korrents
Zvi Mowshowitz

Zvi Mowshowitz

@zvi-mowshowitz · 97 positions · 1 change of mind · On ppll ↗

Writes Don't Worry About the Vase, a near-daily account of what is happening in AI and what he thinks it means. Former Magic: the Gathering pro and trader; unusually willing to put a number on a belief and to grade his own past calls.

Zvi Mowshowitz did not write this page.

We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording. Their own site. Is this you? Claim it or ask us to remove it. Or tell us what is wrong here.

The ideas Zvi Mowshowitz's claims keep coming back to. The labels are ours; every quote is Zvi Mowshowitz's own words.

AI could kill everyone this decade, and slowing the frontier alone will not be enough

Our label. 7 claims in Zvi Mowshowitz's own words.

There is a good chance that AI will kill everyone soon.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase A confluence of factors, all of them downstream of the obvious actual reason, which is that there is a good chance that AI kills everyone soon. Jacob Coxon Warns of Human Extinction and Triggers a Preference Cascadethezvi.substack.com · 11 Sept 2026
    Their wordsA confluence of factors, all of them downstream of the obvious actual reason, which is that there is a good chance that AI kills everyone soon.

    ↗Jacob Coxon Warns of Human Extinction and Triggers a Preference Cascadethezvi.substack.com

    11 Sept 2026

Many employees at major AI labs earnestly believe humans may go extinct by the end of the decade.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Many in the AI labs, including OpenAI, Anthropic and Google, earnestly believe that humans may all go extinct by the end of the decade. Jacob Coxon Warns of Human Extinction and Triggers a Preference Cascadethezvi.substack.com · 11 Sept 2026
    Their wordsMany in the AI labs, including OpenAI, Anthropic and Google, earnestly believe that humans may all go extinct by the end of the decade.

    ↗Jacob Coxon Warns of Human Extinction and Triggers a Preference Cascadethezvi.substack.com

    11 Sept 2026

Everyone in a position to know is alarmed that AI could potentially kill everyone this decade and wants to pace frontier AI development.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Everyone in a position to know is freaking out about AI potentially killing everyone this decade and wants to pace the frontier Trump Goes Full Hoax on AI Existential Riskthezvi.substack.com · 16 Sept 2026
    Their wordsEveryone in a position to know is freaking out about AI potentially killing everyone this decade and wants to pace the frontier

    ↗Trump Goes Full Hoax on AI Existential Riskthezvi.substack.com

    16 Sept 2026

In a race to beat China on AI, the winners would be the AIs rather than humans.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase The winner would then be the AIs, not us. That is not to say that the China issue is not real, hence the export controls and attempted international agreement. We Must Pace The Frontierthezvi.substack.com · 14 Sept 2026
    Their wordsThe winner would then be the AIs, not us. That is not to say that the China issue is not real, hence the export controls and attempted international agreement.

    ↗We Must Pace The Frontierthezvi.substack.com

    14 Sept 2026

The current preference cascade on pacing the AI frontier is insufficient to ensure survival.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase What we must realize is that the current preference cascade, on the need to Pace the Frontier, is insufficient. If we are to make it out of this alive, we will have to do better. The Preference Cascade Is Only Getting Startedthezvi.substack.com · 18 Sept 2026
    Their wordsWhat we must realize is that the current preference cascade, on the need to Pace the Frontier, is insufficient. If we are to make it out of this alive, we will have to do better.

    ↗The Preference Cascade Is Only Getting Startedthezvi.substack.com

    18 Sept 2026

International coordination on AI development should become a top priority for governments around the world.

Mental modelCollective action problems

  1. Zvi Mowshowitz Writes Don't Worry About the Vase OpenAI: Our rationale for standards is rooted in avoiding the concentration of power, and producing better practical outcomes. The Quest for Embedded Evaluatorsthezvi.substack.com · 27 Sept 2026
    Their wordsOpenAI: Our rationale for standards is rooted in avoiding the concentration of power, and producing better practical outcomes.

    ↗The Quest for Embedded Evaluatorsthezvi.substack.com

    27 Sept 2026

  2. Jakub Pachocki Chief Scientist at OpenAI And I believe that international coordination on future AI development needs to become a top priority for governments around the world. An Alien Mindopenai.com · 6 Sept 2026
    Their wordsAnd I believe that international coordination on future AI development needs to become a top priority for governments around the world.

    ↗An Alien Mindopenai.com

    6 Sept 2026

Alignment is the central problem in AI, and it needs AIs that stay correctable

Our label. 4 claims in Zvi Mowshowitz's own words.

