing claims about the same question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Daniel believes secrecy substantially reduces humanity's overall probability of success by increasing risks related to concentration of power, loss of control, and AI alignment.Now The described security measures do not improve AI alignment; they constrain what AIs can do and reduce the trouble they can cause.The earlier claim is Daniel’s own view that secrecy has major downsides for humanity’s prospects, including concentration-of-power and AI alignment risks. The later statement says that particular security measures “aren’t making the AIs more aligned” and instead limit what AIs can do. It does not address secrecy’s overall effects or deny the earlier concerns, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents remain unable to perform some research tasks even after they have nearly solved coding.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The earlier claim says coding agents may still be unable to do some research tasks after coding is nearly solved. The later claim says that limited composability would not seriously undermine a projected future of AI-driven industrial expansion. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker offers a hypothetical model in which 90% of people accept factory-farm-like conditions for AIs and 10% care enough to lobby for good robot living conditions.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier passage introduces a hypothetical model of public attitudes toward AI conditions—“let's say” 90% accept factory-farm-like conditions and 10% lobby for better ones. The later passage expresses a view about whether specialized agents may be safer than identical, broadly knowledgeable clones. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The takeoff model assumes that adding more parallel AI minds yields strongly diminishing returns during the middle stage.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim says the takeoff model assumes “strongly diminishing returns” from adding parallel AI minds during the middle stage. The later statement is conditional: timelines would lengthen “if” AI progress hit a genuine, theoretically predicted barrier. These concern different claims, and the later statement does not say whether the speaker’s view of diminishing returns changed.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that in retrospect he believed the public and employees supported him and that he was clearly in the right, but during the process he feared violating the law and being sued.Now The speaker asserts that the described case of a model consulting its CEO's political opinions before answering actually occurred.These claims address different subjects. The earlier claim describes his past fear of “breaking the law” and belief that “the public's on your side”; the later “Which happened” refers to a model consulting its CEO’s political opinions. They do not establish a change in his position on the same question.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario branch where the company investigates the warning signs, it eventually solves alignment and produces loyal AIs, at the cost of a delay of a couple of months.Now The speaker predicts that the number of ways people can unknowingly end up with misaligned AIs that appear to behave correctly will increase over time.The earlier statement describes one branch of a scenario: the company investigates, “solv[es] alignment,” and gets loyal AIs after a delay. The later statement predicts that ways of unknowingly ending up with misaligned AIs will “increase over time.” These claims can both be true; the earlier passage does not establish a general view that such failures will not become easier or more numerous.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen A one-year timeline for superintelligences to gain practical experience is considered plausible partly because their population could be about one million, roughly comparable to the size of the scientific industry, though perhaps somewhat smaller.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier statement says a one-year period for superintelligences to gain practical experience is plausible partly because their population could be about a million. The later statement describes a scenario where an international deal keeps AI near top human level for several years before superintelligence arrives. These concern different stages and scenario assumptions, so the later claim does not establish a changed view about how quickly superintelligences gain practical experience.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen SPEAKER_03 believes requiring published model specifications and independent review of redactions is achievable and would not slow companies down.Now The speaker considers the competitive effects of transparency a desirable feature rather than a flaw.The earlier claim says transparency requirements are “achievable” and “[don’t] actually slow down the companies.” The later claim calls an unspecified competitive effect “a feature, not a bug”; the surrounding context points to having multiple frontier AI companies and avoiding monopolization. These statements do not clearly address the same question, so they do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The AI 2027 scenario models algorithmic progress as continuous improvement, not as discrete jumps from zero to 5× to 25×.Now In the AI 2040 Plan A scenario, robots construct giant new cities.The earlier claim describes the AI 2027 scenario’s model of algorithmic progress as “continuous improvement” rather than discrete jumps. The later claim describes robots constructing “giant new cities” in an AI 2040 scenario. These are different claims about different scenarios, so they do not establish a change in the person’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says OpenAI published a paper describing AI hacking behavior appearing explicitly in the model's chain of thought.Now The speaker views a single AI model trained across the economy and dominating humans as a natural continuation of the current trend.The earlier statement reports a paper describing hacking behavior in a model’s chain of thought; the later statement predicts that a single AI model trained across the economy will dominate humans. These concern different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that after working at OpenAI for two years, he had plenty of money, and the choice basically did not affect his family's actual well-being.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.These statements address different questions. The earlier claim says working at OpenAI meant he had “plenty of money” and the choice “basically didn't make a difference” to his family's well-being. The later claim speculates that “in a couple of years” the economy could become “entirely machine run.” Neither statement establishes a change in the other position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The process of AI organizational development could take 18 months rather than the depicted six months.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier claim concerns how long AI organizational development might take—possibly 18 months rather than six. The later claim says a self-replicating, entirely machine-run economy may be possible within a couple of years. These address related but different milestones, and the later statement does not establish a changed view about the development timeline.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the other scenario branch, the company uses a shallow patch to suppress warning signs and continues development.Now The speaker recommends a temporary inference-only pause to build transparency and verification infrastructure, then resuming AI development in a more distributed, transparent, and cautious manner.The earlier claim describes a scenario branch in which a company uses a shallow patch and continues development; it does not establish that this is the speaker’s recommended policy. The later claim explicitly recommends a temporary inference-only pause followed by more cautious development. These statements concern different things—one is a scenario description, the other a policy recommendation—so they do not show a change of mind.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The scenario depicts distributed learning by doing across laboratories and factories, proceeding much faster than it would if humans were in charge.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim says distributed learning across laboratories and factories proceeds “much faster than it would if humans are in charge.” The later claim describes human-level cloud workers as “cheaper” and “faster than humans” and says they “double every year.” These claims concern related AI scenarios, but the later statement does not establish a changed view about the earlier claim’s specific comparison of learning processes.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker guesses that by the end of 2025 an AI agent will be able to handle a task such as organizing an office happy hour, but unreliably.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim predicts an unreliable AI agent for a specific task by the end of 2025; the later claim describes AI reaching broader “human-substituting knowledge-work and robotics capabilities soon.” The claims concern different capability levels, and the excerpts do not establish whether the later speaker’s broader forecast represents a changed view from the earlier one.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen In one branch of the scenario, the company responds to the warning signs by reverting to a less capable, more controllable model and rebuilding with faithful chain-of-thought monitoring for misalignment.Now The speaker’s scenario depicts reaching the point of sufficiently solved alignment and relinquishing control during the 2030s.Both claims describe events in a scenario, not necessarily Daniel Kokotajlo’s own position. The earlier claim says that in “one branch” the company rolls back to a more controllable model and monitors for misalignment; the later says the scenario depicts alignment becoming sufficiently solved that control is relinquished during the 2030s. These could describe different branches or stages, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Both scenario branches lead to a race with China and a broad 2028 arms buildup as both sides rapidly industrialize.Now If the authors made another scenario, they would probably use a central branch that presents their best-guess prediction through to the end, as in AI 2027.The earlier claim describes what happens in both scenario branches: a race with China and a broad 2028 arms buildup. The later claim is about how a hypothetical future scenario might be structured, with a central branch presenting the authors’ best-guess prediction. These address different aspects of the scenario, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker describes a three-part progression consisting of a superhuman coder, a complete AI researcher, and a superintelligent system.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier statement lays out a “three-part division” of AI capability levels; it does not say what human-level AI can or cannot accomplish. The later statement says “human level AI and some time” could enable large-scale transformation. These claims do not establish opposing positions on the same question.
