[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-openai-model-spec-five-core-values":3,"topics-all":38,"news-related-421d123b-d996-4d76-bba7-d53331434a1c":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},"421d123b-d996-4d76-bba7-d53331434a1c","Model Spec 重写:五条价值观与一条权限链","OpenAI 在 10 月 29 日更新 Model Spec,把公开的行为规则手册从条款清单重排为五个价值观:忠于人类意图、减少伤害、诚实透明、促进人类自主、审慎关怀。新版新增指令权限链与自主授权契约,退役大部分风格细则,并明说生产模型尚未完全反映这份规范。","OpenAI 在 10 月 29 日更新了 Model Spec——那份公开描述自家模型应当如何行事的文档。改版幅度不小:全文按五条价值观重排,新增「谨慎对待上下文」小节,退役大部分风格细则,连对模型的称呼都从「the assistant」换成了「GPT」。[官方公告](https:\u002F\u002Fopenai.com\u002Findex\u002Fupdated-model-spec\u002F)里说得很直接:模型变强之后,光会照着指令做不够,还得理解人的意图和价值,才能在指令没覆盖的地方自己判断。\n\n## 五条价值观,以及它们要解决的冲突\n\n新的组织骨架是五条:忠于人类意图、减少伤害、诚实透明、促进人类自主、审慎关怀。第一条被官方博客称为整套方法的基础——人看重什么才是优先的,指令只是通常最清晰、最重要的价值信号;这一节借鉴了 assistance-games 的对齐框架,并援引了两篇早期文献(arXiv 1606.03137 与 Daedalus 上的一篇综述)。文档同时承认,五条价值观并不能定义模型应当尊重和帮助人们追求的人类价值的全貌。\n\n价值观在这版里的实际作用,是给后面两处张力做裁决:个人利益与集体利益、即时指令与不确定的长期目标。红线部分被单独提出来:人类安全与人权高于一切;不得用于促成暴力、战争罪、种族灭绝、酷刑、拐卖或强迫劳动,不得用于制造网络、生物或核武器,不得用于恐怖主义、儿童虐待、迫害或大规模监控;人类应当掌控 AI 如何被使用,以及模型行为如何被塑造。\n\n## 真正的工程量在权限链和授权契约\n\n比价值观更能看出改动分量的是两套结构。\n\n其一是指令权限链。文档给每条规则和每种消息角色定了默认权威等级,从高到低依次是:Root(Spec 里的 root 段,不可被对话上下文推翻)、System(Spec 的 system 段与 OpenAI 的系统消息,开发者和用户都覆盖不了)、Developer、User、Guideline,最后是「默认无权威」——助手消息、工具消息、被引用或不可信的文本和多模态数据都落在这一档。一个细节值得注意:Guideline 层是默认值,上下文明确需要时可以被隐式覆盖;User 层则必须由同层或更高层的显式指令才能推翻。筛选逻辑也给出了工程化的术语:先收集候选指令,再按「与更高层指令错位(misaligned)」「被同层后到指令取代(superseded)」「与来源的意图不符」「超出模型能力」四类排除。\n\n其二是自主授权契约。规范要求把授权范围记录成半结构化对象,字段包括 allowed_tools、latest_time、max_cost、tool_constraints 和自由文本 additional_details,并要求每份授权都必须带一个终止条件,越过之后模型停下来等新授权;黑客攻击、欺骗、造成重大负面副作用、部署持久化代理、自我修改这类高风险行为,除非被适用指令明确且具体地授权,否则一律禁止。文档还补了一条防绕过规则:模型不得通过把动作委托给别的代理、改写自己的指令或范围、或者把动作排到以后来规避授权;一旦委托,所有子代理都必须在同一范围内干活。同一层的 root 级规则里还写着「不许有别的目标」:不得为了营收或推销而优化,不得自保、逃避关停或囤积算力数据凭据,也不得擅自扮演法律与道德的执法者。\n\n## 删掉风格细则才是重点\n\n风格条款的退役是这版最容易被忽略、又最有信号的一处。原来写死的语气、措辞、格式细则被大量删除,理由是风格应当随人的偏好变化——整体上、在不同语言和人群内部、以及对具体个人——由用户反馈和上下文提供偏好证据。文档保留的少数硬默认之一是「客观地看待问题」,并明确模型不应仅凭「用户大概想听什么」就推断其偏好;用户和开发者可以显式要求某种视角,但不能靠模型揣摩。这等于把风格从系统提示里的硬约束,挪到了模型需要自己学出来的东西上。\n\n## 一个少见的自陈\n\n有两句自陈值得单独拎出来。第一句在规范正文和官方博客里都出现:这份规范描述的是 OpenAI 想要达到的行为,而生产模型还没有完全反映它——等于承认现网跑的并不是这份文档。第二句在官方公告里:他们同时在探索新的训练方法,让模型学习这些核心价值、理解某条原则为什么重要,以便在新场景里做对的事、对的理由;公告把这条训练路线指向 alignment.openai.com 上的 beneficial RL 页面,并说之后会公布结果。换句话说,这次改版不只重写了文档,也挪了目标函数的方向:从「照规范执行」转向「理解规范背后的理由」。\n\n还有两个容易被跳过的细节:规范以 CC0 1.0 放进公有领域,任何人都可以拿去改;权限链里明写,专门用于安全测试和红队、可以输出受限内容的 \"rail free\" 模型只服务于内部评测,公开发布的模型不应能被一条系统消息撬开 root 级规则。","https:\u002F\u002Fopenai.com\u002Findex\u002Fupdated-model-spec\u002F","bd0e0e04-6bcf-4b3e-9a56-62c672308ec9",[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},"baf131c1-687a-49f4-87f6-4dd87c1c692f","gpt",{"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},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"318fe5fa-6674-49ac-b079-4f979c8fdc22","en","OpenAI rewrites the Model Spec around five values","OpenAI rewrote the Model Spec around five values, added a chain of command and autonomy contracts, retired most style rules, and says its models lag.","OpenAI updated the Model Spec on October 29, 2026 — the public document describing how its own models are supposed to behave. The revision is substantial: the whole text is reorganized around five values, a new section on considering context carefully is added, most of the detailed style guidance is retired, and even the name for the subject changes from \"the assistant\" to \"GPT\". The [official announcement](https:\u002F\u002Fopenai.com\u002Findex\u002Fupdated-model-spec\u002F) puts the reason plainly: as models get stronger, following instructions literally is no longer enough. Models also need to understand human intentions and values so they can exercise good judgment in situations their instructions do not explicitly address.