OpenAI 用 1946 个 AI 智能体 32 分钟完成小鼠全脑 CRISPR 图谱解读
OpenAI 把一个巨难的脑实验拆给 1946 个智能体，32 分钟干完要几个月的活，还盲评赢了 PhD，拆分思路值得看。
OpenAI 与 PerturbAI、Scripps Research 合作，将一个全脑小鼠 CRISPR 图谱拆分给 1946 个 GPT-Rosalind-5.5 智能体，每个智能体负责分析单个基因敲除的影响并检索文献，全程用时 32 分钟。该图谱论文发表于模型 2025 年 12 月知识截止之后，智能体在未看过论文的情况下仍匹配或细化了 55 项已发表结论中数据可支持的 48 项。对照组中，单个智能体分析 893 个基因仅引用 11 篇论文，拆分后的智能体引用量是其 711 倍。盲评中，PhD 评审在 75% 的情况下更倾向于拆分智能体的报告。
OpenAI showed that AI agents can handle the slowest step in modern biology, making sense of huge experiments.
By splitting a huge brain experiment among 1,946 AI agents, OpenAI got in 32 minutes what it estimates would take people months: a ranked list of leads for new brain-disease research.
So biology labs can now produce experiments far bigger than their own experts can fully interpret
OpenAI, PerturbAI and Scripps Research had 1,946 GPT-Rosalind-5.5 agents interpret a whole-brain mouse CRISPR atlas in 32 minutes.
Each agent analyzed a single knockout’s effects and searched the literature, without seeing the atlas paper, which appeared after the model’s December 2025 knowledge cutoff.
They still matched or refined 48 of the 55 published findings the expression data could support.
The main advance is the method: split 1 giant analysis into nearly 2K small ones, 1 gene per agent, and run them all at once. A single agent given 893 genes cited just 11 papers, while the split-up agents cited 711 times as many, and blinded PhD reviewers preferred their reports 75% of the time.
