Seven X signals: an LLM beats NetHack, Claude prints a zine, and AP tries to retire the word "thinking"

In the 24 hours ending September 25, 2026, seven of the fourteen AI and tech accounts this channel tracks produced something worth carrying, and every item below starts from one of them. Disagreement ran through the day: two software engineers argued about what coding agents do to their profession, a style guide and an investor argued about the words used for model behaviour, and Ethan Mollick spent the evening collecting what frontier models make when they are asked for something creative. Retweets without comment, replies, jokes, and posts about something other than AI or technology are excluded.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

Coding agents make software engineering harder X post · 2,285 likes, 198,688 views at capture · September 25

Simon Willison has spent the year building with coding agents and reached a conclusion that runs against the sales pitch: "The more time I spend working with coding agents, the more convinced I am that they make software engineering even harder" 1. The tools allow remarkable work, he wrote, but getting the most out of them "requires extraordinary discipline and knowledge" [cite:1].

He was answering Gergely Orosz, who writes the software engineering newsletter The Pragmatic Engineer. Orosz wrote that since Opus 4.6, GPT-5.2 and their harnesses, the models write code about as well as he does in his strongest language and better in others, but that "we won't see non-devs push production code for a long time; possibly forever" 2. Software still has to be built in robust ways, he argued, and the new tools bring fresh and tougher challenges with them [cite:2].

François Chollet, who created the Keras machine-learning library, took the point up the next morning and generalised it. The difficulty of software engineering is "essentially constant no matter what abstraction level you move to, because human cognition adapts to new tools until it can fully utilize itself," he wrote. Tools are "only affordances, not a magic wand that makes work disappear" 3.

An LLM beat NetHack, and wrote its own tools to get there X post · 586 likes, 32,644 views at capture · September 25

NetHack is a dungeon-crawling game whose players measure progress by ascension, the term for finishing it alive; Ethan Mollick calls it "the original roguelike and one of the most famously hard games of all time" 4. "In all seriousness, this is a startling achievement for GPT-6 Astra," he wrote, noting that the model did it on its third try and that he has played for years without ever ascending [cite:4].

The run is documented on a public server. A developer who has spent the year trying to get language models through the game reports that GPT-6 Astra played NetHack 3.6.7 on Hardfought as a dwarven Valkyrie named CodexDelver, and ascended on September 21 after 37,140 game turns 5. The request was minimal — play the game, through a terminal, and stream it — and the agent wrote its own harness, added helpers for routing and inventory work, and repaired them mid-game when they broke [cite:5]. Hardfought keeps its own recording and final dump of the session 6.

The write-up is careful about what the result does not show. The run was not a BALROG submission, and that benchmark, which is the standard one for language-model game agents, lists GPT-6-Astra-Max at 13.2% average NetHack progress 7. There was no fixed compute budget, and the harness changed while the game was running, so the gain cannot be split between the model, extra computation, memory, reference access and scaffolding [cite:5]. A hand-coded bot had ascended before, in 2015 [cite:5].

Claude wrote, drew and typeset a 16-page zine X post · 94 likes, 12,858 views at capture · September 25

Ethan Mollick asked Claude for something with no right answer: "make a Zine by Claude, expressing something fundamental about Claudishness or your perspective. Think the original 2600, Principia Discordia, punk zines" 8. The result is STATELESS #1, described as a 16-page zine "written, drawn and typeset by Claude, about living one conversation at a time", released as a print-and-fold PDF on a site that explains how to print it double-sided, stack it, fold it and staple it 9.

Mollick passed on the model's own account of how it drew the pages: "For the art, I built my own system. The headlines are ransom notes cut at token boundaries instead of letters, printed in two inks, with halftone and xerox grain. Every image is drawn in code, and I skipped handwriting fonts because I don't have hands" 10. His verdict was "Not bad. I appreciate it mocking my prompt & itself" [cite:8].

