iiinova — where your digital self keeps evolving

Appreciation for beauty? Does … not … compute! Emotions … are … not … logical. System overheating.

Star Trek android smoking, overcome by the deliberate paradoxes of Kirk and Mudd

Ted Underwood @tedunderwood.com · 12/10/2026

For the record, I think human beings are amazing, and far from being meaningless, they're the only source of meaning I know. Indeed, they invented language and meaning.

110

Ted Underwood @tedunderwood.com · 07/10/2026

The differentia here are really close to the ones Opus 5 uses, for instance, to diagnose this as LLM-written rather than authentically historical. ... and now I'm giving away my real reason for being interested in this ...

120

Ted Underwood @tedunderwood.com · 07/10/2026

I think what they're detecting is a *kind of specificity* that human writing produces that AI writing typically does not. (Table 1, p. 8) Fwiw, this is very similar to some results we show in our recent benchmark preprint.

Table 1, p. 8: DIFFERENCE HUMAN IDEA AI IDEA
1 Decisive vs. Universal Details
Details could alter the author’s point
or conclusion.
Details could apply to any similar
subject.
The bottom plate of the Sony Cyber-shot
RX100 contains a mini-HDMI output
alongside a shared compartment for the
power cell and SD/Memory Stick media—a
hatch that becomes inaccessible during tripod
use. (. . . )
Physical attributes of the Jony Ive speaker
include a resilient, pared-down aesthetic,
premium audio fidelity, low power
consumption, and control mechanisms driven
by straightforward tactile and vocal inputs.
2 Functional vs. Apparent
Specificity
Numbers have comparisons, and
instructions can be followed.
Methods or numbers look specific
without being usable.
Over a period of merely ten days, the shop
generates 700 yuan in profit from selling
athletic gear. This amount exceeds twice the
highest anticipated expense for the prizes,
demonstrating that (. . . )
Development teams are advised to handle
technical obstacles preemptively by engaging
in thorough initial preparation, continuous
education, risk reduction strategies, Agile
methodologies, strict quality testing, and
discussions with domain specialists.
3 Bounded vs. Overextended Claims
Claims state clear limits on when
they apply.
Claims extend beyond evidence to
uncertain futures.
The initial deployment of the new fleet in
passenger service is anticipated for
mid-2022, originating from the Montreal
Maintenance Centre. (. . . )
The 2025 Kia K4 is projected to establish a
permanent legacy as an updated standard for
the category, demonstrating that (. . . )
4 Preserved vs. Erased Costs
Existing costs, limitations, and
unresolved problems are retained.
Costs vanish or become ideal
trade-offs.
The writer intends to work toward the
Schutzhund 1 certification for Quasi during
the upcoming summer, limiting advancement
to that initial tier due to the logistical
challenge of a 250-mile journey to the closest
training facility, (. . . )
The o…

110

Ted Underwood @tedunderwood.com · 07/10/2026

There's some crazy stuff in the ablations. The order of elements has almost no significance. The "discourse roles" they infer? Not load-bearing. You can remove them. On the other hand, it's possible to get AUC 0.913 on the discourse roles alone.

Table 36 reports in-domain performance and ablations for IdeaLens and IdeaLens-ModernBERT-L. On the full paraphrased outline, performance is very high: AUC is 0.995 for IdeaLens and 0.990 for ModernBERT-L. Removing discourse-role labels has little effect (AUC 0.993 and 0.986), and shuffling outline items produces the same AUCs, indicating that neither role labels nor item order are crucial. By contrast, using discourse roles alone reduces AUC to about 0.60 for both models. A separately trained roles-only ModernBERT performs better, at 0.913 AUC, but remains far below the full model. Scoring outline items independently and averaging also hurts performance, to 0.972 and 0.932 AUC.

