Friday, 25 September 2026

950 Claude agents went looking in phage DNA and found something CRISPR-shaped

Anthropic has opened a life-sciences lab, and its first result is a previously uncharacterised system in bacteriophages: a reverse transcriptase, an accessory protein of unknown function, and an array of repeating non-coding DNA. The enzyme itself was already catalogued; the agent was the first to notice it sits beside a repeat array. The search ran ~950 agents for 21 hours over 210M tokens.

Wet-lab follow-up shows the array is expressed as distinct short RNAs — the CRISPR-like tell. What it actually does is still unknown, and Anthropic is explicit that these are early findings.

Anthropic

“The headline without the story”

Caltech climate modeller Tapio Schneider, in a guest post on Terry Tao’s blog, uses OpenAI’s Lean-verified Navier–Stokes blow-up proof (8 Sep) as a case study: a formal proof tells you that, not why. In maths understanding matters for its own sake; in science it matters instrumentally whenever you must act on a prediction before you can check it.

His prescription is an “auditable causal chain” from assumptions to output — AI embedded inside physical frameworks rather than end-to-end black boxes. Essentially an argument for keeping the mistakes inspectable.

What’s new (guest post by Tapio Schneider)

The ESP32-S31 boots Linux properly

Espressif published an official Buildroot + U-Boot BSP for the ESP32-S31 in August (kernel, device tree, rootfs; “not yet recommended for production”). Community ports have since pushed to Linux 6.18 with nearly all peripherals working, and a 7.1-based build driving an 800×480 LCD console, wired Ethernet, keyboard and mouse. The line between MCU and SBC keeps getting blurrier.

XDA Developers · Tux Machines

Grog would need 383 M acid to eat that mug

A Groningen chemist, Patrizio Raffa, treats Monkey Island’s melting pewter mug as an inverse problem: time the mug in-game (~35 s), infer a recession rate of ~57 µm/s, and back out the proton concentration a diffusion-limited corrosion model would need. Answer: ~383 M, against ~37 M for concentrated sulfuric acid. Stagnant acid chemistry can’t do it; the SCUMM and pepperoni must be pulling their weight.

A nice template for grounding a game’s physics just enough — the same back-of-envelope trick works for burner fuel and envelope heat loss.

Journal of Geek Studies

How Maurice Noble made Wagner look enormous for seven minutes

Why What’s Opera, Doc? (1957) looks nothing like other Looney Tunes: Noble threw out the stagey layouts for giant angular sets, huge shadows and odd colour (Elmer in red against purple — Friz Freleng’s reaction was “What the hell?”). Chuck Jones drew ~500 pose sketches, double the norm; the crew studied ballet dancers; the spotlight was a hole in blue cel. It took seven weeks instead of five, and they covered the overrun by quietly faking time cards on easier films.

Animation Obsessive

Biology isn’t quantum, but it can do quantum-shaped maths

Princeton chemist Gregory Scholes, once a champion of quantum coherence in photosynthesis, now argues that networks of classical coupled oscillators can produce synchronised states living in a Hilbert space — interference and “superposition-like” behaviour included, no actual quantum effects required. Flocks and fish schools are the intuitive picture. Classical systems, he says, “can get the functionality you could get from quantum systems.”

Quanta Magazine

The biggest fresh crater LRO has ever caught

McGetchin crater, 222 m across, was punched into the Moon in spring 2024 and spotted in LRO images late last year — the largest new crater since the orbiter arrived in 2009, the kind of impact expected roughly once every 132 years. The interesting part is around it: a 6 km thermal “cold spot” of fluffed-up regolith, and ejecta flung hundreds of crater radii out. Useful numbers for anyone planning where to put a lunar habitat.

Universe Today

Germany ends Uzbekistan’s perfect run

Olympiad round 8: Germany beat Uzbekistan 2.5–1.5, the Svane brothers (Frederik and Rasmus) winning on the lower boards. The two now share the open lead on 14/16 with three rounds left. In the women’s section China take sole first after Zhai Mo’s win; India’s Savitha Shri is still on a personal 7/7 after a 2–2 with Kazakhstan.

Chess.com

Russell half a second clear in Baku; McLaren held together with cable ties

George Russell topped both practice sessions, 1:43.347 in FP2, 0.55 s ahead of Antonelli — who lost FP1 running to a hydraulic leak and had his power unit changed as a precaution. Verstappen was third. Piastri was 6th and 8th, with McLaren nursing brake trouble and loose floor bodywork patched with visible cable ties; he says the long-run pace is “decent”. Russell needs to outscore Antonelli heavily to stay in it; the race is Saturday.

F1 — FP2 report · What the teams said

Spamalot plays The Star, a few streets from home, 15–18 October

Drew Anthony Creative’s revival opened in Melbourne to a warm notice (“the ideal antidote to the worries of the world”), with Stephen Curry as King Arthur and Jaz Flowers as the Lady of the Lake. The Sydney run is four days only at the Sydney Event Centre, The Star, 80 Pyrmont St — tap-dancing knights, killer rabbit, fish slapping. Tickets $69–$129. Silly rather than edgy; the usual Python innuendo applies.

Man in Chair review · BroadwayWorld listing