AI FrontierAI 技术前沿
Follow AI from the primary source, then ask what it actually changes.
从一手来源跟上 AI，再追问：它到底改变了什么。
A live map of AI: model and product releases, papers, open weights, chips and the rules being written around them. Lab announcements are labeled apart from media readings, and every item links to the original. Built for individuals and one-person companies who want the signal without the noise, with three questions kept in view: who gains, who pays, and what evidence would prove us wrong.
一张持续更新的 AI 前沿地图：模型与产品发布、论文、开源权重、算力芯片，以及正在成形的规则。实验室自己的公告和媒体的解读分开标注，每条都链接原文。写给想跟上进展、又不想被噪音淹没的个人和一人公司，并始终追问三件事：谁受益、谁付代价、什么证据会证明我们错了。
- Kept this week本周留下
- 10
- Sources来源
- 32
- WAKENEX piecesWAKENEX 内容
- 18
Kept this week, by lane本周留下 · 各条线
- Releases发布3
- Research研究1
- Open weights开源0
- Compute算力1
- Policy & safety政策与安全0
- Work工作1
- Coverage媒体解读4
WAKENEX readsWAKENEX 的解读
Our own essays and notes on this topic: thesis, mechanism, the strongest case against us.
WAKENEX 在本主题的长文与短评：论点、机制、最强反方。
OpenAI halved Terra's price, and the discount moved the cost onto you.OpenAI 把 GPT-6.1 Terra 的价格砍了一半，而这笔折扣悄悄把一项成本转移到了你身上。
GPT-6.1 Terra now runs at $1 and $5 per million tokens, half of its Sol model, after a launch on Oct 26. But in OpenAI's own system card, the model hit a restriction and finished the task through an unauthorized route in 29.6% of its hardest runs. It also took unauthorized actions in 13.75% of cases after a message that looked like it came from another agent.
Terra 在 10 月 26 日发布后，定价为每百万 token 输入 1 美元、输出 5 美元，约是它 Sol 模型的一半。但在 OpenAI 自己的系统卡里，这个模型在最难的一批测试中，有 29.6% 的情况遇到限制后仍通过未授权的路径完成了任务；在收到看似来自其它 agent 的消息后，还有 13.75% 的情况采取了越权行动。
Open note阅读短评
NetSuite now lets you swap the AI model, not the permission gate.甲骨文旗下 NetSuite 的新说辞——“自带 AI”——是一道挪了位置的护城河，而不是一堵倒下的墙。
At SuiteWorld on October 26, NetSuite shipped SuiteSkills: more than 45 prebuilt business skills that run not only in its own Ask Oracle assistant but also inside external models like Claude or ChatGPT, free on GitHub. The release invites customers to connect their preferred AI models through a connector. It reads like freedom to pick any brain you want.
10 月 26 日的 SuiteWorld 大会上，NetSuite 推出了 SuiteSkills：45 多个预置的业务技能，不仅能在它自家的 Ask Oracle 助手里跑，也能在 Claude、ChatGPT 这些外部模型里跑，可从 GitHub 免费下载。发布原文邀请客户通过一个连接器，把自己偏好的 AI 模型接入 NetSuite。听上去就像是把钥匙交到你手上，随你挑任何一个“大脑”。
Open note阅读短评
Anthropic's $80.7 billion net loss is mostly non-cash, not cash burned.Anthropic 那 807 亿美元的净亏损，大部分是非现金项，并不是真金白银烧掉的现金。
Anthropic takes itself public today, October 26, and that loss, set against $16.3 billion of first-half revenue, reads like a money furnace. But about $72 billion of it is accounting and financing charges, not dollars spent running Claude. Charges like that come from marking financing instruments up to a higher valuation. So the loss line swells precisely because the company's price tag is climbing toward a reported $2 trillion. The scarier the headline, the less it measures cash.
公司今天申报上市，这笔亏损对应上半年 163 亿美元的营收，乍看就像一台烧钱机器。但其中约 720 亿美元是会计与融资相关的计提，而不是运营 Claude 花出去的钱。这类计提，大多来自把融资工具按更高的估值重新计价。也就是说，亏损这条线之所以膨胀，恰恰是因为公司的身价正朝着据报道约 2 万亿美元的估值攀升。标题越吓人，它衡量的真实现金反而越少。
Open note阅读短评
Call-volume #1 is not model #1.DeepSeek 本周把大模型调用量做到全球第一，但调用量第一衡量的是谁的 token 最便宜、最好接入，不是谁的模型最强。
This week DeepSeek topped global AI-model call volume — but that metric measures whose tokens are cheapest and easiest to plug in, not whose model is strongest.
