今日头条
刚刚，Anthropic公开335页招股书：史上最大IPO，史上最亏AI公司
Anthropic 正式向 SEC 公开提交纳斯达克 IPO 申请，股票代码 ANT，目标估值约 1.5 万亿至 2 万亿美元；招股书披露对微软采购承诺由 314 亿美元增至 449 亿美元、对谷歌 1145 亿美元、与 SpaceXAI 算力合同最高 1205 亿美元。
首家纯大模型公司冲刺史上最大 IPO
2026-10-29·周四
中文热点点标题展开要点
-
Anthropic巨亏背后，藏着行业什么秘密？ 招股书显示 Anthropic 2026 年上半年营收 162.86 亿美元、同比增长约 1300%，净亏损却达 807.48 亿美元；对比 SpaceX 同期靠星链贡献 85% 营业利润，模型公司成本结构直面推理与刚性算力承诺。
据钛媒体报道SpaceX 于 10 月 20 日发布第三季度财报，营收 145.75 亿美元、净利润 44.04 亿美元，为上市后首次季度盈利。Anthropic 由 Dario Amodei 与妹妹 Daniela 及另外五名 OpenAI 核心成员于 2021 年初创立，Dario 曾参与 GPT-2、GPT-3 研发，是 RLHF 共同发明人之一。
揭示模型公司估值依赖算力供应商
-
DSH测试文件“泄密”，DeepSeek V5就要来了？ DeepSeek Harness 官方 GitHub 仓库测试文件 provider.spec.ts 中出现 deepseek-v5、deepseek-v5-open、deepseek-v5-next 三个未公开模型名，分别对应免更新可用、全量开放、需升级客户端三种发布分支。
据钛媒体报道基础版 deepseek-v5 未声明客户端版本要求，当前内置目录不认识它，调用时返回 MODEL_UNSUPPORTED；deepseek-v5-open 最低版本 0.0.1，任何客户端可用；deepseek-v5-next 要求 99.0.0-alpha.1，而公开 DSH 版本仍为 0.2.x，该版本号用于稳定触发版本不足分支。
V5 发布节奏提前泄露给开发者
-
小米披露灵巧手最新技术探索成果：基于1827个工厂SOP验证，升级21自由度仿生灵巧手 小米基于汽车工厂 1827 个标准作业程序采集分析，称 17~21 自由度区间仿生灵巧手平均稳定抓握成功率较 6~15 自由度提升 15%~38%，据此设计 21 自由度混合驱动加腱绳张力控制方案，并称自由度升至 23 以上边际收益明显放缓。
据量子位报道该灵巧手同时对散热体系和全掌触觉做了系统性优化。小米称数据来自小米汽车工厂柔性制造场景的完整采集与分析验证，21 自由度在工厂操作覆盖度和抓握成功率上提升最大。
为具身智能硬件定下场景化路线
-
腾讯AI压力来到微信小微 腾讯开源混元同声传译模型 Hy-SimulMT，平均响应时间 3 秒；2026 年 10 月 27 日当周混元 Hy4 preview 在 OpenRouter 调用量达 9 万亿 tokens 排名第五，而 2025 年混元系列尚未进入该榜单。
据钛媒体报道在 2026 年 10 月 27 日当周 OpenRouter 榜单上，混元 Hy4 preview 排在 DeepSeek V4.1 Flash 1017、DeepSeek V4.1 Flash、智谱 GLM-5.3 Flash 和 MiniMax M3.1 Flash 之后。腾讯还在推进元宝、WorkBuddy，并已被曝出个人智能体产品 Handy Bot。
腾讯大模型调用量与产品节奏追赶
-
Meta 利用私有处理技术将机密计算扩展至 AI 眼镜 Meta 将私有处理扩展至 AI 眼镜，用云端机密虚拟机执行敏感推理，结合 TEE、远程认证、匿名凭证、OHTTP 与加密存储，支持流式转录、上下文搜索和长期记忆召回，Meta 无法访问明文。
据InfoQ 中文报道CVM 边界覆盖 CPU 与 GPU，混合工作负载全程受 TEE 保护；上下文数据加密存储于云端。通话、消息等基础交互可在眼镜端运行，更复杂负载需更大模型并访问随时间积累的个人上下文。Meta 于 2025 年先面向 WhatsApp 和 Meta AI 推出该功能。
为可穿戴 AI 隐私架构立标杆
-
Instagram 采用原生 AI 架构对用户界面进行了现代化改造，将 token 成本降低了 33% Instagram 将 Android 端 Direct UI 全量迁移至 Jetpack Compose，构建 AI 原生代码库，代码量减少 50%，每个 AI 代理会话 token 成本降低 33%，并强调用架构设计约束 AI 代理走最小阻力路径。
据InfoQ 中文报道迁移采取整体架构评估而非逐组件零敲碎打，重点判断底层架构是否便于生成正确实现。工程师指出 AI 工具倾向走阻力最小路径，可能带来隐蔽 Bug、技术债务和性能退化，因此以架构设计约束代理行为，而非用 AI 弥补不良设计。
给 AI 编码代理的工程化提供样本
-
当 AI 智能体代表用户行动时，企业治理将面临哪些新问题？ ｜ 技术趋势 Snowflake 提出智能体五维控制框架：识别智能体身份、权限最小化、保护敏感数据、限制资源消耗、审计行为，并支持对 Cortex Agents 等主流智能体会话自动识别与上下文透传，指出智能体易跨工具聚合泄露敏感数据。
据InfoQ 中文报道五维框架对应五个问题：智能体是谁、能做什么、能看到什么、能消耗多少资源、做过什么。智能体身份需独立于用户会话识别，权限按最小必要原则动态授予；风险在于智能体可能汇总敏感数据、传给其他工具或呈现给无权访问者。
企业部署智能体前须先解决权限治理
-
Anthropic 工程经理把传统编程面试砍了，“它不再提供有效信号” Anthropic 工程经理 Felix Rieseberg 称本月已将传统编程面试从其团队流程中移除，理由是现场写 React 小挑战与考汇编语言一样失去信号价值，因其管理的 Claude.ai、Cowork、Claude Code Desktop 已能完成此类任务。
AI 编码能力倒逼招聘标准重写
-
字节辟谣"群演数据喂大模型"：严重失实！ 针对“横店将 15 万名群演肖像和声音喂给字节大模型”传言，字节豆包公关负责人刘星回应严重失实：火山引擎仅为横店影视 AI 服务平台技术支持方，数据资产归横店或授权方所有，个人肖像声音使用须本人授权。
据雷锋网报道刘星称项目初衷是建设影视行业公共数字基础设施，将影视基地实景等资产做成高质量数据集，帮助剧组选景、筹备和拍摄，降低中小影视企业使用 AI 的门槛。该项目目前处于中标阶段，将围绕影视行业数据集建设、配套技术平台展开。
数据授权合规是影视 AI 落地红线
English Pickstap a headline to expand
-
Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity Google released Gemini 4 Carbon, claiming significant gains in coding, agentic tasks, and cybersecurity, and outperforming OpenAI and Anthropic models on some benchmarks.
Frontier model race shifts competitive benchmarks for developers and security teams.
-
Microsoft recorded $10 billion in revenue from OpenAI Microsoft reported $10 billion in revenue from OpenAI commercial arrangements, including revenue sharing, nearly 11% of its $92.3 billion quarterly revenue.
据TechCrunch AI报道The $10 billion figure covers the quarter ended September 30 and was disclosed in a regulatory filing on Wednesday. It includes revenue-sharing payments from the OpenAI commercial arrangements.
