Where frontier model developers once poured computational budgets into code generation, Beijing startup Moonshot AI is now redirecting its Kimi language models toward biology, chemistry, and pharma research after encountering diminishing returns in coding.
At the Web Summit in Lisbon, Moonshot AI international director Giacomo Ficari stated that science demands smarter artificial intelligence models than law, finance, or code. As performance gains stall on standard developer benchmarks, foundation model providers are forced to look beyond software engineering to justify continuous infrastructure spending.
Screening Workflows in Laboratory Pipelines
Beijing startup DP Technology is integrating Kimi K3.1 Pro into its research assistant BioMaster to test commercial deployment in laboratory workflows. During one test, the artificial intelligence analyzed gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes in 90 minutes instead of an estimated four hours of human work.
Science demands smarter artificial intelligence models than law, finance, or code.
Yet mechanical efficiency in processing cells has not translated into independent scientific discovery across the sector. On TCSExam, a 100-problem theoretical computer science benchmark, leading models produced verifiable proofs but could not crack any open conjectures alone. United States developers face identical constraints: while Anthropic's Claude found an enzyme system in September, mathematicians criticized OpenAI's claimed math and physics breakthroughs.
Commercial Adoption Meets Enterprise Skepticism
This gap between rapid data sorting and genuine autonomous insight defines corporate procurement in life sciences. A June Citi survey of 50 United States research leaders found that 72 percent were already using artificial intelligence broadly, but only 24 percent expected artificial intelligence to noticeably improve drug development success rates.
Enterprise buyers paying for inference capacity face the structural risk of deploying expensive compute clusters on architectures that accelerate routine screening while remaining incapable of solving open scientific problems.
Against the field
Cutting the screening time for mouse pancreatic cells to 90 minutes highlights the current ceiling of foundation models in hard science: systems take over routine labor rather than make independent discoveries. Across 100 problems on the TCSExam benchmark, leading models failed to solve a single open hypothesis, while mathematicians sharply challenged claims by OpenAI regarding scientific breakthroughs. Foundation model builders are pitching automated grunt work to the biopharma sector as autonomous scientific discovery.