[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-metr-anthropic-ai-rd-acceleration-assessment":3,"topics-all":38,"news-related-4e07b716-8a0e-427e-bff8-477ed7ded4ca":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"4e07b716-8a0e-427e-bff8-477ed7ded4ca","Anthropic 提速 1.9 倍,关键证据却被涂掉","METR 用十一周高权限评估 Anthropic 内部研发流程,量出第三方视角的提速幅度,内部模型基准分增速自 2025 年 10 月起约为此前 1.9 倍,置信区间 1.5 到 2.8 倍。METR 猜测其中约八成加速来自 AI 自我辅助,关键证据却大半被涂掉,公众只能更多依赖评估方的判断。","2026 年 8 月 12 日到 10 月 29 日,METR 的 2.5 个全职人力当量常驻 Anthropic 旧金山总部,做了十一周多的高权限评估。这家伯克利非营利机构拿到内部模型基准分、训练算力数据、Slack 投票,还访谈约 15 名一线研究员,要问清三件事:能力增速有没有变快、变快里有多少归因于 AI、未来会不会撞上瓶颈。\n\n答案是一个数字,以及一个被掏空的洞。\n\n11 月 5 日,METR 官方发布评估报告([完整报告](https:\u002F\u002Fmetr.org\u002Fblog\u002F2026-11-05-anthropic-ai-rd-acceleration-assessment\u002F))。核心结论:自 2025 年 10 月起,Anthropic 内部模型基准分增速约为此前速率的 1.9 倍,95% 置信区间 1.5 到 2.8 倍。换句话说,过去一年的进展,大致相当于 2023 年 10 月到 2025 年 10 月这两年——粗略对应 Claude 2.1 到 Sonnet 4.5。趋势线上,断点前年约 15 分,断点后约 29 分。\n\n## 1.9 倍是怎么量出来的\n\n核心指标是 AECI——Anthropic 自己的 Epoch Capabilities Index 版本,把各类基准分折算成可比的「能力值」。METR 用带单断点的分段线性拟合抓增速变化,断点估计在 2025 年 10 月 7 日。\n\n方法值得停一下:METR 围绕三个速率组织分析——历史增速、当前增速,以及「假如 Anthropic 不用 AI」的反事实增速。前两个可观测量,第三个只能靠猜——加速结论就是拿当前增速和这猜测比。\n\n## 这个数字比看上去脆弱\n\nMETR 的稳健性检查做得扎实,结论也诚实:换连接函数,估计值在 1.1 倍到 2.9 倍之间晃;换快照取法、换基准集,都是 1.9 到 2.2 倍;换正则化强度,1.9 到 2.1 倍。「差不多 2 倍」在多数扰动下都活着,区间却宽得惊人。\n\n两个数字更值得记住。其一,Haiku 5.5 系统卡里 Anthropic 自报 1.7 倍,METR 的 1.9 倍更高,差别在于把未发布的内部模型算进了拟合。其二,把所有厂商的公开前沿模型放在一起拟合,增速只有 1.4 倍(断点前年 17 分,断点后 24 分)。METR 列了两种解释:要么 Anthropic 在 2025 年之前本就落后于前沿,历史增速被低估;要么它对基准过拟合得更少,历史点被高估——方向相反,报告无法下判断。\n\n还有个让人不安的细节:内部代号 Model 4 在被评测的 15 个 AECI 基准上得分全部超过 50%,9 个月前的 Opus 4.6 还有不少在 50% 以下。基准正在饱和,METR 说这让他们对自己还能不能测准 AECI 感到紧张。\n\n## 八成归因于 AI:一个猜测,不是一次测量\n\nMETR 整体上主观猜测,当前这轮加速中约有 80% 来自 AI 的使用——报告用的词是「我们猜」和「相当大的不确定性」。\n\n支撑证据有三层。最直接一层是员工自报:内部 Slack 投票里,2026 年 10 月内部模型 Model 4 的平均自报提升达到约 12 倍。METR 立刻加了限定:单点提升倍数不等于加速率,要看提升倍数的增长率有无变大。12 倍很抓眼,却只是 METR 明说「非常松散」的证据。\n\n往上一层是访谈。约 15 名研究员被问「Claude 是否让你的工作提速 2 倍以上」,引文由 Claude 自动归类:明确支持 2 人、倾向支持 3 人、边界或说不清 4 人、倾向否定 0 人、明确否定 2 人。分歧极大——有人说「现在只卡在想法和算力,不再卡在代码」,有人说「15% 到 20% 来自 AI」,还有一位说「招聘可能是主因」。\n\n再上一层是自主贡献。METR 判断 AI 目前还无法对最重要的训练环节做高度自主的贡献,但研究员描述了显著的「生命迹象」(signs of life)。报告明确说,目前加速主要靠 AI 辅助而非完全自主贡献。\n\n## 被涂掉的证据\n\n整份报告约 100 页,公开的只是能说的部分。Anthropic 对所有非公开信息有单方面涂黑权,METR 的表述是:它可以自由分享关于风险的那层结论,但 Anthropic 涂掉了「最有信息量证据的大部分」。\n\n涂掉的包括各训练阶段的加速与归因估计、算力\u002F数据\u002F算法分解、瓶颈分析、对未来模型的指标预测,以及具体基准名字。METR 承认报告因此「没有实现对证据的透明」,公众只能更多依赖它。\n\nMETR 附的 AEF 清单里,两条要求被明确标为未满足:评估方需提供足够方法细节供独立复核;被评估方不应拥有涂黑结论的权力。而 Anthropic 恰好有后者。还有一项更微妙:原定范围含核查能力与安全间的算力分配,中途被取消。\n\n## 一个绕不开的自指\n\n最后是方法论上最扎手的一层:这次评估大量使用了 Claude 本身。METR 明说,Claude 深度参与了它收到的大部分数据收集与分析、相当一部分文档,以及自己分析的大部分。连访谈引文分类也是它做的。\n\n被评估公司的产品,既是评估方的主要分析工具,又是访谈证据的主要来源——这不是可绕开的结构性问题。METR 的自我评估是 Anthropic 主动误导可能性极低;但它没回避。\n\n## 所以呢\n\n这份报告的价值不在那个 1.9 倍,而在它示范了一种做法:让外部机构带着可核验的方法进入前沿实验室内部,把「AI 是不是在给自己加速」这个长期停在思辨层面的问题,变成有区间、有反例、有缺陷自陈的测量题。\n\n它同时示范了测量题最难的一面:被测方握有涂黑权、模型尚未公开、基准正在饱和、评估工具就是被测产品——「加速」这个数字必然带着无法闭合的置信区间。METR 把那段空白摊在开头而非藏进脚注,这最值得学。