An AI-assisted proof of the Aaronson-Ambainis conjecture, posed in 2011, bears on quantum advantage for a class of problems. See what the preprints claim.
MIT researchers report that OpenAI's GPT-6 Astra can interpret raw weights of other models and edit them to improve maths skills. Experts urge caution on the claims.
OpenAI was due to publish Lean code for its Yang-Mills existence proof on GitHub, with a five-physicist assessment. The mass gap is not covered by the proof.
Google DeepMind announced AlphaGenome 2 on 11 November 2026, a research-only model predicting DNA variant effects by cell type. Claims are the company's own.
Google DeepMind announced AlphaGenome 2, a DNA sequence model that predicts variant effects by cell type. Training data, claimed results and limits explained.
Vals benchmark: DeepSeek, an open-weight model, completed 73% of objectives on simulated corporate networks versus 81% for OpenAI's restricted Daybreak model.
METR's assessment of Anthropic, published 5 November 2026, estimates capability gains on well-specified tasks at about twice the earlier pace, with wide uncertainty.
Mechanism design, or 'reverse game theory', aims to control AI agents through incentives rather than interpretability. What researchers propose, and its limits.