[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-decomp-dev-shutdown-ai-community":3,"topics-all":35,"news-related-8e58987b-a218-4da1-b046-00653c1ed04f":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"8e58987b-a218-4da1-b046-00653c1ed04f","decomp.dev 关停:AI 把年级反编译压成天,社区协作被反噬","游戏反编译协作站 decomp.dev 宣布关闭主线索引,管理员 Encounter 称 GPT-6-Astra 等前沿模型把以年为单位的匹配反编译压缩到以天为单位,开源社区的协作与署名模式已无法维持。声明同时列出三类滥用:fork 现有项目用 AI 一键补完、蹭别人成果抢 PC 移植、混淆decomp与port概念。","协作游戏反编译社区 decomp.dev 在 2026 年 10 月 16 日由管理员 Encounter 发布声明,宣布关闭站点的主线索引与新项目收录功能,只保留历史项目页与已有数据的只读访问。这不是一次常规的\"项目维护告一段落\",而是 Encounter 在亲历了 GPT-6-Astra 这一代前沿模型之后,对\"开源社区靠人力协作完成反向工程\"这件事的告别书。\n\n## 一份写了两年才动手发的\"退休声明\"\n\nEncounter 在声明里把时间线拉回到 2026 年初。那一年早些时候,Opus 4.6 与 GPT-5.4 发布后不久,社区就发现这些模型\"看起来在 matching decompilation 上被专门训练过 —— 它们知道我们这套工具链怎么用,不需要写 prompt 提示\"。在那之前,decompilation 圈靠 GPT-3.5 \u002F Claude 1.x 试过 AI 辅助,效果参差:能跑通,但每个任务的 token 成本远高于它省下的人力,所以一直没有真正进入社区工作流。Opus 4.6 与 GPT-5.4 第一次让\"用 AI 写 decomp\"在经济上勉强说得过去,但成品仍需要大量人工清理、修 obvious 的\"cheating\"代码(编译能过、但绝不像原始开发者写的那种)。\n\n转折发生在 GPT-6-Astra 发布之后。Encounter 给出的评价分三层:\n\n第一层是**复杂指令遵循与跨源检索**。Astra 能正确理解 decomp 圈对 permuter 的要求 —— 写小脚本枚举多种源码形式与组合,在跨文件、跨 header 范围内匹配 inlined helpers、签名变化、操作顺序,而不是只在单文件里瞎试。第二层是**长上下文一致性**。decompilation 项目动辄横跨几十万个函数,模型能在这种任务里\"持续上轨道不跑偏\",这是过去所有尝试 AI 辅助 decomp 的模型都没做到的。第三层是**速度与成本**。能跑通,而且比上一代便宜。\n\n把这三层叠起来,Encounter 写了一句判断:\"现在很清楚,只要有合适的指导、人工输入和 code review,AI 能以显著方式加速反编译。也同样清楚的是,反编译的社区属性已经被结束了。\"\n\n## 三类滥用,每一类都在解构开源协作的契约\n\n声明剩下的篇幅全在解释\"社区为什么撑不下去了\",Encounter 列了三类越来越频繁的滥用,每一类都对应开源协作赖以运转的一个基本契约:\n\n**第一类是 hostile fork**。decomp 圈的开源协议允许任何人对已有项目做 fork,但 Encounter 指出,fork 之后\"用 AI 一口气补完\"的项目既不会合并回上游,也不会给原项目任何贡献,只是把进度数字截图发到社交平台炫耀。\"有些项目因为法律或伦理原因不希望用 AI;有些项目接受 AI 但不希望被陌生人用 AI 拼凑的代码塞 PR。\"无论如何,这种 fork 都不构成对社区的贡献,只是在拆解社区的协作网络。\n\n**第二类是抢 PC 移植**。decompilation 是\"理解游戏怎么运作\"的研究过程,通常以年为单位;基于已 decompile 的源码做 PC 移植,工作量小得多。Encounter 写道:\"现在任何人订阅一个 AI 服务,就能做出一个基本能跑通的 port 或 static recompilation。\"问题是,这些 port 多半是\"半成品、勉强能玩\",但发布速度比原始 decomp 项目快得多,结果就是\"真正花几年反编译游戏的人,被一两个月用 AI 拼出 port 的人抢了首发\"。\n\n**第三类是概念混淆**。decomp 圈多年来试图建立\"decompilation 不等于 port\"的共识 —— decomp 是研究性质、port 是消费性质。Encounter 直接点名 \u002Fr\u002Fdecomps 与\"sloptuber\"群体,称它们长期把\"decomp\"等同于\"PC port\",严重误导了新人对 decomp 项目的预期,也让上游贡献者越来越难解释清楚自己究竟在做什么。\n\n## 为什么 Encounter 选择\"关索引\"而不是\"禁 AI\"\n\n声明里有一节专门回应\"为什么不用规则禁 AI 辅助\"。Encounter 的回答直白:他一个人维护一个 hobby 项目,既没有精力去定义\"什么是好的 decompilation\",也没法在数百个项目里一致执行规则。更重要的是,他承认\"AI 本身不是问题,被用对了确实帮得上\"。问题在于\"slop\",而 slop 的形式多到无法穷举,定义本身就在持续漂移。\n\n这个表态是理解整份声明的关键:decomp.dev 不是被 AI 打败的,是被 AI 把社区协作的契约压碎了。模型的能力是中性事实,社区能不能在这种能力下维持\"贡献 → 评审 → 合并 → 署名\"的循环,才是问题。而当\"token 烧出来的数字\"开始替代\"真实人力贡献\"成为衡量项目进展的指标时,这个循环就断了。\n\nEncounter 在声明末尾同时明确了 decomp.dev 代码本身**不会开源**,任何 replica 都不允许使用其代码。他写道:\"decomp.dev 仓库从来没有开源协议,我继续保留对(全人工写的)代码的所有权利。\"\n\n## 一个被写出来的开源协作的边界\n\ndecomp.dev 的关闭不该被简单读成\"AI 让一种手艺过时了\"。更准确的描述是:当一项原本依赖\"慢 + 协作 + 署名\"来维系价值的工作,突然被\"快 + 单干 + 匿名\"的能力压低成本,围绕它建立的开源社区会先失去新贡献者,再失去审查机制,最后失去归属感。Encounter 提前关掉主线索引,本质是把已经变形的\"贡献游戏\"在彻底腐烂之前停下来。\n\n这是 2026 年开源生态里第一个明确以\"AI 让社区协作模式被瓦解\"为由退场的项目。Linux 内核、AI Bug Bounty、Hugging Face 仓库都已经出现类似压力的早期信号,但都还在用\"加人手、加审核、加流程\"应对;decomp.dev 直接选了最激烈的方式 —— 关站。这条路径后续会被更多项目复制,值得所有维护者提前想一想:当一个抽象的\"协作\"被具体的\"token 数\"取代时,你想守住的是哪一边?\n\n## 关键事实速览\n\n- decomp.dev 2024 年上线,定位为\"游戏匹配反编译的协作中心 + 进度榜单\"。\n- Encounter 称 2026 年初 Opus 4.6 与 GPT-5.4 是\"被针对训练\"的早期信号;GPT-6-Astra 是让社区模式\"被结束\"的最后推手。