AI research & science

8 Sep 2026Google DeepMind published AlphaGenome Atlas, a set of predictions for roughly 9 billion possible single-letter changes to the human genome. It comes with a new summary measure, the AVI score, and the dataset is about 1 petabyte.

DeepMind highlights two early uses. In work involving the University of Exeter and UK Biobank data, researchers found 22% more associations in non-coding DNA, the large parts of the genome that do not code for proteins and are hard to interpret. The atlas also helped identify a variant in the DNM1 gene linked to epilepsy.

30 Sep 2026DeepMind extended its SynthID watermarking idea from images and text to proteins. SynthID Bio embeds a detectable signature in proteins designed by AI. It was tested in the wet lab on proteins including VEGF-A, the SARS-CoV-2 receptor-binding domain and PD-L1, and DeepMind says the methods are open-sourced. The aim is traceability: if an AI-designed protein turns up somewhere it should not, people can tell where it came from.

15 Oct 2026Google DeepMind described an agent-driven search for antimicrobial peptides, short proteins that can kill bacteria, using its ProtNLM protein language model. One candidate is now in wet-lab validation. That is an early stage, and most candidates fail, but it shows how agent workflows are being pointed at real scientific search problems.

21 Oct 2026Microsoft Research says Xinzhi Zhang and Zhao Song have proved the Kannan–Lovász–Simonovits (KLS) conjecture, a question in high-dimensional geometry open since 1995. The researchers directed AI agents that explored more than 100 possible approaches, while the people set the direction and checked the proof. Microsoft says AI runs left to work alone on a similar budget fell short. It also notes other independent proofs and a formalisation in the Lean proof assistant. As with any new proof, the final word belongs to peer review.

3 Sep 2026In its GPT-6 Astra announcement, OpenAI says the model helped improve the bound in the "bounded gaps between primes" result. The bound is the largest gap that infinitely many pairs of consecutive primes are guaranteed to stay within. The previous best, 240 (due to Julia Stadlmann), came down to 186. This is OpenAI's claim, and the formal mathematical record will depend on review by mathematicians.

22 Oct 2026Anthropic and Theorem report that Claude wrote machine-checked proofs, in the Rocq proof language, that 239 entry points across 123 open-source libraries cannot crash on inputs under 64 MB. The proofs total about 180 million lines and turned up hundreds of bugs along the way. The results hold only under stated assumptions, such as single-threaded RISC-V builds, and Anthropic stresses that "formally verified does not mean bug-free". Read my post on AI and software security.