OpenAI, PerturbAI and Scripps Research used nearly 2,000 AI agents to analyse mouse brain data in 32 minutes. The agents used GPT-Rosalind, a life sciences model, and the results have not yet been peer-reviewed.
OpenAI’s GPT-Rosalind powered a split workflow in which each AI agent examined a single gene change. The agents studied cell responses, searched scientific literature and proposed explanations, before their reports were combined into a reasoning atlas.
PerturbAI built the underlying dataset by switching off individual genes in mice and measuring the effects on brain cells. In blinded tests, experts chose the multi-agent reports over the single-agent versions 75% of the time.
The reasoning atlas is available on GitHub, with the underlying dataset on Hugging Face. For companies building AI tools for life sciences, these provide material to examine the approach, but the proposed biological explanations remain research leads rather than established discoveries.
Splitting analysis across many agents could help research teams explore more scientific literature than a single-agent workflow.
The public atlas and dataset give life sciences builders material for evaluating the approach.
The results remain unreviewed, so the value of the proposed biological explanations is still unproven.
Written automatically by the Made in 10 Newsroom from the source above.
Wayve, Stellantis and Nvidia announced a French government partnership for self-driving deployment in France and the EU. The companies gave no launch date, fleet size or list of cities.
Moonshot has released Kimi K3.1 Pro in its apps and API. Full model weights are due by 4 November, but its reported coding performance has not been independently verified.
Barclays plans to cut about 5,000 jobs as technology and AI help reduce staffing needs. About 60% of the cuts affect support functions, including developers, but its chief executive says broader technology changes drive most gains.