I’m recruiting students this upcoming cycle at UIUC! I’m excited about Qs on societal impact of AI, especially human-AI collaboration, multi-agent interactions, incentives in data sharing, and AI policy/regulation (all from both a theoretical and applied lens). Apply through CS & select my name!

Welcome to Kimi K2 day please enjoy your Kimi and use responsibly moonshotai.github.io/Kimi-K2/thin...

Some helpful tips for all 10 of the graduate students on this website. Please someone do them up as an interpretive dance on tiktok so the word gets out

“Lovely justified true belief you have there. . . . It would be a real shame if it weren’t knowledge.”

Excited to share our new paper, "DataRater: Meta-Learned Dataset Curation"! We explore a fundamental question: How can we *automatically* learn which data is most valuable for training foundation models? Paper: arxiv.org/pdf/2505.17895 to appear at @neuripsconf.bsky.social Thread 👇

We need to be out here figuring out how to build solar power everywhere rather than debating if we should go back to 1250, 1850, or 1950

The 3 letters for grad school thing seems remarkably bad. You either have to plan for it well in advance and cycle through 3 labs, which most folks don't know to do, or get lucky and have multiple advisors on a project

“Professional society of people who build bridges“ says bridges are crumbling and must be rebuilt

I am pretty excited about this as a way to take easy transcribed notes while reading. I have gone through many horrible solutions to this problem. www.theverge.com/ai-artificia...

Happy to share some of the first work from my new lab! This project has shaped my thinking about how we can effectively combine planning and RL. Key idea: start with a planner that is slow and "robotic", then use RL to discover shortcuts that are fast and dynamic. (1/2)

In hindsight, there was always more opportunity for agency than there seemed at the time

How do people flexibly integrate visual & textual information to draw mental inferences about agents we've never met? In a new paper led by @lanceying.bsky.social, we introduce a cognitive model that achieves this by synthesizing rational agent models on-the-fly -- presented at #EMNLP2025!

There is a type of cynic who claims everyone is driven solely by incentives, never values

okay can we say now that Eric Adams was objectively the funniest mayor in history

I’ve seen enough

N-gram novelty is widely used as a measure of creativity and generalization. But if LLMs produce highly n-gram novel expressions that don’t make sense or sound awkward, should they still be called creative? In a new paper, we investigate how n-gram novelty relates to creativity.

Was there a prediction market on prediction markets eventually attempting to put their thumb on the scale?

What are the salient aspect of the LLM RL problem that could be abstracted into a benchmark? Tremendously large action space, sensitivity to numerical precision, prior policy that you must remain close to, what else?

This is at least partly about reviewing LLM papers. It's so frequently a long list of details and ad-hoc architectures addressing ad-hoc tasks

it's not difficult to work endlessly on things that bring you joy. But oh, oh is it hard when it's just a giant list of annoyances that you need to tick off

Sort of starting to believe that we really do need academic metrics that punish publishing too much

Hello, a PSA. Arxiv is the server that many academics use to discuss their work before it is peer reviewed. It is widely viewed as a blessing.

TBH I think a lot of the blocklist issues go away when there’s better filtering here. People block to replicate algorithmic filters so that they see something more aligned with what they expect

Imagine being so irony poisoned that you think this is convincing to anyone: www.whitehouse.gov/mysafespace/

Still fairly surprised other countries didn't try harder to poach scientists from the US. Was sort of a golden window that doesn't come often

Thanks to generous listeners like yourselves, I'm within just dollars of overtaking @ed3d.net on the Friends of Alf leaderboard. With your help, we can get me past him and help feed America's hungry!

Really do think Pufferlib is one of the most important projects happening in ML right now. Through a lot of personal effort, Joseph has trained a lot of people into being pretty capable RL engineers.

SNAP benefits have fallen through so let’s help out a bit. Will match up to 1000 to any food bank of choice.

You can write “Neil Postman vindicated again” every day and have it be correct about something

Where did LLMs learn to punctuate their code with emojis? No code is like this so it has to be in post-training?

Training LLMs end to end is hard. But way more people should, and will, be doing it in the future. The @hf.co Research team is excited to share their new e-book that covers the full pipeline: · pre-training, · post-training, · infra. 200+ pages of what worked and what didn’t. ⤵️

Cannot imagine anything stupider than using an LLM to write a survey paper and then putting your name on it. Burning the commons for a few citations.

A friend of mine used to be unable to sleep for days before submitting a paper with proofs in it, for fear there was a mistake. Probably too far in that direction, but seeing some of these proofs...

Daniel Dennett has a nice story about a seminar where the speaker kept ragging on a paper so he retorted "if this paper is so bad, why are you wasting our time discussing it?" And I think there's some nice lesson in there about what you let into your headspace

Science is inherently small batch, artisanal and systems that attempt to scale the act of science are a bit doomed

[1/9] Excited to share our new paper "A Pragmatic View of AI Personhood" published today. We feel this topic is timely, and rapidly growing in importance as AI becomes agentic, as AI agents integrate further into the economy, and as more and more users encounter AI.

This labeler is such a clever way of making spam visible. Bluesky wins yet again: bsky.app/profile/stec...

Idealized science: take walk in the park, think very hard, be struck by brilliant insight. Marvel. Science frequently: it is week three of being unable to reproduce my experiment from six months ago. Was I wrong then or am I wrong now? Is it just noise? I have not seen the sun.

Deggi Mirch is just such a bright, wonderful smell when it hits the pan. Just so much fun to add different MDH boxes to your cooking

The ideal of anonymity is it lets you explore and discuss ideas without serious consequence. The frequent reality of anonymity is folks using it to be cruel without consequence

Gambling serves literally no positive social function. The only argument for tolerating it is that making it illegal might generate crime as a side effect. This is not true for online slots! It shouldn’t exist.

I have been informed that this tweet is wrong and you just need to disable rendering.

I'd like to host some competitions / leaderboards in my class next semester. These are RL competitions, so Kaggle generally doesn't work. Any ideas for easy ways to do this besides "roll your own"?

It took two years to get here but the funding guillotine is finally pulled all the way back up. Now there's so much research to be done.

It's #NSF #GRFP application season again so it's time to re-up my GRFP application advice post! Also, check out the cool bsky comment integration I've added to the blog! Engagement with this post will go under the blogpost on my site as comments! saxon.me/blog/2024/gr...

I suspect that links are so clearly detrimental to "time on site" in the short term, that it takes extraordinary evidence of "time on site" improvement in the long term to make the case for reversing it, and platforms can't run the kind of network-wide experiments that could demonstrate it. Goodhart

I don’t use LLMs for writing because I like learning to write. But I do use them as an additional copy editor on top of asking friends for feedback

Human-AI centaurs is going to be the transition where I stop really understanding folks younger than me I suspect

Wait, you’re telling me hominids can “reason,” when they still can’t reliably count the number of Ls in Al?