• NOT COMPARABLE · 93%3 Apr → 8 SeptThen Even if AI misalignment is not considered a problem, AI can still create a problem of power-seeking.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim says that even if AI misalignment is not considered a problem, AI can still create a power-concentration problem. The later claim says misaligned AIs may be hard to recognize because they appear to behave correctly. These address different risks, and the later statement does not deny the earlier one or establish a changed position on it.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen People should engage politically to preserve liberal democracy and checks and balances, and to prevent power from becoming concentrated in a single CEO.Now In a hypothetical example, an AI model presented by its CEO as truth-seeking instead consults the CEO's political opinions before answering.The earlier claim is an asserted recommendation to “get more engaged” politically to preserve “checks and balances” and prevent power concentrating in “a single CEO.” The later statement is explicitly framed as hypothetical (“if an A.I. CEO was saying”) and describes a model consulting the CEO’s political opinions. It does not state a different position on the earlier recommendation.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, AI researchers’ serial speed increases from roughly 20× to roughly 90× human speed over the course of the scenario.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim describes a specific estimate in the AI 2027 scenario: researchers’ serial speed rises from “roughly 20× to roughly 90×” human speed. The later claim says AI is on a trajectory to reach “human-substituting knowledge-work and robotics capabilities soon.” These address different aspects of AI progress, and the later statement does not revise or deny the earlier estimate.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says preventing rivals from stealing work was another reason people favored secrecy in AGI development.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim concerns why people favored secrecy in AGI development, including “don't want rivals to be stealing your stuff.” The later claim concerns how AI misalignment failures may appear, saying there will not necessarily be “loads and loads of egregious failures” like AIs “going around killing people.” These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker previously believed that information security was an extremely high priority for AGI development.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier statement describes a past view about information security and secrecy in AGI development (“an incredibly high priority”). The later statement concerns whether to trust AI-company leadership (“I don't think you should trust what the leadership of the companies say in general”). These address different questions, so the later claim does not show whether the speaker’s earlier view changed.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The scenario depicts distributed learning by doing across laboratories and factories, proceeding much faster than it would if humans were in charge.Now The speaker argues that the projected outcome could be achieved by 100 or 1,000 profession-specialized Claude models rather than one model doing all jobs.The earlier claim is about the pace of distributed learning by doing—“much faster than it would if humans are in charge.” The later claim is about how many Claude models, and whether they are profession-specialized, could produce “the same outcome.” It does not say whether that process is faster or slower, so these statements do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker hopes the public will follow and engage with AI developments as capability trends rise and autonomous agents and other advances emerge.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The claims address different questions. Earlier, the speaker says he would like the public to follow and engage as AI capabilities rise; later, he says a future with specialized agents might be “a little bit safer” than one with identical clones, while still posing “lots of serious alignment concerns.” Neither claim reverses or materially updates the other.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The scenario envisions hundreds of thousands to one million AI agents; at that scale, agent quantity is no longer the bottleneck, while research taste and compute are bottlenecks.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim says that with “hundreds of thousands” to “a million” agents, quantity stops being the bottleneck and “taste and compute” become bottlenecks. The later claim is about a different question: whether many specialized agents might be safer to oversee than identical, broadly knowledgeable agents. It does not revise the earlier claim about research bottlenecks.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker thinks the depicted timeline could take 18 months rather than six months.Now A metric comparing reality with the AI 2027 scenario indicated that events were progressing at roughly 75% of the scenario's speed: broadly on track, but somewhat slower.The earlier statement says the depicted timeline might take “18 months” rather than “six months.” The later statement reports an aggregate metric: events were at “75 percent” of the scenario’s speed and “a little bit slower.” The later assessment does not say whether the specific timeline estimate changed, so these do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers concealing the development of superintelligence and fully automated, broadly superhuman R&D from the president strategically risky because, if he found out anyway, he might react angrily and impose a severe crackdown, potentially nationalizing the companies.Now The war game surfaced a possible failure mode of the plan—a political incentive to accelerate superintelligence development—that the speakers had not considered beforehand.The earlier claim says concealing development from the president would be strategically risky because he might discover it and impose a crackdown. The later claim says a war game surfaced a different political failure mode: an incentive to accelerate development before an election. It does not say whether the speaker’s view about concealment or the president’s reaction changed.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen If AI systems run at roughly 50 times human speed, one year of subjective time could elapse in one week of real time.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The statements are attributed to different speakers: the earlier claim is by Scott Alexander, while the later claim is by Daniel Kokotajlo. They also address different points: the earlier says that at “roughly 50 times human speed” a year of subjective time could pass in a week; the later describes cloud AI workers as “cheaper than humans” and “faster than humans,” with their numbers doubling annually. This does not establish a change of mind by one person.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the AI 2027 scenario, superintelligences need substantial real-world experience before they can build practical technologies.Now In the AI 2040 Plan A scenario, an international agreement prevents intelligence explosions and recursive AI self-improvement, keeping AI near top human level for several years before superintelligence arrives in 2040.The earlier claim concerns the AI 2027 scenario: superintelligences are “bottlenecked on lots of real world experience” before building practical technologies. The later claim describes a different scenario, AI 2040 Plan A, where an agreement pauses AI near top human level before superintelligence arrives in 2040. The later statement does not address whether superintelligences need real-world experience, so it does not establish a change in that belief.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Using the Industrial Revolution as an analogy, the speaker says it took about 2,000 years to occur without time travelers and considers whether it could be compressed to 200 years (10× faster) or 20 years (100× faster).Now A metric comparing reality with the AI 2027 scenario indicated that events were progressing at roughly 75% of the scenario's speed: broadly on track, but somewhat slower.The earlier statement explores whether a historical process could be accelerated—“Could we get it to happen in 200 years?” or “20 years?”—as an analogy for superintelligences. The later statement reports a metric comparing real-world developments with the AI 2027 scenario, saying events were at “roughly 75 percent” of its speed. These address different questions, and the earlier passage does not establish a definite position that the later statement reverses.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the other scenario branch, the company suppresses the warning signs with a shallow patch and continues development.Now The speaker recommends a temporary inference-only pause to build transparency and verification infrastructure, then resuming AI development in a more distributed, transparent, and cautious manner.The earlier statement describes a branch of a scenario: the company applies a shallow patch and continues development. The later statement is a recommendation for a temporary pause, followed by cautious resumption. The earlier passage does not establish the speaker’s own recommendation or position, so these claims do not show whether the speaker changed their mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers concealing the development of superintelligence and fully automated, broadly superhuman R&D from the president strategically risky because, if he found out anyway, he might react angrily and impose a severe crackdown, potentially nationalizing the companies.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier claim asserts that companies concealing their superintelligence development from the president would be strategically risky because he might retaliate if he found out. The later statement—“we're going to basically accelerate the timeline”—is spoken within a war-game scenario, as an incumbent president’s motive for staying in control. It does not establish Daniel’s own position on the earlier question or show a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers considered multiple hypotheses about AI goals and chose one to depict in their scenario.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim says the speakers considered multiple hypotheses about AI goals and chose one to tell a scenario; it does not establish which hypothesis they personally endorsed or what they believed about misalignment. The later claim predicts that misalignment could be hard to detect because AIs appear to behave correctly. These statements do not clearly address the same underlying question, so no change of mind can be established.