Good guide but would suggest For You and Popular With Friends and ditching Discover bsky.app/profile/did:...

Some TPOT and e/acc folks have started popping in again lately. Hi! Your unhinged posts and chatbot romances are essential to our healthy AI ecosystem, so please check out my old post on constructing your feed to make bsky enjoyable.

Ever wondered how Energy-Based Models (EBMs) work and how they differ from normal neural networks? ☕️ We go over EBMs and then dive into the Energy-Based Transformers paper to make LLMs that refine guesses, self-verify, and could adapt compute to problem difficulty.

might be time

psychology is waaaaaaay more fun to write about when you don't have to give a shit about your h-index anymore

People keep talking about tech firms in relation to the H1-B $100k fee but it’s gonna really sink universities.

This site contains 80% of the remaining academics on social media but for some reason they can be a bit silent. Feed the beast

The people on this site crave papers. You can post a random excellent paper from 1978 on any topic and it'll blow up a little bit. You can fill the gap!

I wish students understood in most empirical AI research there’s a huge scientific advantage from being constitutionally excited by math vs intimidated, but very little additional gain from being actually “good” at math. Maybe they’d be less intimidated if they didn’t feel they had to be “good”.

it's so funny to see America's health insurance companies be like "look. If there's one thing we love it's denying you care. but we ran the numbers - and we cannot BELIEVE we're saying this - and it's cheaper for us to just pay for the vaccines than deal with you going to the ER."

I don’t trust any bird that talks and neither should you. A bird that talks will one day be the bird that takes your job.

a few takeaways: - Musk is serious about building a software competitor to Microsoft, called Macrohard - Grok has about 64 million monthly active users - Musk's efforts to make Grok less woke & more popular triggered the 'white genocide' and MechaHitler incidents - some employees quit over sexbots

Giving a talk at NYU text as data tomorrow (Thursday) at 11! Catch me there buds

Are there conceptual directions in VLMs that transcend modality? Check out our COLM oral spotlight 🔦 paper! We use SAEs to analyze the multimodality of linear concepts in VLMs with @chloesu07.bsky.social, @thomasfel.bsky.social, @shamkakade.bsky.social and Stephanie Gil arxiv.org/abs/2504.11695

Jesus Christ.

Every time I go to SF I get jumpscared by going to a random party and then finding out I’m at the like Mechanistic Interpretability Acid House Rave Sponsored By Sequoia

Our new paper asks: what is the goal of “natural language verbalization” interpretability approaches? If a verbalizer is supposed to tell us something about what’s in the target LM and NOT just what’s in the verbalizer LM, how do we actually evaluate that?

Wouldn’t it be great to have questions about LM internals answered in plain English? That’s the promise of verbalization interpretability. Unfortunately, our new paper shows that evaluating these methods is nuanced—and verbalizers might not tell us what we hope they do. 🧵👇1/8

don't you tell me vendingbench is solved

It brings me no pleasure to report that completing a minor task you've been avoiding (1) is not very hard and (2) makes you feel better afterwards

A really neat paper thinking through what "Identifiability" means, how we can determine it, and implications for modeling. Arxiv: arxiv.org/abs/2508.18853 #statssky #mlsky

My two biggest take-aways are: - Standard testsets are too easy (Figure 1). - We can make testsets that are not easy (Figure 2). 😎

feed is all politics today please send good RL papers you've read recently

When I was placed on the Professor Watchlist in 2021, people sent death threats about my children. I had security officers monitor my 8yo at school. Where is all the outrage for those of us who have been targeted for years? Where is the outrage for our families? My own colleagues are silent.

LLMs introduce a huge range of new capabilities for research, but also make it possible for researchers to "hack" their results in new ways by how they chose to use models for annotation This is a useful pass at quantifying some of the risk, and some mitigation strategies arxiv.org/pdf/2509.08825

1. What does a Cold War-era game theory problem known as the silent duel have to do with high-risk research strategies, publication in Cell/Nature/Science glamor journals, and the academic job market? Kevin Gross and I tackle these questions in our latest arXiv preprint: arxiv.org/abs/2509.06718

A study shows that locality in diffusion models arises from dataset statistics, not network architecture, bridging theory and practice in generative modeling. The new analytical denoiser significantly outperforms traditional methods, enabling novel image generation. https://arxiv.org/abs/2509.09672

Why do LLMs fail at long horizon tasks? Because errors in execution, not planning i.e. an error made early on conditions the LLM into a bad state later. Newer agentic LLMs deal with early errors better arxiv.org/abs/2509.09677

Looked at my data collections for state-backed activity. If the report were a picture:

Comparing AI Labs in the United States and China, by Jingyuan Liu

I do actually think cancel culture was a thing in the 2010s: I would define it as a habit of giving viral outrage more moral gravity than it often deserves. But part of the reason I’m on Bluesky is that we don’t have strong algorithmic incentives for negative engagement.

TLdr: Authors dramatically underpaid, Anthropic paid 200 TIMES more than they would’ve paid if they just asked upfront permission.

last night as i was trying to wind down for bed i had a unified theory of what the fuck just happened, and i wrote it down, and filed it under "well, i can't talk about this on main without sounding insane until and unless the perp gets caught and is provably a groyper" anyways, great news! (1/X)

The Sting (1973)

getting fired from your job because some dipshit student who missed their calling in the Stasi filmed you talking about gender in a class about gender

I thought I wouldn‘t be one of those academics super into outreach talks, but I just put together something about understanding LLMs for laypeople and I get to talk about results that I don’t really focus on in any of my technical talks! It’s actually really cool. I made this lil takeaway slide

The image that inspired my most popular comic.

My toxic personality trait is I get real mad when I see videos of that ASMR guy who mows lawns for free. you’re killin the bugs man. that weird hoarder who can’t afford a mower is like the only sanctuary in town for butterflies

Mathematicians will contract pleurisy before they'll give up chalkboards

kibbles imply the existence of one giant kib

Paper from OpenAI says hallucinations are less a problem with LLMs themselves & more an issue with training on tests that only reward right answers. That encourages guessing rather than saying “I don’t know” If true, there is a straightforward path for more reliable AI through better training.

I came up with a nice paper title. just you wait

Benchmarking the variety of different proposed LLM optimizers: Muon, AdEMAMix, ... all in the same setting, tuned, with varying model size, batch size, and training duration! They find that AdEMAMix is SOTA, the importance of weight decay, or how Muon “prefers” WSD, and more...

wow they’re Literally cooking with this

It’s cool that Massachusetts is getting a better flag but disappointing that we weren’t brave enough to select this entry from the competition

Very similar to the figure that @johannbrehmer.bsky.social and I made in “Flows for simultaneous manifold learning and density estimation” arxiv.org/abs/2003.13913

Google's new image model is so good it has me worried for the future of facts. My latest for @theipaper.com on what it all means inews.co.uk/opinion/than...

On one hand, reducing car use reduces urban pollution. But on the other, everyone likes it and cities become more livable

How can an imitative model like an LLM outperform the experts it is trained on? Our new COLM paper outlines three types of transcendence and shows that each one relies on a different aspect of data diversity. arxiv.org/abs/2508.17669