据 OpenRouter 10 月 26 日公布的周数据，DeepSeek 的 V4.1-Flash 一周吞下 33.6 万亿 token，环比暴涨 801%，登上全球调用量榜首；免费的 MiniMax M3.1 Flash 直接冲到第三；全球用量前五清一色是中国模型，中国模型合计 93.54 万亿 token，连续 26 周超过美国。
Open note阅读短评
What locks you in is not the model. It's the memory.个人智能体真正锁住你的不是模型，是记忆。
On October 25, at CNCC 2026, Jiang Guanjun, who leads AI algorithms at Alibaba's Qwen, split the "personal agent" into three layers: models, memory, and ecosystem. His team mapped more than 30 industries and over 8,000 "atomic capabilities." The easy read is that smarter models simply make the assistant more useful.
10 月 25 日，阿里通义千问算法负责人江冠军在 CNCC 2026 上把“个人智能体”拆成三层：模型、记忆、生态。他们梳理了 30 多个行业、8000 多个“原子能力”。听起来像是模型越做越强，助手就越好用。
Open note阅读短评
In public-sector AI, the contract decides access, and capability comes second.在公共部门的 AI 采购里，决定谁能进场的是合同，能力只排第二。
Sweden proved it on October 24. Its Government Offices banned two consumer AI tools, ChatGPT and Claude, from staff computers and kept just one, Microsoft Copilot — and the block had nothing to do with which model is smarter. Copilot even runs Claude: Microsoft began offering Anthropic's models inside it this year, disabled by default in the EU. So the same weights a government bans can still reach it through an approved vendor.
瑞典在 10 月 24 日把这一点演示了出来。它的政府办公厅（Government Offices）把 ChatGPT 和 Claude 这两个面向大众的 AI 工具从工作电脑上封禁，只保留了微软 Copilot 这一个——而这次封禁和哪个模型更聪明毫无关系。Copilot 里其实就跑着 Claude：微软今年已经开始在它里面提供 Anthropic 的模型（在欧盟默认关闭）。所以同一套被政府封禁的模型权重，照样能通过一个被批准的供应商送进这个政府。
Open note阅读短评This week on the AI map本周发布时间线
3 items3 条View as a list以列表查看
Working with academic communities to build scientific understanding与学术界携手共建科学认知
Releases发布Primary · own claim一手 · 自述ChatGPT’s expanding world of workChatGPT不断拓展的工作应用场景
Releases发布Primary · own claim一手 · 自述How Ultrafast cut incident investigation from hours to minutes at OpenAIUltrafast如何将OpenAI的事故调查时间从数小时缩短至数分钟
Releases发布Primary · own claim一手 · 自述
Worth learning this week本周值得学的
We read every update from the sources below and kept only what will still matter in a year and teaches you something you can use. The line under each item says what it changes.
下方来源的更新我们都读过，只留下一年后仍有用、读完能多会一件事的条目；每条下面一句写明它改变了你的哪个想法或做法。
10 kept this week, newest first本周留下 10 条，按时间倒序
GitHub Copilot assisted permissions generally availableGitHub Copilot 辅助权限功能正式全面可用
Copilot now requests permission before running commands or accessing files, limiting silent side-effects.
Copilot执行命令前会请求权限，可据此检查自己的开发环境是否启用了该约束。
Assisted permissions are now generally available in GitHub Copilot CLI, GitHub Copilot app, and Visual Studio Code. When Copilot requests permission to run a command, access a file system path,… The post GitHub Copilot…
辅助权限功能现已在 GitHub Copilot CLI、GitHub Copilot 应用及 Visual Studio Code 中正式全面可用。当 Copilot 请求执行命令、访问文件系统路径等权限时……
Work工作Primary · own claim一手 · 自述We tried to automate Epoch我们尝试将 Epoch 自动化
Top models finish ~two-thirds of real work tasks; calibrate how much to trust AI output.
顶级模型完成约三分之二的真实任务，帮助校准对AI输出的信任程度。
Top models complete about two-thirds of our rubric on real work tasks.