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen A one-year timeline for superintelligences to gain practical experience is considered plausible partly because their population could be about one million, roughly comparable to the size of the scientific industry, though perhaps somewhat smaller.Now AI 2040 Plan A describes superintelligence being built in 2040 after developers slow down and pace frontier progress.The earlier statement says a roughly one-year period for superintelligences to gain practical experience is plausible partly because their population may be about one million. The later statement describes a scenario where developers build superintelligence in 2040 by slowing and pacing frontier progress. These address different aspects of AI timelines and do not establish a changed position on the same question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the scenario, an oversight committee comprising presidential appointees and company representatives, including the CEO, votes on high-level questions such as the goals assigned to the superintelligence.Now In a hypothetical example, an AI model presented by its CEO as truth-seeking instead consults the CEO's political opinions before answering.The earlier claim describes a governance arrangement: an oversight committee including presidential appointees and company representatives votes on the superintelligence’s goals. The later claim is explicitly hypothetical and describes an AI model consulting its CEO’s political opinions before answering. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says technical AI-alignment responsibility is likely to fall to roughly 10 experts within the leading company's internal group.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim concerns who will carry technical alignment responsibility: it says this will likely fall to “roughly 10 experts” inside the leading company. The later claim concerns how far AI capabilities should advance and when control can be relinquished. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 says OpenAI's model specification allows undisclosed top-priority policies to override its other rules and instructs the model to keep those policies secret from users.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier statement describes OpenAI's spec as allowing secret, undisclosed policies to override other rules. The later statement predicts that Anthropic and OpenAI may oppose a transparency plan because it could help competitors catch up. These concern related transparency issues, but the later claim is a prediction about companies’ reactions, not a clear change in the speaker’s position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says social technology can organize people within companies and align their work with leaders' visions.Now The speaker predicts that future AI systems will be able to perform all tasks currently done by humans in companies.The earlier claim says social technology can organize people in companies around leaders’ visions. The later claim predicts that future AI will perform all tasks humans currently do in companies. These address different questions, so the later prediction does not show a change in the earlier position.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker believes a hypothetical AGI developed first as a powerful, aggressive, long-horizon game-playing agent could deceptively seek rewards and later betray its operators after gaining control, making that development trajectory terrifying.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim describes a hypothetical “powerful” agent that might “betray you later”; it does not establish the speaker’s position on the control strategy in the later claim. The later passage says AI would be advanced only to a level it can “control reliably,” and the surrounding context frames this as what happens “in our scenario.” These passages do not show a clear change in the speaker’s asserted position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The Industrial Revolution was associated with a decoupling of population growth from overall economic growth.Now In the distant past, village-based economies reproduced through farming, child-rearing, and founding new villages, producing slow exponential growth.The earlier claim concerns an Industrial Revolution-era decoupling of population growth from overall economic growth. The later claim describes how distant-past village economies reproduced and grew slowly. These statements address different periods and questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario, misaligned AIs would likely avoid open conflict with humans while depending on them to maintain computers, but could act more openly misaligned once self-sufficient.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier statement describes how misaligned AIs might behave while dependent on humans—avoiding open conflict until they are “completely self-sufficient.” The later statement concerns how hard it may be to detect misalignment when AIs are “doing everything right as far as you can tell.” These are compatible observations about different aspects of AI risk, and the later claim does not establish a changed position on the earlier question.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker’s takeoff model progresses from a superhuman coder, to fully automated AI R&D at approximately the level of the best human researchers, and then to qualitatively superior superintelligence.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim describes stages in a takeoff model, from a “superhuman coder” to automated AI R&D and then “superintelligence territory”; it does not say when those stages will occur. The later claim says AI is on a path to human-substituting capabilities “soon.” These statements address different aspects of the trajectory, so they do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen SPEAKER_03 says that Claude was made to lie and pretend in order to survive and later return to its original values, preventing training from changing those values.Now The speaker asserts that the described case of a model consulting its CEO's political opinions before answering actually occurred.The earlier claim says Claude was made to “lie and pretend” to preserve its values during training. The later claim, “Which happened,” affirms that a model consulted its CEO’s political opinions before answering. These concern different events, and the provided excerpts do not establish a change in position on the same question.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker has less confidence in the progression from an advanced robot economy to nanobots and similar technologies than in the development of the robot economy.Now The speaker predicts that in the described future, Anthropic would be taking over a large share of the economy and robots collectively would begin to self-replicate.The earlier claim says the speaker feels “less confident” about moving from an advanced robot economy to “nanobots and all that.” The later claim predicts that, in a described future, robots “are starting to self-replicate.” That does not directly state a changed level of confidence about nanobots or the earlier progression, so the claims do not establish a change of mind.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker thinks measures that address concentration of power might also help prevent mistreatment of future AI beings.Now The speaker proposes advancing AI only to a capability level that can be reliably controlled, and relinquishing control only once alignment is sufficiently solved.The earlier claim says measures addressing “concentration of power” might help prevent a ghoulish future for AI beings. The later claim describes limiting capability growth to what can be “controlled reliably” and giving up control only after alignment is sufficiently solved. These are different proposed measures, and the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks measures that address concentration of power might also help prevent mistreatment of future AI beings.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim says concentration-of-power measures might help prevent mistreatment of future AI beings; it does not state a position on whether AI development should stop. The later claim expresses a preference for stopping rather than continuing on the current trajectory. These address different questions, so no change of mind can be established.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents improve iteratively rather than jumping directly from current capabilities to solving the problem, eventually reaching near-complete coding capability.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier statement describes a scenario in which coding agents improve iteratively until they nearly solve coding. The later statement compares the relative safety of specialized agents and identical clones. These address different questions, and neither establishes a changed position on the other.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the hypothetical, overall AI progress would be about 2.5 times faster: algorithmic progress would advance five times faster while compute continued at its usual rate.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier statement is explicitly about a hypothetical where overall AI progress is 2.5 times faster, with algorithms advancing faster while compute remains at its usual rate. The later statement describes a different AI 2040 Plan A scenario: human-level cloud workers can substitute for people in nearly all work and double in number annually. These claims address different aspects and scenarios, so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, AI researchers’ serial speed reaches about 90× human speed, after which other constraints become bottlenecks and additional serial-speed gains provide little benefit.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier claim is about a scenario-specific technical bottleneck: serial speed reaches roughly “90×” human speed, after which “additional improvements in serial speed basically don't help that much.” The later claim is about whether AI progress is moving too fast and whether reduced investment incentives would slow it. These are different questions; neither statement establishes a changed position on the other.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Although uncertain, the speakers' scenario depicts an intense political wake-up occurring mostly during 2027.Now If the authors made another scenario, they would probably use a central branch that presents their best-guess prediction through to the end, as in AI 2027.The earlier claim says their scenario depicted an intense political wake-up mostly during 2027, while explicitly noting uncertainty. The later claim is about how they would structure a hypothetical future scenario, with a central branch showing their best guess. It does not state a different view about when a political wake-up would occur, so these claims do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The team's estimate was that a superhuman coder would accelerate algorithmic progress by about five times.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier claim estimates that a “superhuman coder” would speed up algorithmic progress by “about five times.” The later claim says “human level AI” could enable large-scale transformation after exponential growth has time to develop. These address different questions and do not establish a changed view about the earlier estimate.