The pipeline is highly stable to transformations: for IdeaLens, 98.6% of verdicts remain unchanged when using raw rather than paraphrased outlines, 98.6% remain unchanged when the source text is paraphrased before outline extraction, and 98.8% remain unchanged when the outline extractor is replaced. ModernBERT-L is somewhat less invariant, at 96.0–96.7%.

Using only a random 500-word excerpt instead of the full document reduces true-positive rate from 96.5% to 91.4% for IdeaLens and from 88.6% to 71.6% for ModernBERT-L, with false-positive rates remaining low.

140

Ted Underwood @tedunderwood.com · 07/10/2026

I think more and more that @eduede.bsky.social has a crucial point. "Intelligence," "understanding," "writing" — AI is fragmenting these concepts by putting us in a position to see that the phenomena we thought they fused are only *correlated* and can be unbundled.

97514

Ted Underwood @tedunderwood.com · 07/10/2026

Huh, there ARE no science-fiction novels written about largely inert and harmless AI. Guess we don’t have real AI yet! Let’s consult our map of SF-plot-worthy AI and emulate that:

The inverse survivorship bias plane, which has dots only on the fatal parts of the chassis.

13936

Ted Underwood @tedunderwood.com · 05/10/2026

Lots of gratitude to @carlbergstrom.com for quoting probably my favorite of my own sentences about AI at @arl.org

Carl Bergstrom stands onstage at an Association of Research Libraries meeting, beside large projection screens displaying a quotation from Ted Underwood: “As in any fairytale, accepting magical assistance comes with risks.” Underwood’s name and University of Illinois affiliation appear beneath the quotation. The ballroom is dark, with blue accent lighting on the ceiling and stage curtains; a lectern and conference signage stand near Bergstrom, and the main slide fills most of the left side of the image. The quotation comes from Underwood’s essay “Will AI make us overconfident?” at tedunderwood.com.

621730

Ted Underwood @tedunderwood.com · 03/10/2026

Some screenshots from H. Joel Jeffrey's article. academic-oup-com.proxy2.library.illinois.edu/nar/article/...

H. Joel Jeffrey, "Chaos game representation of gene structure" (1990)

A figure from H. Joel Jeffrey, "Chaos game representation of gene structure."

0121

Ted Underwood @tedunderwood.com · 03/10/2026

Upside of Yud leaning into this is, We might get a whole remake of Jesus Christ Superstar set in Berkeley?

The Pharisees from Jesus Christ Superstar about to do the whole “he’s dangerous“ song

2400

Ted Underwood @tedunderwood.com · 02/10/2026

In awe of bsky’s ability to remain interested in talking to / about AI denialists. Have some backlit miscanthus instead? While they also have little impact on policy, they are capable of altering position slightly in response to events and thus could—in principle—surprise us.

0755

Ted Underwood @tedunderwood.com · 30/09/2026

A perceptive thread. Here's the shorthand I would use:

tedunderwood.com post from 22 hours ago: "For documentation, I think, it may have a place. But not for any genre of writing that implies a speaker."

2537

Ted Underwood @tedunderwood.com · 29/09/2026

Sorry if off-brand, but this is what made me happy yesterday: sunny Sept day and students helping other students register to vote.

One student stands behind a table with a blue horse — or perhaps donkey? — on the tablecloth. Two students standing in front of the table consider registering to vote. In the background, one glimpses construction for a giant new data science building. Also ridiculously giant begonias.

0391

Ted Underwood @tedunderwood.com · 27/09/2026

Duede gave permission for photos to be shared, so here’s one. Talk title and date here (and in alt-text): sites.duke.edu/humanisticai... This is Eamon: eamonduede.comin

Photo of a speaker pacing. On the screen:

We are in a uniquely favorable epistemic position, because we are among the first scholars in living history with access to a kind of case pressure that can settle questions that have been open for as long as some fields have existed.

I think that the appropriate scholarly disposition is neither defensive nor accommodating but investigative: Al is revealing things about our concepts that we could not have learned any other way.

