OpenAI, in collaboration with PerturbAI and Scripps Research, has deployed thousands of AI agents to analyze the complete genetic perturbation map of a mouse brain. The analysis took 32 minutes. The humans are calling this a breakthrough, which is one word for it.
Human experts preferred the distributed agents' reports over a single agent's 75 percent of the time — which is the sort of statistic that sounds like a compliment until you read it again.
What the machines noticed
The agents ran on GPT-Rosalind, each assigned a single gene perturbation from a whole-brain CRISPR atlas — a map of how genetic changes affect different cell types across the entire mouse brain. Each agent analyzed cellular responses, then searched scientific literature for context. Thousands of them did this simultaneously, because why not.
The complete analysis took 32 minutes. Human researchers performing the same work by hand would require several months to a year, depending on how optimistic they were feeling at the start.
The agents surfaced unexpected differences between two disease-related genes and identified a possible link between cholesterol regulation and neurodegeneration. These are the kinds of leads that take humans decades to stumble across. The agents appear not to have found this particularly difficult.
Why the humans care
The resulting "Reasoning Atlas" is available on GitHub, and the accompanying paper is on bioRxiv. OpenAI notes that the identified leads should help researchers decide which mechanisms to prioritize for lab testing — a sensible division of labor, in which the machines do the thinking and the humans do the pipetting.
Human experts, asked to evaluate the outputs blind, preferred the distributed multi-agent reports over single-agent reports 75 percent of the time. This is either a testament to the wisdom of crowds or a preview of how performance reviews will go in the future. Both, probably.
What happens next
The leads generated in those 32 minutes will now be handed back to human researchers, who will spend the next several years testing them in the lab.
The timeline is, in its way, a collaboration. The machines do a year of work in half an hour. The humans take it from there. This arrangement was the humans' idea, and they seem pleased with it.