Unlearning the Universe
A 2026 study shows AI trained on the standard model of cosmology can miss new physics — a flaw called negative transfer. Why machines must learn to forget.
AI & Scientific Discovery
The Balloon Forecast: AI Beats the Weather Agencies
WindBorne’s WeatherMesh-6 makes hourly 3-km forecasts and beats Europe’s flagship model on key variables — with 400 balloons and fresher data, not a bigger net. The privatization of weather.
The Day Washington Decided Who Gets Frontier AI
OpenAI built GPT-5.6 Sol, its most powerful model yet, then limited it to ~20 government-approved partners at the White House’s request. Inside the first gated frontier release.
How AI Found a 50-Year Hole in Economic Theory
Axiom Math’s EconLib formalizes economics in Lean — and its verifier already flagged a gap in a foundational proof economists taught for fifty years. The machine-checked future of social science.
The Hacker in the Machine
June 2026: the Five Eyes spy chiefs warn that AI capable of devastating cyberattacks is months away, not years. Inside the offensive-AI threat and the defense.
The Motion Problem
AlphaFold solved the static shape of proteins. The 2026 frontier is motion — AI that predicts how proteins flex, breathe, and bind. Here’s the leap.
The Drug Engine
A 1,016-GPU supercomputer and a $2.5B AI drug deal show how, in 2026, artificial intelligence became the infrastructure of drug discovery itself.
The Hard Horizon
AI agents can work for hours and keep getting longer — but brutal new June 2026 marathon benchmarks show full autonomy is still far off. Here’s the real state of play.
The Fair-Use Frontier
A $1.5B Anthropic settlement, 70+ suits, and the first appellate fair-use argument: inside the court fight over whether AI could train on copyrighted work.
The Crown Changes Hands
Anthropic passed OpenAI, Claude Opus 4.8 topped the leaderboard, Gemini Deep Think landed. Inside the most compressed frontier AI race yet — and where it heads.
Buy me a coffee