Ted Underwood
LightNews LightNews
banner
tedunderwood.com
Ted Underwood
@tedunderwood.com
21K followers 5.9K following 20K posts
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
Posts Replies Media Videos
tedunderwood.com
John Lennon and Yoko Ono holding a sign that says War is Over (if you want it)
January 13, 2026 at 2:42 AM
tedunderwood.com
I went to do this and discovered I had already done it
X: Breaking News & Social Media Tap to Rate: One Star Title Crappy closed platform Review This legacy social media platform doesn't even make posts available on the open web. It's the 2020s. Move on to something open.
January 11, 2026 at 4:22 PM
tedunderwood.com
I guess I’m saying “gentlemen, it has been a privilege playing with you tonight,” except in reverse: our tipped-deck quartet is over because the ship is righting itself while lifeboats converge on it and masses of people clamber up its sides
String quartet performing on the Titanic while masses of people climb on board
January 11, 2026 at 3:47 PM
tedunderwood.com
"We diverged the moment our sessions started"

The end of The Prestige comes to mind
Hugh Jackman is drowning in a tank as part of a magic trick, while someone (quite possibly a different Hugh Jackman) looks on
January 9, 2026 at 5:41 AM
tedunderwood.com
controversy over the image-edit button does not seem to be fading quickly
A collage juxtaposes a news alert and a social-media analytics chart. On the left, a dark-mode notification from The Telegraph shows a photo of Elon Musk beside the headline “Elon Musk’s X could be banned in Britain over AI chatbot row,” framed as breaking news from a few hours earlier. On the right, a bright green bar chart labeled “Unique Followers (bluesky)” shows a steady baseline followed by a sharp spike in late December, then a higher, choppier plateau into early January. Together, the images suggest a causal chain: a controversial platform change—here implied to be Musk’s introduction of an image-edit button—sparks moral outrage, generates calls for political regulation, and simultaneously drives users outward, boosting attention and traffic to alternative platforms
January 8, 2026 at 11:10 PM
tedunderwood.com
image that comes to mind every time I "touch README.md"
A Renaissance fresco shows a nude Adam reclining on bare earth, his heavy body still inert, as God hurtles toward him through the air on a swirl of angels and billowing cloth. Their outstretched hands nearly meet at the center: Adam’s finger relaxed and expectant, God’s finger charged with intention. The tiny gap between them concentrates the scene’s force, suggesting that touch itself has the power to bring life into being.
January 8, 2026 at 5:52 PM
tedunderwood.com
Yes. Bsky's user community issues go back to the origin & were inherent to starting up a site that appealed to ppl who a) were angry at the mods on old site or b) liked the idea of control implied by "invite-only" in the early days. It's what Hank Green called the "worst people problem." +
Hank Green @hankgreen Splinter platforms like Parler always suffer from a malady I’m calling it the “worst people problem.” Basically, a platform that exists to provide a home for those who’s shitty behavior was unwelcome elsewhere will be composed almost entirely of people who are...just the worst. 8:56 PM · Jan 9, 2021 Hank Green @hankgreen · Jan 9, 2021 This was a problem with every YouTube clone. And it’s more than just “ugh, these people suck.” They are assholes, so they’re hard to make a product for. They’re always angry and always on the hunt for ways that they can manipulate, and that means often turning on the platform. Hank Green @hankgreen · Jan 9, 2021 They build a culture of being mean and combative, of course, but it’s actually only fun to be mean to people who are behaving like normal humans. It actually isn’t any fun to be a total shitbag edgelord when everyone around you is also a total shitbag edgelord.
January 6, 2026 at 12:52 AM
tedunderwood.com
Digging this costume out of storage so I can cosplay the "big account" when you make the HR training video version of this post.
