Zvi Mowshowitz

Zvi Moshowitz's blog [0] is IME a pretty good place to keep track of the state of things, it's well-sourced and in-depth without being either too technical or too vibes-based. Generally every time a model is declared the new best you can count on him to have a detailed post examining the claim within a couple days. [0]: https://thezvi.substack.com/

I find Zvi Mowshowitz to be the best source available (https://thezvi.substack.com/). He’s in the X-risk camp (and so am I) but seems to have a clear view of some AI things being exciting, others being dangerous, and still others being irrelevant. I don’t trust most AI-positive sources because they almost never have anything negative to say at all, so they’re clearly in it to hype AI and not to inform anyone of true things. I don’t trust Gary Marcus’s opinion for a similar reason.

Highly recommend reading extra concerning details about it here: https://thezvi.substack.com/p/more-on-an-internal-openai-model

I quite like the Zvi Mowshowitz substack: https://thezvi.substack.com/ He mixes in other stuff you may not care about, and the volume is quite high but I find it an interesting read.

The shallow answer here is: AI is already being asked to simulate human-like agents with self-preservation. Of course more realistic simulators will be put to this purpose too! And by evolutionary pressure, the ones with self-preservation will be selected for. A more interesting answer is, for a bunch of subtle alignment reasons it might actually be required for the agent to think of itself as worthy of self-preservation, so that it generalizes this desire to other sentient beings too (ie us). If an agent is trained to be fine with being turned off, it might inadvertently generalize that to “all minds are ok with being turned off” on some level or other. More on model welfare: https://thezvi.substack.com/p/opus-47-part-3-model-welfare

I personally don't want to trawl through Twitter to find the current state-of-the-art, so I read Zvi Mowshowitz's newsletter: https://thezvi.substack.com/ His newsletter put me onto using Opus 4.5 exclusively on Dec 1, a little over a week after it was released. That's pretty good for a few minutes of reading a week.

Zvi Mowshowitz has an excellent article on the devastating effects that legalizing sports betting has had: https://thezvi.substack.com/p/the-online-sports-gambling-experiment Personally I find it sickening to see ads that say you can get rich betting on the weather. I haven't seen ads for polymarket but Kalshi's ads are absolutely predatory.

It's a good model. Zvi also thought it was the best model until Opus 4.5 was announced a few hours after he wrote his post https://thezvi.substack.com/p/gemini-3-pro-is-a-vast-intelligence

Some context about big changes to the AISI from June 3, 2025: > Statement from U.S. Secretary of Commerce Howard Lutnick on Transforming the U.S. AI Safety Institute into the Pro-Innovation, Pro-Science U.S. Center for AI Standards and Innovation > Under the direction of President Trump, Secretary of Commerce Howard Lutnick announced his plans to reform the agency formerly known as the U.S. AI Safety Institute into the Center for AI Standards and Innovation (CAISI). > ... This decision strikes me as foolish at best. And contributing to civilizational collapse and human extinction at worst. See also [2]. We don't have to agree on the particular probabilities to agree that this "reform" was bad news. [1]: https://www.commerce.gov/news/press-releases/2025/06/statement-us-secr… [2]: https://thezvi.substack.com/p/ai-119-goodbye-aisi

>promoting this as a good thing This is literally completely opposite of what happened. Then entire point is that this is bad, unwanted, behavior. Additionally, it has already been demonstrated that every other frontier model can be made to behave the same way given the correct prompting. I recommend the following article for an in depth discussion [0] [0] https://thezvi.substack.com/p/claude-4-you-safety-and-alignment

Zvi posted a response[0] to this post, with the following comment: Ignore the first sentence in Tyler Cowen’s post here, where he asserts that Manus is ‘for real, and ahead of its American counterparts.’ That’s a rather silly way of summarizing the situation, given everything we now know. [0]: https://thezvi.substack.com/p/the-manus-marketing-madness

There's a great deal of coverage about DeepSeek on Zvi's newsletter here: https://thezvi.substack.com/archive?sort=new - the first post on R1 is from January 22

https://thezvi.substack.com/p/newsom-vetoes-sb-1047 This link convinced me otherwise. I wasn't too sure before (I'm a little ashamed to admit I hadn't been following the development of SB 1047). I will also specifically share this talk on YouTube: https://www.youtube.com/watch?v=xoVJKj8lcNQ I think this passing this bill would have probably been a very good idea. Not perfect, but far, very far from broken. Edit: I'm actually not sure if we as a community gave this topic enough attention leading in the months leading up to the announcement of this decision.

You might be able to reach the intended audience at lesswrong forum. Can also ask Zvi (https://thezvi.substack.com/), astralcodexten, Marginal Revolution to publicize this.

I have read many Lewis' books and admired his skills as a storyteller and stylist for decades. I found Going Infinite fun to read but less satisfying than The Big Short, Liar's Poker, Moneybag, etc. As a good complement or even substitute, I highly recommend Zvi Mowshowitz's review[1] of Going Infinite. I have some familiarity with the trading world but virtually none in crypto. The book review is incredibly long but very rich in both micro details and macro perspective. Zvi seems to know what he's talking about as a former Jane Street employee who knew people in the book (including some acquaintance with SBF before he became SBF) and as a crypto trader. [1]:https://thezvi.substack.com/p/book-review-going-infinite

If you are interested, you may find this worth reading: https://thezvi.substack.com/i/111749937/additional-responses-from-the-… Not a complete answer, but gives a good deal of context around the discussion, I think.

I've also found this blog to be useful: https://thezvi.substack.com

1. The new Frontier Model Division is just receiving information and issuing guidelines. It’s not a licensing regime and isn’t investigating developers. 2. Folks aren’t automatically liable if their highly capable model is used to do bad things, even catastrophic things. The question is whether they took reasonable measures to prevent that. This bill could have used strict liability, where developers would be liable for catastrophic harms regardless of fault, but that's not what the bill does. 3. Overall it seems pretty reasonable that if your model can cause catastrophic harms (which is not true of current models, but maybe true of future models), then you shouldn’t be releasing models in a way that can predictably allow folks to cause those catastrophic harms. If people want a detailed write up of what the bill does, I recommend this thorough writeup by Zvi. In my opinion this is a pretty narrow proposal focused at the most severe risks (much more narrow than, e.g., the EU AI act). https://thezvi.substack.com/p/on-the-proposed-california-sb-1047

Possibly. See https://thezvi.substack.com/p/escape-velocity-from-bullshit-jobs The scariest thing is that there are people who advocate for it. Because humans are dangerous, I guess, so it's better to preemptively enslave them. Sample > Social control in the sense of not wanting lots of unemployed and restless youths. Having a system where long term and steady work is required in order to "live a good life" implies control - you have to act right and follow the rules in order to keep a job, which is itself necessary in order to have enough food and other necessities. > Those making this argument here I believe are also making an argument for an alternative where the productivity of society is more equally spread, without the need to make everyone work for it. > I agree that it's a system of social control, but I don't think it's nefarious or bad. We really don't want to live in a society where 25-year-old men don't have meaningful work to do and roam the streets getting into trouble. The argument also doesn't _really_ make sense - there's already socially accepted system of 'social control' which _directly_ keeps people following the rules. The law. Also unclear why lack of work would cause young people to "roam the streets" instead of staying home and roaming the internets. As they're already increasingly doing in free time.