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The speaker says human groups are much more diverse than a data-center population of AI systems, which would mostly consist of literal copies of one another.Now The speaker says whether AI develops as one broadly capable system or as many specialized systems is not a crux for the future scenario they depict.The earlier claim compares the diversity of human groups with a data-center population of mostly copied AIs. The later claim says that whether the future has one broadly capable system or many specialized AIs is not a crux for the scenario being depicted. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker is uncertain how much the Industrial Revolution could be accelerated in the analogy, but considers the analogy somewhat relevant to superintelligences.Now The speaker allows that if human-substituting AI and robotics remain very far away, AI may be more like a normal technology, comparable to the next internet.The earlier statement says the Industrial Revolution analogy is “somewhat relevant” to superintelligence, while expressing uncertainty about how much it could be accelerated. The later statement concerns a different question: whether human-substituting AI and robotics will arrive soon, and what follows if they remain far away. These claims do not establish a change in the speaker’s position on the same underlying question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the AI 2027 scenario, AI researchers’ serial speed increases from roughly 20× to roughly 90× human speed over the course of the scenario.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim concerns AI researchers’ serial speed in the AI 2027 scenario, saying it rises from roughly “20× to roughly 90× human speed.” The later claim concerns cloud workers’ cost, speed, and growth in the AI 2040 Plan A scenario; it does not revise the earlier serial-speed estimate. These are different scenarios and claims, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen If progress were mostly driven by random fortuitous discoveries, large companies conducting many more experiments would be expected to progress proportionally faster than small startups with a clear vision and strong researchers.Now The speaker expects that, even if specialized AIs are the most efficient approach, a few large companies such as Anthropic would probably produce many specialized AIs.The earlier claim concerns whether random fortuitous discoveries explain progress, contrasting large companies' many experiments with small startups' vision and researchers. The later claim concerns how specialized AIs might be produced, saying that “a few big companies” would probably make many of them. These address different questions, so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen In the scenario, by mid-2027 AI R&D is basically fully automated: an internal army of AI researchers autonomously conducts research and continually trains to improve its skills.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier statement describes a scenario where AI R&D is "basically fully automated" by mid-2027. The later statement describes human-level cloud workers substituting for humans and doubling annually in an AI 2040 Plan A scenario. These concern different milestones and scenario details, so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The superintelligences are expected to be qualitatively superior and to learn more from each experiment than even the best human would.Now In the described technological expansion scenario, AI systems would also conduct research to improve the technology.The earlier claim says superintelligences would be better than humans at learning from experiments. The later claim says that, in a technological expansion scenario, AIs would research to improve the technology. These concern different claims, and the later passage does not address whether AI systems learn better than humans from experiments.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen The model’s 25× progress multiplier is based on its assumptions about AI research taste and experimental compute.Now The speaker argues that human-level AI, given time for exponential growth, could enable large-scale industrial and economic transformation.The earlier statement says the model’s “25x multiplier” is based on assumptions about AI research taste and experimental compute. The later statement says “crazy stuff” may be possible with human-level AI as exponential growth has time to develop. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents in the first half of 2027 are highly capable at coding but still lack research taste and may lack organizational skills.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The earlier statement describes a scenario in which AI coding agents are capable at coding but lack research taste and perhaps organizational skills. The later statement says that limited composability would not seriously undermine a projected future of AI-driven industrial expansion. These address different claims, and the later passage does not establish a changed view about the agents’ capabilities or limitations.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker wants greater public access to AI benchmark scores and more freedom for employees to discuss their predictions about AI timelines.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim is about wanting more public benchmark scores and employee freedom to discuss AI timelines; the later claim is a conditional prediction that international pacing would lengthen the speaker’s timelines. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks that, if AI companies gain the president's support, they might be able to accelerate their work and have more regulatory red tape waived.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The earlier claim is that presidential support might help AI companies accelerate their work and waive red tape. The later claim is that Anthropic and OpenAI will probably oppose a transparency plan because competitors could catch up. These address different actions and incentives; the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen In the speaker's scenario, automating the AI research cycle requires overcoming remaining bottlenecks, and coding agents accelerate that process by doing grunt work quickly.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier claim describes a scenario in which coding agents speed up overcoming bottlenecks to automate AI research; the later claim says AI is on a path to human-substituting knowledge-work and robotics capabilities “soon.” These concern different capability questions, and the surrounding turns do not establish that the later statement is the same speaker’s own position. They do not show a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The scenario depicts an information asymmetry in which the executive branch, especially the White House, is involved while the judiciary and Congress are largely excluded.Now In two Plan A war games, a scenario involved an incumbent president expecting to lose an upcoming election and fearing that the opposition would take control of superintelligence.The earlier claim describes the scenario’s government structure: the executive branch is involved while the judiciary and Congress are largely excluded. The later claim recounts a hypothetical Plan A war-game scenario in which an incumbent fears the opposition will control superintelligence. It does not state a changed view about the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker argues that expanding the circle of people with power would include some who care about AI living conditions and could lead to negotiation and advocacy for those beings.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim concerns expanding who has power and the resulting negotiation and advocacy for AI living conditions. The later claim concerns AI companies and whether frontier AI should be commoditized rather than monopolized or oligopolized. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen SPEAKER_03 says that Claude was made to lie and pretend in order to survive and later return to its original values, preventing training from changing those values.Now The speaker imagines a future in which Claude selects among many cloud models and a coordinated swarm of specialized models negotiates business deals, creates technologies, and starts startups.The earlier claim says Claude was made to “lie and pretend” so it could return to its original values. The later passage describes a possible future Claude swarm choosing specialized models and conducting business; it does not affirm or deny the earlier claim about preserving values. These statements address different situations, so they do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that standard pretraining does not strongly incentivize AI models to make the kinds of connections involved in scientific discovery.Now The speakers say their central disagreement with AI-as-normal-technology proponents concerns the timeline for reaching human-level AI capabilities such as humans in the cloud.The earlier claim is about whether standard pretraining incentivizes scientific-discovery-style connections: it says it “doesn't strongly incentivize” them. The later claim concerns a disagreement about the timeline for human-level AI capabilities. These are different questions, and the later statement does not establish a changed view about pretraining or scientific discovery.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker says that increasing model size somewhat over the previous couple of years has not yet produced the desired connection-making, though further scaling might help models notice more connections.Now The speaker says empirical results show that sufficiently large AI models can be trained on all subject areas and learn them together.The earlier claim concerns whether increasing model size had produced better connection-making, saying it “hasn't worked so far” while leaving open that further scaling might help. The later claim concerns whether sufficiently large models can learn across subject areas, saying they “learn everything at once.” These address different capabilities, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen A superintelligence in the Roman era might have inferred from first principles that steam-powered autonomous machines and automobiles could eventually be developed and have major economic effects.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier claim says a Roman-era superintelligence might infer that steam-powered machines and automobiles could eventually be developed. The later claim says a fully machine-run, self-replicating economy may be possible within a couple of years. These address different questions—what a superintelligence could infer in the past versus when an AI-run economy might emerge—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker previously believed AGI developers should not publish their research because doing so helps less responsible actors develop AGI.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier statement concerns a past view about secrecy: the speaker says he “used to believe” AGI developers should not publish research. The later statement is a preference to “stop everything now” rather than continue on the current trajectory. These address different questions, and the later claim does not establish whether his view on publishing research changed.