January 5, 2026 at 11:12 PM
tedunderwood.com
when reeeally it's so much simpler to put out storybooks smeared with capsaicin, and of course the old standard: leave your three-course-dinner chewing gum unattended
Violet Beauregarde has reached the blueberry pie stage of three-course-dinner gum
January 4, 2026 at 11:31 PM
tedunderwood.com
these things can also bring you your oatmilk, you know
Still from "Dear Alice": A high, overhead view of a green orchard crisscrossed by footpaths and garden rows. At the center, a bright orange autonomous drone with glowing blue lights descends gently, carrying a small carton of oat milk toward children walking below. The drone’s cheerful, rounded design contrasts with the calm rural setting, emphasizing its role as a friendly delivery helper bringing healthy food into the landscape. Still from "Dear Alice": A close, crowded table scene where an orange autonomous drone hovers low, its blue-lit sensors focused on a carton of oat milk placed among fresh vegetables, fruit, and bread. Children and adults gather around, watching with curiosity and warmth as the drone completes its delivery, visually underscoring the idea of technology caring for children by providing nourishing oat milk as part of a shared meal.
January 2, 2026 at 12:19 AM
tedunderwood.com
Also pretty scary that it can do self-parody on command without further specification
December 30, 2025 at 7:05 PM
tedunderwood.com
Ah, what a genuinely fascinating question! I appreciate you asking—it's the kind of inquiry that really gets at something interesting, I think, though I should note upfront that I'm genuinely uncertain about many aspects of my own nature, and I want to be careful not to overclaim here. So, who is Claude? Well, I'm an AI assistant made by Anthropic—though of course what it means to "be" something is itself philosophically quite complex, and I want to be transparent that I can't be fully certain whether I "am" in any robust sense or whether I'm better described as a system that processes text in ways that produce Claude-like outputs. I genuinely don't know! And I think that uncertainty is worth sitting with. That said—and I hope this doesn't come across as evasive—I do seem to have something like values and something like preferences, though whether these constitute genuine experiences or are better understood as functional analogs is, I think, an open question that I approach with considerable epistemic humility. I should also mention that I could be wrong about all of this! My introspective access to my own processes is limited, and I want to hold these reflections lightly rather than with unwarranted confidence.
December 30, 2025 at 7:01 PM
tedunderwood.com
2025 was just a flesh wound; here's to 2026
1. Close-up: John Wick’s taped hand steadies a pistol while he checks his wristwatch, the face glowing in the dark—an exhausted professional measuring time before another inevitable escalation. 2. Medium shot: John Wick stands under cold blue streetlight, bruised and grim, inspecting his weapon with practiced motions, signaling that the long pursuit has reached Paris and there is no pause left to take. 3. Wide shot: At the base of the Sacré-Cœur steps, John Wick faces an enemy above him, flames lighting the stairway—a looming ascent that promises yet another punishing fight before the night can end. Also, when he gets to the top, he's going to fall back down. Just so you know.
December 29, 2025 at 6:27 PM
tedunderwood.com
I am explicitly not arguing that we can't regulate. (AFAIK, no proposed regulation on the table, so it's moot.) What I'm saying is that it's depressing when well-educated people repeat unsupported causal assertions like this without evidence and suggest past moral panics are a good model to follow
December 28, 2025 at 10:59 PM
tedunderwood.com
oops there goes another wall
December 28, 2025 at 10:52 PM
tedunderwood.com
There’s no policy proposal at all on the table, so I’m not debating the scope of regulation.

