AI picks the molecule. The molecule still has to survive the human.

Anthropic announced they’re entering drug discovery, launching Claude Science. They’ll develop drugs internally and sell tools to pharma. The industry pushback is that the real challenge is later — proving drugs work in humans.

I built a research pipeline last year that picked a niche from 177K Dutch e-commerce domains. V1 picked one. The shop died in operations the pipeline couldn’t see. I wrote about it in A research pipeline for the internet-shaped part of the decision. The pattern is similar: AI can optimise for a signal (niche attractiveness, molecular binding) but the bottleneck is the part the model didn’t train on. In drug discovery, that’s clinical trials. You can’t shorten the 6-10 year timeline by throwing more compute at it. Biology doesn’t batch process.

What I’d watch next: whether Anthropic actually funds internal Phase 1 trials. If they just announce a bunch of AI-generated candidates without committing to the clinical spend, it’s a tools play, not a pharma play. The difference determines who eats the failure cost. Also, their partnership strategy — if they align with CROs or large pharma, they might actually navigate the operational mess. If they try to go it alone, I’d bet on the mess.