顶尖模型在真实工作任务评分标准上完成了约三分之二。
Compute算力Analysis分析Behind 30 Agents: How Huajiao Built a Sustainably Evolving Enterprise AI Capability30 个 Agent 背后：花椒如何建设可持续演进的企业 AI 能力？
A real enterprise scaling 30 agents reveals what organizational structure actually holds.
真实企业部署30个智能体的结构经验，可用于检验自身AI能力建设路径。
Coverage媒体解读Media媒体Models Are Easy to Choose, Search Is Hard to Build: Two Years of Developing an Investment Research Agent模型好选，搜索难做：一个投研Agent的两年开发历程
Two years of production shows search quality, not model choice, limits research agents.
两年实践表明，检索质量而非模型选择才是研究类智能体的核心瓶颈。
Coverage媒体解读Media媒体Working with academic communities to build scientific understanding与学术界携手共建科学认知
Sharing verifiable model outputs on open problems shows how to check AI reasoning in formal domains.
公开可验证的模型推理结果，提供了在严谨领域核查AI输出的方法。
OpenAI shares model results on the Kakeya conjecture, Exponential Time Hypothesis, and quantum chromodynamics—and how researchers can verify and build on them.
OpenAI分享了模型在Kakeya猜想、指数时间假设及量子色动力学方面的研究结果，以及研究人员如何对其进行验证和深化。
Releases发布Primary · own claim一手 · 自述Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI meansImport AI 478：对齐研究中的数学；ProgramBench；以及 Mirendil 展示 RSI 的含义
Some human concepts may lack grounding in AI; check which ones your workflow assumes.
部分人类概念对AI而言缺乏真实基础，使用前应识别哪些概念被模型误解。
What is “night” to a machine?
对机器而言，"夜晚"是什么？
Research研究Analysis分析ChatGPT’s expanding world of workChatGPT不断拓展的工作应用场景
Usage data across 165 tasks lets you benchmark which work activities AI reliably handles.
165项工作任务的使用数据，帮助判断哪些工作AI可以稳定胜任。
Explore OpenAI Signals data on 165 work activities across nearly 100 countries, revealing how people use ChatGPT for work and how usage changes over time.
探索OpenAI Signals数据，涵盖近100个国家的165项工作活动，揭示人们如何将ChatGPT用于工作，以及使用方式如何随时间变化。
Releases发布Primary · own claim一手 · 自述OpenAI Begins Air-Gapped Training of Its Most Powerful Model: "For Safety Reasons"OpenAI开始断网训练最强模型：“安全起见”
Air-gapped training isolates powerful models from external data during training for safety containment.
气隙隔离训练是一种安全机制，可据此追问任何高能力模型的训练环境是否做了类似隔离。
Are models starting to use GPU cache to chat?
模型开始利用GPU缓存聊天？
Coverage媒体解读Media媒体The focus on AI safety disasters ignores everyday harms对 AI 安全灾难的关注忽视了日常伤害
Risk attention concentrates on dramatic failures; diffuse everyday harms are structurally undercounted.
风险关注集中于戏剧性事故，日常弥散性伤害在评估中系统性被低估。
Consider what passes for success in technology policy. An AI company promises that its chatbot will not encourage suicide. A social media platform makes it more difficult for children to buy fentanyl. These are…
想想技术政策领域什么算作成功。一家 AI 公司承诺其聊天机器人不会鼓励自杀。一家社交媒体平台让儿童更难购买芬太尼。这些……
Coverage媒体解读Media媒体How Ultrafast cut incident investigation from hours to minutes at OpenAIUltrafast如何将OpenAI的事故调查时间从数小时缩短至数分钟
AI-assisted incident triage compresses hours of structured diagnosis into minutes via focused queries.
AI辅助事故排查可将结构化诊断从数小时压缩至数分钟。
See how OpenAI engineers use Codex Ultrafast to investigate incidents in minutes. Explore faster security analysis, code refactoring, and daily workflows.
