METR Finds Anthropic AI R&D Accelerating 1.9x; Redacted Report Sparks Backlash

METR 发布对 Anthropic「AI 是否正在加速 AI 研发」的独立评估报告：团队自 8 月起花 11 周研读 Anthropic 内部研发数据并访谈核心研究员，估计截至 2026 年 10 月其模型能力增速约为 2024 年 3 月—2025 年 10 月期间的 1.9 倍（95% 置信区间 1.5×–2.8×），相当于把约两年的历史进展压缩到一年。报告引发 AI 社区大规模预期更新，但公开版本遭 Anthropic 大幅删减，引发透明度争议。

已确认

• METR 报告核心结论：Anthropic 内部前沿模型的能力增长快于公开前沿的历史趋势，加速确实发生。约 80% 的加速归因于 AI 自身（@rohanpaulai 转述）。AECI 趋势线从每年 +15 分升至 +29 分（@zephyrz9）。内部未发布模型「Model 4」2026 年 10 月起内部可用，AECI 得 178.6 分（@scaling01、@kimmonismus）。预训练周期据称缩至 45–60 天（@zephyrz9）。
• 报告流程：METR 自 8 月起基于 Anthropic 内部信息评估，历时 11 周，评估框架围绕三问：加速是否发生、是否可归因于 AI、风险如何（@METREvals、@rohanpaulai）。
• 删减事实：METR 提交约 100 页完整报告，公开版中 Anthropic 删减了 METR 认为最有信息量的多数证据；Anthropic 称目的是移除「可能加速其他 AI 开发商」的机密信息（@tszzl）。METR 另获阅非公开版，复审了 Anthropic 2026 年 8 月风险报告的「自动化 R&D 风险」章节（@DKokotajlo）。
• Anthropic 联合创始人回应：承认删减让他人更难核查 METR 的工作，但称公开加速的驱动因素「并不合适」，理由涉及商业敏感与安全（@GaryMarcus）。
• 辅助证据：Anthropic 员工自报效率提升图表随时间超指数增长，从 2025 年初的 3.7 倍升至 Model 4 的约 12 倍，但 METR 强调这仅是 Slack 投票等非正式调查，只能作宽松证据（@scaling01）。

尚未确认

• 具体驱动加速的机制与细节因删减未公开，外部无法核查（社区多位作者对此表达不满，@GaryMarcus、@tszzl）。
• 员工自报 12 倍效率提升缺乏严格度量，仅为参考性数据（@scaling01）。

为什么重要

• 这是首次将「递归自我改进」从理论命题变成可测量对象，多位原持怀疑态度的人士大幅提前 RSI/ASI 时间表预期（@AndrewCurran 称多人改口认为 RSI 不足 6 个月，社区反应被类比为 2022 年 ChatGPT 时刻）。
• @scaling01 转发的观点指出进展正大致沿 AI 2027 报告预测推进（该场景设定 2026 年 12 月前后研发加速约 2 倍），若延续则 ASI 将于 2027 年发生。
• 政策影响开始显现：纽约州众议员 Alex Bores 引用报告呼吁国会今年禁止递归自我改进（@DavidSKrueger）。投资人也公开「认错」并更新时间表（@GaryMarcus 提及 Jake Mendel）。
• 透明度争议成为第二焦点：@GaryMarcus 批评删减辩护，@AndrewCurran 向安全阵营喊话称依赖一年以上窗口期或友好政策的计划已无意义。

2026-11-06 ~ 2026-11-07 · 94 related posts

Full story(10 episodes)→

• Episode 1: Anthropic Says It May Imminently Hit RSP Threshold for AI-Accelerated R&D(2026-11-03, 7 posts)
• Episode 2: AI 2027 Authors Warn Anthropic's 2x Threshold Could Accelerate Superintelligence Timeline(2026-11-05, 2 posts)
• Episode 3: METR Finds Anthropic AI R&D Accelerating 1.9x; Redacted Report Sparks Backlash(2026-11-06, 94 posts)
• Episode 4: Anthropic's Internal Models Reportedly Surpass Mythos 5(2026-11-06, 2 posts)
• Episode 5: Report: Anthropic's Next Model Finished Training, Tops All METR Benchmarks(2026-11-06, 2 posts)
• Episode 6: Insider: 7 of Anthropic's Last 8 Frontier Models Never Publicly Benchmarked(2026-11-06, 2 posts)
• Episode 7: METR Says Anthropic Canceled Compute Allocation Assessment Pilot, Sparking Criticism(2026-11-06, 4 posts)
• Episode 8: AI Safety Community Criticizes Anthropic's Acceleration Strategy(2026-11-06, 2 posts)
• Episode 9: Oxford's Toby Ord Argues METR's 2x RSI Speedup Drops to 1.4x Once Pretraining Jumps Are Counted(2026-11-06, 11 posts)
• Episode 10: Oxford philosopher explains why AI benchmark error bars have widened(2026-11-06, 2 posts)

