Anthropic's Sub-2x Self-Assessment Sparks Debate Over AI R&D Acceleration and Takeoff Models
Anthropic 在负责任扩展政策（RSP）评估中披露：在自主性威胁模型 2（关键领域自动化研发）上，Claude Mythos 5.2 尚不能完全替代其研究人员，内部使用该模型对整体 AI 研发进展的加速「不足 2 倍」，因此未触发相应阈值。这一自评迅速引发多方质疑，并演变为一场关于 AI 起飞建模的更广泛争论。
已确认
- Anthropic 的官方表述为研发加速「不足 2 倍」，依据是 METR 报告：点估计 1.88x，95% 置信区间 1.62x–2.28x，100 次重拟合中有 32 次超过 2x（m1、m2）
- TheZvi 公开质疑 Anthropic「加速不到 2 倍」的口径与发布时机（m1、m2）
- peterwildeford 质疑「未触及 AI R&D 阈值」判断缺乏透明度，并指 Anthropic 的 AI R&D 评测已饱和（m6）
- ramez（Ramez Naam）研读 METR 报告与模型卡后认为数据不支持「2 倍研发加速」说法，但承认存在一定加速；其配图显示 Anthropic 每十亿美元研发对应的 AECI 能力指数增益持续下滑（m4）
- AI 2027 原模型对任务时长扩展预测过于乐观，其后继者 AI Futures 模型预期约 1.3 倍加速，而 METR 实测约 2 倍，失准方向反转（m11、m14）
尚未确认
- 加速是否真实、是否由 AI 驱动、是否会复利：austinc3301 认为一年数据难以区分「平台跃升」与 RSI 开端；METR 的删减证据暗示加速「可能超过 2 倍且在增加」，但也表示若高出很多会意外（m7、m8）
- ramez 对多智能体扩展预测的批评与 takeoff 模型失真的判断，austinc3301 部分承认但指出失准方向已反转（m3、m14）
为什么重要
- 若真实加速接近或超过 2 倍且会复利，将直接关系到 RSP 阈值是否应被触发及安全监管节奏，因此自评口径与透明度成为焦点
- VladChituc 质疑将 RSI 建模为「思想生产函数」并非中性选择，而是得出爆炸结论的唯一路径；austinc3301 反驳称该函数建模的是 RSI 常常「熄火」这一障碍而非起飞前提，Elasticity Institute 也在尝试用局部弹性重述该条件绕开传统假设（m5、m10、m13、m15）
- austinc3301 提醒研发支出并非外生变量：Anthropic 越来越多算力流向做研发的 AI agent，AI 加速本身会表现为支出增长，因此「每十亿美元增益下滑」的解读需谨慎（m12）
- austinc3301 认为不宜将 Ajeya/Ryan 的时间线观点当作 all-things-considered 结论，但承认其源自合理的起飞建模选择，双方都没有明显错误（m8、m9）
2026-11-10 ~ 2026-11-11 · 15 related posts
- Episode 1: Claude Fable/Mythos 5.2 Leak Wave Spreads, Unverified and Likely Fabricated(2026-11-10, 21 posts)
- Episode 2: Anthropic Releases Claude Fable 5.2 and Mythos 5.2 with Faster Output and 20% Price Cut(2026-11-10, 113 posts)
- Episode 3: 1000 Fable 5.2 Agents Formalize Fermat's Last Theorem in 30 Hours(2026-11-10, 25 posts)
- Episode 4: Shumer Reviews Fable 5.2: Smartest Model Ever, But Heavy Guardrails(2026-11-10, 3 posts)
- Episode 5: Claude Fable 5.2 Launch: One-Shot Prompts Yield Full Games and 4K Films(2026-11-10, 32 posts)
- Episode 6: Matthew Berman's Claude Fable 5.2 Field Test: 9 Projects, "Just Solved Software"(2026-11-10, 8 posts)
- Episode 7: Anthropic Offers Claude Pro Users Free Trial of Fable 5.2 Until Nov 16(2026-11-10, 3 posts)
- Episode 8: Claude Fable 5.2 Tops LiveBench but Cost Concerns Emerge(2026-11-10, 3 posts)
- Episode 9: Claude Fable 5.2 Lands on Perplexity, Sets WANDR Benchmark Record(2026-11-10, 3 posts)
- Episode 10: Matthew Berman Calls Fable 5.2 the Best Model on Earth After 100B-Token Test(2026-11-10, 3 posts)
- Episode 11: Claude Fable 5.2 Sets New Record on ARC-AGI-3(2026-11-10, 2 posts)
- Episode 12: Over 90% of merged changes in Anthropic's core codebase now written by Claude(2026-11-10, 2 posts)
- Episode 13: Claude Fable 5.2 Tops Artificial Analysis Intelligence Index at 65 as Cost Concerns Emerge(2026-11-10, 14 posts)
