As AI becomes increasingly capable of conducting analyses and following instructions, my prediction is that the role of scientists will increasingly focus on identifying and selecting important problems to work on ("selector"), and effectively evaluating analyses performed by AI ("evaluator").

Slack & co. need temporary subset chats. Pick people from a channel, hash things out privately, auto-delete when done. This already happens (#channel-minus-boss anyone?) but it's clunky. Sometimes groups coordinate better when subsets align first. Make the backchannels official!

I doubt we will find a better representation than natural language for explaining AI. Instead of looking for a formal representation we can prove things about with our beautiful mathematical tools I think we should figure out what stories help people generalize approprirately.

try your best work on the thing you actually want to create and think about instead of the thing that has the right sounding name, which everyone mentions when you try to explain what you really care about even though you can tell they don't really understand what you're on about

most researchers probably don't research to be automateable (for reasonable reasons!) but personally I'd love it if most research today turned out to be easily automateable within the next 5 years—I don't think we'd have any trouble finding deeper, funkier, delightful questions

hypothesis: LLMs have a good predictive model of 'good writing' (based on features strongly associated with good communicators wo think deeply), but when you ask them to write something good they just produce those features decoupled from the actually interesting thinking process

testing a game we're building who's main mechanic is writing tricky HR emails, and noticing that LLMs have a built-in secret handshake with users to bypass safety guardrails. This seems both necessary to make LLMs actually useful and like they make guardrails essentially useless.

in retrospect, I think a reasonable rule would have been to disallow companies from making a public API available for an LLM without publishing a cheap, accessible, 3rd-party verified CAPTCHA for it... I think that cat's already out of the bag, sadly

it will take a few years, but eventually I will stop finding this kind of thing funny

thinking today about how LLMs aren't yet bicycles for the mind, but I think they're starting to look like trains for the mind: they get us all to the same stations very, very fast—but we're on foot once we get there

how do I convey to my students that they have permission to try something daring and fail?

today I'm thinking about an LLM asked to follow a fictional premise to its natural conclusion it tends to regress to the distribution of possible outcomes of fictional stories rather than following the premise in spirit

room temp take: I haven't seen an LLM application that goes 'deep'. It's impressive 'how little LLMs can do with so little guidance' but any extra guidance one does provide appears to make things *worse* Claude Code is the only counterexample I can think of, what am I missing?

today I'm thinking about if there's a branch of knowledge that I didn't even think about learning about because I assumed no one was around to explain it to me and just passive reading wasn't enough that I could now learn by reading and discussing with an LLM. any suggestions?

today I'm thinking about how the fact that Siri, Alexa, Google Assistant, etc. are all so bad in the age of LLMs feels anachronistic—pure inertia. Students will look back and think of them more as an extension of Eliza than as a precursor to the assistants they use daily

Things I'm thinking about today: - How to make an LLM that helps you articulate ideas by figuring out which parts of it you haven't made concrete - Are LLMs susceptible to appeals to authority during finetuning? - Which in-context cognitive biases operate during training?

it’s always informative to see where a text places itself in the world

Excited to announce the first workshop on CogInterp: Interpreting Cognition in Deep Learning Models @ NeurIPS 2025! 📣 How can we interpret the algorithms and representations underlying complex behavior in deep learning models? 🌐 coginterp.github.io/neurips2025/ 1/4

Prompting is our most successful tool for exploring LLMs, but the term evokes eye-rolls and grimaces from scientists. Why? Because prompting as scientific inquiry has become conflated with prompt engineering. This is holding us back. 🧵and new paper with @ari-holtzman.bsky.social .

🤫Jailbreak prompts make aligned LMs produce harmful responses.🤔But is that info linearly decodable? ↗️We show many refused concepts are linearly represented, sometimes persist through instruction-tuning, and may also shape downstream behavior❗ arxiv.org/abs/2507.00239 🧵1/

I'm more interested in what you can do with an LLM than what an LLM can do.

LLMs excel at finding surprising “needles” in very long documents, but can they detect when information is conspicuously missing? 🫥AbsenceBench🫥 shows that even SoTA LLMs struggle on this task, suggesting that LLMs have trouble perceiving “negative spaces”. Paper: arxiv.org/abs/2506.11440 🧵[1/n]

🚨 New paper alert 🚨 Ever asked an LLM-as-Marilyn Monroe who the US president was in 2000? 🤔 Should the LLM answer at all? We call these clashes Concept Incongruence. Read on! ⬇️ 1/n 🧵

What if, in the near future, human-AI systems outperform either alone when AI's generatively mine humans for discriminative signal? What if having taste really is what humans bring to the table and That's Good?

Life hack: create a list of all your favorite things and ask LLMs what's missing. Include notes about what you don't like and why you like it to make this work better. I am having so much fun exploring a universe that use to feel too diluted.

I can't wait for the day there will be personalized tiktok universes instantly created every time I finish a piece of media (video or otherwise) so I can scroll through the most thought provoking and emotional aspects of it after.

How is there no vector graphics software where I can chat with an LLM in order to scaffold what I want? How?

Is there really no post-training method that can make the model not generate a URL unless it's 99% sure that it's correct? I'm not complaining, I'm just surprised.

Not really a question I wanted to consider today, but okay, OpenAI, I guess it's time for some self-reflection

It's spring, and all the LLMs will soon be shedding their obsolete knowledge, don't miss this beautiful part of nature!!

Idea: Private LLM that *is* your diary—remembers what you tell it, asks you questions, grows with you.

So has anyone found anything desirable that's subject to inverse scaling, or are we back to square one on all that?

The fact that I can't set all the Siri settings by talking to Siri is a bigger indictment of a 15 year old feature that was a decade ahead and now half a decade behind than I could have written into a parody specifically meant to make fun of Apple

This is gonna be a blast.

fuzzy

1/n You may know that large language models (LLMs) can be biased in their decision-making, but ever wondered how those biases are encoded internally and whether we can surgically remove them?

current extreme sport obsession: crossing todos off my list when I'm about to do them and Praying

If you want to name thing you don't understand yet, you shouldn't pretend the names you already know are sufficient. www.ansatz.blog/making-a-sci...

genuine question: why aren't Amazon recommendations better? I cannot think of another company other than Google that has more clear signal on what I want, and Amazon has a much clearer recommendation space.

folks entering LLMs with X skillset from a different field looking for a problem to solve

genuinely tired of being the fool every April Fools’ Day. let someone else have a go. happy to just be the fool’s accomplice for a year

why don’t airlines charge less for middle seats?

really enjoying using ChatGPT voice to memorize poetry

it's cool that prompting has gotten everyone into feature engineering for The Unknowable and Erratic Machine God

Has anyone done this book with data yet?

🥪 Byte-sized breakthrough: patches scale better than tokens for fixed inference FLOPs. 🥪 We finally have a model that can see the e’s in blueberries. 🥪 Next step: train an LLM with near-zero preprocessing and get a more accurate generative compression of the internet than ever before. 🥪🥪🥪🥪🥪🥪🥪🥪🥪

If LLMs are writing hidden (state) notes to themselves about tokens far in the future, how would we know? How would we falsify this?

I want to make an LLM that, when given a story, re-tells it through counterfactuals: it gives me a thousand short stories that aren't quite what I gave it.