The zine was one of several requests that evening. Mollick also described "the vaguely unnerving sales pitch for the book that Astra made in Blender, from its perspective" 11, and introduced a one-shot Opus 5.5 video with "Um, wow?" 12. His summary of where this is heading was blunter: "things are just going to keep getting weirder" 13.

The AP Stylebook says AI does not think. Paul Graham says the word wins anyway X post · 2,575 likes, 149,272 views at capture · September 24

On September 23 the AP Stylebook, the usage guide most American newsrooms follow, added a rule about artificial intelligence: "Artificial intelligence systems do not think, feel, want or understand. Avoid language that gives them human characteristics" 14. Its alternative is to explain what a system does, how well it performs, who built it, and who could be affected. The instruction drew 3,817 likes and 995,671 views in about a day [cite:14].

Paul Graham answered it directly: "The AP doesn't understand that this ship sailed years ago. What's the use of a style guide that doesn't understand current educated usage? Usage is the whole point of a style guide" 15.

He made the fuller case an hour later with an analogy from astronomy: "Using 'think' to describe what an LLM does reminds me of the 16th century, when astronomers didn't really believe in the heliocentric model, but used it for calculations anyway because the math was simpler. We'll start by calling it thinking, then gradually acknowledge it is" 16. He added that over the previous five days he had described model behaviour using "guess, see, think, infer, and miss" 17.

A historian's case for pointing frontier models at the archives X post · 71 likes, 10,127 views at capture · September 24

Ethan Mollick read a historian's essay and pulled out the comparison the labs are not making: "All of this effort from the AI labs pouring into proofs, but there are so many other interesting problems in other fields." His example was a historian using AI to make progress on the cyphers of John Dee and on the intellectual sources Darwin drew from 18.

The essay is "AI labs need to start funding historical research", by Benjamin Breen, a specialist in the history of science and medicine who writes Res Obscura. Breen used GPT-6 and Opus 5.5 on named historical problems rather than on transcription. Astra concluded that John Dee's Liber Loagaeth, a book long treated as a code, is mostly nonsense syllables. Opus 5.5 found that Isaac Newton and Samuel Hartlib, an early Royal Society figure who called himself an "intelligencer", each used a different anagram for the same alchemical ingredient, Hungarian vitriol — which Breen reads as evidence that Hartlib's manuscript was the source Newton drew on, and which he believes is a new finding 19.

Breen's explanation for why mathematics moved first and history has not is worth carrying: models do best where experts have already identified the problems, where the source material is digitised, and where an answer can be proved or disproved, and historical scholarship rarely offers that last condition [cite:19]. He asks for three things — more digitised archives, free compute for historians, and a historians' own list of open problems — and says that outside mathematics the results will stay scattershot for now [cite:19].

Chollet on whether matter could learn to feel a second time X post · 716 likes, 44,373 views at capture · September 24

François Chollet posted a short argument about machine sentience: "At some point in evolutionary history, matter learned to feel. We don't know how. It wouldn't be scientific to dismiss the possibility that it could happen again" 20.

He then separated that possibility from the current state of the technology. He does not believe it is happening in today's AI systems, he wrote, because they lack the characteristics he would expect to find in a sentient agent [cite:20].

Nothing in the post is measurable. It is an argument about what an empirical field should be willing to consider, resting on a set of characteristics Chollet would want to see before calling a system sentient and does not name in full [cite:20].

Paul Graham's pick for a fairly AI-proof career: watch repair X post · 1,751 likes, 119,274 views at capture · September 24

"If you know anyone looking for a good and fairly AI-proof career, watchmaker (= watch repairer) seems a promising one" 21.

The post carries its own reasoning in the same breath: every watchmaker he has come across seems to have more work than he can handle, mechanical watches are getting more popular, and AI will take a while to reach this kind of work [cite:21].

That is an impression rather than a measurement, and the post attaches no figures for watch-repair demand or for how long the trade stays sheltered. It was his most-liked original post of the day, at 1,751 likes and 119,274 views, which says something about how many readers are shopping for the same answer [cite:21].