Cas (Stephen Casper)

@davidad @farairesearch 😂💯nice

@StephenLCasper @farairesearch AI researcher: “I think there's a misunderstanding. The only ‘policy’ we're responsible for is a mathematical function π : S→A. It's just a formal artifact that prescribes decision-making responses to a broad variety of real-world cases. The homonymy is a complete coincidence,”

Do you do technical AI research? In this talk, I argue that you 🫵 should see yourself quite literally as a type of policymaker. Thanks @farairesearch. youtube.com/watch?v=Ekp-eg…

New defense against Emergent Misalignment (EM): train models to recognize their own text. We find that self-recognition finetuning (SGTR) can reverse and prevent EM-induced misalignment 🧵 w/ coauthors: Shawn Zhou, @jiaxinwen22, @ihsgnef

☠️I present to you: my worst tweet ever.☠️ I can't remember a time when I have had a tweet totally bomb -- zero likes. Interesting that this just so happened to be on a post that shared an article referencing a scandal about you-know-who and you-know-what. x.com/StephenLCasper…

Big day for AI (@Kimi_Moonshot) & a good-ish one for US-China relations (see this morning's writeup from @nanjituzhu)! If you're curious about China’s open-source AI strategy and the factors driving its private companies' interest, here’s a talk that I gave last week. The bottom line is: open-source/open-weight AI is becoming a pivotal part of China’s overall AI strategy and will remain strategically important. Towards the end of the talk, I also spotlight various emerging risks from OS AI and conclude with some reasons for optimism. Watch this space for more China AI + open-source AI content! 🔗: youtube.com/live/OxXBSq65G…

The official text isn't out yet & will matter a lot. But Europe may be moving to treat AI nudification similarly to how pirated media or CSAM is treated: as something that, even though it'll always be available, can possibly be made much less accessible. politico.eu/article/eu-gro…

I’m really excited about our new paper! I think we will ultimately need to draw on expertise from both law and AI to get alignment right, and this paper lays out that vision in more detail. As an aside, my PhD thesis was titled ‘Aligning law, policy, and machine learning for responsible real-world deployments’ for a reason. I think this is a very important area, and I’m excited to see so many excellent researchers working together to move it forward.

I'm extremely excited to be working with @BKCHarvard from now through the end of the summer. I'll jointly be an MIT PhD student and a BK Fellow from now til I graduate. Then the BKC will be my institutional home for the summer!

My old team at AISI is hiring! I’d recommend.

Thanks. I see you here. I don’t disagree. I’d just add: - I didn’t discuss this before, but I would emphasize that there is a meaningful difference between donations meant to elect candidates (activism) and donations meant to curry favor with an administration (QPQ). There is sometimes but always a reasonably clear distinction. - It’s easy to say that the system is broken, but simultaneously hard to say what any individual company should do about it, but then again easy to say that OpenAI has been somewhat uniquely corrupt in terms of the amount and nature of these donations.

@StephenLCasper @boazbaraktcs @_NathanCalvin so i’m very very sympathetic to ai lab executives who donate politically according to their views! it seems like the most rational philanthropic action they could engage in

Some thoughts about mass surveillance, red lines, and a crazy weekend. windowsontheory.org/2026/03/03/mas…

@NinaPanickssery @boazbaraktcs @_NathanCalvin There is a separate conversation to be had about how trump’s administration is uniquely corrupt. But that’s not today’s conversation.

@NinaPanickssery @boazbaraktcs @_NathanCalvin Yes. Of course. Damn. I mentioned that the donations were to the inauguration fund and maga inc above. But didn’t say this was the thing that made them corrupt. In my lower posts notice that I stopped describing the donations as being to trump.

@NinaPanickssery @boazbaraktcs @_NathanCalvin …who have not donated to Trump.

@NinaPanickssery @boazbaraktcs @_NathanCalvin Notice above how I even specifically called out Anthropic unprompted for their PAC contributions.

@boazbaraktcs @_NathanCalvin Oops I should also add a fourth point to the list above: (4) it is also in conflict with how some members of the Trump administration have publicly said they interpret the contract language.

Thx. For clarity, I personally see that one tweet as a molehill, not a mountain. My broader concern is with the work you've done across multiple posts and replies to push Altman and OpenAI's broader narrative that OpenAI's deal with the DoW genuinely provides more "guardrails" than what Anthropic was attempting, or at least that "you believe" such. This narrative seems (1) dubious on its face, (2) premature without independent legal analysis, and (3) untrustworthy without public release of the full agreement. To be specific, my claim here is not about outright lies, but about being misleading using selective substance, vagueness, and "legalese." --- Aside, I think your comments here about how democratic threats to AI are serious, while you also see no apparent problem with a 25M PAC contribution, are astounding. At a minimum, please just quit with the doublespeak.

Thanks for the candidness. My thoughts: 1. I do not think that this is a tenable position to take in the context of, e.g., a 25M donation to a super PAC. I cannot help but think that you are crossing a line into dishonesty in saying this. No legitimate definition of "democratic" could accommodate this. 2. Are you saying that you think that Altman and Brockman made these donations for personal reasons and had no expectation of favorable treatment toward OpenAI from the government (e.g., in deals with the DoW)? If so, I think this is clearly wrong. And I think that you think it is too, even if you won't say it. Overall, Boaz, think that in the past few days, the ways that you have chosen to be an apologist for OpenAI have sometimes crossed lines into misleadingness and dishonesty. I think that it's coming into conflict with your role as a professor to tell the truth and responsibly educate people on AI. When it comes to democracy and the use of AI for surveillance/weapons, I think, from what I have seen, that your hands are sometimes doing something different from your mouth. I know you're a contemplative guy. I hope you think more about the strain that your roles as an apologist and professor are putting on each other. I hope you have some red lines in mind that, if crossed, would make you leave. And, if you haven't already, I hope you have an "are we the baddies" moment. I can only imagine how stressful a week it has been for you. And I'm sorry if I am "piling-on." I regret how worked up this week has gotten me as well. But I hope that you will reflect more on this stuff.

The bottom line for me is that: 1) I don’t view political donations to a party that I oppose as undemocratic whether by my company or an executive 2) I do think it makes a difference whether it’s the company or an executive: it is ok if our executives have differing political views and do different things with their money. In making my own decisions I am focused on whether my contribution to AI safety and society is higher from within or from outside. I disagree with your views on social harms and benefits and believe my contribution is currently higher from within.

Thx. My thoughts: - I think all of these donations in the screenshot are bad for democracy, and the system is broken. I think you, too, could admit this. - Not all tech companies are equally corrupt, and Brockman's 25M donation in particular is unprecedented. - To directly answer your question (unlike the way you dodged mine), my personal view here is that the employees at these companies (and even Anthropic, for that matter) are indeed complicit in this type of anti-democratic behavior. I think that the work of many employees at many tech companies is not societally beneficial enough to outweigh the societal harms that they contribute to. Could you answer my original question?

@StephenLCasper @_NathanCalvin So basically everyone who believes in democracy and works in a tech company should have quit?

Later I was also able to recover this screenshot using my phone's cache.

He deleted the post, but you can read an excerpt from it here. news9live.com/technology/art….

I think this further confirms that the government sees the new contract language as just "memorializing" a vague "commitment" rather than drawing any real new lines. See my full thoughts here: x.com/StephenLCasper…