AI & Software
Postgres 19 slips because AI review finds bugs faster than people can fix them
PostgreSQL 19, due around early October, is now expected about 19 November, and eight to ten significant features (including SQL/PGQ graph queries) have been pulled to v20. Peter Eisentraut says some reverts were real architectural defects, but the bigger issue was volume: LLM-assisted review turned up edge-case defects quicker than committers could patch them. He rejects the “vibe coding” explanation, since most of the features predate it.
It’s an odd new bottleneck. Finding the mistakes is now cheap; fixing them still isn’t.
The Register
Practice
The models that sought out the hardest opponents won
StarSkirmish Bench v0.2 gives each LLM one hour to write a C++ Protoss bot for StarCraft: Brood War, then scores it against strong human-written bots. GPT-6 Astra came top, GPT-6.1 Sol second, and Gemini 4 Argon, Claude Opus 5.5 and Claude Fable 5.1 shared third. The best predictor of score wasn’t the model. It was how many test games the agent chose to play against the seven strongest bots. The models that mostly practised against weak opponents, or barely tested at all, did worst.
StarSkirmish Bench
AI
What the benchmarks don’t measure
Mercor mapped 29 agentic benchmarks onto 18,797 O*NET tasks and estimates about 90% of US working hours go to tasks with little or no benchmark coverage. Software engineering is the big exception: it’s so over-represented that they had to cap it.
Separately, Quanta profiles Moritz Hardt (Max Planck), who argues benchmarks were always “the mistake that made machine learning” but worked because rankings stayed stable even when scores didn’t. Training on the test broke that, and his proposed fix is “altitude training”: give every model the same benchmark-relevant preparation so the rankings mean something again.
Mercor · Quanta
AI
“Neural networks are clean”
Nick Cammarata (Goodfire) argues that the messy, polysemantic neurons in early vision models came from fixable architecture problems rather than superposition. Adding residual connections and LayerNorm, with no change in size or data, made InceptionV1’s last layer far more legible, and he reports about 99% of neurons in a comparable ConvNeXt are easy to interpret. Preliminary audits suggest about 90% of Qwen3-32B neuron “poles” have one main job. He presents it as a working hypothesis, but it would make interpretability far more tractable.
Goodfire
Maths & AI
Buzzard, Avigad et al.: how to formalise all of mathematics, properly
A Xena guest post by Jeremy Avigad, Kevin Buzzard and five colleagues warns that “a lousy formalization of existing results is worse than no formalization at all”. They point to sphere-packing AI code that was harder to clean up than to rewrite, and Anthropic’s 13-million-line Fermat formalisation, which they say has little lasting value. The constructive side includes Mathlib (950 contributors), Tau Ceti (~3M adversarially reviewed AI-written lines) and Formal Frontier, plus ten shared values. Their claim: most of mathematics could be formalised within three years, if mathematicians keep control of the tools.
A sharper companion piece is Rikab Gambhir’s satirical guest post on Tao’s blog, which casts AI research output as the Library of Babel.
Xena Project · Terence Tao’s blog
Computing
A signed 16-bit counter, 37 years old, roots OpenBSD
Calif traced an OpenBSD local privilege escalation to shm_nattch, the shared-memory attach count, which has been a short since the original BSD sources. Enough forks push it past 32,767, it wraps negative, and the result is a race into use-after-free. Chained with an out-of-bounds write, it gets root with KARL, SMAP and SMEP all enabled. Errata 018 and 026 fix it. For firmware people, it’s a reminder that a counter’s width is part of its specification.
Calif
Embedded
E Ink Kaleido 4: colour e-paper at 18 fps
The new colour ePaper platform roughly doubles colour gamut volume over the previous generation, with 15% more white reflectance and 50% better contrast. A new oxide-TFT backplane supports higher drive voltages, so video can reach 18 frames per second. The first device to use it is the Hanvon M10C II.
E Ink
Science
Google “cultivates” magic states on Willow
Magic states are needed for universal quantum computing, and by some estimates making them could take up to 95% of a machine’s capacity. In Science, Google Quantum AI shows magic-state cultivation working on its Willow chip: T-states reach 99.99% logical fidelity, with errors about 40× lower than uncultivated states, and they were moved from a distance-3 colour code into a distance-5 surface code. Google says it still has to be shown at larger scale.