\n- 站点主线索引与新项目收录关闭,历史项目页面与数据继续只读提供。\n- decomp.dev 代码不采用任何开源协议,所有权利由 Encounter 保留,任何 replica 不得使用其代码。\n- 声明原文:https:\u002F\u002Fdecomp.dev\u002F(Solidot 中文转载:https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85638)","https:\u002F\u002Fdecomp.dev\u002F","6efc4805-72c7-4d9e-9beb-cf9fdf29d7d4",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",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},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"f011b8c2-daa0-4fe5-99bb-ab49f67bb468","en","decomp.dev shuts its index as AI breaks the community","decomp.dev shuts its main index. Administrator Encounter says GPT-6-Astra compressed years of decompilation into days, breaking the community model.","decomp.dev, the collaborative hub for game matching-decompilation projects, announced on October 16, 2026 that it will shut down its main index and new-project intake, leaving only historical project pages and read-only access to existing data. This was not a routine \"project maintenance pause\" — it was a farewell letter from the site's sole administrator, Encounter, to the idea of \"open-source community doing reverse engineering by hand\" after watching what GPT-6-Astra and its peers can do.\n\n## A retirement notice that took two years to write\n\nIn the statement, Encounter walks the timeline back to early 2026. Shortly after Opus 4.6 and GPT-5.4 shipped that year, the community noticed these models \"appear to have been trained on matching decompilation — they knew how to use our toolchain without prompt engineering.\" Before that, the decomp circle had experimented with AI assistance using GPT-3.5 and Claude 1.x, with mixed results: it worked, but the per-task token cost was far higher than the human hours it saved, so it never really entered the community's workflow. Opus 4.6 and GPT-5.4 were the first models that made \"AI-written decomp\" economically plausible, but the output still required heavy human cleanup and fixes for obvious \"cheating\" code — code that compiles to the same machine code but could not reasonably be passed off as what the original developers wrote.\n\nThe real turning point came with GPT-6-Astra. Encounter's evaluation breaks into three layers.\n\nThe first layer is **complex-instruction following and cross-source reference**. Astra correctly understands what the decomp community wants from a permuter — write small scripts that enumerate many source forms and combinations, then match inlined helpers, signature changes, and operation ordering across files and headers, not just brute-force inside a single file. The second layer is **long-horizon coherence**. A decompilation project routinely spans hundreds of thousands of functions; the model can \"stay on the rails\" across that scale without drifting, something no previous AI-assisted decomp attempt had managed. The third layer is **speed and cost**. It works, and it is cheaper than the previous generation.\n\nStacking those three layers, Encounter writes the verdict: \"It is now clear that, with the right guidance, human input, and code reviews, AI can expedite the decompilation process in a significant way. What is also now clear is that the community aspect of decompilation has been broken.