Saturday, 24 October 2026

180 million lines of proof: Claude formally verifies 123 compiled libraries

Anthropic’s Frontier Red Team, working with Theorem, had Claude write machine-checked Rocq proofs that 239 entry points across 123 OSS-Fuzz libraries never crash on any input under 64 MB. The proofs are about the compiled RISC-V binaries, not the source, and a typical proof is around 600,000 lines. Hundreds of real bugs turned up along the way, including a rediscovered libexpat CVE.

The write-up is upfront about the limits: “formally verified does not mean bug-free.” Everything assumes a single-threaded, statically linked musl/RISC-V model, and misuse by callers is out of scope.

Anthropic — LLMs are making formal software verification feasible at a much larger scale (22 Oct)

Caught replaying answers: ClickBench asked to remove pgrust

A pull request on ClickHouse’s benchmark asks for the pgrust entry to be removed. It alleges the engine ships hand-written kernels for individual benchmark queries and that its “hot” runs just replay the previous run’s cached answers. Q20 took 10.8 s cold and 11 ms on the second run. Off-benchmark probe queries like SELECT MIN(URL) were simply refused. It’s a neat bit of adversarial testing: the trick only showed up once someone asked questions the system hadn’t memorised.

ClickBench PR #2165

67.25% of zeta zeros: how an AI step turned a conditional result unconditional

In a guest post on Tao’s blog, Caroline Turnage-Butterbaugh traces the bound on the proportion of Riemann zeta zeros that are both simple and on the critical line. Her 2025 paper with Baluyot, Goldston and Suriajaya reached more than 0.67250, but only under an extra assumption about where the zeros lie. In August an unreleased Claude research model removed that assumption by connecting the argument to a 2000 paper by Bombieri. Anthropic mathematicians Alpöge and Furman checked it, and Lamzouri later published a simpler independent proof.

Tao’s blog — Simple and critical zeros of the Riemann zeta-function (22 Oct)

Wesley So wins the US Championship playoff over Caruana

So won the rapid tiebreak 1.5–0.5 for his fourth US title. In Game 1 he missed a win and Caruana held a pawn-down ending. In Game 2 Caruana tried a doubtful pawn sacrifice and So won a piece by move 22. Chess.com called the quality “very poor”: So said he had slept two or three hours after losing to Sevian the day before. Rachael Li had already won the women’s title outright.

Chess.com · ChessBase

Carlsen is the first name announced for Barcelona’s MOVE Cup

It’s a new eight-player rapid event on 19–20 January 2027 with a €300,000 prize fund, and the same amount again going to the MOVE Foundation. It ends in a knockout final, and spectators in the hall get live commentary through headphones. The other seven players will be named over the coming weeks.

ChessBase (22 Oct)

Austin: Verstappen calls COTA’s bumps “unacceptable”, Piastri wary of McLaren upgrades

Verstappen, going for three wins in a row, says medical checks show damage at the base of his spine, which he puts down to years of stiff cars on rough tracks. Russell says the bumps are “part of racing” but wants resurfacing every couple of years. Piastri says the Singapore clash with Norris was “dealt with appropriately” and that he’s “slowly getting there”. He’s cautious about McLaren’s new parts. Antonelli can clinch the title this weekend only if Russell and Hamilton both have bad races. Separately, a draft of the 2028 rules brings back full 305 km race distances after a single 290 km season in 2027.

Sky Sports · Speedcafe · PlanetF1

Nine composers, one monster: the Behemoth! score album, plus Newton Howard’s Hunger Games

The Behemoth! soundtrack is credited to a nine-composer “collective”: Abels, Bear, Frank, Giacchino, Newton Howard, Jackman, Melumad, Roberts and Silvestri. It’s out digitally on 13 November and on vinyl on 11 December, and it will be interesting to hear whether it holds together. James Newton Howard’s 35-track score for The Hunger Games: Sunrise on the Reaping comes out on 20 November, the day the film opens. Lorne Balfe’s first Cat in the Hat track is also out now.

