Wednesday, 7 October 2026

A team of Opus 5.5 agents propose two room-temperature magnetic semiconductor candidates

Vals AI set Claude Opus 5.5 agents loose on a hunt for spintronic memory materials: zero net magnetism but spin-sorted electrons, semiconducting, and ordered above room temperature. Using DFT at two levels (PBE+U and the more expensive HSE06) they landed on a newly designed oxide, YBaMnFeO₅ (2.35 eV gap), and a forgotten 1999 Prussian-blue compound, KV[Cr(CN)₆], ordered to 376 K.

The write-up is admirably upfront about the caveats: the oxide’s ordering may scramble at synthesis temperatures, the 1999 sample held water the two methods disagree about, and nobody has measured the spin splitting yet. Inputs, raw outputs and analysis code are all public.

Vals AI

“Ephemeral testing”: build a throwaway app on your library to expose its flaws

Daniel Lemire proposes a QA trick for the agent era: have an AI agent build something real on top of your code, watch where it struggles, then delete it. Hidden state, surprising defaults and thin docs show up as a trail of patches and failed attempts; a clean API lets the agent get something working fast. The disposable app is the instrument, not the product — a deliberate activity whose whole purpose is to make the foundation’s mistakes visible.

Daniel Lemire

Samon: raking a zen garden as a Hamiltonian-path puzzle

A monk walks backwards out of the garden, raking one unbroken stroke over every sand cell. The bare rule is far too loose (1.5 million solutions on an empty 7×7), so instead of adding rules the design scores by turns — and the traditional parallel lines and concentric rings fall out of minimising them. Later worlds add one diegetic constraint each; the “wide rake” (no U-turns in open sand) cuts one garden’s 6.3 million rakings to a single, beautiful answer.

Every claim is checked with exhaustive search and CP-SAT, hints are proofs of what’s still reachable rather than spoilers, and all 148 gardens are playable in the page. A strong worked example of “score, don’t forbid” for puzzle design.

gwern.net

The Erdős problems site stops counting solves

In a guest post on Terence Tao’s blog, Thomas Bloom explains why erdosproblems.com has frozen per-problem comments, stopped showing open/solved status, dropped attribution (“It is known that…”) and now prioritises well-written exposition. The comment threads had filled with unexplained AI-generated proofs and glory-chasing; Bloom wants the site to be about understanding the problems again, in the spirit of mathematics as something done out loud.

What’s new (Terence Tao)

Reflection’s Beam: a 501B open-weight coding model, Apache 2.0

A sparse mixture-of-experts with 23B active parameters, trained with over 100 million RL rollouts aimed at coding and agentic work. Claimed scores: 80.9 on SWE-bench Verified, 77.2 on SWE-Bench Pro v2-Hard, 80.1 on Terminal-Bench 2.1, at 3–4× less inference compute than comparable open models. Weights and tech report are promised “later this month”, so the numbers are unverified for now.

Reflection AI

Dust: training transformers without backpropagation

Instead of perturbing weights, Dust adds noise to each layer’s activations independently per token, so every token becomes a member of an evolutionary “population” within one forward pass. Reward-weighted noise times the layer inputs gives a gradient estimate. Up to 243M parameters it trails backprop at small token budgets, but the gap shrinks with population size — and, against the usual intuition for zeroth-order methods, bigger models are more population-efficient.

Q Labs

George Fenton (1949–2026)

The British composer died unexpectedly at 76. Five Oscar nominations (Gandhi, Cry Freedom, Dangerous Liaisons, The Fisher King), three BAFTAs, and a long partnership with Ken Loach. For a listening list: Groundhog Day, Memphis Belle, Ever After, Anna and the King — and his Attenborough scores, The Blue Planet, Planet Earth and Frozen Planet, which are among the best orchestral writing ever done for television.

Film Music Reporter

Singapore can’t settle it; Austin is Antonelli’s first match point

Antonelli 320, Russell 236, Hamilton 214, with 183 points left (Singapore sprint plus seven Grands Prix). Even a perfect Singapore weekend leaves Russell mathematically alive, so the earliest clinch is Austin on 25 October. Piastri sits 7th on 128. Singapore runs 9–11 October with a sprint, amid reports of unhealthy air quality in the city this week.

GPblog · Grande Prêmio

Flatten SF: every route nothing else beats on both distance and climb

A San Francisco router that shows the whole Pareto frontier between shortest and flattest, with a slider for the trade-off (default: a foot of climb costs 200 feet of walking). It counts cumulative climb rather than net elevation, uses USGS 1 m lidar over 160,000 street segments, and runs entirely in the browser. The same distance-vs-effort framing would make a nice route-planning layer for a balloon over terrain and wind.

Flatten SF · GitHub

Anchoring GPS to the Moon to stop the constellation drifting

If ground stations go quiet, a navigation constellation slowly drifts as a whole, because Earth’s near-symmetric gravity moves every satellite in a shell together and nothing inside the system can see it. A simulation linking 24 BeiDou-3 satellites to four lunar frozen-orbit satellites cut 60-day drift from 7.85 m to 0.35 m — an outside reference point that breaks the symmetry.

Universe Today