What I’m hearing: Instagram’s Trust and Safery org absolutely gutted the last few weeks. ~60% of the org gone - between layoffs and forced reassignments to data labelling. All while “AI maxxing” pushed a bunch of bugs to prod. And hence why today’s massive Instagram account takeover happening.

LOLOL "What stood out most was ꜱᴛʀᴜᴄᴛᴜʀᴀʟ ʙᴀᴄᴋ-ᴇɴᴅ ꜱʜɪᴘᴘɪɴɢ ᴛʀᴀᴄᴋɪɴɢ, which feels like the kind of nuance that naturally sparks layered and honest conversation in our group."

🧵 Public science in the US has taken some heavy hits over the last year, but the White House OMB is proposing a series of changes that will result in a TKO for science as we know it unless we fight back. Here's what's happening, and why it matters. 1/n

Can’t roll an R? Trill issue.

What are the real problems to be solved in continual learning? In my latest post, I tackle this question — reviewing where I think the field went astray in the past, how language models changed things, and where the real challenges remain. infinitefaculty.substack.com/p/what-are-t...

What do you think is your most under-recognised paper and why should people be paying more attention? Go on, take the chance to toot your own horn 🎺

The ‘neat’ and ‘messy’ in task-dependent neural geometry and computation by Cheng Xue @cxue.bsky.social & Gouki Okazawa @gouki-okazawa.bsky.social www.cell.com/trends/neuro...

if you are a PhD student in AI, remember it is in your interests to distract your advisor from how much money they could be making in industry. should be a daily priority.

I've just got hundreds of new citations in Google Scholar Apparently, you can search your name and first letter + last name and find tons of badly parsed versions of your papers. (Yes, citations don't matter, but making camera ready is boring, and I needed an escape,...)

ok the thing about erdos is he obviously loved collaborating with humans. he could have done a lot on his own, but math was how he chose to connect. I'm not sure he would have been very into chatbots?

At @arxiv.bsky.social, we are receiving a new type of paper that I call an "I did this experiment" paper. These papers typically report some experiment with an LLM or LLM "agentic" workflow. They are the kind of experiments an "insider" engineer would run to optimize a system. 1/

YOU’RE LISTENING TO SISYPHUS FM, THE HOME OF NON-STOP ROCK

yeah I watched like 30 minutes of the doge lad depositions, this is human-level performance on this task

I assigned random gender/ethnicity labels to scientific abstracts from the literature and then asked Claude to do a thematic analysis. Claude identified a clinical versus computational split for female/male authors and a DEI focus for Black/URM authors. All in completely random data.

Our new paper reformulates tokenisation as a linear program (LP), which we solve to get SOTA tokenisers 😁 As a bonus, this LP tells us how close to optimal any tokeniser is! Check it out 👇 w/ J. Tempus, @philipwitti.bsky.social, @craigschmidt.com, D. Komm Paper: arxiv.org/abs/2605.22821

I have been thinking about this in light of Anthropic’s recent verbalization interp paper. It had no evidence convincing me that their verbalizations are faithful, but they are convincingly useful. Even wrong output can stimulate human creativity and increase the entropy of exploration.

IIRC the issue is that numbers like this have more representation in the training corpus as bible verses than as numbers, and in that context this is "correct."

When I look at The Rose I think, do I like this because it is big and I am simple? Do I like it because it is nothing like looking at a flat image on a screen? Or do I like it because it carries some piece of Jay DeFeo, who died in 1989 but gave 8 years to painting and scraping and carving The Rose?

Whatever intuitions you formed three years ago when you first heard of ChatGPT (and the people boosting it...), you really should be updating them. I'm but a humble physicist but I'd even say this story is the Deep Blue-Kasparov moment for mathematics

Maybe it is a good day to go to the Whitney in NYC and look at The Rose. It is very big. A human spent 8 years painting it. She made it too big and couldn't get it back out her door. They had to cut a hole in the wall and forklift it out.

my new literary award cannot be won by a commercial frontier LLM because I will require that 10% of each submission is smut

I won't claim this is the most embarrassing social media post I made as a teenager, but it may be the most confusing

I tried to make the theory work out but the computer devil kept lying to me (ChatGPT generated incorrect proofs)

If you liked Anthropic's recent emotions paper, check out our work! We find many similarities: 1) Circular geometry of emotion representations 2) Steering: unlike Anthropic, we steer along circular manifold (at 0°, 30°, 60°...) 3) Steering emotions can affect refusal/sycophancy See Lihao's thread!👇

I will ALWAYS read the youtube comments

I’m all for serious penalties for hallucinated citations. But there’s a reason one of the rules of my lab is “Don’t try to solve Library Science”.