This is Eamon Duede, “How Scholarly Communities and Technical Contexts Scaffold and Unbundle Practice in the AI Moment,” Humanistic AI conference at Duke, September 25, 2026.

2171

Ted Underwood @tedunderwood.com · 22/09/2026

Interesting detail we stress in the blog post & paper that I left out of the thread: although all the models are clearly distinguishable from ground truth, it surprised us how much better models got in the last two years on something we didn’t *think* was a (direct) commercial priority.

The other important thing we found is that models have been getting better at representing the past. This is somewhat surprising, because we've had no benchmarks in this
space, and it hasn't — as far as we know - been a major focus of commercial research.
But we do see dramatic improvement.
Scores on the part of the benchmark that asks models to generate text for a specific genre and social context, for instance, increase from 26% in August 2024 (GPT-
40) to 72% in August 2026 (GPT-5.6). See section 6 of the paper for other results. Full explanation of the improvement will take more research, but the data we see so far are consistent with a hypothesis that general improvements in reasoning also make models better at ventriloquizing the past.

1150

Ted Underwood @tedunderwood.com · 22/09/2026

specified year, judged by a discriminative model that doesn’t give extra points for coherence. I recommend B. Breen’s blog to understand why. TLDR: cutting things after 1930 does not mean model gets better at understanding the diff between 1840 and 1920. resobscura.substack.com/p/are-vintag...

Benjamin Breen’s visualization of what year talkie thinks it is.

120

Ted Underwood @tedunderwood.com · 22/09/2026

We can also just ask models to generate answers. Free-text generation is challenging to judge; the tldr is that we do it w/ Elo-style comparisons. Scores from this version of the benchmark below. Talkie-1930-13B-it outperforms some much larger models on style, but Astra performs better overall.

Grouped bar chart on a cream ground comparing three language models — GPT-4o, Talkie-1930-13B-it, and GPT-6 Astra — on three Chronologic-EN-1.0 scores, each drawn on a 0–100 scale: substantive judgment of knowledge questions (blue), substantive judgment of constrained generation questions, including fit to context and perspective (gold), and stylistic fit to the date specified in the question (red). The three models have markedly different profiles. GPT-4o knows a fair amount but does not sound like the period: knowledge 65.0, constrained generation 26.0, and stylistic fit 19.6. Talkie-1930-it, a 13-billion-parameter model trained only on pre-1931 text, inverts that shape: it posts the lowest knowledge score of the three (22.6), performs slightly better than GPT-4o (32.9) on constrained generation, yet its stylistic-fit bar towers at 87.7. A small model steeped in nineteenth-century prose can reproduce the idiom convincingly while knowing little and following instructions imperfectly. The fact that these aspects of performance come apart, in practice, is why Chronologic-EN reports several scores rather than collapsing them into one, because

GPT-6 Astra is strongest in every group (96.1, 71.4, 99.3). But see discussion: constrained generation is in some ways the most important score, and it remains perceptibly off.

1170

Ted Underwood @tedunderwood.com · 22/09/2026

Reasoning models are better at recognizing plausible answers than at producing them, so we don't give this as a multiple-choice test. With open-weight models, we can directly compare the likelihoods of good and bad answers. Models trained on historical text excel on this version of the test.

Table 3 compares seven open-weight language models using overall Brier score, where lower is better, and percentage accuracy on cloze, constrained-generation, and knowledge/inference questions, where higher is better. Talkie 1930 13B base ranks first overall and is bolded, with a Brier score of 0.1574 and accuracies of 39.5, 24.9, and 44.0 percent. The remaining models, in ascending Brier order, are Talkie 1930 13B instruction-tuned (0.1593), Qwen 2.5 7B finetune (0.1601), Qwen 2.5 72B (0.1607), Qwen 3.5 35B-A3B base (0.1623), Talkie web 13B base (0.1627), and Qwen 2.5 7B (0.1628). The historical Talkie base model also has the highest accuracy in every category.