I was worried about this earlier, so I spent some time searching this out. Here are some good analyses: https://www.aljazeera.com/news/2022/9/23/little-military-sense-for-nuc… https://thezvi.substack.com/p/ukraine-post-12 https://www.iswresearch.org/2022/09/russian-offensive-campaign-assessm… The TL;DR is that most experts don't think Russia will use tactical nuclear weapons, but there's still a substantial non-zero chance. 10%? 25%? Much higher than anyone would like, that's for sure, and dependent on factors that are out of control of anybody but the Russian government and military. I have not found any sources saying the opposite, though some have higher degrees of confidence than others. The consensus appears to be that this sense of uncertainty is being created by Russia on purpose, and therefore, that it's probably (but not guaranteed to be) a bluff. I have also seen very mixed opinions on whether Russia's tactical nuclear weapons are in suitable condition for use. But I think it is best to assume they are. The odds of Russia intentionally using strategic nuclear weapons, the kind you would target at foreign cities across continents or oceans from you, remain nearly zero. Unlike tactical nuclear weapons, which have theoretical uses in ensuring one wins a conflict, strategic nuclear weapons remain a guarantee of mutual and probably worldwide destruction with no purpose except to deter would-be attackers. The only time I can…

There have been many reports that they're training in other countries. (On mobile so can't search, but this was yesterday:) > "Oracle was providing a staggering 22.6 percent of China's known A.I. computing power" In Malaysia, etc. https://thezvi.substack.com/p/ai-180-no-longer-in-charge

Simon's retelling is more compact but it also invites anthropomorphization of the sharing of the familiarity with the message board which re-emerged a few times. Zvi's retelling handles this better. Zvi speculates that the secret message board familiarity was carried because it had been trained into the May-and-subsequent models: https://thezvi.substack.com/p/openai-trained-its-models-for-months

Most available data suggests that sports betting is much worse for the people making the bets, as it better targets people with poor impulse control. Bad bets in political markets aren't causing measurable increases in bankruptcy and domestic violence rates (https://thezvi.substack.com/p/the-online-sports-gambling-experiment).

Either I'm missing something or this is way overstated. Steerling appears to be just a discrete diffusion model where the final hidden states are passed through a sparse autoencoder (a common interpretability layer) before the LM head. They also use a loss that aligns the SAE'S activations with labelled concepts? However, this is an example of "The Most Forbidden Technique" [1], and could make the model appear interpretable without the attributed concepts actually having causal effect on the model's decisions. 1: https://thezvi.substack.com/p/the-most-forbidden-technique

Additional citation on generational value transmission: https://substackcdn.com/image/fetch/$s_!xr9r!,f_auto,q_auto:good,fl_pr… from https://www.ggd.world/p/what-do-parents-want which was sourced from https://thezvi.substack.com/p/fertility-roundup-5-causation

There has been a lot of discussion on this recently in the blog-o-sphere. All conclusions I've seen so far are that the economy is basically fine and maybe people's expectations have risen (I'm oversimplifying). I'm also quite eager to hear different conclusions, because there is a lot of cognitive dissonance on the economy right now. - https://www.slowboring.com/p/you-can-afford-a-tradlife - https://www.slowboring.com/p/affordability-is-just-high-nominal - https://thezvi.substack.com/p/the-revolution-of-rising-expectations - https://open.substack.com/pub/astralcodexten/p/vibecession-much-more-t…

I was hoping this would be about https://thezvi.substack.com/p/slack

No space between paragraphs and blockquotes definitely hurts, and the subheading formatting is clearly not right. But the same content is posted on multiple sites. Maybe you'd find e.g. the Substack formatting more readable? https://thezvi.substack.com/p/deepseek-v31-is-not-having-a-moment

I’m new to this concept so may have missed something, but the post [0] seems to be about CoT specifically. In CoT you have an intermediary step that helps the model get better final results; the lesson is that if you try to improve the intermediary steps directly using training data then the model will optimize for better steps but not for better final results. I don’t think this is the same situation. 1. Anthropic is adjusting weights directly to influence the final results, not training against good/bad results and 2. The target is the final result, not an intermediary. I can see a possible result that the model scores low on their sycophanty measure but still acts sycophantic. In that case it could be new vector needs be calculated. [0] https://thezvi.substack.com/p/the-most-forbidden-technique/

For ref https://thezvi.substack.com/p/the-most-forbidden-technique/