AI研究非营利组织METR的一份新报告称,Anthropic内部对AI的使用可能已使该公司的AI开发速度翻倍。
Anthropic’s R&D Speed May Have Doubled Due to AI, Researchers Find AI use within Anthropic may have doubled the company’s pace of AI development, a new report from AI research nonprofit METR says.
报告指出,截至今年10月,Anthropic模型的改进速度大约是2024年3月至2025年10月期间改进速度的两倍。METR估计,Anthropic研究速度加快的约80%归因于公司内部对AI的使用。
The rate that Anthropic’s models are improving as of this October is roughly twice the rate at which they improved between March 2024 and October 2025, according to the report. METR estimated that around 80% of the acceleration in Anthropic’s research pace is due to internal use of AI within Anthropic.
这份报告发布于CEO达里奥·阿莫代伊承诺引入包括METR在内的第三方研究人员作为嵌入式评估员、给予其“类员工权限”以调查安全实践遵守情况之后。尽管这份特定报告独立于嵌入式评估合作伙伴关系,但METR表示,为撰写报告获得的内部信息访问权限“超出了我们的预期”。METR称,发现AI研究在Anthropic内部尚无法完全自动化,但存在“生命迹象”,表明AI很快可能带来进一步的研发加速。
The report comes after CEO Dario Amodei committed to bring third-party researchers, including METR, into Anthropic as embedded evaluators with “employee-like access” to investigate adherence to safety practices. Although this particular report was separate from the embedded evaluation partnership, METR said it was given access to internal information for the report that “surpassed our expectations.”METR said it found AI research could not yet be fully automated within Anthropic, but there were “signs of life” that it was possible for further R&D acceleration to take place from AI soon.
METR报告称,这最终可能形成一个循环,即AI持续自我改进变得越来越强大,潜在导致“极端且危险的能力”。
That could eventually lead to a loop where AI continuously improves itself to become more and more capable, potentially leading to “extreme and risky capabilities,” the METR report said.
报告还分析了Anthropic自2026年7月以来训练的三个未发布模型。报告引用的最新模型于9月或10月完成训练,与以往模型相比,在Anthropic用于衡量AI能力的一套测试中,基本延续了现有的改进趋势。
The report also analyses three unreleased models Anthropic has trained since July 2026. The most recent model cited in the report finished training in September or October, and largely continues the existing trend of improvement compared to previous models on a suite of tests Anthropic uses to measure AI capabilities.
Anthropic向METR员工提供了模型训练数据、未发布模型的基准测试分数以及内部分析。METR员工还采访了约15名Anthropic员工,以形成研发加速估算。
Anthropic gave METR staff access to data on how models were trained, benchmark scores for unreleased models, and internal analyses for the report. METR staff also conducted interviews with around 15 Anthropic employees to form an R&D acceleration estimate.