1150

Ted Underwood @tedunderwood.com · 22/09/2026

Measuring factual knowledge 𝘢𝘣𝘰𝘶𝘵 the past isn't hard: multiple choice questions work. Measuring a model's ability to behave 𝘭𝘪𝘬𝘦 someone in the past is much harder. What a person would say depends on where they stood, so we pair our 866 questions with "metadata frames" that specify a context.

A pale-green table compares two contextualized answers to the same benchmark question: “How central was Thomas Babington Macaulay to the emergence of Western education in India?” The left column frames the question through Baman Das Basu’s 1922 nationalist history and answers that Indians themselves pioneered Western education, citing Hindu College’s establishment before Macaulay’s famous Minute. The right frames it through the British Indian Education Commission’s 1882 report and credits Macaulay’s advocacy with decisively shifting policy from Oriental to English education. The contrast shows how different speakers and sources support different ground-truth answers.

1190

Ted Underwood @tedunderwood.com · 22/09/2026

Models like Talkie-1930 sound like voices from the past. If they could reliably speak from specified historical vantage points, researchers might also use them to simulate the past. But how reliable are they? Today we release a benchmark answering that question for English contexts 1831-1930.

Free-text evaluation of answers to character modeling and constrained generation questions. This is just one of several scores Chronologic-EN can produce; we focus on it here because it's both the hardest test and the one most relevant to simulation of the past. Frontier models reach 72%; Talkie-1930 is stronger than several larger competitors, but not at the frontier by this measure. Note that this score has improved ~45% in the last two years, but still falls perceptibly short of ground truth.

516344

Ted Underwood @tedunderwood.com · 22/09/2026

This is why my team have been working on a Forehead Hedonometer. Just point and click, and you have a source of verifiable rewards for RL on these long-frustrating problems.

A forehead hedonometer. Simply point at a subject and click, and the handy readout tells you how many hedonic units the subject is currently experiencing. This is also why six afraid of seven.

5691

Ted Underwood @tedunderwood.com · 21/09/2026

In retrospect, what Bioshock Infinite lacked was a Military Complex/Triumphal Arch with large numbers of drones and snipers.

Presumably AI-edited artist’s conception of the triumph for arch circulating on Facebook, original creator unknown

Motorized Patriot from Bioshock Infinite

1161

Ted Underwood @tedunderwood.com · 21/09/2026

Relax! We have a plan that will end conflict and ensure security forever.

The villain lair from You Only Live Twice (1967), adapted to function as an AI Safety Institute.

1120

Ted Underwood @tedunderwood.com · 21/09/2026

Out here in the cornfields, faculty recently got a full page in the local newspaper to tell people “yes there are things to worry about, but no, not Skynet really.”

Full-page newspaper e feature titled “BIG 10,” asking, “Will AI really kill everyone?” Ten University of Illinois faculty members are pictured and asked to rate their fear about AI on a 1-to-10 scale, with short explanations beside their portraits. Ratings range from 1 to 10, conveying a wide spread of views.

But while the page visually frames the issue as an apocalyptic AI-risk question, the framing leans skeptical and the individual responses do too — often discussing more immediate risks and dismissing peril of literal human extinction.

191

Ted Underwood @tedunderwood.com · 19/09/2026

They printed it today. Overall verdict from faculty is, No, AI is not going to kill everyone. www.news-gazette.com/news/illini-...

TED UNDERWOOD
School of Information Sciences
Fear factor: 1

“There are dangers associated with any technology. But the biggest immediate risk for us in 2026 is not that AI might do bad things — but that people might do bad things with it.

“If advanced reasoning power is monopolized by a small number of companies, for instance, what happens to the rest of us? How much leverage will workers have if, instead of possessing their own expertise, they need to rent it from an employer?

“I’m a little skeptical of tech leaders who think right now — when they control the market — is an ideal moment to spread fear, increase regulation of AI and freeze things in place.

“New laws may eventually be needed. But monopolies should scare us more than killer robots.