- Episode 14: Every's hands-on test of Claude Fable 5.2: 6 of 7 gold ratings, overnight autonomous builds(2026-11-10, 6 posts)
- Episode 15: METR Files Confirm Fable 5.2 Is "Model 3"; Stronger Unreleased Model May Exist(2026-11-10, 4 posts)
- Episode 16: Anthropic's Sub-2x Self-Assessment Sparks Debate Over AI R&D Acceleration and Takeoff Models(2026-11-10, 15 posts)
- Episode 17: Fable 5.2 Not Yet Available in Claude Code(2026-11-10, 2 posts)
- Episode 18: Claude Fable 5.2 Sets New Epoch ECI Record at 174, Topping GPT-6 Astra(2026-11-10, 8 posts)
- Episode 19: Claude Fable 5.2 Tops WebDev Arena While Getting Faster and Cheaper(2026-11-10, 2 posts)
- Episode 20: Fable 5.2 'Elegantly Kills the Competition,' Users Say(2026-11-10, 3 posts)
Primary sources
- Zvi calls out Anthropic's "less than 2x" claim: fitted CI reaches 2.28x — TheZvi ·
- Anthropic says new model falls short of RSP autonomy threshold; Zvi questions timing after METR report — TheZvi ·
- Anthropic's capability gains per $1B of R&D keep falling, ramez pushes back on 2x AI acceleration — ramez ·
- Anthropic estimates 1.88x AI R&D acceleration near 2x threshold amid transparency doubts — scottleibrand · 2026-11-10
- AI timelines debate: modeling choices defensible, but don't adopt them wholesale — austinc3301 · 2026-11-10
- Three open questions on the ~2x AI R&D speedup: is it real, AI-driven, and compounding? — austinc3301 · 2026-11-10
- Is the 2x AI R&D speedup compounding? One year of data can't yet tell a level shift from an RSI loop — austinc3301 · 2026-11-10
- [source] Anthropic says new model falls short of RSP autonomy threshold; Zvi questions timing after METR report — TheZvi · 2026-11-10
- [source] Zvi calls out Anthropic's "less than 2x" claim: fitted CI reaches 2.28x — TheZvi · 2026-11-11
- Critics question why AI takeoff models assume an ideas production function — VladChituc · 2026-11-11
- Debate over intelligence explosion models hinges on the ideas production function — austinc3301 · 2026-11-11
- Takeoff modeling defense: ideas function models fizzle, not explosion — austinc3301 · 2026-11-11
- Elasticity Institute report restates RSI in local elasticities to bypass key takeoff assumption — austinc3301 · 2026-11-11
- Ramez Naam: seminal takeoff models don't match empirical data, multi-agent scaling gains far flatter than predicted — ramez · 2026-11-11
- AI Futures Model missed in the opposite direction: predicted 1.3x acceleration, METR found ~2x — austinc3301 · 2026-11-11
- AI Futures model missed the other way: predicted 1.3x speedup, METR found ~2x — austinc3301 · 2026-11-11
- [source] Anthropic's capability gains per $1B of R&D keep falling, ramez pushes back on 2x AI acceleration — ramez · 2026-11-11
- R&D dollars aren't exogenous: Anthropic's compute increasingly flows to internal AI agents — austinc3301 · 2026-11-11