OpenAI ran 1,946 agents on brain cell data with PerturbAI and Scripps

Running hundreds of research assistants at once used to be a thought experiment. OpenAI, PerturbAI and Scripps Research published a joint post on 20 October describing a workflow that fans out across 1,946 AI agents in parallel to analyse millions of brain cell measurements, and the Hacker News thread discussing it has 140 points.

The three parties frame it as a method rather than a result. The agents each take a slice of the data and the hypotheses they emit are collected and ranked downstream.

What the agents produced

The piece names DCTN1, a gene linked to cholesterol handling in cells, as a hypothesis the agents raised for a role in neurodegeneration. The announcement positions this as a lead rather than a finding: PerturbAI and Scripps say the agents produced research directions a human team can now test, and the method is the subject of the publication.

Several Hacker News commenters working in biology pushed back on the specific hypothesis. The DCTN1 link to Parkinson's-like conditions is in existing databases, so flagging it is not new information. Another comment captured the fuller objection: "Hypotheses are cheap. PhD students come up with similar hypotheses every hour." The costly step in biology is not generating ideas, it is designing and funding the wet-lab experiment that proves one.

What the method does change

The useful part of the piece is the shape of the computation, not the specific gene. A single researcher with one slice of data finds one thing; 1,946 agents with 1,946 slices find a thousand things, most wrong and a few interesting, and the work moves to ranking rather than to coming up with the hypothesis. One commenter put the mechanical case plainly: "if you fan-out data analysis work across 1,946 agents instead of just one, the analysis is faster."

The post does not say what the run cost, how long it took, or what fraction of the 1,946 hypotheses passed human review. It does not state which OpenAI model each agent ran, so the compute profile and the token bill are not reproducible from the write-up.

For a team outside biology, the useful question is whether a job already running sequentially on one piece of data is actually a job that could run against 100 slices and have a ranker on the back of it. Most product analysis work in software looks more like that than the write-up suggests.

Source

• Scaling AI research for complex biological data (OpenAI)
• Hacker News discussion