Ted Underwood

I love it. Clair de Lune in the background is a great choice; it helps put me in the right mood to accept that some things can never be fixed or explained

ChatGPT pen test yields unexpected results in 2026 #chatgpt #ai YouTube video by FatherPhi

We're still safe from being completely replaced by AI... for now. 😂
#ChatGPT #AIFail

Everyone here is playing and experimenting with new ways that social media can be useful instead of draining and bad. It might take a while but the future is obviously here

It’s shocking but I know at least one Victorianist who watches it — with a defrocked Romanticist, in fact

terrifying brain injury, but also, secret to a happy marriage

This season of Bridgerton: best approached as an Oliver Sacks story about a man with so little object permanence that he constantly asks himself, with astonishment, “Wait! Could the 5’1” Korean girl I’ve been dating this week be … the 𝑠𝑎𝑚𝑒 5’1” Korean girl I met at the ball last week?”

nice, twitter could never

Ironically, posting this reminded me to respond to the high school students who sent me a seven-part questionnaire.

Maybe I need to train the Underwood model and just send a HF link when I get requests.

that's a pretty clever idea ... I guess

Of course, if we get enthusiastic about it here, someone will email us and ask us to do RLHF to help perfect the Rodriguez and Underwood models

I encourage researchers to just spin up an "Underwood-model" and estimate my responses instead of emailing me. That's why we're printing up all that VRAM right?

jack & @jack •• • In principle, I don't believe anyone should own or run Twitter. It wants to be a public good at a protocol level, not a company. Solving for the problem of it being a company however, Elon is the singular solution I trust. I trust his mission to extend the light of consciousness. 10:03 PM • Apr 25, 2022

nobody else is ever going to attain this combination of visionary genius and stepping on your own dick in a single reply tweet

the Volturi from Twilight, with very high collars and shiny jackets

they've either been chatting with language models or these dudes

Screenshot showing silicon sample responses to a question about increased gun laws in Australia. There are three question variants, and four parties - the main result is that 5.2 tends to have higher levels of support for restrictions than 5.4, especially for right wing parties.

I was really surprised at how different the silicon sample results were between GPT-5.2 and 5.4. Light red is 5.2 and dark red is 5.4 and it's an Australian-setting so we're seeing big differences for right-wing parties.

Now, quite commonly, I'm interested in historical questions ... because I think they'll produce better models.

True, the mechanics of using models to draw inferences about culture needn't be that different.

What seems different to me is that my interest in the models is less instrumental. A logistic regression model was interesting only because I could use it to learn about some historical question. +

Alt-text: A dark-mode smartphone notification screen labeled “TODAY.” A Techmeme alert at 9:20 AM reads: “OpenAI says it is delaying the launch of ChatGPT Adult Mode, originally planned for Q1, to focus on higher priorities, like gains in intelligence (Alex Heath/Sources).” The source line below repeats: “Alex Heath / Sources.” The Techmeme logo—a stylized white “T” inside a circle—appears to the left of the headline.

A neotenous primate says what?

Another potential source bsky.app/profile/nato...

Chinese Open Source: A Definitive History Open source used to be a niche topic.

Must read on Chinese open source from Kevin Xu with the very similarly named substack (story for another time)

interconnect.substack.com/p/chinese-op...

Indeed I have! Promising start.

maybe if humanity outsources all functional, formal, boring writing to LLMs then we get to finally embrace total grammatical anarchy as signs of humans, redo the language in leetspeak and slangs not yet imagined -- watch me produce next token with so much perplexity that oceans evaporate

Check out our special theme: new missions for NLP research!

We finally built Lathe of Heaven, from science fiction classic Don't Build Lathe of Heaven.

/clear your darlings

Not sure if short skirt or long jacket, but a machete is definitely being used to cut through red tape.

There’s something slippery about doing “science of culture” with 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑣𝑒 models … that are also part of culture, and clearly transforming it.

Lady of Shalott? Or Lathe of Heaven?

as a turn away from human culture to a magic lamp or mirror that merely represents it (I’m thinking Lady of Shalott here)

However… alternative framing is quasi-Marx-on-Feuerbach: “the philosophers have merely described culture in various ways; the point however is to change it.”

Machine culture - Nature Human Behaviour Artificial intelligence tools and systems are increasingly influencing human culture. Brinkmann et al. argue that these ‘intelligent machines’ are transforming the fundamental processes of cultural ev...

Well, the cultural evolution people have not gone away; I would say that research program continues to slowly strengthen. And they definitely have that ambition.

There’s also this project (see below): www.nature.com/articles/s41...

But I do see the shift you’re describing. It could be framed +

And I think some things that present as a critique of anti-AI vitriol may actually be frustration that this platform requires a different kind of engagement than X did; it’s designed for conversation rather than “broadcasting” to an audience of 50k subscribers

📢 Call for Papers! The Korean Journal of Digital Humanities (KJDH) Vol. 3, No. 1 is now open for submissions.

🗓 Deadline: April 1, 2026
💸 Free to submit | Fully open access | DOI guaranteed

Research on #DigitalHumanities, AI & humanities, data papers, and more are welcome!

👇 Submit here: