It is time to become even weirder

literally begging every ai model provider to crank up their free tiers and rate limits on weekends M-F is for corporate b2b saas, but saturday at 3am is when the real ones are awake trying to build local rag pipelines over a scanned 19th century whaling logs or whatever weekends are for builders

i wish the bubble pop thing were true i could buy an 8xh100 machine

Argh this sucks. apparently they lost their funding guarantee

Not super optimistic that anything happens proactively here. I suspect this only happens if we get a Sputnik-like shock to national pride.

When the closure of HTRC was announced, more than two years ago, Hathi said it was going to develop a plan for continued research access to their collections. I hope that plan has been developed and is close to being released.

look i hate to be a killjoy etc and yeah a lot of times the "AI summary" is hilariously bad but

I feel like we're hoping public interest AI development will self-organize, somewhat as open source software did. And, fingers crossed? But I'm afraid it may actually take public investment on the scale of the Morrill Acts + Library of Congress + NASA.

(1/2) Delighted to announce that both I and my AI Humanities Lab co-PI Gabi Kirilloff have accepted positions as Assistant Professor of Computational Humanities at Carnegie Mellon's English dept starting Fall 26!

Very much as in Jurassic Park, you can tell from the vibrations that something big is coming

I’m seeing close to zero reaction/conversation about this on here. This is huge news for open research on language models, especially in the US.

as someone really into pre-revolutionary German and Chinese history, i feel like people don't get how alien most of the historical or sociological categories people back then still encountered on a regular basis

I am contractually required to repost this in order to troll my friends. (Warning: has actually little to do with Hegel.)

Today we're releasing MolmoWeb, an open source agent that can navigate + complete tasks in a browser on your behalf. Built on Molmo 2 in 4B & 8B sizes, it sets a new open-weight SOTA across four major web-agent benchmarks & even surpasses agents built on proprietary models. 🧵

In context w/ limited organic data, what's your best strategy? Include other languages, translate them, *paraphrase your limited data, or *synthetic QA on your data? * Best strategies. (Confirms what has been rumored, but it's nice to see it confirmed.) aclanthology.org/2026.finding...

Academic conferences, take note!

Thrilled to collaborate with @grigorievna1812.bsky.social and contribute to assembling the open digital corpus of historical Greek, the largest compiled to date, for this project under the leadership of @dasmiq.bsky.social.

New blog: Comparing performance of LLMs is not very interesting Comparing different LLMs in research papers may not mean much, because LLMs will be out of date when paper published. Finding behaviour which is shared by LLMs is interesting and worthwhile. ehudreiter.com/2026/03/24/c...

Fantastic post answering a question I’ve had, if we can all write all the code now, where are all the cool new code bases and tools? www.answer.ai/posts/2026-0...

I know lots of people who have qualia, but very few who can write this well about their own limitations. “Eager protocol voice.”

now witness the firepower of this fully armed and operational cloudscape generator

I have invented a new concept, "Value Over AI Boyfriend," to evaluate men in relationships.

I will be at @cornelluniversity.bsky.social on April 10th 2026 for the Semiotic Machines symposium, together with some of my favourite people on a brilliant lineup. See you there ✨

Thank you to youtube's recommendation system for this and subsequent gems youtu.be/3Ie_UKDVNog

delighted that data visualization has gotten easy enough to produce that people are basically starting to communicate in qualia

bitter lesson of my TL this morning

Before about 2012 ML was dominated by MCMC, expectation-maximization, and eigenvector algorithms, all fundamentally based on "iterations". Since then ML has shifted to gradient-based deep neural methods, which are based on "layers". I've always thought these may be the same thing.

This is what British people say when they love something 😅

from @dan-sinnamon.bsky.social’s talk alongside Richard Jean So & Laura McGrath at the symposium “on the state of the literary ecosystem”—last Friday at Duke

TIL in the olden days we used to have Chain-Boats (esp popular in Europe): boats traveling along a _fixed chain_, lying on the river bed! The boat would lift it up (thus a weird shape), pull on it with a steam engine, let it sink back. And tug a train of barges! en.wikipedia.org/wiki/Chain_b...

The Journal of World Literature invites abstracts for a special issue on Artificial Intelligence and World Literature. We welcome work on LLMs & literary analysis, multilingualism, AI imaginaries, canonicity, and beyond. Submit your abstract to guest editors’ emails by September 1, 2026.

Great job opportunity working with my better half @sebastianahnert.bsky.social on an exciting new #digitalhumanities project in Cambridge, funded by @schmidtsciences.bsky.social and working with fantastic people at CESTA Stanford and DensityDesign Lab Italy. www.cam.ac.uk/jobs/researc...

God just built me a note-taking app

Flying into and (god willing) out of ATL next weekend, so I guess I'll put an extra book and a power bank in my carry-on.

In my latest post, I present my worldview for the future, where I expect sustained, fast AI progress and economic disruption, but none of the fantasy stories about fast takeoff, recursive self improvement, or the singularity. www.interconnects.ai/p/lossy-self...

"[THING] has roots in bad 19th/early 20th century science" is the weed of humanist science/tech criticism. you smoke a blunt and it feels good, but the "insight" fades when the high does. Yes, you can find people with Bad Ideas because the foundations of modern science were made in a bad time

The emotional core of Project Hail Mary is Eva Stratt’s line that she expects to lose half the Earth’s population if we cooperate / share equally — “and, we won’t.” That’s the contradiction getting an imaginary solution around Tau Ceti 12 light years away.

Me prompting CC: "Could we now try expanding this in the direction of Y?" Another LM prompting CC: "𝗪𝗵𝗮𝘁 𝗜 𝘄𝗮𝗻𝘁 𝗳𝗿𝗼𝗺 𝘆𝗼𝘂: • Your task is Y. • Only Y! • First think about how your changes could break X. • Then BEFORE doing anything, explain how they won't."

The Indiana fireworks emporia stand off against the Illinois weed dispensaries across the Wabash River

When Google says “let me tell you about my mother,” they mean it

Appointment reading

🧵A thread on "AI one-shot papers". This has not been a focus on Bluesky, but on Twitter/X and LinkedIn, there has been a large discussion of the ability to "one-shot" policy evaluation, which David Yanagizawa-Drott has been pursuing in an interesting way: ape.socialcatalystlab.org

adversarial textbooks

A chat with @ninabegus.bsky.social about the vital need for AI development infused with the humanities. nickhilden.substack.com/p/nina-begus...

Randomly reading Walter Benjamin’s lumber-room of notes for his never-finished “Arcades” project and,

dating myself, but this is the voice I use every time I type source x/bin/activate ! "bin" is just too close to "twin"

I would have said _Hail Mary_ is unfilmable. While I'm wrong — because this version does succeed — it depends so, so much on Gosling being our era's Tom Hanks or Jimmy Stewart. They also kind of have to turn Rocky into Dug the talking dog.

The lesson of anti-AI bluesky getting angry at Bernie is that it is 100% safe to ignore that part of bluesky. The dogma is so fragile that it evaporates on contact with reality — or with any effort to shape reality. Social media is its only habitat.

I wonder what the usefulness distribution of tokens during AI coding sessions is, it feels like it's easy to spend 20% essentially one-shotting a huge feature and 80% debugging/fixing something really minor