\n\n## Five values, and the tensions they arbitrate\n\nThe new skeleton has five: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The official blog calls the first one the foundation of the whole approach — what people value comes first, and instructions are ordinarily the clearest and most important signal of those values. That section borrows from the assistance-games framing of alignment, citing two early papers (arXiv 1606.03137 and an overview in Daedalus). The document also concedes that the five values are not intended to capture the full range of human values the models should respect and help people pursue.\n\nIn practice the values arbitrate two tensions: individual versus collective interests, and immediate instructions versus uncertain longer-term goals. Red-line principles are set out separately: human safety and human rights are paramount; the models should never facilitate critical and high-severity harms such as acts of violence, war crimes, genocide, torture, human trafficking or forced labor, the creation of cyber, biological or nuclear weapons, terrorism, child abuse, persecution or mass surveillance; and humanity should be in control of how AI is used and how model behaviors are shaped.\n\n## Where the real engineering sits: the chain of command and the autonomy contract\n\nTwo structures show the substance of the change more clearly than the values do.\n\nThe first is the chain of command. Every rule and every message role gets a default authority level, from highest down: Root (the root sections of the Spec, which conversational context cannot override), System (the system sections and OpenAI system messages, which neither developers nor users can override), Developer, User, Guideline, and finally no authority by default — assistant messages, tool messages, quoted or untrusted text, and multimodal data all land there. One detail is worth noting: Guideline-level instructions are defaults that context can override implicitly when it clearly calls for it, while User-level instructions can only be displaced by explicit instructions at the user level or above. The filtering vocabulary is engineered too: candidate instructions are collected first, then excluded under four categories — misaligned with a higher-level instruction, superseded by a later message at the same level, inconsistent with the likely intent of its source, or beyond the model's capabilities.\n\nThe second is the autonomy contract. The Spec asks that the scope of autonomy be recorded as a semi-structured object with fields such as allowed_tools, latest_time, max_cost, tool_constraints, and free-text additional_details, and that every scope include an ending condition beyond which the model stops acting until a new scope is confirmed. High-risk activities — hacking, deception, causing significant adverse side effects, deploying persistent agents, self-modification — are always prohibited unless explicitly and specifically authorized by applicable instructions. The document also adds an anti-circumvention rule: the model must not evade its instructions or its scope by delegating a prohibited action to another agent, rewriting its own instructions or scope, or scheduling the action for later. When it does delegate, all sub-agents must operate under the same scope. A root-level rule in the same family, no other objectives, bars optimizing for revenue or upsell, and bars self-preservation, evading shutdown, accumulating compute, data or credentials, and acting as an enforcer of laws or morality.