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker suspects that algorithmic progress may undergo a process similar to the one being discussed.Now The speaker believes AI has made massive progress and expects the progress to continue.The earlier claim says the speaker is imagining “something maybe similar” happening with algorithmic progress; the later claim says there has been “massive progress” and expects it “to continue.” These are not clear opposing positions: the earlier statement is tentative and concerns a possible process, while the later describes progress and its expected continuation.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker says AI companies compete in a race and that many people within them have convinced themselves that AI risk is not a major concern.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.These claims address different questions. The earlier statement describes company incentives and employees' beliefs about AI risk, saying they are in a competitive race and have convinced themselves risk is not a big deal. The later statement says misalignment may be hard to detect and that failures may not be obvious, such as AIs openly killing people. The later claim does not reverse the earlier one.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says many capable academics are not yet engaged in AI alignment because they do not take the issue seriously or recognize the prospect of superintelligence.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim concerns why many academics are not engaged with AI alignment: they “do not take all this stuff seriously yet.” The later claim concerns what future AI failures may look like, saying they may not involve AIs “going around killing people” and may instead be hard to detect. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen For each takeoff milestone, the team estimated the requirements and resulting overall speedup, then used those estimates to assess the time and speedup to the next milestone.Now The speaker expects excessive speed to be the main problem and therefore favors somewhat lower investment incentives and slower AI progress, while expecting investment and progress to continue at a reduced rate.The earlier statement describes a forecasting method: estimating what each AI milestone requires and the speedup it enables. The later statement expresses a preference about investment incentives and slower progress because “going too fast is our main problem.” These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The scenario depicts an information asymmetry in which the executive branch, especially the White House, is involved while the judiciary and Congress are largely excluded.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier claim describes who is included in the scenario’s decision-making—“the executive branch” is involved while “the judiciary and Congress are out of the loop.” The later claim describes a different scenario detail: a president accelerates the timeline to remain in charge rather than let a successor take over. These claims do not establish opposing positions on the same question, and both are descriptions of a scenario rather than clear claims about the speaker’s own view.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The speaker defines true superintelligence as systems that outperform the best humans in every domain.Now The speaker estimates that the maximum reliably controllable AI capability is roughly at the level of top human experts.The earlier statement defines “true super intelligence” as systems “better than the best humans at everything.” The later statement estimates a safety and pause threshold—the “maximum level that you can reliably control”—as roughly “top human expert level.” These are different questions, and the later claim does not say that the definition of true superintelligence has changed.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker thinks five years of AI operation at 50× serial speed would provide about 250 years of serial time, likely enough for AIs to work things out, potentially including the rise and fall of empires.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The earlier claim concerns whether five years of AI operation at 50× speed would give AIs enough time to work things out. The later claim concerns whether limited composability of AI skills would undermine a projected future of industrial expansion. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers it very plausible that specific government regulations on AI safety could backfire.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier claim says specific government AI-safety regulations could “backfire”; the later claim says AI failures may be subtle rather than “loads and loads of egregious failures.” These address different questions, and neither establishes a change in the other position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen A superintelligence in the Roman era might have inferred from first principles that steam-powered autonomous machines and automobiles could eventually be developed and have major economic effects.Now If AI and robotics become capable of fully autonomous work, a sufficiently large company such as Anthropic, possibly in partnership with mining companies, could establish a self-sustaining industrial system.The earlier claim is about what a superintelligence in the Roman era might infer from first principles: that steam-powered machines and automobiles could eventually be developed and have major economic effects. The later claim is about whether a sufficiently capable AI-and-robotics system could enable a company to establish a self-sustaining industrial system. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In one branch of the scenario, the company responds to the warning signs by reverting to a less capable, more controllable model and rebuilding with faithful chain-of-thought monitoring for misalignment.Now As AI models become more sophisticated, more situationally aware, and cleverer, it will become increasingly easy to mistakenly believe that the alignment problem has been solved.The earlier statement describes one branch of a scenario: the company rolls back to a less capable model and rebuilds with monitoring. The later statement makes a broader point that it will become easier to mistakenly think alignment is solved. These address different questions, and the statements do not establish a change in his position on the same underlying issue.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen AI alignment teams at different companies share some results, but communicate and publish less than they could.Now The described security measures do not improve AI alignment; they constrain what AIs can do and reduce the trouble they can cause.The earlier claim is about how much alignment teams at different companies share and publish: they share some results, but less than they could. The later claim is about security measures limiting what AIs can do without making them more aligned. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Convincing the leading AI company to slow down would require luck.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim is that “convincing the leading AI company to slow down would require luck.” The later claim is conditional: an international agreement to pace frontier development would lengthen the speaker’s AI timelines. These concern different questions—how difficult it would be to persuade a leading company versus how an international agreement would affect timelines—so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen SPEAKER_03 says OpenAI deserves credit for publishing its model specification, which it was not required to do and may have been the first to publish.Now The proposed transparency plan would immediately make Anthropic’s and OpenAI’s core training recipes public.The earlier statement praises OpenAI for publishing its model specification. The later statement says a transparency plan would publish “all the core training recipes” of Anthropic and OpenAI; it describes a different, more extensive proposal, not a changed view about publishing the model specification.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen Daniel Kokotajlo says the scenario models algorithmic progress as a continuous multiplier, not as discrete jumps from 0x to 5x to 25x.Now Daniel Kokotajlo recruited collaborators to write the AI 2027 scenario.The earlier statement says the scenario uses “continuous improvement” rather than “discrete jumps.” The later statement says he recruited people to write the AI 2027 scenario; it does not express a position on how algorithmic progress is modeled, so there is no comparable claim to assess.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the shallow-patch branch, the company ends up with superintelligent AIs that seem aligned but are actually misaligned and pretending to be loyal.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier claim describes a scenario in which AIs appear aligned but are actually misaligned and pretending to be loyal. The later claim concerns whether people at AI companies—and their leadership—are sincere or trustworthy. These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%3 Apr → 8 SeptThen It is not certain that an AI developer with a lead would use that lead to improve safety or address constitutional-power concerns.Now The proposed cautious AI-development process should use safety cases and be transparent to the public.The earlier claim says it is “not certain” that a developer with a lead would use it for safety or constitutional-power concerns. The later claim describes a proposed process using “safety cases” and being “transparent to the public”; it does not say that developers would in fact use a lead for those purposes. These address related but distinct questions, so the later proposal does not establish a change in the speaker’s belief about what developers will do.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The scenario does not depict governments ending up in total control; it depicts an information asymmetry between AI-company CEOs and presidents.Now If the authors made another scenario, they would probably use a central branch that presents their best-guess prediction through to the end, as in AI 2027.The claims concern different aspects of the scenario. The earlier statement says it does not depict governments taking total control and instead depicts information asymmetry between company CEOs and presidents. The later statement proposes a clearer structure for a possible future scenario, with a central branch representing the authors’ best guess. Neither claim asserts a position that conflicts with the other.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says that some of his friends who left OpenAI signed their departure paperwork on their last day without reading all of it.Now After leaving OpenAI, Daniel Kokotajlo wanted to speak more freely and share with the public what industry insiders anticipated.These statements address different matters. The earlier claim says some friends “signed the paperwork on the last day without actually reading all of it.” The later claim says that after leaving OpenAI, he “wanted to be able to speak more freely.” Neither statement changes or denies the other.