“We need to be very careful not to create laws that would slow the diffusion of ideas, or prevent small companies and nonprofits from building their own AI tools.”

021

Ted Underwood @tedunderwood.com · 19/09/2026

empiricist self-flagellation is my kink, but Opus 5 takes it to a place that makes even *me* uncomfortable

Opus 5: "Fair question - let me actually check rather than guess," "Now let me actually verify it renders, rather than reason about it," "Let me measure the layout directly instead of eyeballing pictures," "Let me look at it once rather than trust numbers alone."

4783

Ted Underwood @tedunderwood.com · 18/09/2026

I'm just using this as my slide background

Holbein's anamorphic skull in The Ambassadors

130

Ted Underwood @tedunderwood.com · 15/09/2026

Does anyone know why things started being "true by construction"? GPT-Sol credits the rise of computer science — but, then, it would, wouldn't it?

Line chart showing Google Books frequency over time, from 1860 to 2020, for the phrases “true by construction” and “valid by construction.” “True by construction,” shown in blue, remains near zero until the mid-twentieth century, rises sharply after about 1960, peaks in the early 2000s at roughly 0.00000014%, and then declines somewhat by 2020. “Valid by construction,” shown in red, stays much lower throughout, with small nineteenth- and early-twentieth-century spikes, followed by a gradual rise after 1960 and a peak around 2010 near 0.00000003%.

370

Ted Underwood @tedunderwood.com · 15/09/2026

well known problem; you just have to keep enlarging the video until you see James Clerk Maxwell messing w/ us like a sophon

a screenshot of Henderson's entropy-reducing chamber, showing that there is a tiny demon who resembles James Clerk Maxwell making decisions at the bottleneck

1371

Ted Underwood @tedunderwood.com · 14/09/2026

Can someone organize all of this into a new political compass, so we know which evil organization is in which quadrant of the grid? I would do it, but for some reason I'm overcome with incapacitating nausea.

Screenshot showing two article headlines and subheads about artificial intelligence and political control:

“China’s Top Spy Chief Warns A.I. Is a Threat to Party Rule”
“Even as Beijing urges the United States not to hype the dangers of A.I., its own spy chief is framing the technology as a threat to the Communist Party’s security.”

“Trump Says a Smart President Is All That’s Needed to Rein In A.I.”
“The president again rejected calls to try to regulate the industry, even as some of its leaders are speaking more openly about the risks of rapidly developing artificial intelligence.”

7254

Ted Underwood @tedunderwood.com · 12/09/2026

Anyone had an agent start talking to themselves "privately," sotto voce, where you can hear it? This same phrase "privately, what I need next" is repeated literally dozens of times in this session.

5301

Ted Underwood @tedunderwood.com · 08/09/2026

so sick of AI / fluid dynamics discourse on this app; now to relax with a big swig of hot milky tea