Google
Game Dev
Godot 4.8 beta 1
Highlights: a native Trail3D node, so trails no longer need hand-built meshes or UVs (balloon wakes, burner sparks); FuzzySearch, so “health potion” also finds “Potion of Healing”; all main screens (2D, 3D, Script, …) are now dockable and floatable; mip-level texture streaming, which cuts VRAM use in big worlds; and reusable node groups in visual shaders.
Godot Engine blog
Design
A Linux desktop on a walnut desk, lit by one lamp
DEsk is a beta wlroots/Wayland desktop in full iOS 6-style skeuomorphism. Windows are sheets of paper, the dock is a drawer, apps live in binders on a bookshelf, and notifications arrive in a post box. The lock screen is a front door that opens when you type your password. The room is GPU ray-traced, with a CPU fallback. The README says Claude wrote much of the code under the author’s design direction and review.
GitHub
Science
Hurricane Simon makes ripples in the night sky
Satellite VIIRS imagery (Suomi NPP, plus a night pass from NOAA-21) caught atmospheric gravity waves spreading out from Hurricane Simon off Mexico’s Pacific coast. Strong updrafts at the eyewall launched the waves upward, and they showed up as rings in the faint airglow. They’re normally invisible from orbit and could be seen this time only because there was little moonlight. It could also serve as a visual reference for how unseen air currents might look in a balloon game.
ScienceDaily / NASA
Astronomy
The Treble Clef is a radio phoenix
VLSS J0318.9+5755 is a radio source nicknamed the Treble Clef, hidden behind the Milky Way’s Zone of Avoidance and long mistaken for a pulsar. Andrea Botteon’s team combined Chandra, eROSITA, LOFAR and UKIRT data to find the obscured galaxy cluster behind it. Their conclusion: it’s old plasma from the cluster’s brightest galaxy, re-energised when two clusters merged a few hundred million years ago.
Universe Today
Astronomy
SPHEREx catches brown-dwarf models out
Using SPHEREx’s 102 channels from 0.75 to 5 μm, Zafar Rustamkulov’s team studied 33 brown dwarfs between 7 and 100 light-years away. None of the five leading atmosphere models fully matched: they missed an opacity window near 4.0 μm, and some size estimates were off by 20–50%. Below 1,000 K, carbon monoxide gives way to methane, and only the Sonora Elf Owl model with weak vertical mixing fits that trend.
Universe Today
Chess
US Championship: four-way tie at the top
In round 6, Caruana beat Shankland and Awonder Liang beat Aronian. That puts them level with Wesley So and Andy Woodward, with Jeffery Xiong half a point back and seven of twelve within a point. In the women’s event, Rachael Li beat previous leader Alice Lee, and Li, Lee and Thalia Cervantes now share first. So and Woodward meet in round 8. The event runs to 22 October.
Chess.com
Sydney
Quiet weekend nearby; Crows Nest Fest and heavy trackwork on Sunday
There’s nothing big near Pyrmont this weekend. On Sunday, Crows Nest Fest closes Willoughby Road and side streets from 10am to 5pm (closures run from 12:01am to 11pm), routes 115 and 263 are diverted, and the M1 to Crows Nest is the easy way in. Sunday trackwork also hits the T1, T2, T3, T5 and T6 lines, so check the Trip Planner before heading anywhere by train. The next Rugby League World Cup match in Sydney is Tonga v Lebanon at CommBank Stadium on Friday 23 October.
Transport for NSW · AusStadiums
Musicals
Spamalot moves to Rooty Hill next week
Drew Anthony Creative’s new staging of Monty Python’s Spamalot finishes at The Star on Sunday and plays the Coliseum Theatre, Rooty Hill, from Thursday 22 to Sunday 25 October. It has killer rabbits, cows and “Always Look on the Bright Side of Life”, and it’s a good pick for a teenager who likes her musicals silly.
Sydney.com · BroadwayWorld