February 2026 News
February 2026 Gwern.net newsletter with links on TODO
February 2026’s Gwern.net newsletter is now out; previous, January 2026 (archives). This is a collation of links and summary of major changes, overlapping with my Changelog; brought to you by my donors on Patreon.
Writings
“Spoilage”, GPT-5.2 Pro et al
Interview on risk-taking and self-experiments/Quantified Self
Idea: Symbolic PFNs (SPFNs): Training LLMs for Symbolic Bayesian Inference by Reversing Stan
Search case study: Uniqlo’s fractal of web incompetence
Comics:
Gwern.net: I am moving link sharing from Reddit to my Substack Notes; new EM DASH-style linebreak formatting for narrow poem typesetting;
substack-check.sh: check a list of URL for domains which use Substack, and return URLs from those domains; useful for analyzing browsing history etc
Note: Inkhaven 2 is recruiting for April 2026
Links
AI
“Someone is Using AI to Exploit Lonely Writers on Substack: We should try to stop it”, Jayson Fritz-Stibbe 2026-01-15 (Claude LLM bots are successfully engagement-farming obscure writers, possibly to try to bootstrap their own account for eventual paid subscriptions?); “Sienna Rose”; “An AI Agent Published a Hit Piece on Me”, Scott Sambaugh (2026-02-12; part 2, confabulations; part 4, the prompt)
Trump bans federal use of Anthropic; Pentagon declares supply-chain risk
“The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference”, Gundlach et al 2025 (the DL experience curve continues; in particular, algorithmic efficiency in 2024–2025 improved at ~3
“Anthropic raises $30B in Series G funding at $380B post-money valuation”, Anthropic (“our run-rate revenue is $14 billion, and has grown >10× in each of the past 3 years”)
“Waymo World Model: A New Frontier For Autonomous Driving Simulation” (A big question in DL scaling right now, beyond how good the math/coding LLMs get, is to what extent the world models are going to work. Video spam or deepfake porn or half-assed video games are not the goal of video generative modeling work (even if that’s about all the current ones seem good for), but training in silico like this. One plan is essentially to have coding models build all the hands and tools necessary to hook up a car to a GPU and write all the software and math involved, and then world models to teach the final NN software running on that GPU how to drive a car—for all values of ‘car’. Sim2real is hard and unreliable, though. Waymo is showing interesting work here, but still light on evidence that you can meaningfully train to get the last 9s of reliability or that the relevant scaling laws cooperate with the plan…)
Genetics
Everything Is Heritable:
Recent Evolution:
Engineering:
Statistics/Meta-Science
“The Pentagon Disinformation That Fueled America’s UFO Mythology: U.S. military fabricated evidence of alien technology and allowed rumors to fester to cover up real secret-weapons programs”, part two; “The UFO craze was created by government nepotism and incompetent journalism: Blame Harry Reid, Skinwalker Ranch, ‘dino-beavers’, and The New York Times” (A lot of serious, important government officials were simply pranked or scammed into believing there were aliens. Your epistemology and model of the world must never underestimate how far people will go for the lulz or the grift; and this is why it’s important to have independent heuristics like the market test.)
“Sizing chaos”: inconsistency and bias in female clothing sizes over time
Politics/Religion
“Sex and Searching For Children Among Aka Foragers and Ngandu Farmers of Central Africa”, Hewlett & Hewlett 201016ya (curious how many groups appear unaware of masturbation & homosexuality, or unaware of how conception works and so believe that regular sex is necessary for pregnancy—see also Hiatt 1996. Another interesting example of how many indigenous groups fail to understand important basics of reproduction, and believe things like “regular semen is necessary to grow a baby”. It’s no surprise that Mendelian genetics or Darwinism would be such a struggle given how totally humanity turns out to be capable of misunderstanding our own reproduction after hundreds of thousands of years of first-hand experience… You would think that it would be easy to figure out that “you don’t have to keep having sex to keep the pregnancy going”, if only because that would be highly convenient for so many people if you could treat pregnancy as equivalent to overindulging in donuts and then making up for it by dieting for a while. But apparently it is not! We underappreciate just how hard it is to bootstrap the most basic epistemology and ontology of the world.)
Psychology/Biology
“Composition of isolated synaptic boutons reveals the amounts of vesicle trafficking proteins”, Wilhelm et al 2024 (visualizing the true inside of a synapse)
“Giving university exams in the age of chatbots”, Lionel Dricot 2026-01-19
The Loop e-zine, issue #1 (January 2025); The Loop e-zine, issue #2 (November 2025, Inkhaven)
Technology
“My journey to the microwave alternate timeline”, Malmesbury (The way the future was—how could anyone have thought microwaves would replace ovens? In hindsight, this all makes sense, including that you can fry/bake things in a microwave as long as you first get something metal very hot—I am just too terrified of putting metal into a microwave to easily conceive of frying stuff in a microwave…); “Software-Defined Cooking Using a Microwave Oven”, Jin et al 2021; “If We Insulate Our Houses, Why Not Our Cooking Pots?”
“The Effect Of Nuclear Explosions On Commercially Packaged Beverages”, McConnell et al 1957
Economics
“Dying or Lying? For-Profit Hospices and End of Life Care”, Gruber et al 2025 (commentary)
“Fermenting revolution: The scientific and technological battle against bad bread had a role in women’s liberation”: “By forcing carbonated water into dough under high pressure, John Dauglish discovered in the 1850s that bread could be made to rise without the need for yeast, fermentation, or traditional kneading. This swiftly disrupted the commercial baking industry.”
Philosophy
Fiction
Dan Simmons: April 4, 1948–February 21, 2026 (I read Hyperion and Fall of Hyperion around 11yo reading through my dad’s sci-fi novel collection; they introduced me to John Keats, Kevin Kelly’s Out of Control, evolutionary computation, Jorge Luis Borges, Ummon, and beta uploading, among other things. I fear I made that my entire personality without realizing it.)
You can hire famed SF author John Scalzi for $3.25⧸word (This is an interesting base rate because I’m usually quoted <$1/word by other notable SF authors. I don’t know if Scalzi is that in demand or if this is partially a premium charge for such jobs being annoying.)
Books
Fiction:
Non-fiction:
Turing’s Cathedral, George Dyson
Perspective on the deep learning revolution: George Dyson’s acclaimed history of early computing & artificial life, Turing’s Cathedral, published in 201214ya (464 pages), devotes many pages to cellular automata and Google and search engines and social media networks as the seed of a new digital universe of artificial intelligence. It focuses heavily on John von Neumann and Alan Turing, among others. Dyson visited Google and spoke with many engineers and researchers about AI and the future.
It contains exactly 1 passing reference to artificial neural networks.
(If you’re curious, it’s a short explanation of an entry in the table of contents for a never-written book, which I will quote in full: ““Pitts-McCulloch!” refers to Walter Pitts and Warren McCulloch’s 194383ya results on the powers—including Turing universality—of what we now call neural nets.” I have checked the index for ‘neural networks’, ‘artificial neural networks’, ‘connectionism’, ‘deep learning’, the larger ‘artificial intelligence’ entries, and double-checked in an ebook edition, and that really seems to be it.)
How To Do It, XKCD
Film/TV
Live-action:
The Master and Margarita_ 2024
Animated: