iiinova — where your digital self keeps evolving

Fortunately, that bridge itself is easy as heck to copy and make free. At this point pretty much everything that was gained from the initial pre-training is in fact, universally shared. Differences between models that people pay for are now all post training.

1231

To avoid subtweeting; I'm responding to this theophite/scalzi convo and suggesting that -- while I agree with Mr Arson -- it might be strategic to take a even harder line by rejecting the tacit premise.

1482

It's okay to learn from everything, and train on everything. That's everyone's birthright, and we shouldn't apologize for it. Prohibitions on copying were a late invention, and a narrowly tailored one. They do not express a general ethical law that all information ought to be priced and paid for.

69618

Leonardo Flores @leoflores.bsky.social · 5h

I like compromises. But one place I don't believe in meeting people halfway, or 5% of the way, is the notion that AI would be better if not based on texts/images that governments have made illegal to copy. "OK, technically legal since you're not copying, but still, you should feel bad." Nope! +

29111

Reposted by Ted Underwood

My current favorite sentence about AI comes— after showing how the environmental impact of beef is so much worse and making a point about moral panics and selective scapegoating— is “Where’s the beef with beef?” You can see the keynote here: leonardoflores.net/posts/blog/p... And I cite Ted, too!

leonardoflores.net

Presentation and Keynote at the ELO 2026 Conference

Leader, Scholar, Educator, Creator.

061

Ted Underwood @tedunderwood.com · 5h

Needs more dancing Milla Jovovich

020

Reposted by Ted Underwood

Mark Riedl @markriedl.bsky.social · 7h

My lab will be busy at COLM 2026! 1. Main conference: - Capability Provenance in Language Models: A Case Study in Social Reasoning arxiv.org/abs/2606.19625 We present a method for answering causal hypotheses about where in a dataset behaviors emerge. We apply our method to social reasoning.

1112

Ted Underwood @tedunderwood.com · 5h

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.

614519

Ted Underwood @tedunderwood.com · 10h

I do have a friend who works in a bank where PowerPoints rather than text are the medium of communication; it seems a rather different skill.

110

A successful environmental politics works hard to decarbonize electricity generation—and spends absolutely no political capital telling voters whether they should use that electricity to watch a season of television, drive to the store, or answer a few thousand LLM queries.

060

That’s because a lot of us have examined the evidence, and don’t believe this technology is more environmentally harmful than other uses of electricity. Nor do we believe “reduce overall electric use” is good environmental politics.

130

I know a human being had that thought.

210

I wonder whether it would be to their advantage to prompt a model to produce a “literal translation.” Because a paper with clear signs of Chinese syntax is 1000% more prestigious right now than one with chatbot voice. I feel real gratitude when “Besides” is used to start a sentence.

1101

I’m still thinking through the ESL/Science thing, but I think that’s a bit different from what I mean in OP. If you write a paper in Finnish, hit “translate,” and then edit a bit, you are definitely the writer of that document. Otoh bullet points -> prose still seems wrong to me for science.

240

Or to to be more fair: they’re doing a kind of writing where it matters that everyone remembers to bring sunscreen to the event, and does not really matter who reminded them.

2130

They really feel that it’s labor-saving to have a model turn a scattered list of bullet points into an email of about the same length that reads more clearly. For me that would not save labor, and the loss of personal voice is so painful that it’s a huge net loss. But they dgaf about voice.

280

norvid_studies @norvid-studies.bsky.social · 19h

If the writing is important, I still think a language model’s role should be advisor / editor / translator — not writer. But what I’m starting to grasp is that there are lots of people for whom writing is not very important, and who find it real work to organize an email. +

160

Reposted by Ted Underwood

generational madeleines

042

Ted Underwood @tedunderwood.com · 18h

Oh I’m 100% okay with AI for search, though the Google AI overviews are admittedly to that market what spaghetti-Os are to pasta

130

Ted Underwood @tedunderwood.com · 20h

I believe strongly in 100% human-written blogs and papers, and bridle even at emails with a canned AI voice. But talking to family friends who don’t write for a living and find writing a chore … I’ve come to accept that ppl in other walks of life will use these tools for writing more than I do.

121384

Ted Underwood @tedunderwood.com · 20h

I have a song lyric dataset that I’ve used for that; it’s not up on the web, but I could email. Invites speculation about music genres they know, though you have to be ready for the rap topics.

120

Ted Underwood @tedunderwood.com · 22h

I’m not giving anything benefit of the doubt, because I’m not doing moral discourse at all. I’m differentiating distinct causal processes, all of which involve AI somewhere. You can judge them all to be pernicious if you like.

160

Ted Underwood @tedunderwood.com · 22h

My guess in this case is that the influence is mediated by the human author’s knowledge — or by the AI doing web search, and discovering discussions of the Magdalena Bay video — more than by direct training.

180

Reposted by Ted Underwood

Samuel @samuel.fm · 04/10/2026

Bluesky, the app for .json files

71689

Reposted by Ted Underwood

Henry Farrell @himself.bsky.social · 04/10/2026

Creating art. If Andy Warhol were still alive, he'd be making silkscreens of this interaction that people could hang on their walls.

2242

True of so many technologies: steam, nuclear, the internet. All can be misused. I think the calculus with AI is actually a shade less painful, because there’s a chance that it makes people more perceptive or better balanced. In which case, use would have some inherent edge on misuse.