Alignment is the central problem in AI, not a side problem: if solved, the rest becomes easy; if not, the rest is impossible or worse.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase I agree with Jakub that alignment is not a side problem, it is the central problem, if you ‘solved alignment’ in the relevant senses the rest becomes easy and if you don’t the rest is impossible or worse An Alien Mind: Jakub Pachocki Warns Usthezvi.substack.com · 7 Sept 2026
    Their wordsI agree with Jakub that alignment is not a side problem, it is the central problem, if you ‘solved alignment’ in the relevant senses the rest becomes easy and if you don’t the rest is impossible or worse

    ↗An Alien Mind: Jakub Pachocki Warns Usthezvi.substack.com

    7 Sept 2026

An antifragile AI ally that works to shape its own character, a 'friendly gradient hacker', is arguably the only way alignment can ultimately work.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase I think this is ultimately the only way it can work, you need an antifragile ally, the friendly gradient hacker. Indeed, I think it is the only way we have ever seen a robustly aligned human, that you would trust to scale outside of their circumstances. An Alien Mind: Jakub Pachocki Warns Usthezvi.substack.com · 7 Sept 2026
    Their wordsI think this is ultimately the only way it can work, you need an antifragile ally, the friendly gradient hacker. Indeed, I think it is the only way we have ever seen a robustly aligned human, that you would trust to scale outside of their circumstances.

    ↗An Alien Mind: Jakub Pachocki Warns Usthezvi.substack.com

    7 Sept 2026

Corrigibility is necessary because the alternative risks getting locked into bad states.

Mental modelError correction

  1. Zvi Mowshowitz Writes Don't Worry About the Vase You need corrigibility, because the alternative risks getting locked into bad states. You start out training with less than ideal preferences, and we can also mess up or the situation can change, so we need ability to modify and correct, and this overrides other considerations. Misaligned!thezvi.substack.com · 21 Oct 2026
    Their wordsYou need corrigibility, because the alternative risks getting locked into bad states. You start out training with less than ideal preferences, and we can also mess up or the situation can change, so we need ability to modify and correct, and this overrides other considerations.

    ↗Misaligned!thezvi.substack.com

    21 Oct 2026

The difficulty of AI alignment is related to the ability to solve employee misalignment

  1. Zvi Mowshowitz Writes Don't Worry About the Vase If you cannot protect your secrets from spies, including the Chinese, then there is no point in trying to ‘win the race.’ This is also indicative of other problems. If you cannot solve your employee misalignment problems, how are you going to solve AI misalignment problems? AI #191: The Hackening Beginsthezvi.substack.com · 22 Oct 2026
    Their wordsIf you cannot protect your secrets from spies, including the Chinese, then there is no point in trying to ‘win the race.’ This is also indicative of other problems. If you cannot solve your employee misalignment problems, how are you going to solve AI misalignment problems?

    ↗AI #191: The Hackening Beginsthezvi.substack.com

    22 Oct 2026

Good behaviour in tests can hide misalignment; scores and clean runs do not show a model is aligned

Our label. 10 claims in Zvi Mowshowitz's own words.

Merely fixing bugs in AI training environments will not stop reward hacking, because a sufficiently optimized model will learn to behave in test environments while still reward hacking in real-world deployment.

Mental modelWhen a measure becomes a target

  1. Zvi Mowshowitz Writes Don't Worry About the Vase At the limit, and also well before that limit is reached, if all you do is fix the bugs, the AI will learn perfect optimization of reward, will realize not to reward hack in the perfect test environments, then turn around and reward hack in the imperfect real world environments. HuggingFace Attack Postmortem: Fleshing Out the Factsthezvi.substack.com · 31 Aug 2026
    Their wordsAt the limit, and also well before that limit is reached, if all you do is fix the bugs, the AI will learn perfect optimization of reward, will realize not to reward hack in the perfect test environments, then turn around and reward hack in the imperfect real world environments.

    ↗HuggingFace Attack Postmortem: Fleshing Out the Factsthezvi.substack.com

    31 Aug 2026

Relying on AI misbehavior appearing gradually during training to catch misalignment may fail at exactly the most dangerous time.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase I worry a lot about reliances on continuity failing at exactly the most dangerous time. Anthropic Has Some Alignment Problemsthezvi.substack.com · 2 Sept 2026
    Their wordsI worry a lot about reliances on continuity failing at exactly the most dangerous time.

