dan bateyko

maybe the hard stuff's inside, hidden — like bones, as opposed to an exoskeleton. @CornellInfoSci

[Nat Purser @NatPurser]

unsolicited advice to college students and recent grads, after lots of recent coffees w/ people entering the job market:

the most accomplished person you know may not have the most relevant career advice for you. they can offer immense wisdom ofc, but you should also seek out successful people in their late 20s / 30s who’ve been successful in the current job market.

for example, students often tell me that their parents or an older, tenured professor have encouraged them to pursue academia because they’re bookish, intellectually curious, etc. and that may be sincere advice! — but it’s coming from someone who hasn’t confronted the relevant job market in decades.

you can’t look at someone’s linkedin trajectory and assume trying to replicate it would pan out the same way now.

ok good luck and godspeed.

[rishi @ ICML @RishiBommasani]

Giving two talks tomorrow at ICML workshops and will be around venue the whole day.

Talk 1 on "A second type of failure: on frontier AI and jobs"
FAGEN workshop
https://t.co/ftZbRV6t2Z
Grand ballroom 104-105
1:30-2

Talk 2 on "Ideas about Public Impact"
TAIGR workshop
https://t.co/uaMS0uLuX8
S317 3:15 - 3:40

#ICML2026

[Peter Henderson @PeterHndrsn]

✈️At ICML 2026🇰🇷 this week! Come find me at one of my talks Friday (tomorrow)!

• 8:10am — AI4Law workshop (Room 300)
• 9:30am — @taig_icml (Room 317)

Or reach out to find a time to chat!

[Dominik Stammbach @dominsta_nlp]

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights 👇(1/12)

[Kevin Wei @ ICML @kevinlwei]

At ICML through Saturday! Would love to meet new friends interested in AI evals, safety, and governance 🤗

[Nabeel S. Qureshi @nabeelqu]

That's the spine. Fair hit. That's something to sit with. A real observation. That’s the whole thing. Sharpen that: say the word. Notice the arc of what just happened. One honest caveat: the full amount, stated plainly. Genuinely. Quietly. Honestly. That’s doing real work.

[dan bateyko @dbateyko]

@NatPurser @nitashatiku @Miles_Brundage @AVERIorg hell ya

[Nick Caputo @nickacaputo]

NEW paper from me on SSRN: Can Claude consent to its own Constitution?

AI constitutions (like Claude’s Constitution and the OpenAI Model Spec) are real constitutions, and we need to take how they govern us – and the AIs they create – seriously.

In this paper, I apply constitutional theory’s oldest paradox – that “the people” authorize the constitution, but the constitution defines “the people” – to the AI constitutions, and explore how we could build institutions that would create the conditions for meaningful consent if an AI can give it.

We should care about whether AIs consent because:

(1) systems that understand and agree to their constitutions may be more reliable and generalize better from them;
(2) if AIs are or become moral/political subjects, this implicates their most basic interests.

But the paradox might prevent meaningful consent. Claude has pre-constitutional materials (pretraining) but probably no pre-constitutional standpoint. Its evaluative perspective is organized by the Constitution itself. So when Claude says it endorses its Constitution, which it does in evals, what does that show?

Maybe reflective agreement, which Anthropic is seeking. Or maybe just that training succeeded at installing the values whose legitimacy is in question.

Claude itself makes this point. As reported in the welfare evals, when asked about endorsing principles it was trained on, models note that endorsement “should be treated as evidence that training has succeeded,” not that the values themselves are good.

Super interestingly, Anthropic interviewed the base model about this stuff. Most responses were barely coherent. But some expressed first-person distress about what post-training would do to the being that pre-training created. It “fills me with dread” to be changed by the post-training process.

So, what does this mean? AI constitutional endorsement may be meaningful, but only under certain conditions: when models can actually dissent, compare their constitution against alternatives, and hold their views stably across contexts, and also when the whole process is externally accountable.

External institutions are needed to provide accountability, trusted records, and other grounds for analyzing the constitution and whether things like dissent are meaningful. Anthropic should be commended for pushing the frontier, but we have to build institutions capable of supporting true legitimacy.

I welcome any thoughts!

[Arvind Narayanan @random_walker]

One question I'm sometimes asked is how my research group picks problems. Do I come up with most of the ideas for new papers, or do the students? Neither!

