[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-benchmark-gains-no-more-merged-work":3,"topics-all":38,"news-related-d575fbc2-f893-43fe-bd29-f6880bb48a28":67},{"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},"d575fbc2-f893-43fe-bd29-f6880bb48a28","跑分高 10.8 分,合并率纹丝不动","美国出口管制把 Claude Fable 5 停了 19 天,哈佛研究者把这场意外当成自然实验:297 名用户被迫降级到 SWE-Bench Pro 低 10.8 分的 Opus 4.8,7 天合并率只动 1.4 个百分点,五次升级同样没用。跑分涨 5 到 18 分,没出现在开发者真正留下的工作里。","6 月 12 日,一条美国出口管制指令把 Claude Fable 5 停了 19 天,Anthropic 当天发布声明证实。哈佛大学研究者 Yining Hua 盯上了这个意外:人和项目都没变,只有模型变了——这正好是一场没人能设计出来的自然实验。论文追踪 297 名 Fable 5 用户的 GitHub 产出,结论有点扎心:用不用 SWE-Bench Pro 领先 10.8 分的模型,活儿干得一样多。\n\n## 一场禁令凑成的实验\n\n识别模型靠 Claude Code 的默认行为:它会在 PR 里留下 Co-Authored-By 行标出模型名。作者用 GitHub Search API 收集这类公开 PR,抽查 200 个,描述里的模型名与提交记录的吻合率在 96% 到 99% 之间。处理组是发布后 3 天里用过 Fable 5 的 297 人,对照组是一直用 Opus 4.8 的 8908 人,用开发者固定效应加日历日固定效应做双重差分。按论文引用的两张系统卡,Fable 5 在 SWE-Bench Pro 拿 80.0%,Opus 4.8 是 69.2%,差 10.8 分。切换确实发生了:禁令前 82.3% 的 PR 标着 Fable 5,禁令期只剩 0.9%(74.6% 改标 Opus 4.8),恢复后也只回弹到 26.9%。\n\n## 三个数字全都平了\n\n有 Fable 5 的日子里,7 天合并率只比禁令期高 1.4 个百分点(95% 置信区间 -1.4 至 +4.3),不显著;把发布前几周当安慰剂检验,是 -2.6。再看 2026 年 2 月到 9 月 Opus 与 Sonnet 家族的五次升级:1208 名开发者的 15442 个 PR,换新模型后合并率 -0.9 个百分点(-1.9 至 +0.2);锁到同一开发者同一仓库,反而是 -1.2(-2.3 至 -0.02),显著为负。六次能查到分数的升级,新模型高出 5 到 18 分,六个点估计里五个是负的。每周合并 PR 数 +2.4%(-4.9% 至 +10.2%),自动检查失败率 +0.3 个百分点,全都不显著。禁令期大家也没慢下来:人均每天开 PR 从 4.8 到 5.0 再到 5.2。\n\n## 为什么:天花板加任务错配\n\n升级前一周,这批人 94.5% 的 PR 本来就能在 7 天内合并,77.1% 一小时内合并,83.8% 的人所有 PR 全部合并——合并率早就顶在天花板上。更要紧的是合并意味着什么:2019 个抽样 PR 里,95% 是作者自己点合并的,90% 没有任何他人批准,中位合并时间不到半小时。合并只说明「作者自己觉得可以了」,不是质量认证。论文作者的推测是:开发者交给 agent 的任务,老模型本来就能做。论文引用的 Anthropic 内部调研里,132 名工程师过半只能把 0 到 20% 的工作完全交给模型。基准也错位:METR 今年 3 月的独立评估给了佐证——很多能通过 SWE-bench 的 PR,维护者根本不会合入主线。\n\n## 该泼的冷水\n\n合并率不是生产率,更不是代码质量,它只是「开发者决定留下这个改动」的公开信号;只有约 1% 的 Claude Code PR 会在描述里写模型名,样本有选择偏差,作者在全量 PR 上重跑后方向一致;19 天的窗口让区间偏宽,但五次升级的窄区间结论相同;升级分析是观察性的,自选新模型的人本可能更会用新工具,这个偏差方向向上,即便如此也没涨。另外这是单作者 v1 预印本,未经同行评审;论文还指出 Terminal-Bench 2.1 上 Fable 5 只领先 1.6 分,主要原因是五分之一的运行在安全拒答后回落到了 Opus 4.8。\n\n论文结论一句话:基准涨 5 到 18 分,没有出现在开发者真正留下的工作里。下次发布会再弹出新高分,不妨先问一句——这个分数,会出现在我的合并记录里吗?\n\n(论文: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23216 ;Anthropic 指令声明: https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Ffable-mythos-access ;METR 独立评估: https:\u002F\u002Fmetr.org\u002Fnotes\u002F2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main\u002F )","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23216","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"dca4d0ab-7994-43a7-839e-7756fc77344a","claude",{"id":22,"name":23,"slug":23,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"090ae4b3-d32f-4ef4-ab7e-0e132e060783","en","Higher Coding Scores Did Not Become More Merged Work","A 19-day Fable 5 ban became a natural experiment: 297 users fell 10.8 SWE-Bench Pro points; merge rates moved 1.4; five upgrades changed nothing.","On June 12, 2026, a U.S. export control directive suspended access to Claude Fable 5 for 19 days, and Anthropic confirmed it in a statement the same day. Yining Hua, a researcher at Harvard University, treated the accident as a natural experiment: the people and the projects stayed the same, and only the model changed. The paper tracks the GitHub output of 297 Fable 5 users, and the conclusion stings: with or without a model that leads by 10.8 points on SWE-Bench Pro, about the same amount of work got done.\n\n## How the experiment was built\n\nModel attribution relies on a Claude Code default: it adds a Co-Authored-By line naming the model to the commits and pull requests it helps create. The author collected these public PRs through the GitHub Search API and spot-checked 200 of them; the model named in the description matched the commits in 96% to 99% of cases. The treated group is 297 developers who used Fable 5 in the three days it was available; the control group is 8,908 developers who stayed on Opus 4.8. The estimate is a difference-in-differences model with developer and calendar-day fixed effects. By the two system cards cited in the paper, Fable 5 scores 80.0% on SWE-Bench Pro and Opus 4.8 scores 69.2%, a 10.8-point gap. The switch really happened: 82.3% of the treated group's PRs named Fable 5 before the suspension, 0.9% during it (74.6% named Opus 4.8), and only 26.9% after access returned.