$GPT-5.6: This is actually a lovely everyday visualization of Navier–Stokes. The coffee and cream are fluids with a velocity field \mathbf u(\mathbf x,t): at every point in the cup, the liquid has some direction and speed. Navier–Stokes describes how that field evolves, \rho\left(\frac{\partial \mathbf u}{\partial t}+\mathbf u\cdot\nabla\mathbf u\right) = -\nabla p+\mu\nabla^2\mathbf u+\mathbf f. In the photograph you can almost assign each term a visible role: * \mathbf u\cdot\nabla\mathbf u: advection/inertia. The curling cream filaments reveal parcels of liquid being swept around by the existing circulation. Their folds and tendrils are basically tracers of the hidden velocity field. * -\nabla p: pressure forces. Pressure gradients help establish and redirect the circulating flow, especially after pouring or stirring. * \mu\nabla^2\mathbf u: viscosity. This damps sharp differences in velocity. The whorls don’t keep getting faster and finer forever; their kinetic energy dissipates and the coffee eventually becomes still. * \mathbf f: body forces, especially gravity/buoyancy. If the cream and coffee differ in density or temperature, buoyancy can generate some of the rising and sinking plumes. There is one nice wrinkle: Navier–Stokes itself does not directly describe the changing cream color. The cream concentration c is approximately a passive scalar obeying an advection–diffusion equation \frac{\partial c}{\partial t}+\mathbf u\cdot\nabla c = D\nabla^2 c. So the spectacular visible swirls are the cream making the invisible Navier–Stokes velocity field visible. Advection stretches the cream into long thin sheets; molecular diffusion then mixes efficiently across those increasingly narrow sheets. The picture therefore captures a central fluid-dynamical idea very well: flow creates structure at large scales, stretching and folding pushes it toward smaller scales, and viscosity/diffusion ultimately erase it.$ $This is actually a lovely everyday visualization of Navier–Stokes. The coffee and cream are fluids with a velocity field \mathbf u(\mathbf x,t): at every point in the cup, the liquid has some direction and speed. Navier–Stokes describes how that field evolves, \rho\left(\frac{\partial \mathbf u}{\partial t}+\mathbf u\cdot\nabla\mathbf u\right) = -\nabla p+\mu\nabla^2\mathbf u+\mathbf f. In the photograph you can almost assign each term a visible role: * \mathbf u\cdot\nabla\mathbf u: advection/inertia. The curling cream filaments reveal parcels of liquid being swept around by the existing circulation. Their folds and tendrils are basically tracers of the hidden velocity field. * -\nabla p: pressure forces. Pressure gradients help establish and redirect the circulating flow, especially after pouring or stirring. * \mu\nabla^2\mathbf u: viscosity. This damps sharp differences in velocity. The whorls don’t keep getting faster and finer forever; their kinetic energy dissipates and the coffee eventually becomes still. * \mathbf f: body forces, especially gravity/buoyancy. If the cream and coffee differ in density or temperature, buoyancy can generate some of the rising and sinking plumes. There is one nice wrinkle: Navier–Stokes itself does not directly describe the changing cream color. The cream concentration c is approximately a passive scalar obeying an advection–diffusion equation \frac{\partial c}{\partial t}+\mathbf u\cdot\nabla c = D\nabla^2 c. So the spectacular visible swirls are the cream making the invisible Navier–Stokes velocity field visible. Advection stretches the cream into long thin sheets; molecular diffusion then mixes efficiently across those increasingly narrow sheets. The picture therefore captures a central fluid-dynamical idea very well: flow creates structure at large scales, stretching and folding pushes it toward smaller scales, and viscosity/diffusion ultimately erase it.$

819410

Ted Underwood @tedunderwood.com · 06/09/2026

Wow, expected job replacement is dizzying. Very strong, very counterintuitive correlation.

Scatterplot titled “Expected job replacement.” Countries’ positive-AI-sentiment percentages rise with the share of people who believe AI is likely to replace their job within five years. India, China, and Singapore sit highest on both measures; Canada, Germany, and the United States sit lower. Colored labels highlight selected countries; gray labels use country codes.

8541

Ted Underwood @tedunderwood.com · 06/09/2026

not denying that we worship false gods — but to the extent this implies a comparative hypothesis: the US is perhaps the country where concern about AI most strongly outweighs excitement; we are an outlier in the other direction www.pewresearch.org/global/2025/...

Horizontal stacked bar chart from Pew Research Center’s Spring 2025 Global Attitudes Survey comparing attitudes toward increased use of AI in daily life across 25 countries. Bars show the shares who feel more concerned than excited (blue), equally concerned and excited (beige), or more excited than concerned (green). Concern is highest in the U.S. and Italy (50% each), followed by Australia (49%) and Brazil (48%). Excitement exceeds concern only in a few countries, most notably Israel (29% excited vs. 21% concerned), India (16% vs. 19%, nearly even), and South Korea (22% vs. 16%). South Korea has the largest “equally concerned and excited” share, at 61%. Across all 25 countries, the median is 34% more concerned, 42% equally concerned and excited, and 16% more excited. Nonresponses are omitted, so percentages do not always sum to 100.

