Ted Underwood

Models like Talkie-1930 sound like voices from the past. If they could reliably speak from specified historical vantage points, researchers might also use them to simulate the past. But how reliable are they? Today we release a benchmark answering that question for English contexts 1831-1930.

I think it's vastly underrated how important org chart and team functionality is to building a good LLM. Over 18 months old, and still one of my favorite all time Interconnects posts. It's why we still often see orgs crumble and need to get rebuilt (e.g. Llama/xAI) & why underdogs can still catch up

If I were RFK I don’t think I would give talks lamenting how a culture that once led the world in science became less productive after takeover by ideological extremists

"If we wish to reduce the phenomenon of AI slop flag planting, the main thing we need is for the alternative strategy (publishing lots of low-effort papers) to further lose out for young people." sophie.huiberts.me/blog/2026/me... by @sophie.huiberts.me on the "top 3-paper-only" hiring policy.

Maybe Bluesky is cringe and the users are awful. But I don't feel I am being systematically lied to here, which is more than I can say about most places.

the Newark Penn Station superstructure gives the illusion of a suspended tunnel

An internal model at OpenAI has produced a claimed refutation of ETH! Russell Impagliazzo, Mohan Paturi, and I have been asked to help organize an open and independent examination of this claim. 1/4

you could read a 75-foot stack of Buddhist scripture and still not learn as much about the necessity of detachment from a world of bottomless suffering and ever-changing illusion as we learn from <gestures widely>

TIL: someone has ported English to Rust github.com/Aaryan-Kapoo...

Yea, verily, for the multitude are deceived. They believe themselves to have written code; they believe that the code executed. But all of that is Bubble. And while Bubble doth glister and shine with the colors of the rainbow, all of that passeth away, for substance and value belong only to Profit.

even our socialists worship capital

everybody believes markets are useless at transmitting information until it it useful for them to believe markets are actually great at transmitting information in all cases

People are celebrating or mourning the Ant IPO, but I'm still trying to understand the worldview where frontier labs not being profitable *would have proven* AI isn't transformative. I think it's maybe a capitalist mysticism where social life is mostly Illusion, and only the Profitable is Real?

intersectional tenure track faculty job @ university of washington research in ethical, social, legal, policy, and/or humanistic dimensions of artificial intelligence (AI) and data-intensive technologies in science, medicine, biomedical research and health care apply.interfolio.com/193732

Very good thread. Bookmark it for when you get replies that want to endlessly rehearse "your presumably naive faith in elections vs my cynical dooming"

19% adjusted operating margin, holy smokes

My latest for National Geographic looks at the 131 new Hittite cuneiform tablets unearthed at Šamuḫa (c. 1400 BCE) & what they tell us about everything from augury to an epidemic that killed Hittite children. Thx to @digkabri.bsky.social, Theo van den Hout, @knpazuzu.bsky.social & ISAC for aid. 🪏

🚨 Student researcher position at Google DeepMind! I am recruiting for a 6-months internship in Paris in 2027, on the topic of discrete diffusion. Application link: www.google.com/about/career...

"are you ready to argue about nature vs nurture?" "pfft, I was *born* ready"

Com 90,35% de urnas apuradas, 3,16% de diferença LULA 51,58% outros 48,42%

LULA 50,45% outros 49,55% 57,65% de urnas apuradas

turns out that voting is a very good way to kick right wing authoritarians out of power

This thread has good info. Basically places where counting is complete show Lula performing well enough to win. bsky.app/profile/math...

If Bluesky was charging for português -> english translation they could clean up in the next couple of hours

I spend a lot of time working from cafés in South Brooklyn and I can't emphasize enough that virtually every conversation I've overheard at a neighboring table in the past week has mentioned Nathan Fielder.

Apparently, my "stupidest thing ever" is ... "Soft money will give me more research freedom." 😭😭😭 I was just looking for a funny roasting Claaaaaaude 💔

Oh, this is good

yeah, fair

In a world where most conferences apps are pretty bad one conference app will actually leverage social media in a useful way Come join us at #EMNLP2026 !

I'm presenting a paper on measuring article engagement with interdisciplinary citations, co-authored with Alex Leto and @mlpacheco.bsky.social at #EMNLP2026! Come discuss the science of science and measuring abstract concepts in discourse! arxiv.org/pdf/2601.17020 Poster in Hall H at 11, Oct 27.

If you've noticed a distinct dialect emerging in frontier AI models and struggled to understand what they're talking about, you're not crazy! We turned to psycholinguistics to investigate what makes the writing style so weird, and why this might be happening 🚨New preprint! 🧵

The poll so far between Twitter, Bluesky and Mastodon. I think it looks less absurd than last time, but still pretty clear that the base is divided. bsky.app/profile/mits...

I think this is probably true - but if 'not surprising' (it's the kind of thing people were denying until recently), it's still significant, and has to be taken into account by relevant disiplines, and probably the rest of us too ultimately.

It's not full immersion if you remember you have a body and self outside of the game. It's not full immersion if you remember the language you spoke before

Talking to students last week and they are really terrified. Some of it is misinformation. But it’s also legitimate uncertainty, which I think hurts more when you have a short résumé and long future to pay for. “What exactly is my value? Will I have value?”

what else is like this

the reason you issue preemptive mass pardons is to reassure junior staff who are shitting themselves that they’re safe and they don’t have to flip on you. going to be very important to push back on their validity, which is nil, preemptively also

Let it not be said the Midwest has no history or culture 🫡

Today, in DD/MM/YY form, is 24/10/26, a Pythagorean triple, the last one this century apparently