What they did
Citrini Research, an investment research firm with no biology background, approached Anthropic with a question: could Claude help motivated non-experts make progress on a real health problem. They settled on scorpion neurotoxins, on the idea that a protein that binds the toxin could one day lead to an antivenom. Per Anthropic's write-up, an internal research model then worked mostly on its own for 64 hours with four follow-up prompts from the team. It generated 318,484 candidate protein designs and selected 180 for the lab. The team added 31 more with Claude's help, and Adaptyv Bio and Twist Bioscience synthesised and tested all 211.
How it went
Seventeen of the 211 designs met Anthropic's criteria for a binder, and the two strongest bind tightly to Cn2, a toxin from a Mexican scorpion. Anthropic says it has not shown that any of them neutralise venom in a cellular environment and does not claim to have made an antivenom. All 211 designs and the data are being released. The write-up does not split the 17 hits between the model's own 180 and the 31 the team added, which a commenter on the thread asked for and the poster could not find either.
Worth knowing
The account comes from the vendor, and the firm that commissioned it sells investment research on AI, so read the framing with that in mind. The Verge spoke to outside scientists who said the result is a set of starting points and nothing close to a drug. Binding is not neutralising: the test that matters has not been run. Commenters also noted the wet lab, not the model, was the bottleneck, and that the token spend behind 318,484 designs is not something a hobbyist reproduces.