I haven't seen anyone share this post here, but it's (1) a very nice guide to interviewing for ML/NLP positions these days and (2) what everyone in my NLP circle wants to talk about in person for the last week, partly out of surprise that coding interviews remain so central.

I'm afraid one thing I've learned lately is that I enjoy being negged. Starting to get gooseflesh whenever Opus tells me my framing needs to be sharpened.

you know sometimes i don't like having an active imagination

I was a grade-school kid for 1976 bicentennial, but I remember it well. It was a massive, nationwide cultural phenom. There was (awful) merch everywhere, and stuff on TV, and communities re-enacting Revolutionary War battles. 250? Crickets. The country is in a deep spiritual funk and no one cares.

Why does using LLMs for coding change your view of them? Bunch of things: 1. Yes, code can be verified, so they can check their work & iterate until they get it right. 2. But also: what makes coding hard is strain on short-term memory. Lots of moving parts. Long context & patience shine here. +

This is a fantastic opportunity: a Fulbright fellowship with the Center for Digital Narrative (@cdn-bergen.bsky.social) at the University of Bergen in Norway. Apps due by September 15!

The upside is, this happening during the current regime will make the stakes of access clear Under a Democratic president people would be a lot less skeptical about the safety rationale

do I have another taker for my next t-shirt order, I hear?

Amazing https://load-bear.ing

100% cosign; strong models being monopolized by oligarchs or nation-states is my actual doom scenario, and would be even if I trusted the nation-state in question. Which. I. Do. Not.

Celebrated Jāņi (midsummer) in Latvia this year! Bonfires and dancing till dawn, flower and oak wreaths, songs about Saule, the sun goddess, that are older than the churches. The Baltics hold a surviving piece of old Europe in living tradition.

This divide is also why half of my friends are "What will we put in the abandoned shells of these data centers when people realize they're useless?" and the other half are "Can you hook me up with some fast Chinese tokens, man? I got a project."

so, using GLM 5.2 on Umans, this is the first time I am experiencing "oh yea, that's the stuff" feeling using these things. like i've had this feeling before about output quality, yes. but i've never had tps this high _with_ good output quality.

"Instead of treating AI models as a finished computational model of the brain, we should treat them as 'model organisms.'" www.thetransmitter.org/artificial-i...

I was wrong about being wrong about being right about being wrong about AI 😔

The goal with my rlhf book is to make the "home on the internet" for the next generation learning post-training. That's why I'm doing all formats (lectures, code, book, discord, model completions... & ofc blog of interconnects). A hub is more lasting than non-fiction writing. Join! rlhfbook.com

In times like these, you might be asking: How hot can Earth actually get? My feature touring Earth history to answer that question for @sciencenews.bsky.social last year feels, uh, timely ... and I just realized that it made the cover! Check it out to be glad you don't live in the Cretaceous. 🧪

Seems like an important finding: “Our results suggest that indiscriminate, group-based sharing is more important for facilitating cumulative cultural evolution than discriminate reputation-based sharing.” Applications for social networks?

You can’t trust them individually (they’ll say literally anything). But if you build a system where they check and verify each other — amazing!

the qanonisation of large parts of the tech left has been one of the saddest things to witness

I think we did mass-scale infrastructural digitization of print a tad too early — in 10-15 years OCR / scanners/ layout analysis / computer vision improved incomprehensibly, but now we’re sitting on a grand pile of Hjaaofeoiou and filtering out everything published pre-1800

As someone else in this boat, I think admitting this is important to avoid pundit-style “I was wrong for the right reasons” cognitive traps. If making predictions that turn out wrong makes you more confident you aren’t doing empiricism.

Anyone commenting on AI should be asked 1) Did you expect few-shot learning? 2) When chat came out, did you say "this will change syllabi"? 3) Did you predict agentic coding? If the answers are no, and they're minimizing the impact of AI in 2026, one has to ask: Can they learn from experience?