Reveals how deeply Microsoft's earnings now depend on OpenAI.
-
Microsoft just shared Azure’s actual revenue in dollars: $33.2 billion Microsoft disclosed Azure quarterly revenue of $33.2 billion in dollars for the first time, showing how close it is to AWS.
据TechCrunch AI报道The disclosure came Wednesday with Microsoft's quarterly earnings. Previously the company shared only Azure growth percentages or described revenue dollar milestones in vague terms.
First hard cloud number reshapes AI infrastructure market comparisons.
-
Google expands AI Search ads globally as core ad growth slows Google expanded AI Search ads to international markets during Alphabet's Q3, when revenue rose 25% year over year to $127.6 billion as core ad growth slowed.
据TechCrunch AI报道The international expansion was announced Wednesday during Alphabet's third-quarter earnings. The company is betting generative AI ads open new revenue as core search ad growth slows.
Signals monetization shift affecting advertisers and publishers globally.
-
Chrome’s AI-powered Autofill can now dig through your Gmail to fill out forms Chrome's upgraded Autofill pulls flight confirmations, tracking numbers, order details, and membership IDs from Gmail with user permission to fill forms.
据TechCrunch AI报道Announced Wednesday, the advanced Autofill suggests data from Gmail including flight confirmation numbers, membership IDs, tracking information, order numbers, addresses, and phone numbers, with user permission.
Raises privacy questions for Gmail users and rival browsers.
-
GitHub Octoverse 2026 GitHub Octoverse 2026: merged pull requests tripled to 220 million monthly, commits topped 7.38 billion monthly, and new repositories doubled to 30.3 million monthly.
据Simon Willison报道EVP Jay Parikh announced the figures at GitHub Universe. Commits more than tripled to 7.38 billion per month, with most of the increase in the second half of 2026; new repository rate doubled to 30.3 million per month.
Quantifies how agentic coding is reshaping software development workflows.
-
Will the real Anthropic please stand up Anthropic filed a 335-page IPO prospectus, roughly 100 pages of which cover risks, laying out its view of how AI will change the world.
据MIT Tech Review报道Cofounder and head of policy Jack Clark is cited for the company's public messaging. The prospectus follows weeks of speculation about the filing.
First look at Anthropic's finances and risk disclosures for investors.
-
Meta says it’ll probably delay its controversial new business terms again Meta may delay its new business terms past November 20th amid publisher backlash over AI training concerns, saying it won't change policy without full industry consultation.
据The Verge报道Meta said it does 'not plan to make policy changes without a full consultation with industry groups and leaders.' The terms were previously set to take effect November 20th.
Affects publishers and content owners whose data trains AI.
-
OpenAI and ServiceNow expand partnership across enterprise workflows OpenAI and ServiceNow expanded their partnership to bring OpenAI models and agents into ServiceNow and ServiceNow workflows into ChatGPT and Codex.
Brings agentic AI deeper into enterprise IT service workflows.
-
Microsoft is ending Copilot support on smart TVs Microsoft is ending Copilot support on smart TVs, nearly two years after LG and Samsung announced CES 2025 integration; an LG TV notification confirms the shutdown.
据The Verge报道A notification on LG TVs states Copilot functionality is going away on all TVs. LG and Samsung announced CES 2025 integration nearly two years ago.
Shows retreat from consumer AI hardware partnerships, affecting TV owners.