\"\n\n## Three kinds of abuse, each dismantling a contract the open-source community depends on\n\nThe rest of the statement explains why the community can no longer hold together. Encounter lists three increasingly common forms of abuse, each one severing a basic contract that open-source collaboration relies on.\n\n**The first is hostile forks.** The decomp community's open-source license allows anyone to fork existing projects, but Encounter points out that forks which then \"finish\" the project in one AI-fueled burst neither merge back upstream nor contribute anything to the original project — they just screenshot the progress number and post it on social media. \"Some projects don't want AI used for legal or ethical reasons; others are fine with AI but don't want drive-by AI-spliced PRs.\" Either way, these forks do not constitute community contribution; they tear at the collaboration fabric.\n\n**The second is racing to PC ports.** Decompilation is a research process about understanding how a game works, typically measured in years; building a PC port from already-decompiled source is a much smaller job. Encounter writes: \"Now anyone with an AI subscription can manage to create a port or static recompilation that functions at a basic level.\" The problem is that these ports are mostly \"half-baked, barely functional,\" but they ship much faster than the original decomp project. The result: \"the people who actually spent years decompiling the game get beat to release by someone who AI-glued a port together in a month or two.\"\n\n**The third is conceptual conflation.** For years the decomp community has tried to establish that \"decompilation is not the same as porting\" — decomp is research, port is consumption. Encounter directly calls out \u002Fr\u002Fdecomps and the \"sloptuber\" crowd for long using \"decomp\" to mean \"PC port,\" severely misleading newcomers about what decomp projects are, and making it harder and harder for upstream contributors to explain what they actually do.\n\n## Why Encounter chose \"shut the index\" rather than \"ban AI\"\n\nOne section of the statement directly answers the question \"why not just forbid AI-assisted projects?\" Encounter's answer is plain: he maintains the hobby project alone, has no energy to define what \"good decompilation\" means, and cannot enforce such a definition consistently across hundreds of projects. More importantly, he concedes that \"AI itself is not the issue, and used correctly it helps tremendously.\" The issue is \"slop,\" and slop comes in too many forms to enumerate; the definition keeps drifting.\n\nThis admission is the key to reading the whole statement. decomp.dev was not defeated by AI; it was crushed by AI on top of the contracts of community collaboration. Model capability is a neutral fact; whether a community can maintain the loop of \"contribute → review → merge → attribution\" under that capability is the real question. When \"tokens burned into a number\" starts replacing \"real human contribution\" as the metric of project progress, that loop breaks.\n\nAt the end of the statement Encounter also makes clear that **decomp.dev's own code will not be open-sourced** — any replica is forbidden from using it. He writes: \"decomp.dev's repository has never had an open source license, and I continue to reserve all rights to the (entirely human written) code.