Monday, 19 October 2026

Golden age or dark age? Two essays on who gets to hold AI maths

A guest post on Terence Tao’s blog by Francesco Maggi argues the outcome depends on governance, not capability. Machine-generated Lean proofs can run to hundreds of thousands of lines and are hard to turn back into human understanding. Inventing the frameworks that make a problem mathematical is still human work. He proposes that universities, societies and funders run publicly governed LLM infrastructure, with disclosure standards for AI-generated results.

Daniel Paleka adds a sharp observation. Statistics, ML theory, optimisation and numerical analysis account for about 20% of arXiv theorems but under 1% of OpenAI’s 700+ released results. His reading is that the commercially useful maths stays in-house.

What’s new (Tao blog) · Daniel Paleka

Anthropic publishes its own numbers on how engineering work changed

By August, agents at Anthropic were logging about six hours of work for every human hour. Commits per engineer are about 5× what they were six months earlier, with no rise in reverted code. About 80% of interactive agent-session time is spent waiting on a human, and agent-wrangling has grown from 10% to 22% of sessions. Juniors fell from 21% to 14% of technical staff. Hiring now weighs autonomy and judgment over domain expertise.

Anthropic

Sphinx: a git hook that won’t let you push AI code you can’t explain

Before a push to a protected branch, a separate reviewer model that never saw your chat reads the diff. It then asks you how the change behaves when something goes wrong. Answer correctly and the push goes through. Answer wrongly and it explains the gap and gives you a fresh scenario. Your answers are never shown to the coding agent. Every bypass via sphinx yolo is logged. It works with Claude Code, Codex, Cursor and others. The hook deliberately exposes your own misunderstanding before production does.

Sphinx · GitHub

Shirriff’s 8087, part next: how the chip adds and subtracts

Addition is a microcode subroutine that aligns exponents, does an unsigned add or subtract, then normalises. Depending on the signs, a high-level “add” can become an unsigned subtraction. Leading zeros are counted in two passes of 8-input detectors, one for bytes and then one for bits, instead of a single 64-bit detector. That trade cost some speed but simplified the circuitry. The special cases for infinities, zeros and NaN go through a run of 18 consecutive conditional branches.

Ken Shirriff

Open-source hardware goes micron-scale

RepRapMicron is a RepRap-style printer for features of a micron and below, built from accessible materials. Its PIKA V0.02 build achieves about 1 µm repeatable positioning in an 8 mm cube. It has deposited 4 µm resin droplets and drawn 9 µm cured lines.

Separately, Open MicroManipulator is a parallel-kinematic XYZ stage that drives cheap steppers in a 30 kHz closed loop. A “magnetic gearing” trick gives its low-cost magnetic encoders about 30× their native resolution, for roughly 50 nm steps. The README stresses that resolution is not accuracy. It also warns that PCB v1.2 shorts the 5 V rail to ground through a misplaced via.

RepRapMicron · Open MicroManipulator

Pele: Mermaid diagrams as tiny, themeable SVG

Obsidian’s kepano released an MIT-licensed, dependency-free renderer that handles every Mermaid 12.1 diagram type. It claims 230 kB gzipped versus 1.6 MB for Mermaid.js, and sub-millisecond renders. Theming is done with CSS variables, so a colour change needs no re-render. On narrow screens it reflows the layout instead of shrinking the diagram.

pele.run

Five Hundred Clocks

Tom Gally tests frontier models by asking for self-contained clock pages, his own version of Simon Willison’s pelican test. Over several sessions, Opus 5.5 produced 500 clocks, packed into one 6.3 MB HTML file with time-warp controls up to 3600×. They include a Pong clock and an interactive-fiction clock whose prose changes with the hour. Others are built around Nine Men’s Morris, Mancala and Scrabble.

Five Hundred Clocks · HN

Duck Pond: a flock of mallards, synthesised live

A browser simulation of three hens, two drakes and two broods of ducklings, with no recordings. Every quack is generated when it happens and placed by distance from where you stand on the bank. There are range rings out to 40 m and a headphone mode. Press p to throw peas. The method isn’t documented, but it’s a nice reference for procedural ambient audio in a game world.

Duck Pond

lpviz: draw a polytope, watch the solvers race

You sketch a feasible region and an objective in the browser, and the page animates simplex, interior-point methods, PDHG and the ellipsoid method as they solve it, with the central path drawn. Simplex walks the vertices while interior point cuts through the middle, and the difference is easy to see.

lpviz

Altermagnetism wins the Europhysics Prize

Jairo Sinova, Libor Šmejkal and Tomas Jungwirth share the 2026 EPS condensed-matter prize for establishing a third kind of collinear magnetism. As in an antiferromagnet, the moments cancel and there is no net magnetisation. Yet the electrons carry strongly spin-polarised currents, as in a ferromagnet. The idea came from a 2022 symmetry classification that predicted hundreds of candidate materials, which ARPES has since confirmed. The likely payoff is fast, stray-field-free spintronics.

ScienceDaily / JGU Mainz

Roman’s wide-field camera is ~30% sharper than spec

Seven weeks after launch, final alignment brought the Nancy Grace Roman telescope’s wavefront error to 63 nm RMS, against a 90 nm requirement. The last correction moved the secondary mirror by about a sixteenth of a hair’s width. Coronagraph contrast tests come next.

ScienceDaily / NASA

A rocket jellyfish over the eastern US

The 13 October dawn launch of SpaceX CRS-35 drew a glowing “jellyfish” across the sky. The ground was still dark, but the exhaust plume high above was already in sunlight, so it lit up as it spread. The same twilight geometry could make striking sky effects for a balloon game set at dawn: a dark landscape below with sunlit high cloud above.

ScienceDaily / NASA

Orionids peak Wednesday night

Debris from Halley’s Comet should peak around 21–22 October at about 20 meteors an hour. The waxing gibbous Moon, five days before full, sets after midnight, so the hours before dawn are darkest. That is also when Orion is highest in Sydney. Slow, bright Taurid fireballs start appearing late in the month, and Encke’s February 2027 perihelion may mean a better show than usual.

Universe Today

Justice are scoring GTA VI

Rockstar has confirmed the French duo Gaspard Augé and Xavier de Rosnay for the game’s original score: the instrumental music under missions and free roam, separate from the licensed radio. The launch-trailer single “Flower in a Hurricane” features Tame Impala (in Sydney this week), SZA and King Gizzard. The game ships 19 November on PS5 and Xbox Series X|S.

DJ Mag · Gfinity

US Championship: So a full point clear, and plays Caruana next

Wesley So beat Andy Woodward in a pawn endgame after Woodward went wrong near the time control. So is on 6/8, with Caruana on 5 after a draw with Sevian. In the women’s event, Alice Lee missed a big advantage against Zatonskih, leaving her tied with Rachael Li on 6. Cervantes beat defending champion Carissa Yip with Black and is on 5.5. Round 9 is Caruana–So and Krush–Lee, Tuesday 4am Sydney time. The event finishes on the 22nd.

Chess.com

Weekend ahead: Zara Larsson, Randwick and a Wanderers double-header

After Tame Impala’s three nights, Olympic Park has Zara Larsson on Friday 23 Oct (finishing around 10:30pm). Saturday 24 Oct is Spring Champion Stakes Day at Royal Randwick, with gates at 11:15 and the last race at 5:55pm. Expect busy eastern-suburbs buses and L2/L3 light rail. The Western Sydney Wanderers play an A-League double-header at CommBank that afternoon, while Parramatta Lanes runs Wed–Sat evenings.

Transport for NSW events