Novo Nordisk hands Anthropic its drug discovery pipeline
A drugmaker with a $15 million-a-day incentive to move faster just bought into frontier models, and Google published the paper that shows where the self-improvement loop actually lives.
Novo Nordisk will test Anthropic's models and its Claude Science workbench inside its own research and development workflows. The Danish maker of Wegovy announced the collaboration on Wednesday, saying the two companies will first pick the scientific problems where their combined capabilities can matter most, and that Novo will also use Anthropic's frontier models to strengthen its AI-driven software development.
The framing from both sides is deliberately unglamorous. Novo CEO Mike Doustdar said the partnership extends an ambition to become "the world's most AI-driven healthcare company," and that beyond R&D productivity gains, AI tools can open "completely new scientific opportunities" and help reasoning about human biology and drug mechanics. Anthropic CEO Dario Amodei framed it as access: giving leading researchers trusted frontier models can shorten research timelines and produce medicines faster.
The interesting part is that Novo is not starting from zero. Anthropic's own case study describes NovoScribe, a documentation platform Novo built with Claude Code and MongoDB Atlas that produces regulatory-grade content — clinical study reports, device verification protocols, patient materials — in minutes rather than the months the old manual process took. Novo says review cycles have been cut in half and that a development team of 11 people now has non-programmers prototyping their own tools. The company says the output has drawn positive feedback from regulators, which is the only validation that counts in a heavily scrutinized industry.
Read that as the real pitch for the deal. Pharma's bottleneck is not idea generation; it is the years of document production, verification and submission between a promising molecule and a patient. Novo is buying time in exactly that stretch, and it is buying the ability for domain scientists to build the tooling themselves instead of filing tickets. The wider industry is following the same track — Bristol Myers Squibb has put Claude in front of tens of thousands of employees, and pharma R&D is now the sector where frontier-lab partnerships get signed fastest. We looked at the other side of this last week — More molecules, same failures: the crack in AI drug design — because faster documents do not fix a validity problem in the science underneath them.
Google published a technical report on recursive self-improvement that turns an agent's own history into a simulator it can practise in. Dream-RSI, from researchers at Google, Google DeepMind, the University of Maryland and the University of Virginia, leaves the underlying coding agent untouched and adds a thin orchestration layer that makes exploration explicit. The trick is what the layer does with past runs: a completed discovery process is a tree of decisions and their real execution outcomes, so an alternative exploration strategy can be replayed against that recorded tree instead of being run online. Candidate policies get drafted, scored against the replay, and the winner is redeployed to collect a new tree — the pool grows every lap, and that is where the recursion sits. The authors call it dreaming in a world built from your own history.
The numbers are the story. Across eight discovery tasks in algorithm engineering, mathematical optimization and GPU kernel engineering, the report says Dream-RSI beat standard Lasso solvers such as sklearn and glmnet while using up to 162 times fewer discovery-agent calls than SimpleTES — a system that used 51,200 generations to reach a comparable result, against fewer than 1,000 here. On the math benchmarks it reports 1.145427 on sum-difference, 2.635983 on circle packing and 1.456375 on autocorrelation, matching or beating strong automated-discovery systems at roughly a fiftieth of the budget. On KernelBench it either hits target execution speeds with 1.79 to 2.43 times fewer generations, or improves kernel performance by up to 2.09 times under the same budget.
Two things make this more than a leaderboard entry. First, almost all of the security debate about recursive self-improvement this month has been about models improving their own weights; Dream-RSI's recursion never touches weights at all — it improves the search strategy that decides what the agent tries next. Second, the efficiency claim is a safety-adjacent claim: the report's premise is that better exploration saves enormous compute, which means the same capability arrives at a fraction of the cost that labs budgeted. That is not a doomsday scenario and it is not a reassuring one either. We covered the political fight around this in September — Google won't confirm the RSI rumor. Its chief scientist just backed a slowdown.
What to watch: whether Novo publishes outcomes from the Claude Science pilots — the timeline improvements are the claim that matters, not the model access — and whether Dream-RSI's replay trick survives contact with tasks where outcomes cannot be re-derived from a recorded tree.
If the self-improvement loop that ships first turns out to be the cheap exploration layer rather than the model itself, does that change how you think about the risk? Tell us in the comments.
Sources: Dream-RSI (GitHub)
