Meta releases open datasets for training AI Co-Scientists

Meta shows how to train AI on 1000s of research problems at once, without having to create costly environments for each task. They show improvements across arXiv domains, and crucially also medicine, where otherwise digital simulators would not be feasible.

"Crucially, subjective judgments regarding scientific novelty and value remain with human researchers, who articulate their objectives and constraints in the research goal."

This might be an alternative path for accelerating science, applicable across more domains, and more easy to scale task diversity for.