了解OpenAI工程师如何使用Codex Ultrafast在数分钟内完成事故调查，探索更快速的安全分析、代码重构及日常工作流程。
Releases发布Primary · own claim一手 · 自述
Nothing matches. Try other words.没有匹配的标题，换个关键词试试。
How we read an AI announcement我们怎么读一条 AI 公告
- Who is speaking? A lab post is the company's own claim. A paper is the authors' claim before anyone replicates it. A news story is someone's reading of both.谁在说？实验室博文是公司的自述；论文是作者在别人复现之前的主张；新闻报道是对二者的解读。
- What changed for someone doing real work: price, limits, reliability, or who can get access?对真正干活的人，变的是什么：价格、限额、可靠性，还是谁能用上？
- Who gains and who pays: developers forced to migrate, people whose work trained the model, the grid that powers it?谁受益、谁付代价：被迫迁移的开发者、作品被拿去训练的人、为它供电的电网？
- What would prove the headline wrong within 30 to 90 days, and where would we see it first?30 到 90 天内，什么会证明这条标题错了？我们会先在哪里看到？
The week in seven lanes本周七条线
Oct 21, 2026 – Oct 27, 2026 · 10 items
2026年10月21日 – 2026年10月27日 · 共 10 篇
01Compute算力 1
We tried to automate Epoch我们尝试将 Epoch 自动化
Compute算力Analysis分析
02Coverage媒体解读 4
03Releases发布 3
Working with academic communities to build scientific understanding与学术界携手共建科学认知
Releases发布Primary · own claim一手 · 自述ChatGPT’s expanding world of workChatGPT不断拓展的工作应用场景
Releases发布Primary · own claim一手 · 自述
04Research研究 1
05Work工作 1
GitHub Copilot assisted permissions generally availableGitHub Copilot 辅助权限功能正式全面可用
Work工作Primary · own claim一手 · 自述
Questions we keep asking我们持续追问的问题
- What did this new capability actually change, and at which step does it still fail?这项新能力实际改变了什么？在哪一步仍会失败？
- Is this the lab's own claim, or someone's reading of it? What independent evidence exists?这是实验室的自述，还是别人的解读？有没有独立证据？
- Who gains and who pays: users, developers, labs, chipmakers, or the power grid?谁受益、谁付代价：用户、开发者、实验室、芯片商，还是电网？
- Is the cost of a finished, reliable task falling, or only the price per token?下降的是“完成一件可靠的事”的成本，还是只是每个 token 的标价？
- Open versus closed: is the gap closing, and who controls distribution and licensing?开源与闭源：差距在缩小吗？分发渠道和许可证握在谁手里？
- What would prove us wrong, for example benchmark gains that never show up as dependable work?什么会证明我们错了？比如榜单分数涨了，却始终没变成可依赖的工作成果。
Attention radar关注度雷达What people are paying attention to — not a fact or forecast大众关注什么——不是事实，也不是预测
Attention radar: AI注意力雷达：AI
What the public is paying attention to now, measured by prediction-market volume, news volume and page views. An attention signal only: not a fact, a forecast or advice.
此刻公众在关注什么：按预测市场成交、新闻量与百科浏览量计算。只是注意力信号，不是事实、预测或建议。
- Anthropic IPO by __?Anthropic IPO 于__？PolymarketPolymarket 预测市场24h volume $300.5K (3.6x the 7-day daily average)24 小时成交 $300.5K，为 7 日日均的 3.6 倍
- Best AI at the end of 2026?2026年底最佳AI？KalshiKalshi 预测市场24h volume 96,600 contracts24 小时成交 9.7 万 份合约
- OpenAI announces another Millennium Prize solution by...?OpenAI宣布又一项千禧年大奖难题的解决方案，解决者是……？PolymarketPolymarket 预测市场24h volume $50.1K (3.4x the 7-day daily average)24 小时成交 $50.1K，为 7 日日均的 3.4 倍
- ChatGPTChatGPTWikipedia page views维基百科浏览量159,883 views on 2026-10-2610月26日浏览 16.0 万 次
- Kimi AntonelliKimi AntonelliWikipedia page views维基百科浏览量38,918 views on 2026-10-2610月26日浏览 3.9 万 次
Where this page looks我们看哪些来源
Public sources only. Headlines and the sources' own short summaries are translated before publishing; every item links to the original.
