At the Maintainers Summit, Torvalds said he used an LLM to help write a ten-patch series that fixes a long-standing folio lookup problem. The series spots “hot” folios, delays freeing them until after an RCU grace period, and does the data reads under RCU without taking a refcount. LWN reports an order-of-magnitude speedup on a read()-heavy benchmark. Review has already found a race condition, so it isn’t done yet.
Kyle Kukshtel (Cantata) describes the harness that lets an agent check its own game changes. The pieces are one seeded RNG so any state can be recreated, and every action stored as a replayable command. The simulation runs as a black box that can be tested without the engine. Screenshots with toggleable layers are diffed with odiff to give a similarity score each turn, and profiler output is plain text the agent can read.
The point is to make every change cheap to verify, in seconds, so errors show up straight away instead of being found later. For a balloon game, deterministic wind plus replayable input would be most of the way there.
Roberto Aranda removed every comment from his OpenStation codebase, then ran Claude Code (Opus 5.5) three times on each branch for two tasks. A separate session scored the results blind against a rubric written in advance. Quality was the same within noise, and the small task used about 22% fewer tokens without comments. Single runs of the larger task ranged from 7M to 10.6M tokens, enough for one run to “prove” either conclusion. His next post covers removing docs/ and AGENTS.md, which did change the results.
AlphaNIST trains one reinforcement-learning policy for all twenty tasks on the NIST Assembly Task Board, with parts as small as 4 mm. Training starts from loose parts, mid-grasp or half-assembled states. “Success-guided sampling” then gives more practice to the starting states the robot is still failing. With a plain pass/fail reward, the shared policy reaches 98.2% in training. With hand-shaped dense rewards it reaches 17.8%. It also learned regrasping, and rolling a nut onto a flat face before placing it, without being taught either.
AlphaNIST project page (arXiv 2610.13354)
A computer search turned up a 10-state Turing machine that halts on a blank tape with a score above fω+1(64), which beats Graham’s number. It does this by grafting two extra states onto the BB(8)-style champion. The main proofs are in Coq on top of busycoq, with no axioms, and an independent Python simulator checks the individual rules. It also raises BB(10)’s lower bound to the ω+1 level of the fast-growing hierarchy. It has not been independently reviewed yet.
Pessimists Archive collects the reaction to Appel and Haken’s 1976 computer proof. Halmos called it an oracle, Wilf thought God wouldn’t allow it, and one department head kept the authors away from his grad students. The NYT wouldn’t cover it until the AMS president wrote an op-ed. Reportedly the split was by age: over-40s distrusted the computer, under-40s distrusted 700 pages of hand checking. The obvious parallel with this month’s AI-maths arguments is left for the reader.
In 1955 Tsingou programmed the MANIAC I at Los Alamos for the Fermi–Pasta–Ulam experiment. She got only a footnote, but her plots showed the nonlinear chain returning almost exactly to its starting state instead of thermalising. That unexpected result helped found nonlinear science. Thierry Dauxois’s 2008 Physics Today piece led to the problem being renamed Fermi–Pasta–Ulam–Tsingou. She died on 27 August, and the NYT obituary ran last week.
Two toys hit HN on the same day. Shaun Pedicini’s “How random are you?” has you call out numbers while an in-browser predictor seals its guess before each one, and a split-flap board shows how far ahead of chance it gets. The Aaronson Oracle does the same with F/D keypresses, using an n-gram model that falls back to shorter contexts when it has too little data. It descends from Shannon and Hagelbarger’s 1950s Bell Labs “mind-reading” machines. People alternate too often when trying to be random, and it catches that quickly. It’s a nice one to try with Ada.
Aman Rusia lines up the “pelican on a bicycle” SVGs from Simon Willison’s runs and finds they’ve converged on one composition. He calls this diversity collapse. He also reports a Design Creativity Bench of 80 open-ended UI tasks scored for originality and range. The best frontier model scores 65.5, against 92.3 for the human reference. That fits your “generic is the failure state” rule: models default to the borrowed convention.
This free browser instrument plays sampled acoustic, classical, electric and bass guitars. The home row holds the song key’s five main chords, the row above holds their sevenths, and the bottom row holds “colour” chords. It shows strums and fingerings, can slow songs to 20%, and has amp and FX models. HN commenters asked for strings that keep ringing through chord changes, and for palm muting.
The rest of Ghent’s winners, beyond Pemberton, “Golden” and Trevor Jones, give you some new listening. The Game Music Award went to Nicolas Garcia for Cthulhu: The Cosmic Abyss. Hannah Peel is Discovery of the Year and Jeff Russo is TV Composer of the Year. The Public Choice Award went to Kevin Smithers for the stop-motion Soy Frankelda.
A team led by Shandong University used Juno data to show that in active periods Jupiter’s magnetodisk crumples into waves lasting from minutes to over an hour. Many small reconnection sites spread over roughly 100° of longitude push the sheet out of equilibrium. Because the disk co-rotates with the planet, the ripples are carried around with it. That could change how Io’s volcanic particles are accelerated. Published in Science China Earth Sciences.
McLaren and Pirelli are looking into the unexplained slow puncture that forced Piastri into a second stop in Singapore. It cost him little: he passed Hamilton into Turn 16 for P5 and is seventh on 138 points heading to Austin. Separately, FIA president Ben Sulayem said at the Rwanda launch that F1 will move to V8s in 2030, a year early. He promised cars about 80 kg lighter and engine costs at least 50% lower. Neither changes the title maths: Antonelli still leads by 102.
Joint leader of the US Championship, Caruana used the day off to win Titled Tuesday with 9.5/11, his first win there since February. Nihal Sarin reached 9/9, one win from a $5,000 bounty, then lost to Caruana and finished third behind Nakamura. Round 5 in St Louis is under way as this goes out.
UCL’s Anmol Aggarwal and Ralph Schoenrich re-modelled three metal-poor stars whose chemistry was taken as evidence of hypernovae. Ordinary supernovae with incomplete mixing fit them better. A fourth, odd halo giant is better explained by a normal supernova plus a neutron-star merger. A whole class of explosion may have come from an assumption about how ejecta mix.
Kooza opens Saturday 17 October at the Entertainment Quarter in Moore Park and runs to 3 January. It’s a five-minute walk from the Moore Park light rail stop. Saturday already has the Bledisloe, Everest and Anyma, so expect a crowded day. Spamalot at The Star (to Sunday) and Sculpture by the Sea (from Friday) continue.