The timeline for quantum computers breaking real-world encryption just collapsed. Two groups — one at Caltech, another at a Google–Princeton collaboration — have independently shown that quantum low-density parity-check (qLDPC) codes can dramatically reduce the qubit count needed to factor RSA keys and crack elliptic curve cryptography. The Caltech team puts the ECC number at roughly 26,000 physical qubits, down from prior estimates in the millions. They’ve formed a company, Oratomic, to build the machine using neutral-atom arrays. This is Quanta at its best: patient exposition of the error-correction breakthrough that made these numbers possible, and a clear-eyed look at what remains hard.

Stellar theory has long predicted a “forbidden zone” in black hole masses: stars between roughly 50 and 130 solar masses should blow themselves apart via pair-instability supernovae before they can collapse, leaving a gap in the black hole mass spectrum. Previous gravitational-wave catalogs hinted at it but couldn’t nail it down. Now, analysis of LIGO–Virgo–KAGRA’s fourth transient catalog (GWTC-4) has found clear evidence of this gap — but only in the secondary (lighter) black hole of each merging pair. The primaries show no gap at all.

The implication is striking: the heavier partners in these mergers are likely “second-generation” black holes, themselves products of earlier mergers, which can leapfrog the gap. The lighter partners, formed directly from stellar collapse, respect the theoretical boundary. The lower edge lands at ~45 M&sun;, broadly consistent with predictions but now measured rather than assumed. This is the kind of result where the mass spectrum of black holes becomes a laboratory for nuclear and stellar physics at once.

The Nature paper is technical but the result is clean. If you want the astrophysics context, the Bioengineer.org summary is serviceable; for the real thing, see the paper in Nature.

Simon Willison — the most consistently useful voice on practical AI tooling — lays out his current mental model in a long interview with Lenny Rachitsky. The core claim: November 2025 (GPT-5.1, Opus 4.5) was the inflection point where coding agents went from “mostly works” to “actually works.” He walks through three agentic engineering patterns he uses daily (red/green TDD, templates, hoarding), then describes the next leap: the “dark factory” pattern, where AI writes code, tests it, and QAs itself with no human review. His prediction — 50% of engineers writing 95% AI-generated code by end of 2026 — is aggressive but argued carefully. He also flags prompt injection as the unsolved “lethal trifecta” that could produce an AI Challenger disaster. Worth the full read or listen.

Harvard physicist Matthew Schwartz gave Claude Opus 4.5 a real HEP-theory problem — resumming the Sudakov shoulder in the C-parameter for e+e− collisions — and supervised it through the entire calculation without ever touching a file himself. 110 drafts across seven stages (kinematics through matching), 36 million tokens, 50–60 hours of human oversight, and the result was a published paper with genuine physics contributions. The honest assessment: Claude is at the level of a strong second-year grad student — capable but sloppy, prone to fabricating justifications, and requiring constant expert correction. Schwartz calls it “the most important paper I’ve ever written — not for the physics, but for the method.” If you sit at the intersection of physics and AI, this is the most concrete case study yet of what supervised AI research actually looks like in practice.