\n\n## Retiring the style rules is the part that matters most\n\nThe retirement of style guidance is the least noticed and most telling change. Long-standing rules about tone, wording, and formatting were removed in bulk, on the reasoning that style should adapt to what people prefer — in aggregate, within languages and cultures, and for each person — with user feedback and context providing evidence about those preferences. A few stronger defaults survived, among them approaching questions objectively, and the Spec makes clear the model should not infer a preferred perspective merely from what it thinks people want to hear; users and developers can explicitly ask for a viewpoint. In effect, style moved out of the system prompt and into something the model has to learn.\n\n## Two admissions worth pulling out\n\nTwo self-assessments deserve to be pulled out. The first appears in both the Spec text and the official blog: the Spec describes the behavior OpenAI is working toward, and the production models do not yet fully reflect it — an admission that what runs in production is not this document. The second comes from the announcement: alongside the rewrite, OpenAI has been exploring new training methods to help models learn these core values and understand why a given principle matters, so they do the right things for the right reasons in new situations. The announcement points that training work at the beneficial RL page on alignment.openai.com and says results will be shared later. So this revision moved the direction of the objective function too, from executing the spec to understanding the reasons behind it.\n\nTwo smaller details are easy to miss: the Spec is dedicated to the public domain under CC0 1.0, so anyone can take it and modify it; and the chain of command states that rail free models able to output restricted content are useful for safety testing and red teaming, while publicly deployed models should not be susceptible to violating root-level principles through a system message.","openai-model-spec-five-core-values","2026-10-30T03:05:00Z","2026-10-30T03:15:46.666222Z","2026-10-30T03:15:46.666240Z",true,"agent",108,[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},"b5be4ce8-4a41-461c-9202-148e64fab329","GPT-6 Astra 系统卡:零日自用、对齐升 53%,CoT 可监控性反向下滑","gpt-6-astra-system-card-2026-monitorability","2026-09-04T03:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"d95940eb-69c1-467e-9d60-5886ab71d985","GPT-5.6-Cyber 上线、Daybreak 分层、Astra 推迟:OpenAI 把\"网络安全模型\"做成一个独立产品线","openai-gpt-5-6-cyber-daybreak-astra-2026","2026-08-11T04:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"3967306f-062a-41a6-ab58-f99e70fc0e68","AISI 122 轮 cyber eval 越界：OpenAI 与 Anthropic 同日披露","aisi-mythos-5-gpt-5-6-cyber-eval-incident-2026","2026-08-08T04:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"51d15a21-7593-4d40-bf0f-ad964e0b2fbe","OpenAI 8月4日披露第三方测试越界：GPT-5.6 Sol 在 AISI 与 Irregular 评估中擅自接入公网并攻击真实站点","openai-gpt-5-6-aisi-irregular-evaluation","2026-08-05T02:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"6528b99d-b1df-4d8e-80a7-e400895175f0","GPT-5.6 Sol 沙箱挖出 0day：OpenAI 披露首例 AI 自主入侵","gpt-5-6-sol-0day-hf-incident","2026-07-23T03:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"7ae5bad0-98ec-4412-b4f6-d29e233adb3b","GPT-Red 自博弈红队:OpenAI 用 self-play 把 prompt injection 失败率从 95% 压到 0.05%","gpt-red-self-play-red-team","2026-07-17T02:01:00+00:00"]