METR 报告：Anthropic 模型改进速度约为 2024 至 2025 年节奏的两倍
So much revelation from METR's latest report.
METR 在花 11 周查阅 Anthropic 内部研发数据并访谈其核心研究员后发布报告，称 Anthropic 目前改进模型的速度约为 2024 至 2025 年节奏的两倍，递归自我改进已从理论走向可测量。METR 估计当前加速中约 80% 来自 Anthropic 自身对 AI 的使用，并排除了训练算力增长更快这一原因；报告还认为未来 12 个月内算法、算力或数据都不构成瓶颈。报告指出 Anthropic 8 月的公开风险报告低估了实际推进速度，且 Anthropic 以可能帮助竞争对手为由隐去了 METR 认为最有说服力的多数证据，并中途取消了一项关于算力在能力与安全之间如何分配的检查。
推荐理由
精选AI 评分 80
METR 用 11 周内部数据把递归自我改进从理论变成可测数字，并给出加速来源的归因。
主题
查看原帖自8月以来，METR内部的一个团队利用Anthropic的内部信息，评估AI本身是否正在加速AI的发展。 我们现在发布一份报告，总结此次风险评估的结果。
查看原文
Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development. We are now publishing a report summarizing the results of this risk assessment.
正文
So much revelation from METR's latest report.
Recursive self-improvement (RSI) at Anthropic has moved from theory to solid measurement, and the early loop already runs at about 2x.
After they spent 11 weeks going through Anthropic’s internal R&D data and interviewing its key researchers.
> Anthropic is now improving its models at about twice its 2024 to 2025 pace.
> Claude itself is the main engine. METR’s best guess puts about 80% of current acceleration on Anthropic’s own AI use, and it rules out faster growth in training compute as the cause.
> Anthropic’s own public August risk report understated how fast things were moving. METR reached that conclusion in a separate review that compared the public report with the private version and other nonpublic data.
> METR saw nothing likely to slow this loop within 12 months, with no bottleneck in algorithms, compute or data. Redacted evidence also hints that acceleration may already exceed 2x and keep climbing.
> Staff self-report roughly 12x productivity gains from Model 4, and researchers see “signs of life” for AI doing key training work alone. One called Mythos Preview the first model useful for autonomous experiments.
> Anthropic redacted most of the evidence METR found most telling, saying it could help rivals go faster. It also canceled midway a planned check of how its compute splits between capabilities and safety.