Weekly Arc: Anthropic now runs its own biology lab
Last week's issue put Novo Nordisk at the center of this thread: a top-five pharmaceutical company placing drug discovery work on Anthropic's Claude, which made a frontier vendor's model a production dependency inside a large pipeline. The open question was what the vendor gets from sitting that deep in someone else's research. This week answered it in a way the arc had not anticipated. Anthropic reported the first results from its own life sciences research lab: Claude agents found a novel enzyme system, and nobody yet knows what it does.
The supplier started doing the science
The finding itself is modest. An enzyme system with no assigned function is a lead, not a result, and it arrived company-reported with no outside group confirming it. What changed is the position of the company that announced it. For two years the working assumption in this field has been a division of labor: AI companies build models, biology organizations use them to find things. Anthropic now sits on both sides of that line. It sells Claude to Novo Nordisk for discovery work, and it runs discovery work of its own with the same model. A pharmaceutical company weighing whether to standardize on a vendor's model has to price in that the vendor is also, in a small way, a competitor in the same search space.
Buying the part that cannot be downloaded
The second move that week points at the bottleneck. Anthropic and Adaptyv launched a protein design competition whose entries get expressed and assayed rather than scored in a computer, sending more than 5,000 designs to the bench. Adaptyv runs the protein characterization service; the competition is a way of buying a large block of real measurement and attaching it to model output. The reason this matters was argued directly by Polyphron, which is building a tissue foundry on the premise that physical throughput sets the ceiling on how fast biological models improve. Coding models got better quickly because a machine can generate code, test it, score it and try again without a person in the loop. Biology's version of that loop runs at the speed of wet-lab capacity. Weights can be copied and compute can be rented, but measurement throughput has to be built, and it is now something a model company is willing to pay for. The counter-position appeared the same week in WetRobo, a lab robot released as a buildable kit that coding agents can drive. Most agent-in-the-lab demonstrations have run on custom automation nobody outside the group can reconstruct. A reproducible kit is the version an ordinary university lab can own.
Who pays for the check
The third move closes the circuit. Anthropic and Accenture announced a partnership on independent evaluation of frontier models, with both sides expecting to invest at least $1 billion over five years to build the capacity. External evaluation becoming a real industry is what this column has been asking for since July, when the recurring complaint was that safety records exist but only the company that produced them can read them. The complication is the funding direction. Evaluation at that scale needs capital, and the organizations with the capital are the ones whose models get evaluated. Other fields have run this arrangement before. Credit rating agencies are paid by the issuers they rate, and most clinical trials are funded by the company that makes the drug. Neither arrangement is fraudulent and both required decades of separate rules about disclosure, independent monitoring and registration before their outputs could be relied on. Biomedical AI has none of that yet, and the money is already flowing.
What a lab still owns
Put the three moves together and the arc's question has shifted. It was about access terms: what happens to a lab when a model gets retired, repriced, or moved behind a verification program. It is now about whether any independent party holds a piece of the loop at all, because one company this week supplied the model, ran the experiments, bought the assay capacity, and co-funded the evaluation of its own systems. The answer readers reached for most was the specialist one. The week's most-clicked story was Inductive Bio's Beacon-2, which predicts human dose directly from chemical structure, launched with the argument that most AI drug-discovery systems optimize endpoints that do not decide whether a compound reaches a patient. That is how a smaller company stays load-bearing: own an endpoint the generalist does not measure, and own the data that defines it. The open question for the weeks ahead is whether Anthropic's enzyme finding gets confirmed by a group that did not run the agents, and whether anyone outside the company can obtain enough of the experimental record to try.
- Agents as bench scientists: thousands of agents now staff a simulated biotech, PRAXIS-VirtualCell adds a specification layer over agent-run experiments on cell models, and Cell published a single-cell dataset an agent assembled; still no headline result confirmed by a group that did not run the agents.
- Benchmarks as audits: AI systems now sit inside coverage decisions at UnitedHealth, Cigna and Humana while the tests keep arriving after the deployment, from MentalHealthBench to TopU-LBVS; watch whether the audit method that falsifies a cell model's claimed inputs gets run on a model somebody is actually using.
- The open stack and its liabilities: an identity check at the door binds only at the moment of access, which is the problem posed by research agents that rewrite the code running their own experiments and MedRSI's self-evolving clinical agent; meanwhile watermarking output changes what agents do downstream, so the record costs something to keep.
Reply with what you're seeing. A human reads them. Forward freely.
|