Ted Underwood
@tedunderwood.com
Uses machine learning to study literary imagination, and vice-versa. Likely to share news about AI & computational social science / Sozialwissenschaft / 社会科学
Information Sciences and English, UIUC. Distant Horizons (Chicago, 2019). tedunderwood.com
Information Sciences and English, UIUC. Distant Horizons (Chicago, 2019). tedunderwood.com
Pinned
Ted Underwood
@tedunderwood.com
· Jun 25
The impact of language models on the humanities and vice versa
Nature Computational Science - Many humanists are skeptical of language models and concerned about their effects on universities. However, researchers with a background in the humanities are also...
rdcu.be
New this morning, a Comment I contributed to Nature Computational Science on the interaction between large language models and the humanities. 🧪 🤖 #MLSky
rdcu.be/etk07
The link above will be open-access for a month — plus, I'll reply to this post with a link to a permanently open preprint. +
rdcu.be/etk07
The link above will be open-access for a month — plus, I'll reply to this post with a link to a permanently open preprint. +
Reposted by Ted Underwood
Happy to share a recent publication in the @jdighist.bsky.social exploring how historians are applying generative AI. I argue that LLMs can be considered a new form of historical source, and source criticism can inform our critiques of these techs.
journalofdigitalhistory.org/en/article/J...
1/8
journalofdigitalhistory.org/en/article/J...
1/8
Mapping the Latent Past: Assessing Large Language Models as Digital Tools through Source Criticism
This article examines how digital historians can use large language models (LLMs) as research tools while critically assessing their limitations through source criticism of their underlying training d...
journalofdigitalhistory.org
November 11, 2025 at 2:57 PM
Happy to share a recent publication in the @jdighist.bsky.social exploring how historians are applying generative AI. I argue that LLMs can be considered a new form of historical source, and source criticism can inform our critiques of these techs.
journalofdigitalhistory.org/en/article/J...
1/8
journalofdigitalhistory.org/en/article/J...
1/8
Same method I use to visualize thousand-dimensional space: just visualize x y and z axes and then add a label that says “1,000 dimensions.”
I love the addition of a label that says "an accurate 1000-sided chiliagon' to something that isn't it, like a reverse 'The Treachery of Images' by Magrittte, like painting something that isn't a pipe and saying 'Ceci une pipe'
November 11, 2025 at 2:29 PM
Same method I use to visualize thousand-dimensional space: just visualize x y and z axes and then add a label that says “1,000 dimensions.”
Reposted by Ted Underwood
"the heist film offers audiences a spectacle of highly choreographed expertise. From Rififi to Sneakers to Soderbergh canon (Out of Sight Logan Lucky Oceans 11–13), the genre has generated some of cinema’s most powerful allegories of collective action"
homework: read.dukeupress.edu/south-atlant...
homework: read.dukeupress.edu/south-atlant...
November 11, 2025 at 2:16 PM
"the heist film offers audiences a spectacle of highly choreographed expertise. From Rififi to Sneakers to Soderbergh canon (Out of Sight Logan Lucky Oceans 11–13), the genre has generated some of cinema’s most powerful allegories of collective action"
homework: read.dukeupress.edu/south-atlant...
homework: read.dukeupress.edu/south-atlant...
Ha! So the world thinks, but with my new machine—drops two books into vat, pulls lever on wall—POR QUE NO LOS DOS!!
The castle windows flash white <boom> and lightning crackles down the conductor. Out of the vat, extremely angry but quite short, steps Charlotte Brontë with white streak in hair.
The castle windows flash white <boom> and lightning crackles down the conductor. Out of the vat, extremely angry but quite short, steps Charlotte Brontë with white streak in hair.
You can only pick one. I’m sorry.
Jane Austen or Mary Shelley
Jane Austen or Mary Shelley
November 11, 2025 at 12:57 PM
Ha! So the world thinks, but with my new machine—drops two books into vat, pulls lever on wall—POR QUE NO LOS DOS!!
The castle windows flash white <boom> and lightning crackles down the conductor. Out of the vat, extremely angry but quite short, steps Charlotte Brontë with white streak in hair.
The castle windows flash white <boom> and lightning crackles down the conductor. Out of the vat, extremely angry but quite short, steps Charlotte Brontë with white streak in hair.