My spicy ghost pepper hot, but 100% genuine take: scholars in my specific subfield (sociology of culture) have been using generative AI responsibly in research since at least 2013. I'll even put a date on it: Mohr and Bogdanov, 2013. + www.sciencedirect.com/science/arti...

Self-Compacting language agents This feels like an incredibly obvious (in hindsight) improvement. Rather than compact arbitrarily every N=1M tokens, (1) let the model decide when and (2) how to compact arxiv.org/abs/2606.23525

Have been telling friends about this since the talk version blew my mind. See e.g. this recurring story about Meredith Lexington, an editor who wants to go back in time to change history. Appropriately enough the prompt writer goes back (in time??) to edit Meredith's prompt — more than 30 times.

Preprint with @mariaa.bsky.social and @mellymeldubs.bsky.social on how real people are writing fiction with LLMs based on the Wildchat dataset here arxiv.org/abs/2606.22748.

Using AI to support peer review seems unavoidable, but what quality checks should AI implement? We can take some lessons from metascience on the hard reward design problem that is AI review. I wrote a paper synthesizing a few points the emerging lit seems at risk of confusing. 1/

How can you not paint sunny tree canopies when listening to Nick Drake? . 11”x14” Oil on cradled wood panel

Thanks, Ted! Unfortunately, I won’t be able to attend the workshop in Seoul, but I’d be happy to discuss any aspect of the paper with anybody interested. What are HLLMs actually good for, and when should we use them? Preprint on HAL and PhilPapers. philpapers.org/rec/POIWDH

If you're an undergrad who wants to volunteer in a lab, writing a well-informed email does nothing now. You have to spill coffee on the professor in an elevator or push them out of the path of a truck.

It's interesting that Bluesky is extremely academic, full of subject matter experts who do socials, but the 2010s Academic Influencers who mostly talked about Being An Academic rather than their expertise--those people are on Threads.

well there you go xcancel.com/claudeai/sta...

story shapes, but make it interpretability

great list and sorely needed. things are improving, but it’s incredible how often you still see reporting (and even research) that doesn’t explain what they actually did! llm-checklist.com/checklist/

My newsletter is back, with a new focus! I'm covering the systems, instincts, and incentives that shape what's discovered, what's believed, and what happens when science hits the real world. Subscribe so you don't miss the first new edition on Monday! 🧪

keyboard, not an oracle

I know they are pivoting to health care(?!) but there is still nothing like Midjourney for making strange and atmospheric images and short animations in ways no other AI image generator can do. Here are some strange cities I made with similar prompts but very different styles.

Whenever someone says “no one asked for this” wrt LLMs — I always wish I had papers on hand about natural language search behaviors. Across companies, pre LLMs, there’s always a bucket of queries your engine might fail to handle because…the query writer had high hopes of what could be handled

as a writer, the amount of new technical understanding i have acquired in the last 18 months is wild, especially since i started from a 'the terminal is scary and i wont touch it' baseline

Very welcome cautious take on what historical language models actually do, from @tpoibeau.bsky.social. Section 4, on the ladder of validity, is particularly worth a look: philpapers.org/archive/POIW...

Chatbots were a sealed technology: the conversational interface was separate from things under the hood Agentic coding feels more like a kit: the model can help you build things that contain other models (or pieces of them) as moving parts Much easier to see how you teach this to empower students.

The price to pay for persistent context is being permanently silo-ed in one's random life path, living in the prison of truncated possibilities.

Inevitability discourse cannot be stopped

Everyone would just.

This is the way inevitability discourse works: We all agree (individually) that AI is not inevitable. If anyone said it was inevitable, we'd go "dude! you're not the boss of me!" Also, no one has offered a policy proposal that would do anything to change the (collective) upward slope of adoption. +

Big Sadness doesn't want you to know this, but just quietly being in the presence of someone you love and who loves you in return is very nice, actually.

Please share! Please apply! @culturalanalytics.bsky.social cdh.princeton.edu/opportunitie...