Artificial Intelligence
OpenAI paper uses thousands of agents on brain data
A paper from OpenAI and PerturbAI said distributed AI agents could analyse a whole-brain Perturb-seq atlas in minutes, recover the study's conclusions, and help rank follow-up leads.
- The authors said they built a distributed biological reasoning method that split an experiment into many parts for separate AI agents to examine.
- Applied to a whole-brain Perturb-seq atlas, the system used thousands of parallel agents and was said to finish in minutes rather than the months the authors estimated for human review.
- The paper said the agents reconstructed the study's conclusions without seeing the source manuscript, predicted withheld perturbation effects, and identified perturbation pairs with similar or opposite outcomes at different effect strengths.
- The distributed setup retrieved hundreds of times more references than the single-agent version, according to the abstract.
- Blinded human experts reportedly preferred the distributed-agent interpretations and supported the ranking of the discoveries.
The paper described a distributed biological reasoning approach that assigned many AI agents to different parts of an experiment and combined their outputs. It said the method was used on a whole-brain Perturb-seq atlas.
According to the abstract, thousands of agents reviewed the dataset in minutes. The authors contrasted that with an estimate that human analysis would have taken months.
The paper said the agents reconstructed the study's conclusions without access to the manuscript. It also said they predicted withheld perturbation effects and found relationships between perturbations with matching or opposing outcomes at different effect strengths.
The authors said the distributed mode found hundreds of times more references than a single-agent mode. They also said blinded human experts preferred the distributed interpretations and backed the discovery rankings.
The competing interests statement said several authors worked for OpenAI and PerturbAI. It also said two PerturbAI authors were co-founders of that company.
Named in this story
Companies
- OpenAI
- employed several authors named in the competing interests statement
- PerturbAI
- employed authors named in the competing interests statement and had co-founders among them
How the source tells it
Technical abstract with a promotional emphasis on scale and performance gains.
- boosterism performance is framed through comparisons with human analysis and a single-agent baseline