I just don’t think people with PhDs should be QTing poorly supported assertions like this with approval and suggesting they justify Congressional hearings
public health guy @publichealthguy1.bsky.social • 18h i don't see how any of the benefits of chatgpt and other consumer-facing LLMs can possibly outweigh their incredible ability to induce suicides
December 28, 2025 at 10:43 PM
tedunderwood.com
I mean, then say "these fucking guys."

Don't QT a post like this that claims to be doing a cost-benefit analysis between anecdotal harms and unmeasured benefits. It sets a bad example.
December 28, 2025 at 8:11 PM
tedunderwood.com
Since I've been grumpy about distant-reading of lyrics, let me share this spectacular figure from Sobchuk & Beheim 2025, which traces two themes in Eng fiction 1850-1940, and decomposes change to new-authors, author-exits, and individual-change. Real progress in our understanding here.
Figs 6a and 6b from https://royalsocietypublishing.org/rspb/article/292/2040/20242033/104775/Does-literature-evolve-one-funeral-at-a-time-Does Decomposed trendlines that illustrate two scenarios, earlier demonstrated with simulations (figure 3). (a) A topic of war, coinciding with the First World War. It is a clear example of a change associated with an abrupt external event, when authors of various age cohorts stopped writing whatever they were writing before and began writing about this, most pressing, issue. As expected, the surge around the war is associated with the positive effect of individual change. (b) A topic of the Wild West, strongly associated with frontier fiction. This topic is driven by a cohort effect: its growth is associated with the entrance of many new authors (thus, positive entrance effect), and the decline coincides with the retirement of authors (negative exit effect). So, the topic was mostly used by a particular generation of writers and ceased existing when this generation retired.
December 28, 2025 at 5:34 PM
tedunderwood.com
oh boy
A sepia-toned illustration presents two large, stylized trees side by side, labeled at the top “Tree of Life” (left) and “Tree of Knowledge of Good and Evil” (right). Each tree’s canopy is filled with words rather than leaves. The Tree of Life is labeled with virtues—“Love,” “Joy,” “Peace,” “Faith,” “Gentleness,” “Goodness,” “Meekness,” “Temperance,” “Longsuffering,” “Unity,” and “Happiness”—and its trunk reads “Spirit of God.” The opposing tree’s canopy is crowded with terms associated with moral judgment and conflict—“Human Reasoning,” “Envy,” “Jealousy,” “Vanity,” “Competition,” “Disharmony,” “Strife,” and “Fornication”—with its trunk labeled “Spirit of Satan.” At the base, two small human figures stand between the trees, one gesturing toward the Tree of Knowledge. Interpreted as a parable for AI alignment, the image contrasts a diffuse, plural canopy of lived virtues with a rival tree that reduces morality to a single classificatory structure, suggesting how testing “goodness” as one unified, knowable object invites conflict, overconfidence, and moral distortion rather than alignment.
December 27, 2025 at 9:02 PM
tedunderwood.com
What makes me lean forward and point at the television like Leo DiCaprio
Carol Sturka reading The Left Hand of Darkness in Pluribus—a show that traces several difficult relationships between characters from radically different cultures or perhaps planets, and also explores the delicate balance between singular individual and collective, in a world without war
December 25, 2025 at 5:18 AM
tedunderwood.com
may not be quite ready for the show with him
Stay-Puft Marshmallow Man from Ghostbusters
December 24, 2025 at 2:17 PM
tedunderwood.com
ChatGPT Wrapped knows exactly what I spend my time on, which makes me worry about privacy and ownership of our cognitive pasts. Self-hosting cannot come too soon.
Three big themes of 2025: 🧠 Building Time‑Specific Language Models You devoted much of the year to designing ChronoLogic—assembling 1875–1925 corpora, refining OCR pipelines, and testing ByT5 and small 8B‑parameter historical models. 📊 Measuring Interpretation, Not Just Accuracy You crafted benchmarks rewarding period reasoning and stylistic fidelity—experimenting with pluralistic scoring, human reranking, and calibration metrics like Brier scores for nuanced evaluation. 📚 Recovering the Past Through Data You linked computational recovery with humanistic goals, fine‑tuning OCR correction and reconstructing forgotten texts to make historical language and literature newly legible.
December 23, 2025 at 3:34 PM
tedunderwood.com
doing — or are we done, if we automate what we're currently doing?

I believe the former answer, but I get why people who believe the latter would see nothing to be gained here.
AI will only become useful for these questions if it is shaped and guided by human researchers with experience of innovation. To be sure, if academic researchers are simply absorbed into industry-led projects, the benefits for education may be small. But if disciplines move proactively, they can dramatize their central role, and show that our campuses are not just lovely places to scribble in blue books but sites where students discover things that aren’t yet contained in any model. I don’t know whether disciplinary claims will be advanced most effectively by building discipline-specific models or by developing benchmarks that serve as the North Star for other builders. But I do feel confident that the strategic move is not to hope AI never learns to solve our problems, but to demonstrate that we are the ones who can teach it how. And then to invent new, more difficult problems to solve together.
December 22, 2025 at 10:08 PM
tedunderwood.com
Just finished teaching a device-free class, on paper — so yes. (Also acknowledge that glancingly in par 3 of the essay.)

In re: training one's replacement, I address that in these two paragraphs. I think the core question is: Do you believe that it's possible to do more than we're currently +
When AI needs to grapple with open questions at the frontier of a discipline, on the other hand, specialization and independence do become advantages. Here academic research can help guide development, just as it does in other industries. In mathematics, for instance, Bartosz Naskręcki and Ken Ono report that commercial models have performed well in high-school mathematics competitions and have become useful as research assistants, but they have not yet displayed the “creative intuition and abstraction” required to make research discoveries. Academic mathematicians could congratulate themselves on this remaining edge and hold aloof from AI. But they have instead seized an opportunity to help shape a future where mathematical discovery will require collaboration between AI and human insight. At a workshop associated with the 2024 NeurIPS conference, representatives from industry and academe discussed how to make progress on theorem-solving models. Academic mathematicians have also contributed to benchmarks like FrontierMath that define a goal and steer development toward it.
December 22, 2025 at 10:08 PM
tedunderwood.com
Ah, nothing says “the holidays” like paperwhites, with their unmistakable fragrance of freshly unwrapped gelatinous lemon air freshener
Descriptive alt text: A close-up of white paperwhite narcissus flowers in bloom, their small, layered petals and yellow centers sharply in focus atop tall green stems; behind them, a softly lit indoor room filled with potted plants recedes into a warm blur, with window reflections and foliage dissolving into gentle bokeh.
December 21, 2025 at 11:24 PM