If lectures are bad, the obvious replacement is active learning - getting students doing things, collaborating. But if they prefer solitary learning... don't lecture but don't make them talk to each other either? So if the desire is to watch videos and do individual work, then... why come to class?

In the latest “This paper changed my life,” Maria Geffen shares how John Hopfield’s neural networks taught her to always ground research questions in biology. #neuroskyence www.thetransmitter.org/this-paper-c...

Like humans, LLMs can be right for the wrong reasons. They can also be wrong for the wrong reasons. Anthropocentric bias has gone largely unexamined, posing a serious obstacle to the objective assessment of LLM capacities. Read more at doi.org/10.1162/COLI... #NLProc #CLJournal @raphaelmilliere.com

Diffusion models are slow at inference. Drifting Models fix that but then training becomes the bottleneck. Is it possible to slash the training cost of drifting models without sacrificing quality? Our answer: DriftXpress. 🧵 1/6 (thread from Twitter by first-author Ali Falahati, porting here)

I don't really understand this. If you just want a job---at any career point as an AI academic, student, faculty whatever---you can make ~10x the money and work 9-6 without meandering chats with colleagues.

Somewhat late but thrilled to share that our paper on Algorithmic Primitives & Compositional Geometry of Reasoning in Language Models was accepted to ICML 2026. Huge thanks to incredible coauthors, esp @sflippl.bsky.social @eberleoliver.bsky.social @thomasmcgee.bsky.social. Camera ready coming soon.

It was great to present and discuss our paper, “Rethinking Intrinsic Dimension Estimation in Neural Representations,” at #AISTATS 2026 last week🚀 📄 Paper: arxiv.org/abs/2604.20276 Joint work with @davidruegamer.bsky.social

This is really interesting research I have to imagine this type of interruption and stream processing is also relevant to robots and autonomous vehicles too The implicit sense of time is also interesting

Any papers come to mind as really good examples of qualitative evaluation/analysis in NLP?

This week's #PaperILike is "Human-Guided Complexity-Controlled Abstractions" (Peng et al., NeurIPS 2023). Selecting the right levels and kinds of abstractions remains important and open for many forms of human-AI / human-robot interaction. PDF: arxiv.org/abs/2310.17550

I had intended to see The Sheep Detectives (2026) (Rated PG) until Jill Lepore panned its inaccurate portrayal of animal cognition.

📣 The difference between replicable and not replicable is not itself scientifically replicable. 📣 New work with Erkan Buzbas, showing that verdicts such as "X% of results replicated" are based on an inferential machinery that doesn't work. arxiv.org/abs/2604.26268

A lot of AI criticism makes me feel like someone says “the automobile has been disastrous for society” and I’m nodding along like “yeah I actually really agree, I think that car centric urban design has-“ and then they add “also it’s a scam, cars break down after more than ten miles anyway”

In a #ICML2026 position paper we argue a dynamical systems perspective is needed to drive time series models forward: arxiv.org/abs/2602.16864 For TS, we need to move away from transformers that do not respect a system’s dynamical structure, esp. if out-of-domain generalization & insight is sought.

If this works as advertised it might be the most important deep learning paper in a while. People have been working on MoE type models for distributed training for a long time without meaningfully enabling better distributed training. True modularity has been very hard to achieve.

Fields medalist @wtgowers.bsky.social on how large language models are improving, and the dangers this poses. "I would judge the level of the result that ChatGPT found in under two hours to be that of a perfectly reasonable chapter in a combinatorics PhD." gowers.wordpress.com/2026/05/08/a...

happy grading season, from the Institute for the Study of Ancient Cultures in Chicago

this is a very neat initiative

coding agents are not compilers from English to programs. and it is not because they are non-deterministic. gist.github.com/yoavg/b2454c...

Goodfire released a megapost of all the random feature geometry stuff they're finding, and it's worth a read

Unbelievably honoured to read Tatiana Engel's (@engeltatiana.bsky.social) wonderfully written Preview on our work "Linking neural manifolds to circtuit structure in recurrent networks" (with @lpezon.bsky.social & @gerstnerlab.bsky.social) in this issue of Neuron www.cell.com/neuron/fullt... 🙏