Thursday, 8 October 2026

OpenAI dumps 377 AI-generated maths results on GitHub at once

An unreleased internal model was pointed at roughly 4,000 open problems, averaging about three hours of ChatGPT Pro-level compute each; what came back is 722 manuscripts in ~370 result families across 16 areas. The headline claims include the Unique Games Conjecture, matrix multiplication exponent ≤ 9/4, and integer multiplication in O(n (log n)^1−κ) on a multitape Turing machine with κ = 2⁻¹⁸² — beating the n log n bound Schönhage and Strassen conjectured optimal, by a margin you would never measure.

Many but not all proofs are Lean-checked (the multiplication one isn’t); several claims are narrower than their titles; nothing is peer reviewed. IAS’s advisory group had asked labs a week earlier not to do exactly this with proprietary models.

OpenAI · openai/math · Gizmodo · explainx (claim-by-claim)

Seven years on a proof, then a race against the machine

Dor Minzer (MIT) and students Yumou Fei and Shuo Wang cracked a version of the 2-to-1 games conjecture in April after years of failed attempts; when rumours spread on 11 Sep that OpenAI had the full Unique Games result, they rushed out an unpolished 95-page paper in three days. Minzer’s worry is the interesting part: the long stretch of failing is where the understanding gets built, and tools that skip it may discourage anyone from starting ambitious projects at all.

Quanta Magazine

Hexagon: a new home for maths papers, labelled by how much a human understands

A new repository for mathematics and theoretical CS, run by a foundation whose board includes Bryna Kra and Lauren Williams (advisers include Kevin Buzzard and Ravi Vakil). No peer review; instead every deposit carries an authorship label (human-written, mixed, primarily AI-generated) and a “human understanding” label — all, some, or no parts. Timely, given arXiv’s new two-a-month cap and today’s OpenAI dump.

Ben Antieau (guest post, Terence Tao’s blog) · hexagonmath.org

Mistral Large 4: a 1T-parameter European MoE, weights by month’s end

49B active parameters, reasoning and instruction-following in one multimodal model, trained from scratch on 3,800 Grace Blackwell GPUs. Mistral reports 61.7% on DeepSWE and second place in a blind coding-quality evaluation (3.74/5, behind Claude Opus 5 at 4.22). Preview API now at $1.36/$4.18 per million tokens; open weights promised by the end of October, licence not yet stated.

Mistral AI

A Pi-hole on a $2 ESP32-C3 with no PSRAM

The trick is not to store domain names at all: the blocklist becomes sorted fixed-width FNV-1a hashes in flash, binary-searched per query (and per parent suffix). About 130k domains fit in ~1 MB of flash using ~40 KB of RAM; a single-app partition table stretches to 537k 40-bit hashes at the cost of firmware OTA. mDNS dashboard, per-client stats, scheduled list pulls; a RAM Bloom pre-filter is next.

GitHub: esp32-c3-adblock

openTPU: RTL tuned overnight by a tournament of LLM agents

An open LLM-decode accelerator (SystemVerilog, ISA, bit-exact simulator, kernel compiler) for a Kintex-7 card. Its tourney tool has agents propose and implement hardware changes, keeping one only if it passes lint, RTL-vs-simulator equivalence, cycle counts and yosys synthesis: one night tried 96 changes, kept 38, and took the logic from 41 to 106 MHz while shrinking 132K to 82K LUTs. Caveat: simulation only — no Vivado timing, no bitstream, no hardware yet.

GitHub: openTPU

random() % N is quietly unfair — one team banned it

Mapping 0–9 onto three choices gives the first a 40% chance, not 33%. A team doing deterministic fuzz testing found modulo-biased weighted choices skewing what their tests explored, ripped them out, and put a moratorium on new ones in favour of a random_choice() taking explicit integer weights. A small one, but exactly the kind of bias that hides the bugs you were trying to surface.

ersc.io

Russell to start Singapore from the back after Sepang engine failure

Mercedes confirmed George Russell will exceed his power-unit allocation and take a grid penalty of at least 10 places for Sunday’s race (not Saturday’s sprint) — in practice, the back. He is 84 points behind teammate Kimi Antonelli with a sprint and seven Grands Prix left, so a poor Singapore for Russell makes Austin on 25 Oct an increasingly live match point.

Formula1.com

Quarks leave wakes in the Big Bang’s primordial soup

CMS physicists used Z bosons as a clean tag: one flies off opposite a quark and barely touches the quark-gluon plasma, so any disturbance on the quark’s side is that quark’s doing. Sifting ~13 billion heavy-ion collisions for ~2,000 such events showed ripples and swirls behind the quark — the first direct evidence the plasma responds to a fast particle like a fluid, as a duck’s wake.

ScienceDaily (MIT)

David Shire’s rejected Deadfall score gets its first release

Caldera Records is issuing Shire’s unused score for Christopher Coppola’s 1993 Deadfall (the film went out with Jim Fox’s music instead), paired with his TV-movie score I Married a Monster. Both are premieres on CD, with samples on the label’s site — a chance to hear the road not taken from the composer of The Taking of Pelham One Two Three.

Film Music Reporter