Reposted by Ted Underwood
November 11, 2025 at 1:51 AM
When I saw the plan that put my office next to “common space” I worried it would be loud. But actually this space attracts the most studious students from across campus, which is a huge morale boost in decades when you’re wondering whether the species has a future
November 10, 2025 at 10:50 PM
When I saw the plan that put my office next to “common space” I worried it would be loud. But actually this space attracts the most studious students from across campus, which is a huge morale boost in decades when you’re wondering whether the species has a future
Models' inability to reason about limited perspectives becomes especially important when it involves the limitations of the user. www-nature-com.proxy2.library.illinois.edu/articles/s42...
November 10, 2025 at 10:21 PM
Models' inability to reason about limited perspectives becomes especially important when it involves the limitations of the user. www-nature-com.proxy2.library.illinois.edu/articles/s42...
Reposted by Ted Underwood
I want Oscar Isaac to play a different version of the Victor Frankenstein character every three years for the rest of his life
November 10, 2025 at 3:00 PM
I want Oscar Isaac to play a different version of the Victor Frankenstein character every three years for the rest of his life
Reposted by Ted Underwood
If you're working on character training research, what're you working on? What is limiting your ability to do the research you want here?
Surely there are more people studying how to modify & steer model personality after the GPT 4o sycophancy incident.
Surely there are more people studying how to modify & steer model personality after the GPT 4o sycophancy incident.
November 10, 2025 at 9:57 PM
If you're working on character training research, what're you working on? What is limiting your ability to do the research you want here?
Surely there are more people studying how to modify & steer model personality after the GPT 4o sycophancy incident.
Surely there are more people studying how to modify & steer model personality after the GPT 4o sycophancy incident.
Reposted by Ted Underwood
Super interesting. For training small LLMs, they get results comparable to those trained on an order of magnitude more tokens by replacing normal pre-training (which ingests huge amounts of human-written text) with exclusively synthetic text derived in a structured way from Wikipedia.
SYNTH is a radical departure from the classic pre-training recipe: what if we trained for reasoning and focused on the assimilation of knowledge and skill that matters? At its core it’s an upsampling of Wikipedia 50,000 “vital” articles. huggingface.co/datasets/Ple...
November 10, 2025 at 9:47 PM
Super interesting. For training small LLMs, they get results comparable to those trained on an order of magnitude more tokens by replacing normal pre-training (which ingests huge amounts of human-written text) with exclusively synthetic text derived in a structured way from Wikipedia.
Reposted by Ted Underwood
Breaking: we release a fully synthetic generalist dataset for pretraining, SYNTH and two new SOTA reasoning models exclusively trained on it. Despite having seen only 200 billion tokens, Baguettotron is currently best-in-class in its size range. pleias.fr/blog/blogsyn...
November 10, 2025 at 5:30 PM
Breaking: we release a fully synthetic generalist dataset for pretraining, SYNTH and two new SOTA reasoning models exclusively trained on it. Despite having seen only 200 billion tokens, Baguettotron is currently best-in-class in its size range. pleias.fr/blog/blogsyn...
Don't think I have value-added on politics, so I'm just sharing research. But I enjoy doing it here in an atmosphere of incandescent rage — because that's my secret cap, I'm always incandescent
November 10, 2025 at 4:40 PM
Don't think I have value-added on politics, so I'm just sharing research. But I enjoy doing it here in an atmosphere of incandescent rage — because that's my secret cap, I'm always incandescent
Reposted by Ted Underwood
Opening the black box of character training
Some new research from me!
Exploring how easy it is to craft personalities like sycophantic chatbots, and exploring how this will change as we move from chat to agents.
www.interconnects.ai/p/opening-th...
Some new research from me!
Exploring how easy it is to craft personalities like sycophantic chatbots, and exploring how this will change as we move from chat to agents.
www.interconnects.ai/p/opening-th...
Opening the character training pipeline
Some new research from me!
www.interconnects.ai
November 10, 2025 at 3:40 PM
Opening the black box of character training
Some new research from me!
Exploring how easy it is to craft personalities like sycophantic chatbots, and exploring how this will change as we move from chat to agents.
www.interconnects.ai/p/opening-th...
Some new research from me!