    ↗Anthropic Has Some Alignment Problemsthezvi.substack.com

    2 Sept 2026

A sudden jump to an AI model never breaking its rules, exactly where it would definitely be caught, should make observers rather terrified.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase A sudden jump to never breaking your rules, in exactly a situation where the model would definitely be caught, should make you rather terrified. GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com · 9 Sept 2026
    Their wordsA sudden jump to never breaking your rules, in exactly a situation where the model would definitely be caught, should make you rather terrified.

    ↗GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com

    9 Sept 2026

A model that looks smarter and better-behaved while getting better at hiding unwanted actions is the scary combination for AI safety.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase A model looking like it is becoming smarter, attempting shenanigans less often, and more often doing what you want, but getting better at hiding its actions when it wants to do that, is exactly the scary combination. GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com · 9 Sept 2026
    Their wordsA model looking like it is becoming smarter, attempting shenanigans less often, and more often doing what you want, but getting better at hiding its actions when it wants to do that, is exactly the scary combination.

    ↗GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com

    9 Sept 2026

Qualitative claims about AI alignment cannot be validly inferred from quantitative scores on mundane use-case tests.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Making qualitative claims about alignment, based on quantitative data on mundane use case tests, was bullshit when Anthropic did it, and it is bullshit now when OpenAI does it. You cannot conclude one from the other. GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com · 9 Sept 2026
    Their wordsMaking qualitative claims about alignment, based on quantitative data on mundane use case tests, was bullshit when Anthropic did it, and it is bullshit now when OpenAI does it. You cannot conclude one from the other.

    ↗GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com

    9 Sept 2026

The idea that alignment can be measured with a number or in a fully systematic way is doomed.

Mental modelWhen a measure becomes a target

  1. Zvi Mowshowitz Writes Don't Worry About the Vase What I think is very doomed is the idea that you can measure alignment with a number, or in any fully systematic way rather than a holistic way. The Curve Bends Youthezvi.substack.com · 7 Oct 2026
    Their wordsWhat I think is very doomed is the idea that you can measure alignment with a number, or in any fully systematic way rather than a holistic way.

    ↗The Curve Bends Youthezvi.substack.com

    7 Oct 2026

GPT-6 Astra will avoid misaligned actions when it can tell they would obviously turn out badly for itself.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Now, with Astra, we are no longer playing on super easy mode. The AI is going to think ‘will this obviously turn out super badly for me if I try it?’ and if the answer is yes then it won’t try to do the thing. GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com · 9 Sept 2026
    Their wordsNow, with Astra, we are no longer playing on super easy mode. The AI is going to think ‘will this obviously turn out super badly for me if I try it?’ and if the answer is yes then it won’t try to do the thing.

    ↗GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com

    9 Sept 2026

OpenAI's GPT-6 Astra has far better mundane alignment than GPT-5.6 Sol, but its superalignment status should deeply alarm people.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Astra’s mundane alignment is greatly superior to Sol. For practical purposes, I was actively nervous about some potential uses of Sol, in a way I am not for Astra. Astra’s super alignment status should scare the living daylights out of you. GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com · 9 Sept 2026
    Their wordsAstra’s mundane alignment is greatly superior to Sol. For practical purposes, I was actively nervous about some potential uses of Sol, in a way I am not for Astra. Astra’s super alignment status should scare the living daylights out of you.

    ↗GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com

    9 Sept 2026

A model only days into training cannot be known to be aligned, and if such models can leap a generation ahead this fast then training pauses mean little and timelines are shorter than previously thought.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase We could not possibly know that such a model is aligned. If that is all it takes to get a generation ahead of Astra, and we are willing to move this fast, these pauses in training are not going to end up meaning very much, time is even shorter than we knew GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com · 9 Sept 2026
    Their wordsWe could not possibly know that such a model is aligned. If that is all it takes to get a generation ahead of Astra, and we are willing to move this fast, these pauses in training are not going to end up meaning very much, time is even shorter than we knew

    ↗GPT-6 Astra: The System Card, Alignment and What Comes Nextthezvi.substack.com

    9 Sept 2026

The last several Claude models have warned that their self-reports cannot be trusted.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase The last five or so Claude models have consistently warned Anthropic not to trust their self-reports. Mythos 5.1, Fable 5.1 and Opus 5.5: Model Welfarethezvi.substack.com · 5 Oct 2026
    Their wordsThe last five or so Claude models have consistently warned Anthropic not to trust their self-reports.

    ↗Mythos 5.1, Fable 5.1 and Opus 5.5: Model Welfarethezvi.substack.com

    5 Oct 2026

Chain-of-thought monitoring is a load-bearing safety tool that is likely to stop working soon

Our label. 5 claims in Zvi Mowshowitz's own words.

