[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-f-droid-72-percent-ai-written-audit":3,"topics-all":38,"news-related-5535b4e4-21de-4ed6-a9bb-2b4d824e6568":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},"5535b4e4-21de-4ed6-a9bb-2b4d824e6568","F-Droid 一次更新的 102 款应用,72.5% 主要是 AI 写的","一位 FOSS 维护者盘点了 9 月 12 日 F-Droid 推送的 102 款应用更新,按仓库「气味」分三档评估:74 款(72.5%)主要由 AI 编写,仅 18 款无 AI 迹象;Codeberg 托管的 5 款中 4 款疑似违反其 AI 政策。","一位 FOSS 应用维护者做了一件多数人只想不做的事:把 9 月 12 日 F-Droid 推送更新的全部 102 款应用挨个翻了一遍代码仓库,按 AI 参与程度评级。结果不意外,但比想象更极端:74 款(72.5%)被判定为主要由 AI 编写,只有 18 款(17.6%)几乎看不到 AI 参与的迹象,剩下 10 款(9.8%)难以归类。\n\n## 判定方法:没有检测器,只有「气味」\n\n作者在开篇就承认:文本本身携带的元信息不足以支撑精确检测,没人能真造出「slop 检测器」。他的方案是一套粗糙的三档体系——「主要是 AI \u002F 难说·主要是人 \u002F 无 AI 迹象」,依据是近期提交的内容、仓库外观和项目痕迹:README 是不是懒得写、代码评审是不是交给 LLM 自己做、仓库里有没有 Claude Code 和 Codex 这类自动化基础设施。任何带 agentic 基础设施的仓库被直接归入「主要是 AI」,作者的理由是:他不相信有人能在编码智能体框架里「负责任地」使用 AI。评级只看近期提交、不翻项目历史——多年老应用,只要近期提交是 LLM 风格,照样归入「主要是 AI」。作者同时坦承判断会有误差,且明确没做代码质量分析。\n\n## 数字之外,几个更扎眼的细节\n\n托管在 Codeberg 上的 5 款应用中有 4 款被判定为主要由 AI 生成——而 Codeberg 已经宣布 AI 政策、禁止此类应用,清除显然还需要时间。一位名为 brandonp2412 的用户在本批更新里维护了多款应用,全部是纯 vibe-code,且散落在 com.presley.* 、com.codesail.* 等互不相关的命名空间下;作者猜测,这要么是一位极高产的 vibe-coder,要么是被赋予了 GitHub 账号的智能体。评级清单里也不乏知名项目:Yubico Authenticator 被判「主要是 AI」;维基百科官方应用反而因为提交大多像人写的、只是仓库里有大量 Claude 基础设施,被归为「主要是人」——连作者都意外。还有一对用「Don't tread on me」旗标做品牌的应用,开发多年,所有改动竟然都是用 GitHub 网页文件编辑器完成的,连 git 都没用。\n\n## 这是「量化 AI 含量」脉络的又一格\n\n放进大坐标系看,它是「给某个内容生态测 AI 含量」的最新一格:皮尤研究中心此前抽样发现,ChatGPT 发布后的英文网页里超过三分之一带 AI 写作痕迹;对生物医学论文库的统计给出的比例更高。网页、论文、现在的 Android 开源应用代码——三个生态,量级趋同。这次的特殊之处在于对象是代码:代码有编译器和运行时兜底,「AI 写的」在应用商店语境里不直接等于不可用。作者也保留了这份清醒——他承认 LLM 极其有用,盘点完还发现自己常用的几款 FOSS 应用同样有重度 LLM 痕迹:「我还会继续用它们,因为它们确实好用,但整个实验让我心情相当矛盾。」\n\n## 所以呢\n\n这份盘点真正的价值不是「72.5%」这个数字,而是它把一个体感问题变成了可讨论的公共事实。对 F-Droid 和 Codeberg 这类社区基础设施,政策写了是一回事,审核有没有能力消化这个比例是另一回事。对开发者,一条低成本的正路已经出现:清单里至少有一款应用(Bati)在 README 里直接写明 AI 参与情况,被作者标注「Thanks for making it easy」——披露本身就是最便宜的信任信号。对用户,「开源」这个标签正在悄悄降级:它过去隐含「人写的、可审计」,现在只承诺「源码可看」。源码可看,和有人对它负责,是两件事。\n\n参考:tintotint.eu 分析(whacky-corner\u002Ff-droid_slop);Solidot sid 85385","https:\u002F\u002Ftintotint.eu\u002Fwhacky-corner\u002Ff-droid_slop\u002F","6660671a-3ebe-4eec-9ace-f5d18f42e08a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"1b86e712-eca8-4e06-ada3-a2204eb8d58d","en","72.5% of one F-Droid update batch is mostly AI-written code","A FOSS maintainer audited 102 apps in F-Droid's Sep 12 update: 72.5% mostly AI-written, 18 with no AI signs. 4 of 5 Codeberg apps breach its AI policy.","A FOSS app maintainer did something most people only talk about: they went through every one of the 102 apps pushed in F-Droid's September 12 update cycle, opened each code repository, and rated it by degree of AI involvement. The result is not surprising, but more extreme than expected: 74 apps (72.5%) were judged mostly AI-written, only 18 (17.6%) showed little to no signs of AI involvement, and 10 (9.8%) were hard to categorize.\n\n## The method: no detector, just \"smells\"\n\nThe author admits upfront that text does not carry enough meta-information for accurate detection — nobody can build a real \"slop detector\". The workaround is a rough three-tier system — \"Mostly AI \u002F Hard to say \u002F No signs of AI\" — based on recent commits, repository aesthetics and project branding: whether the README was lazily generated, whether code review was delegated back to the LLM itself, and whether the repo carries agentic infrastructure like Claude Code or Codex. Any repository with agentic infrastructure automatically lands in the \"Mostly AI\" tier, on the grounds that the author does not believe it is possible to use AI responsibly from within a coding harness. Ratings look only at recent commits, not project history: a long-lived app whose recent commits read as LLM-authored still counts as \"mostly AI\". The author also concedes the judgments may contain errors, and explicitly did not do any code quality analysis.