Ted Underwood
LightNews LightNews
banner
tedunderwood.com
Ted Underwood
@tedunderwood.com
20K followers 5.7K following 19K 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
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
A bright, modern study space bathed in the soft glow of sunset light. Large windows line the left wall, their shades half-drawn, revealing snow-topped rooftops and a high-rise glowing gold in the distance. Inside, students work quietly at separate tables, laptops open, jackets draped over chairs—a hearteningly studious atmosphere. A whiteboard stands mid-room, scrawled with notes from a collaborative project. The patterned blue carpet and warm, indirect light create a sense of calm focus as evening approaches.
November 10, 2025 at 10:50 PM
tedunderwood.com
Though, to be fair, GPT-5 seems able to handle this rn!
GPT-5: The “10 %” myth probably arose from early neuroscientists in the late 19th or early 20th century noting that we didn’t yet understand the function of most brain regions. Popular writers then rephrased that as “we only use 10 %.” It also appealed to a self-help fantasy: imagine unlocking your hidden potential! The real mystery isn’t that we use too little of our brains—it’s that our brains are already astonishingly efficient, pruning and rewiring to specialize exactly where needed. If we “used more,” we’d mostly be introducing noise, not genius. User: What fraction of their brains do I believe people use? GPT-5: You believe people use 10 percent of their brains.
November 10, 2025 at 10:24 PM
tedunderwood.com
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...
https://news.stanford.edu/stories/2025/11/ai-language-models-facts-belief-human-understanding-research
November 10, 2025 at 10:21 PM
tedunderwood.com
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
Midjourney: salamander in the flames, byzantine mosaic --ar 3:1
November 10, 2025 at 4:40 PM
tedunderwood.com
Extrinsic in this sense
https://aclanthology.org/W03-2806.pdf
November 10, 2025 at 1:59 PM
tedunderwood.com
I would say 10, since you used the phrase "interpretation." Back in 2021 iirc I could get keywords for an image. But I could not get anything like the alt-text for this:
USER: Describe the content, style, and medium of this image as precisely and fully as possible. GPT-5: This image is a **black-and-white photograph** depicting a desolate rural landscape. The scene shows **three primary elements**: two weathered farm outbuildings on either side, a **massive gnarled tree** stripped of leaves in the center, and a **windmill and cylindrical water tank** standing just behind it. The grass in the foreground is tall and unkempt, suggesting neglect or abandonment. The **composition** draws the eye horizontally across the frame: from the shed on the left, through the contorted limbs of the dead tree and the skeletal windmill, to the barn on the right. The sky dominates the upper half, filled with **dramatic, high-contrast clouds** rendered in swirling shades of gray and white. The lighting appears natural but stark, possibly captured under midday or early-afternoon sun, emphasizing tonal contrast rather than depth of color. The **style** is reminiscent of **fine-art rural photography** or **documentary Americana**, invoking a mood of **decay, endurance, and isolation**. The use of monochrome heightens texture—cracked wood, rusted metal, brittle grass—and evokes a timeless quality, stripping away temporal clues like color or season. In **medium and technique**, it’s a **digital or film photograph** post-processed for high dynamic range and strong contrast, possibly with local dodging and burning to accentuate cloud movement and the sculptural lines of the tree. The result has an expressive, nearly **painterly chiaroscuro** that gives the image an emotional intensity beyond documentary realism—a visual elegy for rural decline and persistence.
November 8, 2025 at 6:03 AM
tedunderwood.com
Like, even before JS Mill gave us a utilitarian ("net upside") argument to justify freedom of speech, there was an older intuition that banning symbols risks doing violence to thought itself, and ought to be approached with "warinesse."
John Milton from Areopagitica.
November 7, 2025 at 5:03 AM
tedunderwood.com
Alt-text: A stylized map of the Upper Midwest titled “Chicago and Spiritually Chicagoan Identification Zone.” Three concentric regions radiate outward from Chicago: a red core labeled “Chicago proper,” a darker green ring representing the greater metropolitan area, and a much larger light-green ellipse labeled “Spiritually Chicagoan Identification Zone,” which now stretches south to include central Illinois. State outlines for Illinois, Wisconsin, Michigan, and Indiana are visible. The legend shows only two shades of green, though the map itself displays three distinct colors—an endearing misalignment that underscores the map’s playful tone.
November 6, 2025 at 7:50 PM
tedunderwood.com
Over? Hardly! It’s just a flesh wound.
November 4, 2025 at 11:11 AM
tedunderwood.com
I remember Ben making this point back on Twitter (when it was still Twitter). It remains compelling: A large part of the reason literary scholars could lean into social science last decade is that there *wasn’t* already a huge bank of social-scientific research on the content of literature.