• [source] METR launches AI R&D acceleration assessment: are capabilities outrunning historical trends? — METR_Evals · 2026-11-06
• Anthropic shared extensive nonpublic AI R&D data with METR, incl. algorithmic progress estimates — METR_Evals · 2026-11-06
• METR: capability acceleration not driven by faster training compute growth — METR_Evals · 2026-11-06
• METR publishes risk assessment on whether AI is accelerating AI R&D — RyanGreenblatt · 2026-11-06
• METR publishes AI-accelerating-AI risk assessment; Anthropic redacted most key evidence — Miles_Brundage · 2026-11-06
• [source] METR says Anthropic redacted most key evidence from its 100-page safety assessment report — tszzl · 2026-11-06
• METR: Anthropic models' capability growth hits 1.9x the earlier rate — dfrsrchtwts · 2026-11-06
• METR finds superexponential fit as AI drives current R&D acceleration, staff self-report gains — ricklamers · 2026-11-06
• METR estimates AI is already accelerating AI development by roughly 2x — harris_edouard · 2026-11-06
• Anthropic Cofounder Sam McCandlish Responds to METR Report on AI Accelerating AI Development — edwinarbus · 2026-11-06
• METR: Anthropic likely hit its 2x RSP acceleration threshold, maybe 3x, ~80% AI-driven — daveholtz · 2026-11-06
• METR finds Anthropic packing nearly 2 years of AI progress into 1 year, no clear bottlenecks — 141_1337 · 2026-11-06
• METR report reveals Anthropic's internal Model 4: trained Oct 2026, scores 178.6 on ECI — legit_api · 2026-11-06
• METR chart shows AI benchmark progress nearly doubled, Anthropic Model 4 hits 178.6 AECI — scaling01 · 2026-11-06
• METR says Anthropic redacted most of the key evidence from its safety assessment — YashasGunderia · 2026-11-06
• METR chart shows unreleased Anthropic "Model 4" at ~12x self-reported coding speedup — YashasGunderia · 2026-11-06
• METR: Anthropic redacted most key evidence from public version of 100-page safety audit — Miles_Brundage · 2026-11-06
• METR report: AI is accelerating AI capability gains inside Anthropic — mattbeane · 2026-11-06
• METR report: RSI already accelerating AI progress ~2x, matching AI 2027's March 2027 mark — RyanGreenblatt · 2026-11-06
• METR: Anthropic's new internal Model 4 tops all benchmarks, may trigger 3x scaling threshold — scottleibrand · 2026-11-06
• 'Recursive Self-Improvement Likely Began October 2025,' Citing METR Data — ben_j_todd · 2026-11-06
• METR Backs Anthropic's ~2x R&D Acceleration Claim, Sees ~80% AI-Driven, ASI by 2028 — ben_j_todd · 2026-11-06
• Report: AI is doubling Anthropic's capabilities progress, a big update toward very short timelines — EvanHub · 2026-11-06
• METR's Greenblatt: AI R&D acceleration already at AI 2027's March 2027 levels — ResultBackground2450 · 2026-11-06
• METR: Anthropic's model capability gains have ~2x'd, no bottlenecks observed — corbtt · 2026-11-06
• METR: majority of capability acceleration likely driven by AI doing AI R&D — nathanbenaich · 2026-11-06
• AI 2027 author: METR data shows progress even faster than forecast — DKokotajlo · 2026-11-06
• METR report on AI accelerating AI R&D flips a skeptic: >50% on superintelligence in 18 months — GaryMarcus · 2026-11-06
• METR report finds AI is accelerating AI R&D at Anthropic; researcher now puts >50% odds on superintelligence within 18 months — EvanHub · 2026-11-06
• METR report: Anthropic is speeding up, not pacing, and redacted the key parts — asusarla · 2026-11-06
• METR-style fit shows benchmark gains accelerating 1.9x, but sensitive to measurement — jessi_cata · 2026-11-06
• METR chart: Anthropic employees self-report ~12x productivity uplift from Model 4 — scaling01 · 2026-11-06
• METR chart: permanent acceleration fits Anthropic's capability curve better (RMSE 1.9 vs 3.9) — scaling01 · 2026-11-06
• METR: AI capability acceleration not driven by faster training compute growth — BenBajarin · 2026-11-06
• METR chart shows Anthropic self-reported uplift hitting 12x with mystery 'Model 4' — maksym_andr · 2026-11-06
• METR's first deep assessment of AI accelerating AI R&D interviewed 15 staff — but skipped a key question — ronbodkin · 2026-11-06
• Anthropic cofounder defends METR report redactions; Gary Marcus unimpressed — GaryMarcus · 2026-11-06
• METR's Anthropic assessment triggers community-wide AGI timeline updates, veteran compares it to ChatGPT moment — AndrewCurran_ · 2026-11-06
• Gary Marcus blasts Anthropic cofounder's AI-safety defense of METR report redactions — GaryMarcus · 2026-11-06
• Anthropic chief architect Sam McCandlish's first post hints at internal-acceleration report, possibly hitting 2x RSP threshold — inductionheads · 2026-11-06
• Andrew Curran to AI safety camp: plans needing a year-plus window are now pointless — AndrewCurran_ · 2026-11-06
• METR's independent assessment of Anthropic's AI R&D acceleration spurs RSI-in-6-months calls — AndrewCurran_ · 2026-11-06
• Anthropic report's key estimates are "guesses" — framing hides massive uncertainty, researcher says — daveholtz · 2026-11-06
• Researchers clash over Anthropic report: conclusions rest on subjective staff assessments — danielrock · 2026-11-06

6 near-duplicate retellings: Recoil42 · kimmonismus · scottleibrand · scaling01 · maksym_andr · RishiBommasani