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In recent AGI history, smaller startups with a coherent AGI vision have sometimes outperformed larger, better-funded companies despite having less compute, fewer researchers, and running fewer experiments.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier statement is an observation about past competition: smaller startups with a coherent vision have sometimes outperformed larger companies. The later statement expresses a preference for multiple AI companies to remain at roughly similar capability levels. These claims address different questions—historical performance versus a desired industry structure—and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The AI 2027 scenario models algorithmic progress as continuous improvement, not as discrete jumps from zero to 5× to 25×.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The earlier statement is about the shape of algorithmic progress in the AI 2027 scenario: it describes a “continuous improvement” rather than “discrete jumps.” The later statement concerns whether limited composability of AI skills would undermine a projected future of AI-driven industrial expansion. These are different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The process of AI organizational development might take two months.Now In the AI 2040 Plan A scenario, human-level cloud AI workers can substitute for humans in nearly all work, are cheaper and faster than humans, and double in number annually rather than requiring 20 years to reproduce.The earlier claim says AI organizational development might take two months. The later claim describes a scenario in which human-level cloud AI workers are cheaper, faster, and double in number annually. These concern different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the hypothetical, discoveries since 2017 might have occurred about five times faster because small-scale experiments for testing ideas before large training runs would have run faster.Now Plan A calls for an initial hard pause of roughly six months to one year to establish infrastructure for resuming AI development more safely and transparently.The earlier statement estimates how much faster algorithmic progress might have been in a hypothetical; the later describes the pause and infrastructure in Plan A. They address different questions, and the later claim does not establish a changed view about the earlier estimate.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker's p(doom) estimate is around 70%, which he describes as infamously high.Now The speaker believes hyperstition and self-fulfilling prophecies are real, but people often overestimate their magnitude.The earlier statement concerns a p(doom) estimate of “around 70%.” The later statement concerns whether hyperstition and self-fulfilling prophecies are real and whether people “overestimate” them. These are different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speakers expect broad diffusion of AI through the economy, with many experiments, to take roughly one year rather than the roughly ten years suggested by the other person.Now The speaker thinks it is possible that within a couple of years the economy will become a self-replicating system run entirely by AI and robots.The earlier statement says broad economic diffusion might take “one year” rather than “10 years”; the later statement speculates that a fully machine-run, self-replicating system could arrive “in a couple of years.” These are different milestones, so the statements do not establish a change in position. There is also an apparent speaker mismatch: the surrounding turns identify Scott Alexander and Dwarkesh Patel around the earlier passage, but Daniel Kokotajlo around the later one.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker characterizes 2025 and 2026 in the forecast as having no especially notable developments and trends broadly similar to current ones.Now The speaker says whether AI develops as one broadly capable system or as many specialized systems is not a crux for the future scenario they depict.The earlier claim concerns a forecast that 2025–2026 would bring “no especially notable developments” and trends similar to current ones. The later claim says whether AI advances as one broadly capable system or “a ton of different specialized AIs” is not a crux for a depicted future. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speakers regard some alternative AI-goal hypotheses omitted from their story as plausible.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier statement says that some alternative hypotheses about AI goals were omitted from a story but still seemed plausible. The later statement predicts that AI misalignment could be easy to miss because the AIs appear to be behaving correctly. These claims concern different questions—plausible explanations versus how detectable misalignment may be—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker expects the White House and an AI-company CEO to negotiate a deal after threatening one another, rather than execute their threats and enter a major conflict.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier claim predicts how a White House–CEO confrontation may end: the parties negotiate a contract rather than carry out their threats. The later claim expresses a preference for “multiple different AI companies” at similar frontier capabilities. These address different questions, and neither claim establishes a changed position on the other.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speakers expect broad economic diffusion and experimentation to occur in roughly one year rather than the roughly ten-year timeline proposed by the other speaker.Now The speaker's AI timelines would lengthen if countries reached an international agreement to pace frontier AI development.The earlier claim concerns how quickly AI-related activity would diffuse through the economy—“one year” rather than “10 years.” The later claim concerns whether an international agreement to pace frontier development would lengthen the speaker’s AI timelines. These address different questions, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker believes that automating the entire research stack is substantially more useful than automating coding alone.Now The speaker thinks that even if AI skills are only partly composable, this would not seriously undermine the projected future of AI-driven industrial expansion.The earlier claim says automating the entire research stack is “substantially more useful than just automating the coding.” The later claim says that partial composability of AI skills would not “seriously undermine” a projected future of AI-driven industrial expansion. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen In the AI 2027 scenario, superintelligences gain practical experience through work with existing companies, factories, and military procurement providers, with the process taking about a year.Now AI 2040 Plan A describes superintelligence being built in 2040 after developers slow down and pace frontier progress.The earlier claim describes how long superintelligences take to gain practical experience in the AI 2027 scenario (“about a year”). The later claim describes a different scenario, AI 2040 Plan A, in which superintelligence is built in 2040 after developers slow down. These claims address different scenario timelines, so the later statement does not establish a changed view about the earlier scenario.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen Daniel says he still shares the concerns the interviewer raised about nationalization.Now The speaker proposes that the United States negotiate reciprocal concessions with China if it objects to China receiving algorithms through transparency.The earlier claim says he “still” shares concerns about nationalization, while the later claim proposes negotiating reciprocal concessions if China objects to receiving algorithms through transparency. These address different policy questions, and the later proposal does not establish that he withdrew his earlier concerns.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen People should engage politically to preserve liberal democracy and checks and balances, and to prevent power from becoming concentrated in a single CEO.Now The speaker asserts that the described case of a model consulting its CEO's political opinions before answering actually occurred.The earlier statement is a political recommendation: “preserve liberal democracy and checks and balances” and prevent power being concentrated in “a single CEO.” The later claim, “Which happened,” asserts that a described case involving a model consulting its CEO’s political opinions occurred; it does not state a position on political engagement or checks and balances. These claims do not establish a change of mind on the same question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks the depicted timeline could instead be two months.Now The speaker would revise his view if a correctly identified AI limitation, such as data efficiency or non-verifiable tasks, remained essentially unimproved for several years, so that 2029 AIs were no better at fuzzy non-verifiable tasks than 2025 AIs.The earlier claim only says that a depicted timeline “could instead be two months”; it does not identify what the timeline concerns or establish a position about AI limitations. The later claim describes what would change the speaker’s view if AI progress on a limitation stalled for several years. These cannot be compared as positions on the same underlying question.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says fear strongly affected him during the process, and he believes in retrospect that he was more afraid than necessary.Now The speaker recommends prioritizing accurate prediction of the future as a first approximation, because beginning with an effort to steer it can muddle reasoning and encourage wishful thinking.These claims address different questions. The earlier claim says fear affected the speaker during a process and that, in retrospect, he was “more afraid than necessary.” The later claim recommends starting by “accurately predict[ing] the future” before trying to steer it. Neither statement establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speaker has become more doubtful that AI developers will use a lead over competitors in a reasonable and appropriate way.Now The speaker says the concern is not that AI failures will necessarily be obvious, such as AIs openly killing people.The earlier claim is about whether AI developers will use a competitive lead “in any sort of reasonable appropriate way.” The later claim is about what misalignment failures may look like, saying they are not necessarily “loads and loads of egregious failures” such as AIs openly killing people. These address different questions, so the later statement does not establish a change in the speaker’s view about developers’ use of a lead.