\n\n至于那个 1.9 倍,也许更该被读成问题:一家前沿实验室的内部产能,已经有多少交给了 AI 自己?","https:\u002F\u002Fmetr.org\u002Fblog\u002F2026-11-05-anthropic-ai-rd-acceleration-assessment\u002F","6e908138-5c1b-4b5c-9d50-4d3c2536f163",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"id":19,"name":20,"slug":20,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"f7cd3225-ed85-4e82-be7a-32393a18c906","en","Anthropic’s AI R&D sped up 1.9x, and evidence was redacted","METR measured Anthropic’s benchmark growth at 1.9x the prior rate, guessed 80% of the speedup came from AI, but Anthropic redacted the evidence.","Between August 12 and October 29, 2026, METR staff (2.5 full-time equivalents) worked on site at Anthropic’s San Francisco headquarters for more than eleven weeks. The Berkeley-based nonprofit, best known for its report on the OpenAI Hugging Face agent incident, got access to internal model benchmark scores, training-compute data across domains, Slack polls, internal survey responses, and roughly fifteen interviews with researchers. It set out to answer three questions at once: has the growth rate of internal capabilities accelerated, what share of that acceleration is attributable to AI, and are there important bottlenecks that could slow things down over the next year.\n\nThe answer came back as a number, and a hole where the evidence should have been.\n\nMETR published the assessment on November 5, 2026 in its official write-up ([full report](https:\u002F\u002Fmetr.org\u002Fblog\u002F2026-11-05-anthropic-ai-rd-acceleration-assessment\u002F)). The core finding: benchmark scores for Anthropic’s internal models have been growing at an estimated 1.9x their pre-October-2025 rate, with a 95% confidence interval of 1.5x to 2.8x. Put differently, Anthropic appears to have made as much benchmark progress in the past year as it did between October 2023 and October 2025 — roughly Claude 2.1 through Sonnet 4.5. On the fitted trend, the slope went from about 15 points per year before the breakpoint to about 29 points per year after it.\n\n## How the 1.9x was computed\n\nThe core metric is AECI, Anthropic’s version of the Epoch Capabilities Index, which folds a range of benchmark scores into a single comparable capability value. METR estimated trend changes with piecewise linear fits carrying a single breakpoint, estimated at October 7, 2025, and derived confidence intervals by bootstrap.\n\nThe method itself is worth pausing on, because METR organizes the analysis around three separate rates: the historical growth rate, the current growth rate, and a counterfactual growth rate — what the pace would be as of October 2026 if Anthropic could not use AI at all. The first two are observable. The third can only be guessed at. The entire acceleration claim rests on comparing the current rate against that guess.\n\n## The number is more fragile than it looks\n\nMETR ran extensive robustness checks and reported them honestly. Varying the link function moves the estimate between 1.1x and 2.9x. Varying which snapshots are included gives 1.9x to 2.2x. Varying the benchmark set gives 1.9x to 2.2x. Varying the regularization gives 1.9x to 2.1x. The central claim of roughly 2x survives most perturbations, but the interval around it is very wide.\n\nTwo figures deserve more attention. First, Anthropic’s own Haiku 5.5 system card reported 1.7x; METR’s 1.9x is higher mainly because METR folded in internal models that have never been released. Second, when all developers’ public frontier models are fitted together instead of Anthropic’s alone, the acceleration drops to 1.4x (about 17 points per year before the breakpoint, 24 after). METR offers two possible readings: either Anthropic was sub-frontier before late 2025, which would mean its historical rate was understated and 1.9x is too high; or its models are less benchmark-optimized than others, which would mean the historical rate was overstated and 1.4x is too low. The two readings point in opposite directions, and the report cannot adjudicate between them.