Maths & AI
The AI “quasi-Riemann” result, two weeks on
Quanta looks at the strongest item in OpenAI’s 6 October batch. It proves that every nontrivial zero of the zeta function has real part between 1/12 and 11/12, a fixed strip narrower than any known before. An earlier, bolder 1/8–7/8 version has been set aside. The 11/12 paper is the only one in the batch a human edited, and it isn’t formally verified yet. Jared Duker Lichtman calls it “the most significant AI-for-math achievement so far, by quite a margin.” Alexander Dunn calls it “a surprising synthesis of well-known ideas.” Peter Sarnak wants humans to check the details first.
Quanta Magazine (Konstantin Kakaes)
Maths
The Bourgain Project: a community effort to digest 500 papers
Lukas Liehr, João Pedro Ramos and Mitchell Taylor are inviting anyone to write up “digestions” of Jean Bourgain’s papers, which are famously dense. The goal is to map the themes that link the papers and to try formats other than the journal article. A Zulip channel and bourgain-pedia.com host the work. On the same day, Tao posted a sketch of Ben Green’s streamlined argument against Siegel zeroes.
Tao blog: The Bourgain Project · Siegel zeroes sketch
AI & Software
tsrs: TypeScript’s compiler, ported to Rust by agents
Cloudflare, VoidZero and Owner.com’s Max Schwenk are building tsrs, an experimental Rust port of the TypeScript compiler, inside the Oxc toolchain. It isn’t a fork, and the TypeScript team still sets the language’s direction. The authors argue that frontier models can now do a port this size in days. Agents will keep it in step with upstream through nightly compatibility runs. Early numbers: up to 8× faster than TypeScript 7.0 with about 70% less memory. Rust also avoids Go’s garbage collector, which matters for WebAssembly.
Cloudflare blog
AI & Software
Agentic monitoring: specs written by agents, run by a dumb scaffold
Ben Congdon argues that systems are now “grown more than designed”, so their operators understand less of the code. His alternative to a free-running agent loop: an agent reads the service code and writes a monitoring spec. A deterministic harness executes the spec, and an investigator agent decides what to escalate to a human. Most of the value comes from noise-reduction heuristics and a labelled eval set. He estimates a big fleet would cost about $100 a day.
benjamincongdon.me
Practice
The 1991 outage that ran silently on batteries for six hours
On 17 September 1991, a rectifier tripped at AT&T’s 33 Thomas Street hub in Manhattan. Batteries quietly carried the load until they ran flat, and New York’s air traffic control lost radar, radio and phone links. Four small faults hid the failure: a burnt-out indicator bulb, a missing alarm strap, a muffled bell and no remote monitoring. Matt Horn draws three lessons. Big buffers hide upstream faults. An alarm should fire when the healthy signal stops. Recovery has to be engineered in, not left to vigilance. In other words, make the failure visible.
matthorn.io
Practice
Penmentor: Claude marks your handwritten calculus like a tutor
You work in xournal++ with a pen tablet and press a key. Claude then annotates the page itself: it circles the first mistake and gives a hint rather than the answer. It can also generate a 3Blue1Brown-style manim animation of the idea you’re stuck on. It’s an open-source (MIT) Linux tool built on Claude Code. It’s a neat case of making errors visible during practice without handing over the fix.
GitHub: penmentor
Chess
Ten months building the most human-like chess bot
Marcus Buffett’s Chessbook-1, a 199M-parameter model, aims to play like a typical Lichess opponent at any rating and time control. It roughly ties Maia-3 at predicting human blitz moves. It stays better calibrated at high ratings, keeps its blunder rate close to that of real humans, and does better at predicting moves and think times under time pressure. Allie 2.0 predicts moves better, but it is about 28× larger.
mbuffett.com
Chess
US Championship R9: So holds Caruana, keeps a full-point lead
Round 9 had five draws, including the leaders’ head-to-head, where Wesley So defended against Fabiano Caruana. The only decisive game was Ilia Oparin’s win over Sam Shankland. Rookie Rachael Li leads the women’s event by a point. Two rounds remain, and the championship ends on 22 October.
Chess.com
Patterns
Bloom filters with one hash, via coprime partitions
A normal Bloom filter computes k hashes per key. The one-hash version (from Lu et al.) splits the bit array into k partitions whose sizes are pairwise coprime. It then takes a single hash modulo each partition’s size. Because the sizes are coprime, those residues are independent, which is the Chinese Remainder Theorem doing the work of extra hash functions. Jim Calabro’s Go implementation, gloom, lets Bluesky skip most database lookups with about 1% false positives.
calabro.io
Game Dev
Civilization rebuilt byte-for-byte with 1991 tools
civd reconstructs C source that compiles to the original CIV.EXE (474.05), byte for byte. The build uses Microsoft C 5.1, MASM 5.10 and LINK 5.01 running in DOSBox-X. The author used models from OpenAI, Anthropic and Google to draft the code and documentation, then edited it. CI with pinned hashes proves the match. The project says it is a reconstruction, not MicroProse’s source: the names and comments are inferred.
GitHub: civd
Design
IKEA instructions, animated step by step in 3D
Ryan Sael put IKEA’s wordless manuals for BILLY and LACK beside 3D builds that follow each drawing in order. The builds show every dowel, cam lock and nail. You can watch a half-turn of a cam pull two panels tight, and see why the nailed-on back is what keeps the bookcase square. (It isn’t affiliated with IKEA.)
sael.net/flatpack
Astronomy
The first green star: a white dwarf coloured by eaten planets
Mark Hollands and colleagues at Warwick found WD J1330+6435, a white dwarf that looks green. It has swallowed rocky debris from its old planetary system. Sodium from that debris absorbs orange and red light, so the star glows green. The same study shows that other atmospheric mixes can make white dwarfs look pink, purple, yellow or cyan.
ScienceDaily (University of Warwick)
Space
Euclid finds a comet it wasn’t looking for
ESA’s dark-energy telescope has discovered its first comet, C/2026 S1 (Euclid), in its Deep Field North images from September. Mikko Pöntinen of the University of Helsinki wrote software that flagged a fast-moving streak. He then followed it across five image sets, and ground telescopes confirmed it was an active comet.
Universe Today
Science
The Moon is still shrinking, and its faults may shake landers
Thomas Watters’ Smithsonian team used Lunar Reconnaissance Orbiter images of thousands of young thrust-fault scarps to estimate how fast the Moon contracts as it cools. The answer is 6–11 cm per million years. That is tiny, but enough to trigger shallow moonquakes and landslides near the faults. The team argues for a lunar seismic network before crewed bases go in.
Universe Today
Formula 1
Can Antonelli clinch in Austin? Only in an unlikely scenario
Antonelli leads Russell by 102 points with six races left. A win in Austin would take him to 365, and Russell would then have to finish eighth or better to keep the title alive. Toto Wolff says it would “probably need the other two, George and Lewis, to retire”. In his words: “This is not a likely scenario.” Mercedes brings a planned W17 upgrade. Norris is gloomy about McLaren’s Austin prospects, and Piastri is still seventh.
PlanetF1
Film Music
Roger Goula scores Idris Elba’s undersea astronaut film
Roger Goula (Next Goal Wins, All That Breathes) is scoring Above the Below, Idris Elba’s first feature as director, co-directed with Martin Owen. Its premise is that three astronauts crash-land and end up stranded beneath the sea, which leaves plenty of room for an atmospheric score. Lionsgate releases it in cinemas on 4 December.
Film Music Reporter