Exploring how easy it is to craft personalities like sycophantic chatbots, and exploring how this will change as we move from chat to agents.
www.interconnects.ai/p/opening-th...
Reposted by Ted Underwood
Not sure whether James is being ironic here but this unironically is one of my favorite parts of the deal
I know everyone is upset about the deal, but don’t worry. We’ll be doing this again in a month.
November 10, 2025 at 2:57 PM
Not sure whether James is being ironic here but this unironically is one of my favorite parts of the deal
Reposted by Ted Underwood
⚠️ New paper! Why do words sound so similar? In an agent-based model + communication game, we show that production/comprehension pressures trade off to shape lexicon structure.
In @cognitionjournal.bsky.social w/ @simonkirby.bsky.social & Jenny Culbertson.
www.sciencedirect.com/science/arti...
In @cognitionjournal.bsky.social w/ @simonkirby.bsky.social & Jenny Culbertson.
www.sciencedirect.com/science/arti...
The lexicon adapts to competing communicative pressures: Explaining patterns of word similarity
Cross-linguistically, lexicons tend to be more phonetically clustered than required by the phonotactics of the language; that is, words within a langu…
www.sciencedirect.com
November 10, 2025 at 11:59 AM
⚠️ New paper! Why do words sound so similar? In an agent-based model + communication game, we show that production/comprehension pressures trade off to shape lexicon structure.
In @cognitionjournal.bsky.social w/ @simonkirby.bsky.social & Jenny Culbertson.
www.sciencedirect.com/science/arti...
In @cognitionjournal.bsky.social w/ @simonkirby.bsky.social & Jenny Culbertson.
www.sciencedirect.com/science/arti...
Recent convos with Deger Turan and @xiaoningwang.ca have converged to persuade me that interpretability could be where LLMs outdo older NLP tools for cultural analysis.
I know that seems exactly wrong. Everyone knows interpretability is the *problem* with LLMs: they’re black boxes. But, maybe not?+
I know that seems exactly wrong. Everyone knows interpretability is the *problem* with LLMs: they’re black boxes. But, maybe not?+
November 10, 2025 at 2:20 PM
Recent convos with Deger Turan and @xiaoningwang.ca have converged to persuade me that interpretability could be where LLMs outdo older NLP tools for cultural analysis.
I know that seems exactly wrong. Everyone knows interpretability is the *problem* with LLMs: they’re black boxes. But, maybe not?+
I know that seems exactly wrong. Everyone knows interpretability is the *problem* with LLMs: they’re black boxes. But, maybe not?+
Reposted by Ted Underwood
So many nonsense ad hoc pipelines could be prevented by requiring that they work on synthetic data.
I tend to think of experiments as special cases of inference, since most of the problems I work on cannot be studied in experiments. But I get that many researchers see experiments as base analogy.
I tend to think of experiments as special cases of inference, since most of the problems I work on cannot be studied in experiments. But I get that many researchers see experiments as base analogy.
"Validate With Simulated Truth: A first habit is to test whether an analytical pipeline can recover known conditions."
Very good advice below. So much COVID nonsense (e.g. 'immunological dark matter') basically came down to a non-identifiable model that hadn't been properly tested.
Very good advice below. So much COVID nonsense (e.g. 'immunological dark matter') basically came down to a non-identifiable model that hadn't been properly tested.
Modelling Like an Experimentalist
Dahlin et al. (2024) apply experimental thinking to a model of mosquito-borne disease transmissions.
onlinelibrary.wiley.com
November 10, 2025 at 12:41 PM
So many nonsense ad hoc pipelines could be prevented by requiring that they work on synthetic data.
I tend to think of experiments as special cases of inference, since most of the problems I work on cannot be studied in experiments. But I get that many researchers see experiments as base analogy.
I tend to think of experiments as special cases of inference, since most of the problems I work on cannot be studied in experiments. But I get that many researchers see experiments as base analogy.
Reposted by Ted Underwood
The new Rosalía album is astonishing. Run to your nearest source of music
November 10, 2025 at 3:31 AM
The new Rosalía album is astonishing. Run to your nearest source of music
Reposted by Ted Underwood
Some interesting stuff here on measuring writing quality and improving on qualitative tasks:
www.dbreunig.com/2025/07/31/h...
www.dbreunig.com/2025/07/31/h...