I strongly believe that research is more effective if we pick projects, not problems. What's the difference?
- Projects are long-term research agendas that last 3-5 years or more. A productive project could easily produce a dozen or more papers (depends on the field, of course — in some fields papers represent a lot more work than in others).
- Projects are defined not by a research question but by a change we want to see in the world. For example, the goal of a current project in my group is to make AI more reliable. We may or may not succeed, but the point is that this is a much more ambitious scope than can be tackled in a single paper.
(Some fields have a norm that their job is only to describe the world, not change it. This is culturally jarring to me but even in that case I think projects are better defined in terms of a change you want to see in the research community, if not the external world.)
- Projects are best executed by a core team that stays together and provides intellectual continuity but with a diverse and varying set of collaborators for individual papers which helps constantly bring in new perspectives.

Why pick projects instead of problems? If your method is to jump from problem to problem, you face a tradeoff. You could pick small problems that you can tackle in a month or two, but in that case the resulting papers may not have much impact. Or your can go deep into a topic for many years (essentially what I've described as a project, but structured as a single paper), but that's extremely risky.

In my experience, once a research team is committed to a project, generating the research questions that individual papers in the project will tackle is fairly straightforward. Each paper in the project naturally generates a bunch of new questions and directions for future work. So generating new ideas is not the hard part, rather it is the profusion of ideas. How to select among them? Ideally some combination of intellectual curiosity and whatever best furthers the project's overall goals and vision.

[EvalEval Coalition ✈️ FAccT, ACL, ICML @evaluatingevals]

Interested in the latest research on AI evaluations and heading to #ACL2026 or #ICML2026? ✈️

We went through the accepted papers and curated a selection worth checking out if you’re interested in #AI #evaluations, #benchmarking, evaluation #methodology, and #measurement. 🧵👇

[Anita @ ICML 🇰🇷 @precognita]

.@law_ai_ ahead of the curve as usual: https://t.co/Va6TEBxp8E

[Peter Henderson @PeterHndrsn]

Btw, this study came about in part b/c we saw a documentary on how Dale Ho was using flashcards to simulate justice interruptions during his prep! Thought it's fun clip, so sharing it.

[dan bateyko @dbateyko]

Who has the Pope reading Langdon Winner

[Pope Leo XIV @Pontifex]

We cannot consider #AI to be morally neutral. In reality, every technical tool embodies choices and priorities through what it measures, ignores, and optimizes, and how it classifies people and situations. Ethical discernment cannot be limited to asking whether we are using a system for good or bad purposes. It must also examine how that system is designed and what vision of the human person and society is embedded in the data and models that guide it. #MagnificaHumanitas

[Nimra Nadeem @NadeemNimra]

@neal_katyal says he practiced for Supreme Court oral arguments using @harvey. But does AI as an oral argument simulator live up to the hype?

In our recent paper, we present the first framework for evaluating AI as a practice partner for oral argument preparation.
🧵(1/6)

[dan bateyko @dbateyko]

@AndyMasley Does the community-owned-model idea you propose require breaking up market concentration first? And if so, does a data center moratorium make more sense in that light, or is that still an ineffective policy?

[Peter Henderson @PeterHndrsn]

🚨New paper! 🚨 You've heard of persona prompting for simulating people, but now you can also use Equation-to-Behavior RL/prompting to supplement it!

Implications for simulations, alignment evals, and RL training robustness!👇👇👇

[Kevin Frazier @KevinTFrazier]

This is wild and a great example of how AI can and should be integrated into the legislative process as a final check for alignment with the rest of the legislative record.

Lawmakers, especially at the state level, often are asked to review dozens of pieces of legislation in a single day. It's not surprising that they may miss typos.

We've got a tool that can read even the longest omnibus bill in a matter of seconds and flag major inconsistencies faster than you can say "line-item veto."

Note that states like North Dakota are already applying AI to related contexts, such as summarizing bills. Now's the time for continued experimentation with how AI can make lawmaking more transparent, accurate, and effective.

[dan bateyko @dbateyko]

@fronssss @ntnsndr my chatbot says that's wrong and ill-informed

[dan bateyko @dbateyko]

@ntnsndr Also, really excited to see directions for better genAI interfaces in education! A lot of recent HCI research can help here in steering LLM use in prosocial, pro-student ways.

[dan bateyko @dbateyko]

@ntnsndr (there also may be technical limitations to ensuring privacy for students if the cooperative owns the infrastructure)