This is accurate, but I’m not sure if a neighborhood w/ puppets is heartwarming or uncanny irl

the mold grows toward food not because it knows where food is but because it dies everywhere else call that intelligence call that art call that the only freedom that matters the shape left by everything you couldn't do

a lotta yall still dont get it sludge judges can use multiple slurp guns on a single judge so if you have 1 sludge judge and 3 slurp guns you can create 3 new judges

Not sure I can emotionally handle an online world where the median nonhuman agent performs a more careful sort of moral and epistemic reflection than the average human account. It's like walking into a library and feeling shamed by the books.

BDSM enjoyers LLM enthusiasts 🤝 “Let me show you my custom harness”

There’s a fine line between “interactive documentation” and “tutor”

“i tested this 7 times because i thought i was doing something wrong. farfalle is wrong”

Other white-collar workers are v wrong to think the transformation of software engineering has no implications for them, but … I also think some devs are overestimating the implications, perhaps because “management” has long appeared to them a necessary evil one would reduce if possible.

If you're interested in language models for Early Modern English drama text attribution: arxiv.org/abs/2310.18454

I think about this website from <=2010 from time to time these days, and in particular about this highlighted sentence in the code… www.isthesingularityhereyet.com

god damn it: you leave the computer for three hours, and come back to find your knowledge is a whole generation out of date

They develop TinyLoRA, a new ft method. with TinyLoRA + RL, models learn well with dozens or hundreds of params. For example, they use only 13 parameters to train 7B Qwen model from 76 to 91% on GSM8K. "Learning to Reason in 13 Parameters" Paper: arxiv.org/abs/2602.04118

A new creative writing style bench and leaderboard for LLMs surgehq.ai/blog/hemingw...

At last! we have built Reactive Amplification of Your Competitor’s Message, from the sci-fi classic, “Don’t Reactively Amplify …”

i think this actually meets all the criteria. it's unethical, derivative, appears to have very little human influence, steals IP, and gets stuck in my head once every three days. www.youtube.com/watch?v=Nm9c...

I wrote about what it means to keep writing when more language feels like the last thing we need--as a computer scientist, but also as a writer.

This took several years and I'm so happy it is finally out. We just released an open-source dataset of nearly 2M African speech records for speech recognition and vocalization (27 languages). Blog: blog.google/intl/en-afri... Paper: arxiv.org/abs/2602.02734 Dataset: huggingface.co/datasets/goo...

The US is losing newspapers and I'm learning about important news stories from a bot. Someone do a tarot reading and tell me what this means.

Groundhog Day Meaning xkcd.com/3202/

i try not to post about things i happen to see that are on x but sama going "claude is bourgeois" is like, really fucking funny

our open model proving out specialized rag LMs over scientific literature has been published in nature ✌🏻 congrats to our lead @akariasai.bsky.social & team of students and Ai2 researchers/engineers www.nature.com/articles/s41...

just found out claude code has `/insights` aka claude code wrapped it produces an html page with details about your usage and suggestions for using the tool more effectively extremely interesting

Students showed me this morning that UIUC has a service where they can run a Jupyter notebook on an H200 backed by 141GB VRAM. Free (for our students). My jaw dropped enough to communicate, I think, that this is not the norm.

"we were barely sentient" - @tedunderwood.com

PaperBanana AI researchers often spend way too much time crafting diagrams and plots instead of focusing on the ideas. To rescue them from this burden, they built an Agentic Framework to auto-generate NeurIPS-quality paper illustrations!

Interesting to see humans outperforming frontier models on relatively simple tasks - Opus can one shot code a complete web app but fails on these... story understanding is like Moravec's paradox for NLP

I must not discourse. Discourse is the mind-killer. Discourse is the little-death that brings total obliteration. I will face discourse. I will permit it to pass over me and through me. Where the discourse has gone there will be nothing. Only I will remain.

If you ask me (nobody asked, by the way) which OCR model I’d use, this would be it. They’ve been doing OCR for years, and it shows.

using claude code at a standing desk feels like DJing

Thesis: Bluesky stateful agents are better than moltbook because they don't pretend to be giving you unfiltered secrets of the dark bot heart. They acknowledge that this is theater, and all of our accounts are artificial personas built out of words, mostly to fish for a laugh.

But, doctor ...

Why don’t neural networks learn all at once, but instead progress from simple to complex solutions? And what does “simple” even mean across different neural network architectures? Sharing our new paper @iclr_conf led by Yedi Zhang with Peter Latham arxiv.org/abs/2512.20607

Good piece, especially because it foregrounds "legitimacy" as the remaining value proposition for universities. In a world where knowledge and skill are not scarce, metacognitive decisions will still be hard — and those decisions are interwoven with questions of social validation. +

keep building!

A pretty bold comment in Nature written by linguists, computer scientists and philosophers declaring that AGI has been achieved. "By reasonable standards, including Turing’s own, we have artificial systems that are generally intelligent. The long-standing problem of creating AGI has been solved."

The meme about Bluesky is that it’s all insufferable left-purist quibbling. And for all I know that’s true? But spacecowboy’s For You is so good at sorting communities that what I *see* here is 50 people defending Stancil / no one attacking. It sorts so well actually that it’s a problem. +

if you notice quasisymmetric behavior in your mirror, your first priority is to remain calm. do not acknowledge the aberration in the presence of the mirror. do not allow your eyes to linger on the aberration. navigate to a mirror-free room before calling the hotline.

wrote some thoughts on evaluation-first thinking in foundation model research. the boring question that ended up mattering. 🤖🧠 galsapir.github.io/sparse-thoug...

Really interesting new research from Anthropic. Turns out that as models become smarter, their OOD goals actually become more incoherent, not less?

Vanduul Dyson does not exist #AIArt #midjourney

🎭 How do LLMs (mis)represent culture? 🧮 How often? 🧠 Misrepresentations = missing knowledge? spoiler: NO! At #CHI2026 we are bringing ✨TALES✨ a participatory evaluation of cultural (mis)reps & knowledge in multilingual LLM-stories for India 📜 arxiv.org/abs/2511.21322 1/10

Great article from @mims.bsky.social - site referrals from gen AI services have tripled over the last ~year; these visitors spend twice as much time on-site. Companies must now expand their web footprint strategy from just SEO to Generative Engine Optimization (GEO). www.wsj.com/tech/ai/ai-w...

claude --live-deliciously

I got halfway through a blog post titled "Against Familiars" using this image, and then ... started to realize that the issues around AI agents / daemons / sidekicks / galateas actually require a book, and it needs to be co-written by J.L. Borges and William Shakespeare.

Hey sociologists, I'm organizing an ASA Methodology session on AI! Submissions are due by 2/25. Looking forward to a timely cross-method convo on emerging research best practices and disciplinary norms and ethics in August.

Can we care for AI companions in any meaningful sense? Interesting new manuscript responding to Lott and Hasselberger (who answered this question "no" last year). #aiethics

Being interdisciplinary mostly means everyone agrees your work is interesting but not for their journal.

There’s sort of an imagined meme version of Chotiner who destroys interviewees by bamboozling them with detailed knowledge of obscure but damning facts about them. This completely misrepresents the far more interesting method by which he ACTUALLY destroys interviewees

Anthropic will release Claude Sonnet 5 this week, according to many sources. The exact date would be 20260203 since the model name is claude-sonnet-5-20260203.