The 750-task benchmark shows leading models can help with life science, but still fall far short of independent research work.
OpenAI's LifeSciBench contains 750 expert-written life-science tasks and 19,020 rubric points. The best model, GPT-Rosalind, fully passes only about one third of tasks.
The signal is not just the headline. It is how capital, regulation, infrastructure and user trust are being rearranged around AI.
Sources: MarkTechPost, StartupHub.ai, CocoLoop.