140

Never a bad moment for that passage

010

Never a bad moment for that quote

010

Laura K. Nelson @lauraknelson.bsky.social · 04/10/2026

Frontier labs are essentially running A/B tests on human intellectual preferences right now: agreement/pushback ratios, explore/exploit ratios, how much background to provide or presume. Would be insanely cool to write up.

3514

Reposted by Ted Underwood

Re-reading Sam Lucas who has for his entire career insisted that non-probability samples are useless at best for understanding the social world. The vast majority of studies I read do not use probability samples. With big data etc., where are we with sampling theory? Do folks still teach it or nah?

7123

Reposted by Ted Underwood

Shahab Bakhtiari @shahabbakht.bsky.social · 03/10/2026

I had a feeling recently that seems to be corroborated by chats with colleagues, so I decided to share it here. A lot of systems neuroscientists driven by algorithmic-level questions are going through what looks and feels like a research depression. 🧵 (1/n) 🧠📈 #NeuroAI

138923

Reposted by Ted Underwood

Benedek Kurdi @benedek.bsky.social · 24/09/2026

📊 New resource: ISAAC, an open corpus for studying social group attitudes 💬 527M+ Reddit posts (2007–2023) 🧠 Covers race, sexuality, age, ability, and skin tone 🗺️ Validated against Gallup and Google Trends 🔍 No-code or Python/SQL: isaac.psychology.illinois.edu 📖 Preprint: arxiv.org/abs/2609.27059

We introduce the Illinois Social Attitudes Aggregate Corpus (ISAAC), an open, modular, and accessible corpus of 527 million+ English-language Reddit posts selected for relevance to six key social group distinctions based on race, sexuality, age, ability, body weight, and skin tone, covering the 17-year period from 2007 to 2023. A multi-step, human-audited filtering pipeline was used to keep irrelevant content in the curated dataset below 10%, both overall and for each social group distinction. Each post was then algorithmically annotated with the user's estimated home region, along with a suite of validated off-the-shelf and custom semantic labels including moralization, sentiment, emotion, and linguistic generalization. We confirm the validity of the resulting corpus through convergent evidence linking ISAAC to macro-level societal trends, such as online search behavior, temporal spikes during major societal events (both nationally and regionally), and long-term shifts in public attitudes. By offering a unified, public infrastructure, ISAAC eliminates research fragmentation and enables seamless replication while supporting diverse empirical workflows at scale. Specifically, ISAAC allows investigators to perform cross-category comparisons, conduct high-precision tracking of long-term temporal shifts in social group discourse, and map spatial variation onto localized public opinion and policy outcomes. ISAAC's fully public, modular pipeline facilitates easy extension of the corpus to new platforms, languages, and social categories. To accommodate various research needs, ISAAC is accessible both without coding through a point-and-click website and labeler web-apps, and programmatically via an SQL playground, a Python package, and HuggingFace.

31911

Reposted by Ted Underwood

hardmaru @hardmaru.bsky.social · 04/10/2026

Recently learned that Schmidhuber’s Formal Theory of Fun and Creativity, covering compression progress and intrinsic motivation and their relationship to beauty, curiosity, art, science, music, and jokes, was published in a Japanese scientific journal in 2009. www.jstage.jst.go.jp/article/sice...

Simple Algorithmic Theory of Subjective Beauty, Novelty, Surprise, Interestingness, Attention, Curiosity, Creativity, Art, Science, Music, Jokes

Jürgen SCHMIDHUBER

Abstract

In this summary of previous work, I argue that data becomes temporarily interesting by itself to some self-improving, but computationally limited, subjective observer once he learns to predict or compress the data in a better way, thus making it subjectively more “beautiful.” Curiosity is the desire to create or discover more non-random, non-arbitrary, “truly novel,” regular data that allows for compression progress because its regularity was not yet known. This drive maximizes “interestingness,” the first derivative of subjective beauty or compressibility, that is, the steepness of the learning curve. It motivates exploring infants, pure mathematicians, composers, artists, dancers, comedians, yourself, and recent artificial systems.

6565

no, they just look similar :)

270

brennan @brennan.computer · 04/10/2026

also had no clue tbh that @cameron.stream played guitar for them

190

Reposted by Ted Underwood

absolutely gooooorgeous out tonight

41141

I found the Tao et al. response compelling in the aftermath of controversy around Navier-Stokes, and still think it makes sense to hold publication to a high standard. But Buzzard’s emphasis on the infinite scope of our remaining ignorance is where my heart lives most days.

0150

Not in math, but I found this very worth reading. I suspect most academics are going to confront similar questions before the 20s are out.

1255

I mean, maybe I’m projecting … but I took a lot of the early “uhhh …” and “waow” reactions to mean both a) this is very accomplished and b) I didn’t know the “manic pixie who saves you” dial *could* go up to eleven.

040

the little pop up "new hardware" notification when grass is touched, and the way the original title of the Windows XP photo ("Bliss") gets used lyrically ... it's clever

070

About this. Don't get me wrong, it's quite good. Brilliant at moments. I would just like to see a gender-flipped version, or a slightly less elfin version.

3120

I'm sure you're right. I know Opus, and Opus is much more interested in being genuinely uncertain than in performing gender.

050