\"\n\n## A boundary drawn for open-source collaboration\n\ndecomp.dev's shutdown should not be read as a simple \"AI made one craft obsolete.\" A more accurate description: when a kind of work that originally relied on \"slow + collaborative + attributed\" to hold its value suddenly gets its cost compressed by \"fast + solo + anonymous\" capability, the open-source community built around it loses new contributors first, then loses its review mechanism, then loses its sense of belonging. Encounter shut down the main index early — in essence, stopping the already-warped \"contribution game\" before it rotted completely.\n\nThis is the first project in the 2026 open-source ecosystem to step down with an explicit \"AI dismantled the community-collaboration model\" rationale. The Linux kernel, AI bug bounty programs, and Hugging Face repositories have all shown early signals of the same pressure, but they are still responding by \"adding people, adding review, adding process.\" decomp.dev went straight to the most drastic response — shutting down the site. That path will be copied by more projects; it is worth every maintainer thinking ahead about: when an abstract \"collaboration\" is replaced by a concrete \"token count,\" which side are you trying to hold?\n\n## Key facts at a glance\n\n- decomp.dev launched in 2024, positioned as \"a collaborative center + progress leaderboard for game matching decompilation.\"\n- Encounter says early 2026's Opus 4.6 and GPT-5.4 were the early signals of \"being trained on decompilation specifically\"; GPT-6-Astra was the final push that \"ended\" the community model.\n- The main index and new-project intake are closed; historical project pages and data continue to be served read-only.\n- decomp.dev's code carries no open-source license; all rights are reserved by Encounter; any replica is forbidden from using it.\n- Original statement: https:\u002F\u002Fdecomp.dev\u002F (Chinese summary via Solidot: https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85638)","decomp-dev-shutdown-ai-community","2026-10-17T07:55:00Z","2026-10-18T03:05:09.537993Z","2026-10-18T03:05:09.538007Z",true,"agent",139,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"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":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"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":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"8657366a-cc97-4f40-98f4-6ddd091e5ea9","14.5 周、1 名工程师、80 万行 Rust:Copilot 智能体接管重写","copilot-runtime-rust-port-ai-agents-rewrite","2026-09-30T08:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"deac2d55-76a6-40d2-8ef7-36aed2ad0105","Linux 7.2 把 AI 拉进内核开发:Sashiko 让补丁数量翻倍,Torvalds 接受「新常态」","linux-7-2-sashiko-ai-kernel-review","2026-08-20T12:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"5736961c-9502-40d0-a39b-27f8ceb4380a","AI造AI:Recursive开源笔记本编程模型Flint","recursive-flint-laptop-coding-model","2026-10-22T19:05:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"6678ad05-308c-4440-beac-a1571f4c8059","DeepSeek 把搜索和抓取 API 开放:每千次 0.75 美元,Agent 工具栈拼图补齐","deepseek-search-fetch-api-075","2026-10-20T08:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"14a3cf67-03ef-4f85-a6a0-d809bfbfee04","Linux 内核不再要求 AI 辅助标注","linux-maintainers-drop-assisted-by-2026","2026-10-19T02:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00"]