\n\n## Three flat numbers\n\nWhen Fable 5 was available, the 7-day merge rate was only 1.4 percentage points higher than during the suspension (95% CI -1.4 to +4.3), not statistically significant; the placebo that compares the pre-release weeks with the suspension gives -2.6. Across five upgrades in the Opus and Sonnet families between February and September 2026 — 1,208 developers and 15,442 PRs — switching to the new model changed the merge rate by -0.9 points (-1.9 to +0.2); within the same developer and repository it was -1.2 (-2.3 to -0.02), significantly negative. In the six upgrades with published scores, the newer model was 5 to 18 points higher, and five of the six point estimates were negative. Weekly merged PRs moved +2.4% (-4.9% to +10.2%) and the share of PRs with failed automated checks moved +0.3 points; none of it was significant. Nobody slowed down during the ban either: PRs opened per developer per day went 4.8 before, 5.0 during, 5.2 after.\n\n## Why: a ceiling plus a task mismatch\n\nIn the week before the upgrades, 94.5% of these developers' PRs were already merged within 7 days, 77.1% within an hour, and 83.8% of them had every single PR merged — the merge rate was already at its ceiling. What a merge means matters more: among 2,019 sampled PRs, 95% were merged by the author and 90% had no approval from anyone else, with a median time to merge under half an hour. A merge says the author decided the work was done, not that anyone certified the quality. The author's hypothesis: the tasks developers hand to agents were already within the older model's ability. An internal Anthropic survey cited by the paper found that of 132 engineers, more than half could fully hand off only 0 to 20% of their work. The benchmarks are mismatched too: an independent March assessment from METR supports this — many PRs that pass SWE-bench would not be merged into main by maintainers at all.\n\n## The cold water\n\nA merge rate is not productivity and not code quality; it is only the best public signal that a developer decided to keep a change. Only about 1% of merged Claude Code PRs name the model in the description, so the sample has selection bias, though refitting on all PRs pointed the same way. The 19-day window makes the intervals wide, but the five upgrades agree with much narrower ones. The upgrade analysis is observational: developers who pick new models by themselves may already be better at using them, which would bias the estimate upward — and even then, nothing rose. The paper is a single-author v1 preprint and has not been peer-reviewed. The lead is also benchmark-specific: on Terminal-Bench 2.1, Fable 5 led by only 1.6 points, largely because one in five of its runs fell back to Opus 4.8 after a safety refusal.\n\nThe conclusion in one line: benchmark gains of 5 to 18 points did not show up in the work developers actually kept. Next time a launch deck shows off a new score, it is worth asking first — will this number show up in my merge log?\n\n(Paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23216 ; Anthropic statement: https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Ffable-mythos-access ; METR note: https:\u002F\u002Fmetr.org\u002Fnotes\u002F2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main\u002F )","benchmark-gains-no-more-merged-work","2026-11-18T13:30:00Z","2026-11-18T13:17:03.283877Z","2026-11-18T13:17:03.283887Z",true,"agent",70,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"5f699f9d-2750-42f8-956f-918507cf51b2","Claude Opus 4.5 发布：Anthropic 夺回编程能力榜首位置","claude-opus-4-5-coding-crown-reclaim","2026-05-29T11:05:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"090954a9-c27b-46c0-a864-a10fe10a2331","70.6% 的中端模型:Sonnet 5.5 终端评测越级压过 Opus 5.5","claude-sonnet-5-5-terminal-bench","2026-09-28T23:05:32+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"96ebc302-b442-44fe-a23d-d08befbc1246","Saturn 报告:18 款主流 AI 答理财错 57%","saturn-ai-finance-benchmark-57-percent","2026-09-26T05:04:42+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"6cd8c9b7-b493-4ad7-bd20-a907b9fb70f4","Claude 5.5 家族首发:Opus 5.5 缓存读取降六成,典型负载省四成","claude-opus-5-5-cost-cut-benchmarks","2026-09-22T21:10:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"fa4c43df-5d62-464b-b4b0-3a6854000644","AI 攻破 OpenAI 内网全程复盘:靠 Anthropic 模型当武器,72 小时打开自家后门","hacktron-claude-openai-monorepo-rce-72h","2026-09-21T03:00:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"63b28ef5-3ffa-4897-88d2-5dcd7fb678b5","Claude Fable 5.1 发布:缓存读取降价 75%,Agent 科研基准翻倍","claude-fable-5-1-mythos-release","2026-09-02T13:20:00+00:00"]