Major progress on AI scientists this week in Nature

Nature published three AI scientist systems on the same day this week: FutureHouse's Robin, and Google's Co-Scientist and ERA.

The headline is (again) agentic discovery. The detail that matters for pharma is what they run on.

Read the results and the same thing holds across all three. Robin proposed ripasudil for dry AMD and analysed the follow-up biology, while humans ran the assays. Co-Scientist surfaced AML repurposing candidates like KIRA6, then human experts chose what to test. ERA wrote software that beat the field on single-cell analysis and public-health forecasting. Every one of those results came from recombining knowledge that is already public: open literature, public datasets, open benchmarks.

The reasoning is extraordinary, and the inputs are open to everyone. Whatever a public-data engine surfaces for you, the same engine surfaces for your competitor.

Peer review commoditised the reasoning this week. It changed nothing about the data.

In pharma, the edge sits in what these systems cannot reach: the proprietary trials and molecules behind your firewall, and the lab throughput to keep up with what they generate.

That is the part you cannot rent. Now, if only pharma data were clean and immediately AI-usable. But that's a different story.