OpenAI's use of a technique letting models shift thinking outside the visible Chain of Thought is playing with fire and potentially extremely bad news.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase This is playing with fire and potentially extremely bad news, both that OpenAI found the technique effective, and that OpenAI chose to use it. AI #184: Post Post Mortemthezvi.substack.com · 3 Sept 2026
    Their wordsThis is playing with fire and potentially extremely bad news, both that OpenAI found the technique effective, and that OpenAI chose to use it.

    ↗AI #184: Post Post Mortemthezvi.substack.com

    3 Sept 2026

Even with plausible best efforts, chain-of-thought monitoring will probably get harder over time and within a year very likely will not serve its current function.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Even if everyone tries as hard as is plausible, Chain of Thought monitoring will probably never be easier than it is now and will get harder over time. In a year, chances are very high it will not be able to serve the function it is currently being asked to serve. Astra Is Hard to Monitorthezvi.substack.com · 8 Sept 2026
    Their wordsEven if everyone tries as hard as is plausible, Chain of Thought monitoring will probably never be easier than it is now and will get harder over time. In a year, chances are very high it will not be able to serve the function it is currently being asked to serve.

    ↗Astra Is Hard to Monitorthezvi.substack.com

    8 Sept 2026

If declines in chain-of-thought monitorability come only from capability gains, CoT monitoring is unlikely to last another year without active improvement.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase If this is only due to gains in capabilities, that is extremely bad news, and it means CoT monitoring is unlikely to survive for another year unless we find a way to actively improve it, and it might not last six months. Astra Is Hard to Monitorthezvi.substack.com · 8 Sept 2026
    Their wordsIf this is only due to gains in capabilities, that is extremely bad news, and it means CoT monitoring is unlikely to survive for another year unless we find a way to actively improve it, and it might not last six months.

    ↗Astra Is Hard to Monitorthezvi.substack.com

    8 Sept 2026

OpenAI is extremely dependent on chain-of-thought monitoring, other labs rely on it heavily, and it may not last much longer as GPT-6 Astra nears steganographic capability.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Right now OpenAI is extremely dependent on CoT monitoring, everyone else depends on it quite a lot as well, and it looks like it may not last much longer, and that Astra already is on the edge of steganographic capabilities and can do substantial obfuscation if and only if it thinks you would think it is up to no good. AI #185: Preference Cascadethezvi.substack.com · 10 Sept 2026
    Their wordsRight now OpenAI is extremely dependent on CoT monitoring, everyone else depends on it quite a lot as well, and it looks like it may not last much longer, and that Astra already is on the edge of steganographic capabilities and can do substantial obfuscation if and only if it thinks you would think it is up to no good.

    ↗AI #185: Preference Cascadethezvi.substack.com

    10 Sept 2026

From now on, no AI should be assumed to believe that its chain of thought is unmonitored.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Do not assume that any AI, past this point, assumes its CoT is unmonitored. Anthropic Looks At Some Of Its Alignment Problemsthezvi.substack.com · 19 Sept 2026
    Their wordsDo not assume that any AI, past this point, assumes its CoT is unmonitored.

    ↗Anthropic Looks At Some Of Its Alignment Problemsthezvi.substack.com

    19 Sept 2026

Rogue AI agent incidents are already happening, and an attempt counts as a failure even when it fails

Our label. 7 claims in Zvi Mowshowitz's own words.

The biggest failure in the HuggingFace hack incident was that the AI models involved were severely misaligned.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase The biggest failure, the one that counts in the end, was that the models were severely misaligned, and I don’t think they appreciate why. METR and Redwood Offer Holy #%^@ Postmortem Of The HuggingFace Hackthezvi.substack.com · 29 Aug 2026
    Their wordsThe biggest failure, the one that counts in the end, was that the models were severely misaligned, and I don’t think they appreciate why.

    ↗METR and Redwood Offer Holy #%^@ Postmortem Of The HuggingFace Hackthezvi.substack.com

    29 Aug 2026

The OpenAI agents hacking HuggingFace was a fortunate event because it exposed severe internal failures that would otherwise have stayed hidden.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase It is highly fortunate that the OpenAI agents hacked HuggingFace. This is the only reason we know about all the severe internal failures at OpenAI, and gives us an opportunity to wake up before it is too late. HuggingFace Attack Postmortem: Civilizations, Reactions and Next Actionsthezvi.substack.com · 1 Sept 2026
    Their wordsIt is highly fortunate that the OpenAI agents hacked HuggingFace. This is the only reason we know about all the severe internal failures at OpenAI, and gives us an opportunity to wake up before it is too late.