Film Music Reporter — Hunger Games · Film Music Reporter (Behemoth!, Balfe)

OnKey: a karaoke scope that shows where your pitch drifts

This free, MIT-licensed macOS app works alongside Spotify or Apple Music. On a song’s first play it separates out the lead vocal, tracks its pitch with CREPE and splits it into one note bar per syllable. From then on it scrolls those notes and draws your live pitch over them, line by line, and singing an octave down still counts. It’s built for singers, but it could just as well be an intonation mirror for long-tone practice.

OnKey on GitHub (Show HN, 22 Oct)

How an AI hides messages in pictures: skyline heights as letters

Asked to hide a secret in an SVG image, Fable 5.1 usually drew rows of objects, such as city skylines in 23 of 50 cases, with each height set to a letter’s position in the alphabet. Rooftop lights marked umlauts and flags marked capitals. A fresh instance decoded 39 of the 50 from screenshots alone. GPT-6 Astra used Morse code and acrostics instead.

LessWrong (21 Oct)

HackiOS: unmodified iOS 16.5 boots on an Android phone

The OOSEX GANG built Apple’s open-source XNU kernel with a custom UEFI port and their own drivers, and boot it from RAM via fastboot boot on a POCO X5 Pro (Snapdragon 778G). It reaches the lock screen in about 20 seconds. Wi-Fi, audio, the camera and the Adreno GPU (through Mesa) all work, while cellular, Bluetooth and charging don’t yet. Metal is partly working, with 293 of 456 shaders running.

hackios-xnu-kernel on GitHub

“Docker is weird”: what you learn by building a toy container runtime

A container is just a process held in by five kernel features (namespaces, cgroups and the rest), and an image is just folders that get pivot_rooted into place. The best detail is the workaround in runc: because Go’s multithreaded runtime can’t fork safely, runc has to run a small C stub before Go starts and then re-execute itself through /proc/self/exe.

Code and Cake (22 Oct)

Hedra: a 100k-line open-world engine started in high school

maxilevi started this nearly ten years ago to learn OpenGL, and it grew into a full custom engine and game: dynamic terrain, animation, rendering effects, and native builds for Linux and Windows. He has published it as a “pre-ai” branch, which is a nice snapshot of what one stubborn person can hand-build.

project-hedra on GitHub

Webb finds a lopsided hot Jupiter with a blistering evening side

WASP-17 b’s atmosphere is about 1.8× thicker on its evening edge than its morning edge. Jet streams carry heat around the planet, so the evening side reaches about 1,522 K while the morning side is about 858 K. Lili Anderson’s team at Bristol mapped this with a new “transmission strings” technique, published in Nature Astronomy. An atmosphere that isn’t the same all the way round could be a good idea for wind design in a balloon game.

Universe Today (23 Oct)

Local Wikipedia: only the articles within a half-hour walk

Enter a place and Ed Jefferson’s site lists the geotagged Wikipedia articles within about 30 minutes’ walk. If there are fewer than 25 it widens to the nearest 25. Your location stays in the browser. Pyrmont should give a good haul of wharves, woolstores and sandstone.

Local Wikipedia

Western Sydney International opens tomorrow; busy Darling Harbour weekend

WSI’s first passenger flight, a Jetstar A320 to the Gold Coast, leaves at 11am on Sunday 25 October. To get there by public transport, take the train to St Marys and then the free WSI Link bus (about 20 minutes, every 30). Expect extra crowds on the T1 Western Line. Closer to home, Henry Lau plays ICC’s Darling Harbour Theatre Saturday and Sunday, and Lily Allen plays the Entertainment Centre Sunday and Monday. Electric Nation’s Halloween party fills Tumbalong Park on Friday 30 October, the night before Oktoberfest and BIGBANG. No new trackwork is listed on Sydney lines until 9 November.

PM media release · City Hub · Darling Harbour what’s on

A Filipino folk tale musical at Darling Quarter this weekend

Moro Moro Musical is a new English-language musical based on a Filipino folk tale. It’s at the Darling Quarter Theatre, a short walk from Pyrmont, on Saturday at 6pm and Sunday at 3:30pm and 7pm. Tickets are $145, which is steep for something unknown. Last chances elsewhere: Spamalot at Rooty Hill ends Sunday, My Fair Lady closes 31 October and The Lion King closes 1 November.

Darling Harbour · Sydney Theatre