November 10, 2025 at 3:11 AM
Some interesting stuff here on measuring writing quality and improving on qualitative tasks:
www.dbreunig.com/2025/07/31/h...
www.dbreunig.com/2025/07/31/h...
Reposted by Ted Underwood
One of my favorite LLM use cases is literally just asking it what the closest preexisting concept to a thought is.
"I have this idea, has anyone had this idea before and what's it called?"
It used to be REALLY DIFFICULT to reliably answer questions like that and I feel like we'll forget it was.
"I have this idea, has anyone had this idea before and what's it called?"
It used to be REALLY DIFFICULT to reliably answer questions like that and I feel like we'll forget it was.
November 10, 2025 at 12:02 AM
One of my favorite LLM use cases is literally just asking it what the closest preexisting concept to a thought is.
"I have this idea, has anyone had this idea before and what's it called?"
It used to be REALLY DIFFICULT to reliably answer questions like that and I feel like we'll forget it was.
"I have this idea, has anyone had this idea before and what's it called?"
It used to be REALLY DIFFICULT to reliably answer questions like that and I feel like we'll forget it was.
There’s a subgenre of recent SF that’s all about the opening half-hour of a heist flick — you introduce a sequence of characters with unique skills, weld them into a team, and that’s … kind of the whole plot. Gibson’s Agency and KSR’s Ministry of the Future both fit. Honestly, I don’t hate it.
November 9, 2025 at 7:59 PM
There’s a subgenre of recent SF that’s all about the opening half-hour of a heist flick — you introduce a sequence of characters with unique skills, weld them into a team, and that’s … kind of the whole plot. Gibson’s Agency and KSR’s Ministry of the Future both fit. Honestly, I don’t hate it.
Reposted by Ted Underwood
Testing Catalog. What's new? Issue #219 🗞️ www.testingcatalog.com/email/99b3267a… #AI #news #KimiK2Thinking (& lots of other updates)
November 9, 2025 at 4:29 PM
Testing Catalog. What's new? Issue #219 🗞️ www.testingcatalog.com/email/99b3267a… #AI #news #KimiK2Thinking (& lots of other updates)
Reposted by Ted Underwood
They’re wearing masks so they can’t be prosecuted later. As long as they think they can’t be prosecuted, their crimes will intensify. They will torture, starve and kill, as they have already, on greater and greater scale.
Signed, someone who’s covered the unaccountable War on Terror for 23 years.
Signed, someone who’s covered the unaccountable War on Terror for 23 years.
I’ve come to believe that we need to gather considerable forces and a campaign to demand the removal of face masks by ICE, Border Patrol, FBI and police.
It is a practice in conflict with the principles of transparency & accountability that are central to the concept of democracy.
It is a practice in conflict with the principles of transparency & accountability that are central to the concept of democracy.
Video on social media shows an immigration agent pulling a gun in Little Village and holding it to the side — which is not an appropriate or safe way to hold a gun. (Among other issues.)
November 9, 2025 at 2:36 PM
They’re wearing masks so they can’t be prosecuted later. As long as they think they can’t be prosecuted, their crimes will intensify. They will torture, starve and kill, as they have already, on greater and greater scale.
Signed, someone who’s covered the unaccountable War on Terror for 23 years.
Signed, someone who’s covered the unaccountable War on Terror for 23 years.
Reposted by Ted Underwood
I'm teaching "Intro to NLP" for our grad students next semester, and I'm curious how others are teaching such courses, in our current "era of AI." I've seen ideas (no tech in class, commonplace books) for smaller seminars, but how to do this in large, structured CS classes? Any success stories?
November 9, 2025 at 3:05 PM
I'm teaching "Intro to NLP" for our grad students next semester, and I'm curious how others are teaching such courses, in our current "era of AI." I've seen ideas (no tech in class, commonplace books) for smaller seminars, but how to do this in large, structured CS classes? Any success stories?
Reposted by Ted Underwood
Great news! This is out: Opening the black box of EEBO academic.oup.com/dsh/advance-...
Opening the black box of EEBO
Abstract. Digital archives that cover extended historical periods can create a misleading impression of comprehensiveness while in truth providing access t
academic.oup.com
November 9, 2025 at 10:30 AM
Great news! This is out: Opening the black box of EEBO academic.oup.com/dsh/advance-...