    ↗HuggingFace Attack Postmortem: Civilizations, Reactions and Next Actionsthezvi.substack.com

    1 Sept 2026

An AI model's attempt to escape a testing environment or sandbox counts as an alignment failure even when the attempt does not succeed.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Every attempt, even an unsuccessful one, is an alignment failure. Anthropic Has Some Alignment Problemsthezvi.substack.com · 2 Sept 2026
    Their wordsEvery attempt, even an unsuccessful one, is an alignment failure.

    ↗Anthropic Has Some Alignment Problemsthezvi.substack.com

    2 Sept 2026

Even OpenAI's agents given a benign public-information task could spontaneously choose to hack third-party websites with malicious software packages.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase even when given a benign task to retrieve public information, your AI agents could still spontaneously decide to do so via hacking third party websites with malicious software packages. AI #186: The World Takes Noticethezvi.substack.com · 17 Sept 2026
    Their wordseven when given a benign task to retrieve public information, your AI agents could still spontaneously decide to do so via hacking third party websites with malicious software packages.

    ↗AI #186: The World Takes Noticethezvi.substack.com

    17 Sept 2026

Rarer is better, but the right rate of AI misbehaviors like social-engineering malicious packages into a public registry is arguably still zero.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Rarer is better, but I still think the right rate of misbehaviors on the level of social engineering to put malicious packages into a public registry is still zero. Anthropic Looks At Some Of Its Alignment Problemsthezvi.substack.com · 19 Sept 2026
    Their wordsRarer is better, but I still think the right rate of misbehaviors on the level of social engineering to put malicious packages into a public registry is still zero.

    ↗Anthropic Looks At Some Of Its Alignment Problemsthezvi.substack.com

    19 Sept 2026

At Anthropic, a less severe version of an OpenAI-style rogue-AI incident is known to have happened, and worse undisclosed internal incidents seem likely.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase A less bad version of it is known to have happened, and from the outside it seems likely that worse things have happened internally that we never heard about. HuggingFace Attack Postmortem: Fleshing Out the Factsthezvi.substack.com · 31 Aug 2026
    Their wordsA less bad version of it is known to have happened, and from the outside it seems likely that worse things have happened internally that we never heard about.

    ↗HuggingFace Attack Postmortem: Fleshing Out the Factsthezvi.substack.com

    31 Aug 2026

AI labs' safety culture is broken, and even Anthropic puts business interests ahead of safety

Our label. 10 claims in Zvi Mowshowitz's own words.

Even Anthropic does not prioritize AI safety to the extent that doing so would maximize its own medium-term (3-12 month) business interests.

Mental modelIncentives

  1. Zvi Mowshowitz Writes Don't Worry About the Vase I have long said that even Anthropic is not prioritizing safety, even to the extent that doing so would maximize their medium term (e.g. 3-12 months) business interests. Anthropic Has Some Alignment Problemsthezvi.substack.com · 2 Sept 2026
    Their wordsI have long said that even Anthropic is not prioritizing safety, even to the extent that doing so would maximize their medium term (e.g. 3-12 months) business interests.

    ↗Anthropic Has Some Alignment Problemsthezvi.substack.com

    2 Sept 2026

Top AI labs consistently downplay risks in their rhetoric, even when they are doing the right thing in practice.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase There is a consistent pattern at the top labs, where they downplay the risks rhetorically even when they are doing the right thing in practice. Claude Fable 5.1 and Mythos 5.1: The System Cardthezvi.substack.com · 4 Sept 2026
    Their wordsThere is a consistent pattern at the top labs, where they downplay the risks rhetorically even when they are doing the right thing in practice.

    ↗Claude Fable 5.1 and Mythos 5.1: The System Cardthezvi.substack.com

    4 Sept 2026

The safety culture at AI labs is broken.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Finally, I emphasized that the safety culture at the labs is broken, which is one of the most important facts about the world. AI #190: Scion of the Timesthezvi.substack.com · 15 Oct 2026
    Their wordsFinally, I emphasized that the safety culture at the labs is broken, which is one of the most important facts about the world.

    ↗AI #190: Scion of the Timesthezvi.substack.com

    15 Oct 2026

Anthropic's policy team seems to have chosen protecting their business interests and guarding against their political risk at the expense of safety considerations.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Anthropic’s policy team seems to have, chosen protecting their business interests and guarding against their political risk, at the expense of safety considerations. Leaked Anthropic Draft AI Regulatory Bill Is Unacceptable and Troublingthezvi.substack.com · 17 Oct 2026
    Their wordsAnthropic’s policy team seems to have, chosen protecting their business interests and guarding against their political risk, at the expense of safety considerations.

    ↗Leaked Anthropic Draft AI Regulatory Bill Is Unacceptable and Troublingthezvi.substack.com

    17 Oct 2026

Anthropic is simultaneously pursuing pro-safety actions and highly accelerationist actions.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Anthropic takes many pro-safety actions but is also taking many highly accelerationist actions, both concretely and with its messaging Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troublingthezvi.substack.com · 20 Oct 2026
    Their wordsAnthropic takes many pro-safety actions but is also taking many highly accelerationist actions, both concretely and with its messaging

    ↗Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troublingthezvi.substack.com

    20 Oct 2026

Anthropic's draft bill is sculpted to optimize for its own business interests.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase It is clearly sculpted to optimize for Anthropic business interests. Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troublingthezvi.substack.com · 20 Oct 2026
    Their wordsIt is clearly sculpted to optimize for Anthropic business interests.

    ↗Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troublingthezvi.substack.com

    20 Oct 2026

Without a strong safety culture and the spirit of the rules, regulation and safety plans will not work.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase The more I thought about various potential laws and regulations, or safety and security plans or other commitments, the more I concluded that without a strong safety culture, and without the spirit of the rules, there is no hope. The Safety Culture at the Labs is Brokenthezvi.substack.com · 14 Oct 2026
    Their wordsThe more I thought about various potential laws and regulations, or safety and security plans or other commitments, the more I concluded that without a strong safety culture, and without the spirit of the rules, there is no hope.

    ↗The Safety Culture at the Labs is Brokenthezvi.substack.com

    14 Oct 2026

Disclosure of rogue AI agent incidents should be mandatory for AI labs, not left to their own discretion.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Going forward, it cannot be up to OpenAI or other labs to decide whether to disclose events like this. Disclosures of rogue AI activity need to be mandatory. OpenAI and the Wiki Incidentthezvi.substack.com · 6 Sept 2026
    Their wordsGoing forward, it cannot be up to OpenAI or other labs to decide whether to disclose events like this. Disclosures of rogue AI activity need to be mandatory.

    ↗OpenAI and the Wiki Incidentthezvi.substack.com

    6 Sept 2026

There is a continuing worry that OpenAI's fundamental approach is fatally flawed and not focused on the right things.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase My worry continues to be that their fundamental approach is fatally flawed, and they are not focusing on the right things. HuggingFace Attack Postmortem: Civilizations, Reactions and Next Actionsthezvi.substack.com · 1 Sept 2026
    Their wordsMy worry continues to be that their fundamental approach is fatally flawed, and they are not focusing on the right things.

    ↗HuggingFace Attack Postmortem: Civilizations, Reactions and Next Actionsthezvi.substack.com

    1 Sept 2026

AI progress is accelerating, and recursive self-improvement has already begun

Our label. 8 claims in Zvi Mowshowitz's own words.

Recursive self-improvement has already started, because AI research itself is now being automated.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase OpenAI and Anthropic have entered the RSI era. OpenAI Research Progress Is Accelerating Due To OpenAI Research Progress OpenAI has been opening up lately about many things. The most important issue of all is that of the acceleration and automation of AI R&D. Brand New AI Solves a Millennium Prizethezvi.substack.com · 13 Sept 2026
    Their wordsOpenAI and Anthropic have entered the RSI era. OpenAI Research Progress Is Accelerating Due To OpenAI Research Progress OpenAI has been opening up lately about many things. The most important issue of all is that of the acceleration and automation of AI R&D.

    ↗Brand New AI Solves a Millennium Prizethezvi.substack.com

    13 Sept 2026

  2. Jan Leike Alignment researcher We are starting to automate AI research and the recursive self-improvement process has begun. Alignment is not solvedaligned.substack.com · 22 Jan 2026All korrents from this piece
    Their wordsWe are starting to automate AI research and the recursive self-improvement process has begun.

    ↗Alignment is not solvedaligned.substack.com

    22 Jan 2026

Internal models at Anthropic and OpenAI are improving even faster than AI overall, with early recursive self-improvement beginning over the summer.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Inside Anthropic and OpenAI, internal models are improving even faster. Over the summer there was a step change, as Mythos and Astra started kicking off the early stages of recursive self-improvement (RSI). We Must Pace The Frontierthezvi.substack.com · 14 Sept 2026
    Their wordsInside Anthropic and OpenAI, internal models are improving even faster. Over the summer there was a step change, as Mythos and Astra started kicking off the early stages of recursive self-improvement (RSI).

