A rice blast experiment in Anthropic's Fable 5 safety report shows who AI can't replace
Fable 5 的安全报告里埋着一个稻瘟病实验,暴露了谁才是绕不过去的人
Anthropic ran a rice blast experiment in Fable 5's safety report: six biology PhDs paired with LLM experts used Claude Mythos 5 to design an agricultural pathogen defense in 16 hours. Two generalist teams beat all specialist teams—work the experts estimated would take 2–3 months manually. AI matched experts at literature search and cross-domain synthesis but repeatedly failed at judging whether an answer was correct or when to stop. The one variable never controlled for was the LLM expert—someone who knows the model fabricates citations, overestimates feasibility, and won't self-correct, and who stays in the loop to calibrate every output. Anthropic's conclusion that Fable 5 hasn't crossed the bioweapons risk threshold hinges on the assumption that ordinary users lack this person.
Why it matters: Anthropic had six biology PhDs compete against its general team using Claude to design a rice blast defense; the general team won in 16 hours.