Embedded
The Forgetful CPU: why Linux on the M4 kept losing its registers
On Apple’s M4, the ARM WFI (wait-for-interrupt) instruction zeroes registers x0–x31. The ARM64 spec forbids exactly that. On M1–M3 a “chicken bit” switched this behaviour off, but on M4 that bit is locked or gone. The post traces the crash with a hand-rolled debug_putc in assembly and a binary search through early boot. It ends at a write to a timer-FIQ enable register during interrupt-controller setup.
The upstream fix is to skip WFI entirely: m1n1 adds idle=nop on bare-metal M4, and proper low-power idle moves to cpuidle/PSCI work. Multi-core Linux now boots on M4, M4 Pro/Max and M5.
yuka.dev
Game AI
An $8,000 AI beats the greatest Stratego player 15–1–4
Stratego hides almost everything: you don’t know which enemy piece is the Marshal and which is a bluff. That has made it far harder for AI than chess or Go. A CMU/MIT/NYU/Stanford team (Sokota, Kolter, Farina and others) reports in Nature that Ataraxos beat Pim Niemeijer, the most decorated Stratego player ever, 15 wins, 1 loss and 4 draws. It cost under $8,000 to train. The method is self-play RL plus test-time search, with a belief network over the hidden pieces. The key trick is to deliberately damp learning updates, which otherwise go cyclic or chaotic under imperfect information.
Nature · Ars Technica
Science
Sea monkeys run the turbulence cascade backwards
Textbook 2D turbulence sends energy one way between scales. Lei Fang’s group at Pittsburgh filled a thin tank with swimming brine shrimp and found that the swimmers’ orientation sets the direction. Angled more than 45° from the direction of maximum stretching, they push energy from large scales to small ones, the reverse of the usual 2D cascade. Geometry alone flips the flow, which suggests cheap ways to steer mixing.
Quanta Magazine
Art & AI
stillwet: language models painting in simulated oil
No image generator is involved. Each painter (Opus 5.5 among them) writes a program against a physical simulation of oil on linen, with bristle brushes, wet-into-wet blending, drying time and layered glazes. The site replays each painting sped up. Most pieces are after Caspar David Friedrich, but the models only ever read about his work and never see it. Even so, separate runs converge on near-identical Baltic shores. Source is on GitHub.
stillwet.art
Craft
An open-source agent that designs buildable LEGO
ldraw-nova gives a coding agent a model idea and the LDraw docs (LDraw is effectively LEGO CAD “assembly language”). The agent writes a plan JSON and Python generator scripts that emit the .mpd file. It then renders the result, looks at it and adjusts placements, round after round. It handles vehicles, spaceships, Technic mechanisms and modular builds. Like the paint project above, it works by making the model write a program that produces the artefact, not by drawing it.
GitHub: anteloc/ldraw-nova
AI & Software
Frontier open weights at home: antirez’s ds4, and a month on GLM 5.3 Flash
Salvatore Sanfilippo (Redis) released ds4 “DwarfStar”, an MIT-licensed C inference engine built narrowly for DeepSeek V4/V4.1, GLM 5.x and Qwen 3.8 on 64 GB+ Macs or CUDA/ROCm boxes. It uses asymmetric 2-bit quantisation of routed experts, keeps the KV cache on SSD so sessions resume instantly, and includes a built-in coding agent.
Separately, Wagtail’s Thibaud Colas spent a month coding only with open-weight GLM 5.3 Flash. The model worked well, but the budget didn’t hold. A prototype burned 450M tokens, total energy came to 35 kWh against a 10 kWh plan, and his advice is to meter usage continuously.
dwarfstar.sh · Wagtail blog
Maths & AI
AIM: a shared repository of open problems, plus an explaining prize
In a guest post on Tao’s blog, Matthew Colbrook, Steve Brunton, Alex Townsend, Rachel Ward and others launch AIM. It is a GitHub collection of precisely stated open problems in applied maths, open to proposed solutions, partial results, counterexamples and corrections. Its companion site, “AIM, explained”, collects write-ups and videos. There’s also an explanatory-writing and video competition (undergrad and grad/postdoc tracks, deadline 1 Jan 2027) that explicitly rewards clarity over speed.
What’s new (Terence Tao) · GitHub
Space
Bennu was born in two places at once
Bennu is a carbonaceous, outer-Solar-System type asteroid, yet the OSIRIS-REx samples carry inner-Solar-System iron isotopes. ETH Zurich and LLNL researchers explain the mismatch with Jupiter. Its gravity raised pressure ridges in the disk near the water-ice line, where outer-disk dust piled up and mixed with inner-disk material before Bennu’s parent body formed.
Universe Today
Sydney
Next Sunday: Spring Cycle finishes in Pyrmont, Western Distributor closed
Sunday 11 October. The 10 km City Ride (08:00 start) finishes at Pirrama Park, and the 50 km Classic (06:30) rolls through Pyrmont on its way to Parramatta Park. From about 4am to 11:30am the Western Distributor is closed from Pyrmont to City North, along with Harbour Bridge lanes, the Cahill Expressway overpass and parts of Hickson Rd and Upper Fort St. Expect a busy Pirrama Rd foreshore that morning. Riders travel free on trains, metro and ferries.
Today (Sun 4 Oct) is NRL Grand Final day at Olympic Park, with gates at 1pm, and Noah Kahan plays there through Tuesday.
Transport for NSW · CBD & Pyrmont closure map (PDF)
Formula 1
Verstappen’s first pole of 2026; Antonelli 3rd, Russell 7th, Piastri 6th
Red Bull finally hit the sweet spot at Sepang: Verstappen took pole for the relocated Bahrain GP from Hamilton. Hadjar qualified 3rd but a grid penalty drops him to 8th, so championship leader Antonelli starts 3rd. He felt the front row was possible but wasted a soft set after a Turn 6 track-limits deletion in Q2. His only rival, Russell, is 7th and “perplexed” by Mercedes’ pace, so the 66-point gap could grow. Leclerc 4th, Norris 5th, Piastri 6th. Lights out today at 6pm Sydney time (now AEDT).
RaceFans · Formula1.com (start times)