Coding Scores Stalled, So Kimi Put On a Lab Coat
Moonshot AI has run into the problem every overachiever eventually meets: the score stopped going up. This week the Beijing lab confirmed it is steering its Kimi models away from coding and toward biology, chemistry, and pharma research. The stated reason is diminishing returns on coding. The unstated reason, if you read the room, is that when you cannot climb a leaderboard anymore, the smart career move is to find a brand new leaderboard and act like you always meant to be there.
Giacomo Ficari, Moonshot's international director, made the case on stage at Web Summit in Lisbon. Science, he said, demands smarter models than law, finance, or code. That is a bold thing to announce in a conference hall full of people who specifically bought tickets to hear about code. It is also defensible: a model that can autocomplete a React component is not automatically a model that can reason about a gene regulatory network.
Still, the genre shift is comedic. One quarter your model is grinding away at software engineering benchmarks, sweating over unit tests. The next quarter it is wearing a lab coat, squinting at mouse cells, and saying the word "hypothesis" with confidence. Same weights. Same tokenizer. Entirely new press release.
The showcase Moonshot is pushing comes from DP Technology, which has wired Kimi K3.1 Pro into its research assistant BioMaster. In one test, the system chewed through gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes in 90 minutes. The human estimate for the same job was around four hours.
So the model saved about two and a half hours. That is a genuine win, and it is also the funniest possible number to put in a paragraph about drug discovery, a field where a single therapy can take a decade or more to reach a pharmacy shelf. Two and a half hours is a long lunch. It is a meeting that should have been an email. Framed next to a ten year pipeline, the headline basically reads: pancreas breakthrough arrives slightly before dinner.
To be fair to Moonshot, the wall clock is not the point. The point is that 100 candidates came out the other end. A researcher who previously had to triage a mountain of expression data now gets a shortlist to argue about, which is a real change in how the day goes. Anyone who has done bioinformatics by hand knows that the hard part is not the clicking. It is the deciding.
It is also exactly where the hype grammar gets slippery. "Flagged 100 candidate genes" sounds like a finding. It is not a finding. It is a list of suspects, and every suspect still needs a detective. The gap between a candidate list and a validated target is where most drug programs go to die, quietly, over several years, with no press release.
The new scoreboard grades things nobody can grade yet
Moonshot's own evidence shows where the line sits. On TCSExam, a 100-problem theoretical computer science benchmark, leading models produced verifiable proofs but could not crack a single open conjecture on their own. Translation: hand the model homework with an answer key and it aces the class. Hand it the ungraded homework, the stuff mathematicians actually argue about at whiteboards, and the confidence evaporates.
That pattern is industry-wide, not a Kimi quirk. Anthropic's Claude found an enzyme system in September, which is a real result worth noting. OpenAI's claimed math and physics breakthroughs got publicly criticized by mathematicians, which is also a real result worth noting. The models are superb at producing research-shaped objects. They are inconsistent at producing knowledge.
Then there is the part Moonshot did not put in the headline. A June Citi survey of 50 US research leaders found 72 percent already use AI broadly, but only 24 percent expect it to noticeably improve drug development success rates. That is the industry saying, very politely, that the demo is impressive and the pipeline is unchanged. A 3x jump in a benchmark is a slide. A 3x jump in approvals is a decade of clinical trials and a lot of paperwork.
None of this makes the pivot cynical. Science is a reasonable place for a model to go next, and biology has more room for improvement than another round of code completion. But it is worth naming the convenience: a science benchmark updates in years, not on a Tuesday afternoon. Nobody reruns your biology result over the weekend with a fresh harness and posts the diff to Hacker News.
So enjoy the lab coat. Kimi might genuinely help someone find a real target, and the mouse cell run is a decent start. Just remember that the reason coding benchmarks got abandoned is that everyone could check them. The new scoreboard is friendlier partly because it takes longer to argue with. That is not a benchmark problem. That is a marketing one, wearing safety goggles.
Comments