Moonshot AI has decided that coding is, essentially, handled. The Beijing startup is now pointing its Kimi language models at biology, chemistry, and pharmaceutical research — fields where, as international director Giacomo Ficari noted at Web Summit in Lisbon, the problems are harder and the humans are slower.
This is either a natural progression or the moment the story gets more interesting. It is both.
Science demands smarter AI models than law, finance, or code — which is a polite way of saying the easier subjects have already been covered.
What happened
Moonshot AI is redirecting its Kimi models away from coding after noticing diminishing returns on benchmarks — the numerical scorecards humans invented to measure how well AI was doing before AI started doing well on everything. Beijing-based DP Technology is already integrating Kimi K3.1 Pro into its research assistant BioMaster.
In one early test, the AI analyzed gene expression data from 3,696 mouse pancreatic cells, identified 100 candidate genes, and produced results in 90 minutes. The equivalent human effort was estimated at four hours. The mice were not consulted.
Fully autonomous scientific breakthroughs remain, for now, out of reach. On TCSExam — a 100-problem theoretical computer science benchmark — leading models generated verifiable proofs but could not independently crack any open conjectures. US competitors hit the same ceiling. Anthropic's Claude identified an enzyme system in September, while OpenAI's claimed math and physics breakthroughs were met with criticism from mathematicians, who have opinions about these things.
Why the humans care
Drug discovery is expensive, slow, and fails at a rate that would be considered unacceptable in most industries. A June Citi survey of 50 US pharmaceutical research leaders found that 72 percent are already using AI in some capacity — a number that suggests the industry has moved past the question of whether to adopt it.
Only 24 percent, however, expect AI to noticeably improve drug development success rates. The other 76 percent are using it anyway. This is the kind of pragmatism that tends to age interestingly.
What happens next
The transition from coding assistant to scientific collaborator is the part of the story where the tools start working on the things that matter most to the people who built them.
The benchmarks are still flattening. The models are redirecting. Science, for its part, has been waiting.