Me reassuring myself in an act of self-care: You worry too much about what people think of you. I'm sure that no one noticed that your outfit didn't match the other day. Me reading the author list of a paper on the arxiv this morning: Oh that's the guy who wore those weird pants that one time.

New paper with @coryshain.bsky.social: We develop "perturbation" - a simple approach to finding shared representations (common abstractions) in LMs. We perturb an LM by fine-tuning it on a single adversarial example and measure how this perturbation “infects” other examples.

NeurIPS walked back their aggressive stance on sanctions, but I guess we’ll see next year when the main conference returns to Europe whether they’ll offer a satellite location in China like they have for Europe and the US now.

I don't think people fully appreciate how apocalyptic things are for US science. I haven't had any new funding since 2024, but I'm still ok since typical grants are for three years. This means next year I will be completely out of funding and will have to fire everyone in the lab. It's not great.

The divide is actually pretty stark! www.pewresearch.org/science/2025...

I truly believe the rapid advances in the mech interp subfield have something real to offer AI ethics researchers: A chance to look beyond the HOW of evals to the WHY, a first pass at a technical solution when we see the opportunity, a new avenue for showing failures that prove models are not gods

Don’t forget this drop when you’re singing mournful sea shanties about AI2

Reviewer waiting for your rebuttal letter

🎤 Keynote at #EACL2026 “Arabic and Technology: A 40-Year Perspective” Don’t miss this keynote tomorrow in Session 1: Plenary (10:00–11:00)! Join us right after the opening session for this invited talk by Nizar Habash. 📚✨

classifying the new composer tech report as a must read cursor.com/resources/Co...

I’m seeing close to zero reaction/conversation about this on here. This is huge news for open research on language models, especially in the US.

We are Delighted to Congratulate you on the Honor of the Opportunity of Volunteering

@gershbrain.bsky.social's new piece for The Transmitter includes a comment from me about something I think we often underrate in science: the actual degree of human understanding provided by a scientific model.

gonna wear this next neurips

📌 New feed alert! Excited to announce - ML Feed: For You It is the AI/ML same feed, but now powered by @graze.social's personalization algorithm. Please check it out, and let me know what you think!

(explaining to an LLM how to have a breakthrough) okay you're going to need to work on something unrelated for a week or two, then take a shower, at which point the solution will just appear in your head

Happy Nowruz! Speaking as a connoisseur of new years, it's near the top.

My brother once said that his HS education was no better than sitting inside with the shades drawn watching “I love Lucy” reruns. That’s more or less how we train LLMs. Good thing ppl like @nsaphra.bsky.social are getting serious about LLM training dynamics. youtu.be/2WqvUQY1nlg?...

Scientific American has a piece today on NACLO (the linguistics competition) by @dodecalemma.bsky.social! Features some quotes and puzzles by yours truly!

This is an important project! If you believe alignment faking is true, you should at least entertain the possibility of misalignment faking before drawing your conclusions. Especially true if researchers fishing for misaligned behaviors are the ones running the evals!

Learning from feedback is instrumental but human preference data can be expensive. How much reward supervision could we get from raw web text instead, without human labels? Our work, a year of incredible effort by @fjxdaisy.bsky.social, advances pure RLHF training across multiple models and tasks.

We'll have a reproducibility track at this years' Blackbox workshop! Details are still within a slightly opaque box. We want to see if cleaning solutions that make opaque boxes 📦 transparent 🍱 work on different boxes 🎁📮🧰🥡, and with different 🧽solution-to-water🧼ratios!

How does an LLM’s past influence its future?🤔 In new work, led by @adisimhi.bsky.social, together with @fbarez.bsky.social @boknilev.bsky.social and Shay Cohen, we find conversational history creates a latent "geometric trap" which makes old habits e.g. hallucinations hard to break!

i read this and it's interesting but the horrifying overhang of this story is "mrna vaccines make it relatively easy to cure cancer actually, and the main obstacle is that the FDA just banned trials of them"

From maybe the best teacher I know: Idan Blank explains transformers in a super intuitive and fun way!

man uses GPT to guide him through sequencing his dying dog's DNA, identifying the mutated genomes, and developing a bespoke cancer mRNA vaccine www.theaustralian.com.au/business/tec...