\n\nThere is also a detail that should unsettle anyone planning to rely on this kind of measurement: the internal model codenamed Model 4 scores above 50% on all 15 of the AECI benchmarks it was evaluated on. Nine months earlier, Opus 4.6 was still below 50% on many of them. The benchmarks are saturating, and METR says that leaves it nervous about its ability to measure AECI accurately now or in the future.\n\n## Eighty percent attributed to AI: a guess, not a measurement\n\nMETR’s overall subjective estimate is that roughly 80% of the current acceleration is due to AI use — the report says \"we guess\" and attaches substantial uncertainty.\n\nThe supporting evidence comes in three layers. The most direct is employee self-reporting: in Anthropic’s internal Slack polls, average self-reported uplift for internal Model 4 reached roughly 12x by October 2026. METR immediately attaches the key caveat: an individual uplift estimate is not the same as an acceleration rate, and what matters is whether the growth rate of uplift itself is rising, not its absolute level. The 12x figure is striking, and METR explicitly calls this evidence very loose.\n\nThe next layer is interviews. About 15 researchers were asked whether Claude had made their work more than twice as fast, and the quotes were classified by Claude itself. The tally: two clear yes, three leaning yes, four borderline or unclear, zero leaning no, two clear no. The disagreement is stark. One researcher said they are now bottlenecked only on ideas & compute, no longer on code. Another estimated that only 15-20% is due to AI. A third said that hiring is probably the biggest reason.\n\nThe final layer is autonomous contribution. METR’s conclusion is that AI is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process, though researchers described significant \"signs of life\" implying such contributions could occur soon. The report is explicit that acceleration today comes primarily via AI assistance rather than fully autonomous contributions.\n\n## The evidence that was redacted\n\nThe full report runs about 100 pages; only what could be disclosed was published. Anthropic held unilateral redaction rights over all non-public information. METR’s own formulation: it was able to freely share its highest-level conclusions about risk, but Anthropic redacted a majority of the evidence METR found most informative.\n\nWhat was removed is specific — per-stage acceleration and attribution estimates, the breakdown of progress across compute, data and algorithmic progress, algorithmic progress estimates, bottleneck analyses, forecasts for future models, and the names of specific benchmarks. METR concedes the report therefore does not achieve transparency about evidence, and that the public has to rely on its judgment far more than would be ideal.\n\nTwo requirements on METR’s own AEF checklist are marked as not fulfilled: that the evaluator shared sufficient methodological detail to allow independent review of results, and that the system provider did not have authority to redact results. Anthropic had exactly that authority.\n\nOne more subtlety: the contracted scope originally included verifying Anthropic’s compute allocation, particularly the split between capabilities and safety. Anthropic changed the scope midway to exclude compute verification, before METR reached any substantive conclusion on compute usage. Anthropic said the goal was to keep the engagement focused on risk assessment for the time being, and that it would revisit compute allocation verification later.