This week: OpenAI sent thousands of AI agents into a giant map of mouse brain cells. They finished before most of us finish lunch.

Elsewhere: Why one large agent lost to a crowd of small ones, plus where scientists can explore the results themselves.

The New Prime Signal

Modern biology has a strange problem: labs can now produce data much faster than anyone can read it. On October 21, OpenAI, biotech startup PerturbAI, and Scripps Research described one way around that. In a joint post, they say they "used thousands of AI agents powered by GPT‑Rosalind to analyze and interpret existing experimental data from millions of mouse brain cells in just 32 minutes." Two words in that sentence matter a lot: "existing" and "mouse." The agents didn't run new experiments, and these weren't human cells. What they did was read a huge pile of results and point to the parts worth a closer look.

What happened: The data came from PerturbAI's whole-brain CRISPR atlas, a dataset showing how switching off individual genes changes different types of cells across the mouse brain. The team split the work so that each agent took one gene disruption, analyzed how the cells responded, searched the scientific literature, and proposed an explanation. OpenAI calls the approach "distributed biological reasoning." All those reports were then stitched into a "reasoning atlas" that links genes, cell effects, published papers, and possible biological mechanisms. OpenAI's head of life sciences, Joy Jiao, put the crowd at "nearly 2,000 agents in parallel" in the PerturbAI announcement. For comparison, the team "estimated that comparable human analysis and literature review, conducted one perturbation at a time, would take months to a year." That's their own estimate, not a stopwatch race against real postdocs, but even a generous discount leaves a big gap.

Why it matters: According to OpenAI, the agents reproduced findings from earlier human analyses and turned up extra leads. Two stand out. First, Tsc1 and Tsc2, two genes that normally work as a team. They acted alike in several groups of neurons, but "in the cerebellum, which helps coordinate movement, Tsc2 disruption altered the activity of dozens of genes involved in neuronal communication, while Tsc1 disruption produced little change." That could be relevant to tuberous sclerosis, where cerebellum problems in people are strongly linked to TSC2 mutations. Second, a gene called Dctn1, whose human version, DCTN1, causes Perry syndrome when mutated—a neurodegenerative disease with "Parkinson-like symptoms." In the mouse data, "disrupting Dctn1 increased the expression of genes involved in making cholesterol." The agents connected that to research on cholesterol transport and raised a question: could problems moving cholesterol around inside neurons contribute to neurodegeneration? Nobody has answered it yet. It's a lead for the lab, not a finding about patients.

The fine print: OpenAI's post opens with a big hook: "What if scientists could uncover new clues about Alzheimer's and Parkinson's in minutes instead of months?" Note the question mark. The work doesn't report new clues about either disease. It reports leads from mouse data that still need experiments. OpenAI says as much, describing the results as a way to "help prioritize further experiments needed to test the proposed biological mechanisms," and adding that "Scientists remain at the center of discovery." Most of the claims come from the companies involved, and PerturbAI and OpenAI are expanding the partnership into joint drug-target discovery, so they have a stake in its success. The details are in a bioRxiv preprint, which hasn't been peer-reviewed.

The flashy number is 32 minutes. The more useful idea is the division of labor: give each agent one small question, let it read deeply, then combine the answers. It won't cure anything on its own, and nobody involved says it will. But if it helps a lab pick a better next experiment, that's a decent day's work for a few thousand interns who never ask for coffee.

The team asked an obvious question: could one agent handed the whole dataset do just as well? They tested it with the model and resources held fixed, and the answer was mostly no.

How it works: As the workload grew, "the single agent searched less of the literature, citing only 11 references at the largest workload, while the distributed agents cited more than 700 references." Quality held up too, at least by expert taste. "In blinded comparisons, human experts preferred distributed-agent reports over single-agent reports 75% of the time." The logic will feel familiar to anyone who's been handed too much reading. One agent with thousands of gene disruptions skims. Thousands of agents with one each can actually read. The underlying atlas is big: PerturbAI's dataset card lists 7.7 million cells and nearly 2,000 disease-linked genes.

Try it: If you work in biology, PerturbAI's brain atlas page links to the new preprint and OpenAI's write-up, and the raw whole-brain CRISPR atlas is on Hugging Face (fair warning: about 205 GB). If you don't, the transferable lesson is free: when an AI tool struggles with a giant job, try breaking it into many small, identical jobs and merging the answers. It works for brain atlases, and it works for a 400-row spreadsheet too.