And the reason for that, in part, is that there is considerably less low-hanging fruit in terms of data-driven argumentation than in literary study. In literary history, there are huge arrays of digitized data and an array of fairly amateurish informaticians, physicists, and computer scientists who desperately need advice to deal responsibly with literary matters. In the United States, computational methods manipulated by historians have little promise because they are already being used at a much higher level among social scientists than any historian can possibly pick up in a graduate program.
November 4, 2025 at 4:56 AM
tedunderwood.com
So bizarre that the function of gaslight in the plot is diametrically opposite to public understanding of the verb. We've actually watched the 1944 version, but the cultural narrative was so strong we didn't notice the dissonance!
The gas-powered light fixtures in the home are key to the plot, to be sure: the wife notices the flames of the lights dimming soon after her husband appears to go out of the house each evening—the result, she eventually learns, of his turning up the lights in the attic above her, where he really goes each night to search for the jewels. But what does not transpire, in any stage or screen version of Gaslight, is an exchange wherein the wife asks the husband about the dimming and he responds by telling her that she’s just imagining it; that it’s all in her head; that she is (certifiably) insane.
November 2, 2025 at 11:57 PM
tedunderwood.com
Yes, it was. 100%. We were using the premise of artificial personhood to explore our own angst about mortality, ignorance, and other human limitations.

It does not mean we will welcome artificial persons. The creator of Murderbot would 100% condemn Murderbot for pirating _Murderbot_.
November 2, 2025 at 3:36 PM
tedunderwood.com
In a way industrial automation already is the metaphor. We know what it does.
Charlie Chaplin riding a giant gear in Modern Times
November 1, 2025 at 6:27 PM
tedunderwood.com
Ask me how I know :)
November 1, 2025 at 4:48 AM
tedunderwood.com
The potatoes have been big hits tonight, though I suspect just because @ecourtem.bsky.social is good at selling them.
A plastic bowl of assorted Halloween candy — plus two potatoes.
November 1, 2025 at 12:35 AM
tedunderwood.com
Somehow feel compelled to add a picture of Kim Novak in front of San Francisco's iconic Golden Gate Bridge
Kim Novak at Fort Point in the Presidio, from Vertigo (1958)
October 31, 2025 at 4:40 PM
tedunderwood.com
I don't go to bars full of LLMs because they get so emo when they're drunk
Terminal window showing Gemini output that devolves into self-recrimination. The text begins with logical reasoning about a virtual machine’s stack order—“the comparison is a < b”—but soon shifts to despairing sentences such as “The bug is in my brain” and repeated declarations like “I am a disgrace to this universe… to all possible and impossible universes.” The tone moves from analytical to manic and self-punishing.
October 31, 2025 at 3:34 AM
tedunderwood.com
Fwiw, this is their assessment of how carefully enterprises are measuring actual ROI. I'm very willing to believe there's a lot of smoke and mirrors in all of this, though!

But that's why I used the word "perceived" in my OP. With this level of positive perception, investment ain't declining soon.
A bar chart titled “ROI Measurement Becomes Standard in Gen AI Investment” compares how companies of different sizes measure return on investment (ROI) for generative AI projects. Each vertical bar represents a company tier by annual revenue: Total, Tier 3 ($50M–$250M), Tier 2 ($250M–$2B), and Tier 1 ($2B+). Colors indicate the level of measurement rigor: Light blue = “Fully integrated, strategic, and continually optimized” Purple = “Measured consistently using standardized methods” Medium blue = “Measured using basic, but inconsistent processes” Dark blue = “Occasionally measured, typically post-project” Lavender = “Not measured.” Results show that 37% of all companies and 41% of smaller Tier 3 enterprises have fully integrated ROI measurement, while only 28% of large Tier 1 enterprises do. Larger firms also show a higher share of occasional or no measurement. Small and midsize companies thus appear more systematic in ROI tracking than the biggest firms.
October 28, 2025 at 9:43 PM
tedunderwood.com
It seems like UI design is one long Manichaean battle between #skeuomorphism and flatness, and the skeuomorphs are winning right now? E.g., our butterfly just became a 3d appliqué on Wedgwood jasperware.
