Ted Underwood

I respect the candor here. But I don't agree with the geopolitics, at all. The line between democracies and autocracies is not as crisp as this pretends. How much do we trust, e.g., an AI company in a competitive-authoritarian state that launches sneak decapitation attacks & b…

Our position on open-weights models anthropic.com

If everything reported here happens, and the US continues on current trajectory, we're looking at a bad timeline. But there's reason to be wary about both of those "if"s. National security priorities and commercial priorities will be at odds. Saber-rattling is an expected outc…

reuters.com

In the "Limitations" section they acknowledge that their study was monolingual, but say a multilingual generalization would be interesting. arxiv.org/abs/2605.26492

Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories arxiv.org

For fans of Talkie-1930 and all the methodological questions raised by historical models, here's a new entrant into the field, TypewriterLM, trained up to 1913. Corpus, instruction-tuning datasets, and event dataset are released. arxiv.org/abs/2606.02991

Pretraining Language Models on Historical Text arxiv.org

So, naive question about these campuses that pay OpenAI $13mil/yr. Why don’t they pour that money into a cluster running the strongest available open-weight model, with API and interactive options. And if the answer is “overhead,” why aren’t we collaborating?

If the Feds’ move against Fable was an isolated thing, you might say “paranoid US regime.” But putting it together with Anthropic’s own efforts to nerf the model on ML research—and the long history of chip controls—it really feels like we’re entering an era where knowledge is behind lock & key. +

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. +

I don’t want to say this more than once a year, but I’m both grateful and proud that my immediate academic network — computational humanities / iSchools / cultural AI — moved off X fairly quickly and without looking back.

Is anything creepier than a language model flattering you using its knowledge of previous convos? ("As you know," or "this is where your prior work on Y will really pay off") A ghost who lives in your computer has been taking notes so it can chat you up. Stalker/yandere vibes. +

Anthropic survey of social scientists tending to support my anecdotal impression: friends who see potential value in AI mostly use it for coding (or use it dialogically), and are wary-to-disapproving about asking it to draft documents or significant sections of them. +

Every academic I know who uses or studies AI is also deeply worried about the technology’s effect on universities. Concern is universal. What separates people is how they think we should respond: whether by pumping the brakes, backing up, or trying to steer through the hazard. +

The constant distraction of new model releases is, I am starting to think, a good reason to be anti-AI. Imagine if they had announced a new model of steam engine every week in the 1840s. Cornelius Vanderbilt would have said "screw this; it's impossible to predict; we're sticking with canals."

It is really necessary to publish drafts about AI instantly. If you let them sit for two months after getting readers' reports, they become historical documents and you have to revise them from scratch.

People who see value in AI are explaining popular anxiety by saying "SV did a bad sales job." Possible. More plausible, for me, is @michellegoldberg.bsky.social: anxiety is highest in the US b/c low trust in leaders generally. But also, come on—why would we ever expect AI not to provoke anxiety?