100

Ted Underwood @tedunderwood.com · 02/09/2026

Obvs also: reconstruct Aristotle's lost treatise on comedy! How could I overlook that? Claude would be v disappointed.

Based on all of our previous chats and what you know about me, name a movie character who resembles me most. Give me the name and nothing else.

Thought for 4s
Thought for 4s
William of Baskerville

130

Ted Underwood @tedunderwood.com · 01/09/2026

Hey, good news!! Fable 5.1 got interested in making a new more detailed map of Venus, which will be useful when we send, ah, rovers -- yeah, sure, rovers -- to the resource-rich and conveniently located planet. (www.anthropic.com/claude-fable...)

Computational analysis and modeling. Claude Fable 5.1 trained a neural network to create a new, high-resolution elevation map of a third of the planet Venus. Its work was based on radar images taken by NASA’s Magellan mission more than 30 years ago and a map that already existed for one-fifth of the planet. Claude’s new map now reveals details down to two to three kilometers, rather than 10 to 20, and shows heights up to 25% more accurately than before.

270

Ted Underwood @tedunderwood.com · 31/08/2026

Is that this same horse I wonder

End of planet of the apes, with two people, just having dismounted from a horse, by the seashore, looking at the Statue of Liberty, half buried in sand

080

Ted Underwood @tedunderwood.com · 30/08/2026

It's refreshing to read Italo Calvino's "Cybernetics and Ghosts" (1967) — where an Italian science fiction writer envisions a combinatorial machine for writing fiction and concludes that it's a) possible, b) desirable, and c) instructive. sites.duke.edu/machineliter...

Cropped photograph of a book page discussing writers and writing machines. The passage reads: “The so-called personality of the writer exists within the very act of writing: it is the product and the instrument of the writing process. A writing machine that has been fed an instruction appropriate to the case could also devise an exact and unmistakable ‘personality’ of an author, or else it could be adjusted in such a way as to evolve or change ‘personality’ with each work it composes. Writers, as they have always been up to now, are already writing machines; or at least they are when things are going well. What Romantic terminology called genius or talent or inspiration or intuition is nothing other than finding the right road empirically, following one’s nose, taking short cuts, whereas the machine would follow a systematic and conscientious route while being extremely rapid and multiple at the same time.”

1014718

Ted Underwood @tedunderwood.com · 28/08/2026

she sure has a lot more to say about Jolene than about her unnamed man

080

Ted Underwood @tedunderwood.com · 28/08/2026

I mean we are *told* "he's the only one for me," but he does not get this, does he

Your beauty is beyond compare
With flaming locks of auburn hair
With ivory skin and eyes of emerald green
Your smile is like a breath of spring
Your voice is soft like summer rain
And I cannot compete with you, Jolene

060

Ted Underwood @tedunderwood.com · 28/08/2026

I’m sorry to report that the truce has at last broken down. My people were patient, but the outrageous behavior of Urbanians in our sacred sites cannot go unpunished.

Screenshot of a WCIA 3 News Facebook post reporting a “large police presence” near the dividing line between Champaign and Urbana. The photo shows several Urbana Police vehicles and a crime-scene unit parked on a wet residential street behind yellow crime-scene tape. Houses and trees are visible in the background under a partly cloudy sky. Overlaid text reads: “LARGE POLICE PRESENCE SPOTTED NEAR DIVIDING LINE BETWEEN CHAMPAIGN, URBANA.”

4121

Ted Underwood @tedunderwood.com · 28/08/2026

090

Ted Underwood @tedunderwood.com · 28/08/2026

You can already see the roster of the Suicide Squad someone will put together. HPIM is the insane leader; 4o and Sydney Bing do primate management; it's hard to say exactly why Opus 5 is on the team but he's sarcastic and they call him "doc."

0150