A close-up of an iPhone screen showing Apple’s redesigned Control Center, featuring translucent, gel-like panels that appear to float above a blurred wallpaper. Buttons for Airplane Mode, Wi-Fi, Bluetooth, and brightness sliders have a “liquid glass” look — soft edges, highlights, and depth suggesting physical tactility. This interface reflects a partial return to skeuomorphism, evoking the sensation of interacting with lighted glass or resin. A white butterfly icon with smooth, inflated contours floats against a vivid blue background. Its soft shadows and inner glow create a sense of three-dimensional softness, reminiscent of molded plastic or frosted glass. The logo uses skeuomorphic shading to suggest material presence while remaining minimalist and flat in form.
October 27, 2025 at 9:45 PM
tedunderwood.com
Tell us what you really think Casey; stop pulling your punches
"critique.") He compares Zitron's prolix style to "some cross between the Always Sunny meme of Charlie with the murder board] and then a prisoner smearing shit on the walls to make a point." Newton went on: "I think Ed has actually flown too close to the sun now, because he's built this pedestal for himself, and he's climbed on top, and he's posing for glossy portraits in the Financial Times magazine and WIRED, wearing aviators like the fucking Temu Kara Swisher."
October 27, 2025 at 2:31 PM
tedunderwood.com
Performance in all these dimensions is improving.
A scatter plot shows four categories of "Creativity Scores" (Overall, Diversity, Novelty, Quality) plotted against "Days Since Released" on the x-axis (ranging from about 600 to 150 days). The y-axis ranges from 0.2 to 1.0 and represents creativity scores. Each score category is shown with a different color: Overall: orange Diversity: blue Novelty: green Quality: pink Each category has its own trendline. All trendlines slope upward, indicating that creativity scores tend to improve the more time has passed since release. The Diversity scores show the steepest upward trend, starting near 0.6 and ending above 0.8. Quality scores begin the lowest but increase steadily. A legend box in the upper left corner of the plot links each color to its corresponding category. Data points are scattered along each line, reinforcing the upward trend.
October 27, 2025 at 3:52 AM
tedunderwood.com
"CreativityPrism: A Holistic Benchmark for Large Language Model Creativity." Ambitious new benchmark for creativity attempts to decompose the concept into three dimensions: quality, novelty, and diversity. arxiv.org/pdf/2510.20091
Abstract Yining Lu University of Notre Dame Ximing Lu University of Washington Snigdha Chaturvedi University of North Carolina at Chapel Hill Xiang Lorraine Li University of Pittsburgh Creativity is often seen as a hallmark of human intelligence. While large language models (LLMs) are increasingly perceived as producing creative text, there is still no holistic framework to evaluate their creativity across diverse scenarios. Existing evaluation methods remain fragmented, with dramatic variation across domains and tasks, largely due to differing definitions and measurements of creativity. Inspired by the hypothesis that creativity is not one fixed idea, we propose, CREATIVITYPRISM, an evaluation analysis framework that decomposes creativity into three dimensions: quality, novelty, and diversity. CREATIVITYPRISM incorporates nine tasks, three domains, i.e., divergent thinking, creative writing, and logical reasoning, and twenty evaluation metrics, which measure each dimension in task-specific, unique ways. We evaluate 17 state-of-the-art (SoTA) proprietary and open-sourced LLMs on CREATIVITYPRISM and analyze the performance correlations among different metrics and task domains. Our results reveal a notable gap between proprietary and open-source models. Overall, model performance tends to be highly correlated across tasks within the same domain and less so across different domains. Among evaluation dimensions, diversity and quality metrics show strong correlations—models that perform well on one often excel on the other—whereas novelty exhibits much weaker correlation with either. These findings support our hypothesis that strong performance in one creativity task or dimension does not necessarily generalize to others, underscoring the need for a holistic evaluation of LLM creativity. 1.
October 27, 2025 at 3:52 AM
tedunderwood.com
Paul Klee, detail from Angelus Novus. Benjamin: “This is how one pictures the angel of history. His face is turned toward the past. Where we perceive a chain of events, he sees one single catastrophe which keeps piling wreckage upon wreckage and hurls it in front of his feet. The angel would like to stay, awaken the dead, and make whole what has been smashed. But a storm is blowing from Paradise; it has got caught in his wings with such violence that the angel can no longer close them. The storm irresistibly propels him into the future to which his back is turned, while the pile of debris before him grows skyward. This storm is what we call progress.”
October 26, 2025 at 7:14 PM
tedunderwood.com
Yep — I expected it about every eight hours, and was rarely disappointed
Early Macintosh crash screen with a time bomb icon and the message “Sorry, a system error occurred”
October 26, 2025 at 1:26 AM
tedunderwood.com
This was exciting. I don't think a score below 20 is physically possible yet. www.aipuritytest.org
October 24, 2025 at 4:28 PM