Meta's Fundamental AI Research team hosted researchers at its Menlo Park, California, headquarters for the Open Source AI for Science Summit, announcing updates to its scientific AI models and tools.
Advancing Medical and Chemical Research
Meta announced an expansion of the Digital Brain Project, which was launched with the Rothschild Hospital Foundation and the University of Montreal. The company is increasing the project's data target from 15,000 hours of de-identified brain data to 25,000 hours to help researchers study human intelligence and neurological health.
At the summit, Meta also released UMA Nitro, a faster version of its Universal Model for Atoms. Studies with Carnegie Mellon and UC Riverside showed UMA models identifying crystal forms of drug molecules, while work with the University of Chicago simulated entire enzymes atom by atom.
Educational Tools and Mathematical Verification
Discovery Education announced that UMA Playground, a tool that lets users build molecules and explore chemical reactions, is now available to high school chemistry teachers with classroom guides. RealChem, a college chemistry course from Arizona State and Carnegie Mellon, also plans to integrate the tool.
FAIR researchers shared open source projects to support mathematical verification using Lean, an open source language. These projects include ATLAS, an open library formalizing 26 open-access textbooks and over 45,000 computer-checked definitions and results.
Key facts and where they come from
- Meta increased the Digital Brain Project's data target to 25,000 hours.
We’re increasing the project’s data target from 15,000 hours of brain data to 25,000
- Meta released UMA Nitro, a faster version of its Universal Model for Atoms.
At the summit, we released UMA Nitro, a faster version of Meta’s Universal Model for Atoms.
- Discovery Education made UMA Playground available to high school chemistry teachers.
Discovery Education announced that the tool is now available to high school chemistry teachers, with classroom guides.
- FAIR shared ATLAS, an open library formalizing 26 open-access textbooks.
ATLAS, an open library that formalizes 26 open-access textbooks and more than 45,000 computer-checked definitions and results
The catch
- The announcement does not say when the expanded Digital Brain Project data target of 25,000 hours will be fully reached.
- The announcement does not detail performance metrics or benchmarks for the new UMA Nitro model compared to its predecessor.
The open question: How will researchers ensure data privacy and security when expanding the Digital Brain Project to 25,000 hours?
Tell us what you think · Discuss it in the AI room
Reader poll: Should open-source AI models for scientific research be subject to additional safety evaluations?
A show of hands from readers, not a scientific poll. Say why in the comments.
