Claude runs a protein design project almost on its own

What happened: A research group (named on air as Trini Research) worked with Anthropic on an internal Claude model that designed 318,484 proteins from one prompt and four follow-ups, in about 64 hours of design work. Outside labs tested 211 designs, and 17 latched onto their target scorpion toxin. Nobody has shown these proteins actually stop venom.
The claim: Anthropic says Claude picked the approach and chose most of what got tested. It says it does not claim an antivenom.
Austin's take: Austin says the real breakthrough is second order: AI running its own experiments could speed up science and medicine. He points to Amodei saying he has held the team back in biology for safety, while admitting we don't know exactly what that caution means in practice.
Why it matters: This is AI doing the science, not just assisting, which could compress years of research. It also raises the stakes for biology safety guardrails.
What to do: Read claims like this carefully: binding to a toxin in lab tests is not a working treatment. Watch for follow-up results before treating it as a medical advance.
Not confirmed yet: No one has shown the proteins neutralize venom, and Anthropic explicitly does not claim an antivenom.

The Loop