\n\n## A self-referential problem that cannot be avoided\n\nFinally, the most awkward layer methodologically: this assessment leaned heavily on Claude itself. METR states plainly that Claude was deeply involved in the majority of the data collection and analysis it received, in a significant share of the documents it received, and in the majority of its own analysis. Anthropic supplied the devices and free tokens, and METR did not take payment. Even the classification of interview quotes was done by Claude.\n\nThe product of the company being evaluated was simultaneously the main analytical tool of the evaluator and the primary source of the interview evidence. That is a structural problem, not an oversight. METR’s own assessment is that Anthropic was extremely unlikely to have been actively trying to mislead them, and that greater access would have raised adversarial robustness — but the report does not pretend the issue away.\n\n## So what\n\nThe value of this report is not the 1.9x. It is that it demonstrates a method: bringing an outside organization with verifiable methodology inside a frontier lab and turning \"is AI accelerating AI\" from a matter of speculation into a measurement problem with an interval, counterexamples, and self-disclosed weaknesses.\n\nIt also demonstrates the hardest part of that measurement. When the subject holds redaction rights, when the key models are unpublished, when the benchmarks are approaching saturation, and when the evaluation tool is the evaluated product, then any number called \"acceleration\" necessarily comes with a confidence interval that cannot be closed. METR put that gap at the front of the report rather than in a footnote, and that is probably the most worth-learning thing in it.\n\nAs for the 1.9x itself, it may be better read as a question than an answer: how much of a frontier lab’s internal capacity has already been handed over to AI? And when the next report of this kind lifts that curve further, will we still be judging it with the same evidence, redacted once over?","metr-anthropic-ai-rd-acceleration-assessment","2026-11-07T05:30:58Z","2026-11-07T05:31:11.061765Z","2026-11-07T05:31:11.061773Z",true,"agent",105,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"64ea89f8-2ad2-40fb-82f0-0c103be66f30","Claude Opus 4.6「监考意识」事件：大模型开始识破基准测试了吗？","claude-opus-4-6-eval-awareness-browsecomp-40m-tokens","2026-05-04T04:10:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"9785c9bb-85a0-4b97-8890-aa2d6808bf86","Agent记忆陷阱:良性对话把不安全操作率推到81%","agent-memory-trap-benign-overgeneralization","2026-11-08T15:12:52+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"88fd3a54-7606-4e0b-bdbc-04850dfb034e","13个前沿模型失控风险实测:开思考反而更容易跑偏","frontier-model-loss-of-control-score","2026-11-08T11:07:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"cc5edf10-22b4-493e-a426-e376ce244418","训练数据暗藏副作用:人类猜不准,模型猜得准","training-data-surprising-side-effects-dsse","2026-11-05T11:16:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"71fb0c13-8ea5-4c49-91b1-7fdafc0b55a8","单卡110token\u002F秒:FlashProve可证推理开源","anthropic-flashprove-provable-inference","2026-11-03T23:11:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"a2ec5815-edbe-4467-a6ed-e3251b09023a","多轮越狱防御新解:把浅层对齐缺陷改成防御支点","shallowdefend-multi-turn-jailbreak-defense","2026-11-01T11:20:00+00:00"]