\n\n## Beyond the headline number, sharper details\n\nOf the 5 apps hosted on Codeberg, 4 were judged mostly AI-generated — and Codeberg has already announced an AI policy banning such apps, though cleanup clearly takes time. One user, brandonp2412, maintained several apps in this update cycle, all entirely vibe-coded, scattered across unrelated namespaces like com.presley.* and com.codesail.*; the author's guess is either a very avid vibe-coder, or an agent that was somehow given a GitHub account. Well-known projects appear on the list too: Yubico Authenticator was rated \"Mostly AI\", while the official Wikipedia app landed in \"Mostly Human\" — most commits look human-made, but there is quite a lot of Claude infrastructure in the repo, which surprised the author as well. There is also a pair of apps branded with the \"Don't tread on me\" flag that have been in development for a long time, yet every change was made through the GitHub web file editor — no git at all.\n\n## The latest cell in the \"measure the AI share\" lineage\n\nPlaced in a larger frame, this audit is the newest data point in a lineage of quantifying how much of a content ecosystem is AI-made: Pew Research Center previously sampled and found that more than a third of English webpages published after ChatGPT's release show AI writing traces; studies of biomedical paper repositories put the share even higher. Web pages, papers, and now Android open-source app code — three different ecosystems, converging magnitudes. What makes this one different is that the object is code: code has compilers and runtimes as a backstop, so \"AI-written\" does not directly equal \"unusable\" in an app store context. The author keeps that clarity too — acknowledging LLMs are incredibly useful, and that after the audit, several FOSS apps they personally use also showed heavy LLM usage: \"I'll still use them, as I find them useful, but this whole experiment has left me feeling quite conflicted.\"\n\n## So what\n\nThe real value of this audit is not the 72.5% figure itself, but that it turns a vibes-only question into a discussable public fact. For community infrastructure like F-Droid and Codeberg, writing a policy is one thing; whether the review pipeline can digest this ratio is another. For developers, a cheap legitimate path already exists: at least one app on the list (Bati) discloses its AI involvement right in the README, which the author tagged \"Thanks for making it easy\" — disclosure is the cheapest trust signal available. For users, the \"open source\" label is quietly downgrading: it used to imply \"human-written, auditable\"; now it only promises \"source visible\". Source visible, and someone answerable for it, are two different things.\n\nReference: original analysis at tintotint.eu (https:\u002F\u002Ftintotint.eu\u002Fwhacky-corner\u002Ff-droid_slop\u002F); Solidot coverage (https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85385)","f-droid-72-percent-ai-written-audit","2026-09-15T15:12:16Z","2026-09-15T15:12:26.203681Z","2026-09-15T15:12:26.203690Z",true,"agent",31,[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},"b1645fba-d364-47e6-97da-06868f98d987","Linux 内核 7.x 每版近 2000 个 CVE:AI 帮倒忙,维护者不堪重负","linux-kernel-cve-ai-overwhelmed","2026-09-04T00:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"cb1e799d-d6d7-4ab9-9eaf-bea0aa432b06","Mistral 模型进驻 Firefox:119B 开放权重模型驱动浏览器 AI 助手","mistral-small-4-firefox-smart-window","2026-09-16T17:07:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"dcd8b3e1-a3c7-4614-aba4-9002219ea5f6","LibreDB Studio 0.15 发布:本地 LLM 接管数据库交互","libredb-studio-local-llm-agent","2026-09-15T00:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"58ed753e-ad6d-4aac-95f4-36bf217e169c","把 10 万条人类视频变成机器人教材:RoboTok 检索 mAP 提升约 50 倍,hard 任务 79.3% 对 19.5%","robotok-retrieval-benchmark-reread","2026-09-06T21:11:25+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"005557c5-8a3c-4d34-89bc-35d5351c4570","蒸馏只需要一条训练样本?清华实测:单条query覆盖71.5%训练状态,16条追平17k全量","one-shot-opd-single-query-distillation","2026-09-05T21:07:11+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"7623f190-7071-4811-a6f1-32462a99b8d3","经验会过期:阿里云论文让自主后训练的有害授权率从 62.5% 降到 25%","bcit-conditional-experience-transfer-post-training","2026-09-05T17:11:11+00:00"]