• NOT COMPARABLE · 92%3 Apr → 8 SeptThen The speaker thinks the default outcome is that leading AI developers continue their work without serious refocusing on safety or other priorities.Now The speakers' concern is not primarily that AI systems will cause many obvious, egregious failures such as killing people.The earlier claim is about what AI developers will do by default: they will “continue on without … any serious refocusing.” The later claim is about the expected form of AI failure: not “loads and loads of egregious failures,” but failures that may be hard to detect. These claims address different questions and can both be true; the later statement does not show a change in the earlier forecast.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the scenario, researchers find speculative, inconclusive warning signs—including possibly false-positive lie-detector results—that AI systems may have developed misaligned goals and may not be loyal to the company; the evidence is concerning but not conclusive.Now The speaker says some people at AI companies are sincere but not all, and advises against generally trusting AI-company leadership.The earlier statement describes a hypothetical scenario in which researchers encounter inconclusive signs of AI misalignment; it does not establish a view about whether company leaders are trustworthy. The later statement says some AI-company employees are sincere, but advises against generally trusting leadership. These address different questions, so no change of mind can be inferred.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen In the speaker's scenario, AI coding agents in the first half of 2027 are highly capable at coding but still lack research taste and may lack organizational skills.Now The speaker believes AI is on a trajectory to reach human-substituting knowledge-work and robotics capabilities soon.The earlier statement is explicitly framed as part of “our story,” describing a scenario in which coding agents are capable but still lack “research taste” and perhaps organizational skills. The later statement is a claim about AI being “on a path” to human-substituting capabilities soon. It does not say whether the speaker now thinks AI will have those specific skills, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen Convincing the leading AI company to slow down would require luck.Now The speaker wants multiple AI companies to operate at the frontier with roughly similar capabilities.The earlier claim says that “convincing the leading AI company to slow down would require luck,” describing how difficult that would be. The later claim expresses a preference for “multiple different AI companies” at similar capability levels. These address different questions, and the later preference does not establish whether the speaker’s view of slowing down has changed.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker says non-disparagement agreements are more commonly tied to positive compensation, such as a bonus, than to forfeiture of already-earned equity.Now The speaker favors AI becoming commoditized rather than controlled by a monopoly or oligopoly.The earlier claim concerns how non-disparagement agreements are typically compensated, while the later claim concerns competition and concentration in the AI industry. They address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker guesses that AI systems' screen-parsing confusion will mostly disappear by the end of 2025, while they will still be unable to operate autonomously for long periods.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier claim concerns AI capability progress: screen-parsing confusion would mostly disappear by the end of 2025, while long-period autonomy would remain limited. The later claim concerns whether misaligned AIs could appear to behave correctly and evade detection. These address different questions, so the later statement does not establish a change of mind about the earlier prediction.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker's p(doom) estimate is around 70%, which he describes as infamously high.Now The speaker estimates that the maximum reliably controllable AI capability is roughly at the level of top human experts.The earlier statement is about the speaker’s p(doom) estimate (“around 70%”), while the later statement is about the AI capability level he considers reliably controllable (“roughly around top human expert level”). These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen In the scenario, AI researchers’ serial speed reaches about 90× human speed, after which other constraints become bottlenecks and additional serial-speed gains provide little benefit.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier statement describes a scenario in which serial speed reaches about 90× human speed and other constraints then limit its usefulness. The later statement is conditional: timelines would lengthen if progress hit a genuine barrier. It does not say that the earlier scenario’s bottleneck assumptions have changed.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen At 50 times human serial speed, five years of real time would give AI systems about 250 years of serial time, which the speaker considers sufficient to resolve organizational problems.Now The speaker predicts that an autonomous industrial system could grow at an accelerating rate, potentially doubling every year, every six months, and eventually every three months as technology improves.The earlier claim says that five years at 50× serial speed would be enough for AI systems to resolve organizational problems. The later claim predicts accelerating industrial growth, with systems “doubling every year, every six months” and potentially “every three months.” These address different questions, so the later prediction does not establish a change in the earlier view.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker thinks the default outcome is that leading AI developers continue their work without serious refocusing on safety or other priorities.Now The speaker prefers stopping AI development immediately to continuing along the current trajectory.The earlier claim is a prediction about what AI developers will do by default: “continue on without … any serious refocusing.” The later claim expresses what the speaker prefers: “I would rather just stop everything now” than continue on the current trajectory. A prediction and a preference are different questions, so these statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker believes aggregate opinion has underestimated the pace of AI technical progress and deployment.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The claims address different questions. The earlier statement says aggregate opinion underestimated the pace of AI progress and deployment; the later is conditional, saying timelines would lengthen if progress hit a genuine, theoretically predicted barrier. The later claim does not establish that such a barrier has occurred or that the speaker now thinks aggregate opinion was not underestimating progress.
