月之暗面将 Kimi 模型从编程转向科学研究,编程基准收益趋缓
Moonshot AI pivots Kimi models from coding to science as benchmark gains flatten
- 1月之暗面认为编程领域收益递减,将 Kimi 模型转向生物、化学和制药研究
- 2Kimi K3.1 Pro 进入 DP Technology 的研究助手 BioMaster,测试中 90 分钟标记 100 个候选基因,人工估计需 4 小时
- 3领先模型在 100 题理论计算机基准 TCSExam 上能产出可验证证明,但无法独立解开任何公开猜想
- 4Citi 6 月调查 50 名美国研究负责人,72% 已广泛使用 AI,仅 24% 预期能明显提升药物研发成功率
大模型厂商在编程赛道收益趋缓后转向科研场景寻找差异化,但目前的能力定位是提速助手,离独立科研突破仍有距离。
- 影响:生物和制药企业多了可量化提速的研究助手选项,90 分钟对 4 小时的对比提供了测算基准;药研负责人仅 24% 预期提升成功率,提示采购应聚焦具体环节而非押注整体成功率;模型厂商涌入垂直科研领域,此类供给会增多。
- 参考:可考虑让研发团队先算出基因分析或文献筛选等环节的实际耗时,再与 BioMaster 这类助手做对照试点;值得把 72% 已用 AI 与 24% 看好结果的落差纳入内部预期管理。
- 依据:判断依据为 Web Summit 上月之暗面国际总监的表述、DP Technology 公布的一次测试数据,以及 6 月 Citi 对 50 名美国研究负责人的调查,测试数据暂无独立验证。
Because Moonshot AI sees diminishing returns in coding, the Beijing startup is redirecting its Kimi language models toward biology, chemistry, and pharma research. International director Giacomo Ficari said at the Web Summit in Lisbon that science demands smarter AI models than law, finance, or code.
Beijing startup DP Technology is integrating Kimi K3.1 Pro into its research assistant BioMaster. In one test, the AI 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.
Fully independent breakthroughs are still a long way off. On TCSExam, a 100-problem theoretical CS benchmark, leading models produced verifiable proofs but couldn't crack any open conjectures alone. US rivals hit similar limits. Anthropic's Claude found an enzyme system in September, but mathematicians criticized OpenAI's claimed math and physics breakthroughs. The pharma industry isn't sold yet either. A June Citi survey of 50 US research leaders found 72 percent already using AI broadly, but only 24 percent expect it to noticeably improve drug development success rates.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
来源:The Decoder:AI News · the-decoder.com