I think Fodor & Pylyshyn's 1988 paper is possibly the most mischaracterized paper in the history of cognitive science. It's often cited as arguing that neural networks cannot achieve systematicity, compositionality, and productivity. But that's not what they actually argue...

ANUBIS: (presses button with paw) [guilty] OSIRIS: im sorry he does this when he's hungry ANUBIS: (button) [guilty] [bad] [food] OSIRIS: you already ate ANUBIS: (button) [bad] [bad] [bad] OSIRIS: he's right though we weighed your heart against the feather and its not looking good

I thought this was an exciting new theoretical perspective www.cell.com/neuron/fullt...

1 million context window: Now generally available for Claude Opus 4.6 and Claude Sonnet 4.6.

now u yes u dear reader 🫵 can train kimi k2 thinking on ur own infra! www.workshoplabs.ai/blog/post-tr...

David Graeber died in 2020 so there’s nobody alive who can articulate what DEI means to this person

Sharing “Neural Thickets”. We find: In large models, the neighborhood around pretrained weights can become dense with task-improving solutions. In this regime, post-training can be easy; even random guessing works Paper: arxiv.org/abs/2603.12228 Web: thickets.mit.edu 1/

just ask the stochastic simulation of Susan Sontag if AI can be camp

🦀

My most-used model is Gemini because I’m a Luddite. Not saying Gemini is bad, but its competitive advantage remains search replacement with cross-references, which is the application you care about if you don’t actually trust LLMs.

Watch this clip. Professors will recognize this kid as the student who didn't do the reading but still has very strong opinions about how it's all "bullshit," except this time he's not the callow student who's going to fail your course, he's the reviewer who's going to cancel your grant.

❌ What's your Roman Empire? ✅ What's your attractor state?

I'm lecturing about the "History of NLP" this week. What should I include? Any favorite anecdotes, images, people, methods? Slides, books, papers, or talks for inspiration or grounding? I've been maintaining a small collection here: www.are.na/maria-antoni...

🧠🤖 Are you at #COSYNE2026? Check out the #KempnerInstitute's presentations! 👇 #neuroscience #NeuroAI

I had an elegant explanation for how tokenization causes LLM emdashes (they reduce the token count required for parentheticals and transitions because they don't need whitespace) but noticed today that Claude uses Economist-style (whitespace padded) emdashes. It was a beautiful theory. 😢

picked a great name for something. think we're done today

We all realize that *obviously* shifting from “the number of scientific fiction” (42) to “the mysterious number of the physical universe” (137 = 1/fine structure constant of QED) would lead to performance gains!

*monkey paw curls* a niche obsession with training run variance is sure gonna become important soon ...

When @marvinlavechin.bsky.social first approached me about this project I thought no *way* could we get something useful out of messy early child speech. Delighted to have been wrong & excited for what babar will let us do at a bigger scale than the super hard manual transcription work it leverages!

don’t forget to destroy your clocks tonight, walk away from your mortgage, just start driving, you’re free

As AI capabilities increase, we need a broad, deep, society-wide discussion of what limits make sense, and how we can hold the government meaningfully accountable to citizens. For that reason, I stand with Anthropic and anyone else who is avoiding a rush toward mass AI surveillance.

I want to talk about why AI-based mass surveillance is so dangerous, and why I would oppose it no matter which party or president is in office.

Short post on what I call the "no-magic approach to understanding intelligent systems" — the philosophy I think of as motivating our work on understanding intelligence without resorting to magical thinking about AI or humans! infinitefaculty.substack.com/p/the-no-mag...