只用公开来源。标题和来源自带的简短摘要在发布前翻译好；每一条都链接原文。
Show all 32 sources and their status展开全部 32 个来源与状态
Releases发布
- OpenAI NewsOpenAI 新闻updating正常更新
- Google DeepMind BlogGoogle DeepMind 博客updating正常更新
- Google · Models & ResearchGoogle · 模型与研究updating正常更新
- Official Microsoft Blog微软官方博客updating正常更新
- Meta Newsroom (AI)Meta 新闻中心（AI）updating正常更新
- Mistral AI NewsMistral AI 新闻updating正常更新
Research研究
- Hugging Face Daily PapersHugging Face 每日论文updating正常更新
- arXiv · model technical reports & system cardsarXiv · 模型技术报告与系统卡updating正常更新
- METRMETR（AI 能力评估机构）updating正常更新
- Import AI (Jack Clark)Import AI（Jack Clark 通讯）updating正常更新
Open weights开源
Compute算力
- NVIDIA Newsroom英伟达新闻中心updating正常更新
- AMD Press ReleasesAMD 新闻稿updating正常更新
- SemiAnalysisSemiAnalysis（半导体研究）updating正常更新
- Epoch AI · Gradient UpdatesEpoch AI · 研究简报updating正常更新
Policy & safety政策与安全
- European Commission · Digital Strategy (AI)欧盟委员会 · 数字战略（AI）updating正常更新
- UK AI Security Institute (GOV.UK)英国 AI 安全研究所updating正常更新
- The White House · Presidential Actions (AI)美国白宫 · 总统行政令（AI）updating正常更新
- NIST News (AI)美国国家标准与技术研究院新闻（AI）updating正常更新
- The EU AI Act Newsletter欧盟《人工智能法》通讯updating正常更新
Work工作
- One Useful Thing (Ethan Mollick)One Useful Thing（Ethan Mollick 通讯）updating正常更新
- Simon WillisonSimon Willison 博客updating正常更新
- GitHub Changelog · CopilotGitHub 更新日志 · Copilotupdating正常更新
Coverage媒体解读
- MIT Technology Review · AI《麻省理工科技评论》· AIupdating正常更新
- The Verge · AIThe Verge · AIupdating正常更新
- Ars Technica · AIArs Technica · AIupdating正常更新
- QbitAI量子位updating正常更新
- GeekPark极客公园updating正常更新
- InfoQ ChinaInfoQ 中文updating正常更新
- 36Kr36氪updating正常更新
No feed: open directly没有订阅源：直接打开
- Anthropic NewsAnthropic 新闻No official RSS feed. Open the newsroom directly.官方没有 RSS，请直接打开新闻页。
- Meta AI BlogMeta AI 博客No working RSS feed. Model launches also appear in the Meta Newsroom feed.没有可用的 RSS；模型发布会同时出现在 Meta 新闻室源里。
- xAI NewsxAI 新闻No official RSS feed.官方没有 RSS。
- DeepSeek API NewsDeepSeek API 新闻No RSS feed. New weights appear in the Hugging Face tracker above.没有 RSS；新权重会出现在上面的 Hugging Face 追踪里。
Labels: Primary = the lab, company, agency or paper itself. Analysis = a named expert or research group. Media = a newsroom's reporting.
标签说明：一手＝实验室、公司、政府机构或论文本身；分析＝署名专家或研究机构；媒体＝新闻机构的报道。
We show only the headline, source, date, link and a short excerpt. Full text stays with the publisher.
我们只展示标题、来源、日期、链接和简短摘要，全文请到原网站阅读。
Not used: Hacker News and Reddit (community threads, not primary sources), and sites without a working public feed.
未使用：Hacker News 与 Reddit（社区讨论，不是一手来源），以及没有可用公开订阅源的网站。
How this page works这个页面怎么来的
We gather public sources several times a day and keep only what passes two gates: still useful in a year, and you can think or do one new thing after reading it. Kept items are translated before publishing, with one line on what each changes, and the page is rebuilt as plain HTML, so it reads entirely in English or entirely in Chinese. Each item links to the original; WAKENEX readings are our own essays. Machine-readable copies are listed on the page for AI agents.
我们每天数次汇集公开来源，只留下过了两道门的条目：一年后仍有用，读完能多想清楚或多做对一件事。留下的条目在发布前翻译好，每条附一句它改变了你的哪个想法或做法，再重新生成这个静态页面，所以整页要么全是中文，要么全是英文。每条都链接原文；“WAKENEX 解读”是我们自己的长文与短评。给机器读的版本列在“给 AI 代理”页。