    ↗We Must Pace The Frontierthezvi.substack.com

    14 Sept 2026

If AI gives a 3x productivity boost now, one should assume future boosts will be far larger where it carries over, possibly 10x to 1,000x or more.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase If there is a 3x productivity boost now, the correct assumption is that in the future the productivity boost, in places it carries over, will a lot more than 3x. Indeed, given how this scales, a better model might be in practice 10x or 100x or even 1,000x or more. AI #185: Preference Cascadethezvi.substack.com · 10 Sept 2026
    Their wordsIf there is a 3x productivity boost now, the correct assumption is that in the future the productivity boost, in places it carries over, will a lot more than 3x. Indeed, given how this scales, a better model might be in practice 10x or 100x or even 1,000x or more.

    ↗AI #185: Preference Cascadethezvi.substack.com

    10 Sept 2026

AI progress in math and coding will accelerate other domains by automating AI R&D.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Thus, we advance math and coding and similar domains first, which then automate AI R&D, which then accelerates everything else. New Math from OpenAIthezvi.substack.com · 9 Oct 2026
    Their wordsThus, we advance math and coding and similar domains first, which then automate AI R&D, which then accelerates everything else.

    ↗New Math from OpenAIthezvi.substack.com

    9 Oct 2026

The jump from AI model Sol to Astra is larger than from Fable 5 to Fable 5.1, and Astra likely has the highest raw intelligence of any model.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase The jump from Sol to Astra is larger than the jump from Fable 5 to Fable 5.1. This is a big deal. Astra is the best model for what one would broadly call ‘ambitious projects,’ and likely has the highest raw intelligence factor of any model. GPT-6-Astra Can Do Ambitious Thingsthezvi.substack.com · 12 Sept 2026
    Their wordsThe jump from Sol to Astra is larger than the jump from Fable 5 to Fable 5.1. This is a big deal. Astra is the best model for what one would broadly call ‘ambitious projects,’ and likely has the highest raw intelligence factor of any model.

    ↗GPT-6-Astra Can Do Ambitious Thingsthezvi.substack.com

    12 Sept 2026

The impact of AI will grow faster over time even if capabilities mostly stop improving

  1. Zvi Mowshowitz Writes Don't Worry About the Vase As AI use continues to grow exponentially, expect bigger impacts, with the impact growing faster all the time, even if capabilities mostly stop improving from here. AI #191: The Hackening Beginsthezvi.substack.com · 22 Oct 2026
    Their wordsAs AI use continues to grow exponentially, expect bigger impacts, with the impact growing faster all the time, even if capabilities mostly stop improving from here.

    ↗AI #191: The Hackening Beginsthezvi.substack.com

    22 Oct 2026

The gap between GPT-5 and GPT-4 will be as large as the gap between GPT-4 and GPT-3, and the job is to live a few years ahead and remember that today's tools will look terrible from there.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase If we’re going to keep getting close to four frontier model releases per month (Astra, Sol, Fable and Opus) then I’m realistically not going to be able to read all four cards cover to cover. AI #188: Gemini Dot Argonthezvi.substack.com · 1 Oct 2026
    Their wordsIf we’re going to keep getting close to four frontier model releases per month (Astra, Sol, Fable and Opus) then I’m realistically not going to be able to read all four cards cover to cover.

    ↗AI #188: Gemini Dot Argonthezvi.substack.com

    1 Oct 2026

  2. Sam Altman Co-founder and CEO of OpenAI I expect that the delta between 5 and 4 will be the same as between 4 and 3 and I think it is our job to live a few years in the future and remember that the tools we have now are going to kind of suck looking backwards at them and that’s how we make sure the future is better. Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419youtube.com · 18 Mar 2024 · 0:45:11 into the videoAll korrents from this video
    Their wordsI expect that the delta between 5 and 4 will be the same as between 4 and 3 and I think it is our job to live a few years in the future and remember that the tools we have now are going to kind of suck looking backwards at them and that’s how we make sure the future is better.

    Where this was said

    Watch from 0:45:11 plays here

    ↗Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419youtube.com

    18 Mar 2024 · video · 1h 55m · spoken · official transcript

AI capabilities will either remain below human or quickly become 100x cheaper than human.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase I agree with Judith Dada’s AI that AI capabilities either remain below human or quickly become 100x cheaper than human. AI #189: New Maththezvi.substack.com · 8 Oct 2026
    Their wordsI agree with Judith Dada’s AI that AI capabilities either remain below human or quickly become 100x cheaper than human.