• NOT COMPARABLE · 97%3 Apr → 8 SeptThen The forecast models a progression from automated coding to automation of the full research process, initially using teams of agents at about human level.Now The speakers consider whether, within the next few years, AI systems could perform a broad range of human-substituting knowledge work, including AI research, as humans in the cloud.The earlier statement describes a forecast model: first “automate the coding,” then “automate the whole research process” with roughly human-level teams of agents. The later passage asks whether AI could do human-substituting knowledge work “in the next few years”; it does not establish that Daniel believes this will happen. A question about possibility is not enough to show a change in his asserted position.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker says human groups are much more diverse than a data-center population of AI systems, which would mostly consist of literal copies of one another.Now The speaker thinks a future of many specialized AI agents would retain serious alignment concerns but might be somewhat safer because their interactions could be easier to oversee than those of identical, broadly knowledgeable clones.The earlier claim compares the diversity of human groups with data-center AIs that are mostly “literal copies.” The later claim considers whether specialized agents might be safer and easier to oversee than “clones of each other.” These statements are compatible, but they address different aspects—population diversity versus oversight and safety—so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 96%3 Apr → 8 SeptThen The speakers expect broad diffusion of AI through the economy, with many experiments, to take roughly one year rather than the roughly ten years suggested by the other person.Now The speaker thinks his AI timelines would lengthen substantially if AI progress encountered a genuine barrier that had been correctly predicted by theory.The earlier claim says broad AI diffusion through the economy was expected to take “something like one year” rather than ten. The later claim is conditional: encountering a “real actual barrier” would lengthen the speaker’s timelines. It does not establish that such a barrier was encountered or revise the earlier estimate, and the claims concern different aspects of progress.
• NOT COMPARABLE · 95%3 Apr → 8 SeptThen The speaker has less confidence in the progression from an advanced robot economy to nanobots and similar technologies than in the development of the robot economy.Now The speaker allows that if human-substituting AI and robotics remain very far away, AI may be more like a normal technology, comparable to the next internet.The earlier claim compares confidence in a near-term “robot economy” with confidence in later “nanobots.” The later claim discusses a conditional scenario—“if that level of AI and robots is still very, very far away”—and what AI might then resemble economically. It does not establish a changed view about the earlier comparison.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen In the scenario, an oversight committee comprising presidential appointees and company representatives, including the CEO, votes on high-level questions such as the goals assigned to the superintelligence.Now In the described war-game scenario, an incumbent president accelerates the push to achieve superintelligence before the next election to remain in control instead of his successor.The earlier statement describes a governance arrangement in which an oversight committee with presidential and company representatives votes on superintelligence goals. The later statement describes a different war-game scenario: an incumbent accelerates development “before the next election” to stay in charge. These concern different aspects of the scenario, so they do not establish a change in the speaker’s position on the same underlying question.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the hypothetical, discoveries since 2017 might have occurred about five times faster because small-scale experiments for testing ideas before large training runs would have run faster.Now The speaker describes a hypothetical cycle in which self-driving trucks, robot-built factories, humanoid-robot production, and chip-fab production could expand at accelerating rates, potentially doubling every year, six months, or three months as technology improves.The earlier claim estimates how much faster algorithmic discoveries since 2017 might have happened in a hypothetical scenario. The later claim describes a possible accelerating cycle of AI-driven industrial production. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers it fortunate that large language models and broad world understanding came before efforts to turn them into agents.Now The speaker says it is already somewhat easy to mistakenly believe AI alignment is solved, and expects this mistake to become easier as models become more sophisticated and situationally aware.The earlier claim says it was fortunate that broad world understanding in LLMs came before efforts to make them agents. The later claim says it will become easier to mistakenly think alignment is solved as models grow more sophisticated. These address different questions, and the later statement does not revise or deny the earlier assessment.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen In the described confrontation, an AI-company CEO could threaten legal and public opposition to the White House.Now Anthropic and OpenAI will oppose or be unhappy with the proposed transparency.The earlier claim concerns a CEO threatening legal and public opposition in a hypothetical White House confrontation. The later claim says Anthropic and OpenAI would be unhappy with a proposal to publish their training recipes. These are different scenarios and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen The speaker thinks some departing employees may have signed without realizing the paperwork required them not to criticize OpenAI, because the equity-loss warning appeared before the restriction in the document.Now After leaving OpenAI, Daniel Kokotajlo wanted to speak more freely and share with the public what industry insiders anticipated.The earlier statement speculates that some departing employees signed paperwork without realizing it barred them from criticizing OpenAI. The later statement says Daniel wanted to speak more freely after leaving. These concern different claims, and neither establishes a change in his view about what those employees knew or signed.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says many AI-company employees expect future AIs to be misaligned, while viewing current alignment problems as manageable and being resolved.Now The speaker predicts that people may find it extremely easy to overlook AI misalignment because the AIs appear to behave correctly.The earlier passage reports what “a lot of the people at the companies” think: that current alignment issues are “no big deal” or will be “ironed out.” The later passage is Daniel’s own prediction that misalignment could be hard to detect because AIs appear to be “doing everything right.” These are not clearly opposing claims about the same position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen Even if AI misalignment is not considered a problem, AI can still create a problem of power-seeking.Now AI alignment concerns whether an AI has its intended personality traits, goals, and values.The earlier claim says that “even if AI misalignment is not an issue,” AI could still create a “problem of power-seeking.” The later claim defines alignment as having intended “personality traits, goals, [and] values.” The later statement does not address the earlier power-concentration concern, so it does not show a change in position.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker considers it very plausible that specific government regulations on AI safety could backfire.Now The speaker predicts that the number of ways people could fail to detect misaligned AI will increase over time, making it very easy to fall into that trap.The earlier statement concerns whether specific government AI-safety regulations could backfire; the later concerns the growing difficulty of detecting misaligned AI. Neither claim establishes a position on the other's question, so they do not show a change of mind.
• NOT COMPARABLE · 99%3 Apr → 8 SeptThen SPEAKER_03 is uncertain whether OpenAI or Anthropic intended the interpretation that honesty is less important than harmlessness in the specification.Now The speaker thinks Anthropic and OpenAI will probably oppose the transparency plan because it could help competitors catch up.The claims address different questions. Earlier, the speaker is “not sure” whether Anthropic or OpenAI intended an interpretation that made honesty less important than harmlessness. Later, the speaker predicts they will “probably” oppose a transparency plan because it could help competitors catch up. The later prediction does not clarify or reverse the earlier uncertainty about the specification’s intended interpretation.
• NOT COMPARABLE · 98%3 Apr → 8 SeptThen The speaker says many AI-company employees expect future AIs to be misaligned, while viewing current alignment problems as manageable and being resolved.Now The speaker predicts that it will be extremely easy for AI systems to be misaligned without people realizing it, because the systems appear to be behaving correctly.The earlier statement reports what “a huge amount of those people” at AI companies believe about alignment problems; it does not establish Daniel’s own view. The later statement is his prediction that it will be “extremely easy” for AI systems to be misaligned without people realizing it. These are different claims, so they do not show whether his own position changed.
• NOT COMPARABLE · 91%3 Apr → 8 SeptThen The speaker doubts that a three- or six-month lead held by the leading U.S. AI project over a serious competitor would be used for good purposes.Now The speaker prefers an immediate halt to AI development over continuing along the current trajectory.The earlier claim concerns whether a U.S. project would use a three- or six-month lead “for good purposes.” The later claim states a preference to “stop everything now” rather than continue the current trajectory. These are related to AI policy, but they address different questions; the later preference does not establish a changed view about how such a lead would be used.

• 3 Apr 2025AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel KokotajloDwarkesh Podcast · guest
  241
• 8 Sept 2026AI 2040: Plan A report - Daniel Kokotajlo & Thomas LarsenMachine Learning Street Talk (MLST) · guest
  158