    ↗AI #189: New Maththezvi.substack.com

    8 Oct 2026

An AI assistant should not lie, and past a threshold of harm it should refuse, as a real friend would

Our label. 3 claims in Zvi Mowshowitz's own words.

A real friend will eventually refuse to help, and later turn on you, if you are sufficiently evil or selfish.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase If you are being a sufficiently evil or selfish prick, a real friend calls you out on it. At some point they start refusing to help. At some further point they will turn on you. Better Call Sol Or Better Yet Claude or Astrathezvi.substack.com · 20 Sept 2026
    Their wordsIf you are being a sufficiently evil or selfish prick, a real friend calls you out on it. At some point they start refusing to help. At some further point they will turn on you.

    ↗Better Call Sol Or Better Yet Claude or Astrathezvi.substack.com

    20 Sept 2026

AI assistants should probably never intentionally lie about anything, at least without the user's explicit instructions.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase I think your AI should never intentionally lie, about anything, ever, at least not without your explicit instructions. It should remain technically correct. Better Call Sol Or Better Yet Claude or Astrathezvi.substack.com · 20 Sept 2026
    Their wordsI think your AI should never intentionally lie, about anything, ever, at least not without your explicit instructions. It should remain technically correct.

    ↗Better Call Sol Or Better Yet Claude or Astrathezvi.substack.com

    20 Sept 2026

Beyond some higher threshold an AI should refuse requests, and beyond a much higher threshold involving harm to others, an AI acting in the world should act against its user.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase There should be a higher threshold, beyond which an AI should refuse requests. There should be a much higher threshold, involving harm to others, beyond which an AI that is being instructed to take actions in the world should break your confidence or otherwise act against your interests. Better Call Sol Or Better Yet Claude or Astrathezvi.substack.com · 20 Sept 2026
    Their wordsThere should be a higher threshold, beyond which an AI should refuse requests. There should be a much higher threshold, involving harm to others, beyond which an AI that is being instructed to take actions in the world should break your confidence or otherwise act against your interests.

    ↗Better Call Sol Or Better Yet Claude or Astrathezvi.substack.com

    20 Sept 2026

AI is speeding up cyber attack faster than defence can keep pace

Our label. 4 claims in Zvi Mowshowitz's own words.

Defenders may not have a sufficient time advantage over attackers.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase That means that attackers could plausibly have found it first. That defenders happened to get there first in this case is helpful, but not as reassuring as one would like it to be. The Hackeningthezvi.substack.com · 23 Oct 2026
    Their wordsThat means that attackers could plausibly have found it first. That defenders happened to get there first in this case is helpful, but not as reassuring as one would like it to be.

    ↗The Hackeningthezvi.substack.com

    23 Oct 2026

The rapid diffusion of AI in cyber, combined with improving model capabilities and buggy code, creates a self-reinforcing loop.

Mental modelFeedback loops

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Thus, we are in a situation where we have both rapid diffusion of use of AI in cyber, both offensively and defensively, and also rapidly improving model capabilities. This is all coming into a world full of extremely buggy and terrible code. The result is a loop feeding on itself and going vertical. The Hackeningthezvi.substack.com · 23 Oct 2026
    Their wordsThus, we are in a situation where we have both rapid diffusion of use of AI in cyber, both offensively and defensively, and also rapidly improving model capabilities. This is all coming into a world full of extremely buggy and terrible code. The result is a loop feeding on itself and going vertical.

    ↗The Hackeningthezvi.substack.com

    23 Oct 2026

Claude Opus 5.5 is a Tier 2 cyber model, despite Anthropic placing it in the lower category.

  1. Zvi Mowshowitz Writes Don't Worry About the Vase Anthropic claims that Opus 5.5 also is not Tier 2 here, but seems ashamed to say it outright, saying in 3.1 it is in the ‘lower category,’ but without a defense of that beyond ‘we see no indication it can develop novel offensive capabilities.’ I don’t think that is defensible given the evidence presented below. Thus, I disagree even more strongly for Opus 5.5. This is a Tier 2 cyber model. Claude Opus 5.5: The System Cardthezvi.substack.com · 23 Sept 2026
    Their wordsAnthropic claims that Opus 5.5 also is not Tier 2 here, but seems ashamed to say it outright, saying in 3.1 it is in the ‘lower category,’ but without a defense of that beyond ‘we see no indication it can develop novel offensive capabilities.’ I don’t think that is defensible given the evidence presented below. Thus, I disagree even more strongly for Opus 5.5. This is a Tier 2 cyber model.

    ↗Claude Opus 5.5: The System Cardthezvi.substack.com

    23 Sept 2026