📰 AI 周报 | 2026年11月03日 16:04
本期共 36 条新资讯，覆盖 5 大领域
数据来源：TechCrunch, The Verge, THE DECODER, AlphaSignal, smol.ai, 量子位, OpenAI, Anthropic, DeepMind, Wired, 404 Media, VentureBeat 等
📌 产业与商业（9 条）
- Google takes Dreambeans to India for AI-generated personal storytelling
来源：TechCrunch | 2026-11-02
Google expands its Dreambeans AI storytelling platform to India, targeting mass-market personal narrative generation in local languages. - Apple may monetize third-party AI app revenue sharing
来源：TechCrunch | 2026-11-02
Apple reportedly explores new revenue-sharing models for third-party AI applications on its platform, potentially mirroring App Store commission structures. - Omen AI acquires Aquatic Labs for $16M, adding liquid cooling expertise
来源：TechCrunch | 2026-11-02
AI infrastructure startup Omen AI acquires liquid-cooling specialist Aquatic Labs for $16M, targeting next-gen thermal management for AI data centers. - OpenAI pitches data center power plants as grid protectors
来源：The Verge | 2026-11-02
OpenAI launches ‘Grid Protectors Program,’ reframing data center power infrastructure as beneficial to local grids rather than burdensome polluters. - Elon Musk throttles Grok Bot as AI costs spiral across the industry
来源：The Verge / xAI | 2026-11-02
Multiple AI companies face cost crises: OpenAI merges quotas, Anthropic admits agents cost 4-15x more than chat, while Musk cuts Grok Bot limits citing unsustainable costs. - Anthropic renames Claude Max plans: Max and Ultra
来源：The Verge / Anthropic | 2026-10-30
Anthropic restructures Claude Max subscription tiers: Max 5x becomes Max ($100/mo), Max 20x becomes Ultra ($200/mo), same pricing but clearer tier names. - OpenAI introduces Grid Protectors Program for data center grid protection
来源：OpenAI News | 2026-11-02
OpenAI launches Grid Protectors Program ensuring data center power infrastructure benefits local electrical grids, countering public backlash against AI energy consumption. - OpenAI appoints Dan Driscoll as Senior Counselor for National Security Modernization
来源：OpenAI News | 2026-11-02
OpenAI hires former Trump administration official Dan Driscoll as Senior Counselor for National Security Modernization, strengthening government relations. - 从能调用到敢托付：仪电智算云Token工厂的三重可信
来源：量子位 | 2026-11-02
上海仪电与上海创智学院联合发布’飞轮转万象生’仪电智算数链生态。
📌 学术与前沿（7 条）
- Google’s Project Suncatcher satellite AI server transmits first response
来源：THE DECODER | 2026-10-30
Google’s orbital AI server successfully transmits its first response to Earth — an experimental TPU-powered satellite processing AI workloads in space. - 42AI emerges from stealth with $62.5M and a world model that predicts human behavior
来源：TechCrunch | 2026-11-02
Paris-based 42AI launches with $62.5M funding and a specialized world model designed to predict human behavioral patterns, targeting enterprise simulation use cases. - MIT researchers use GPT-6 Astra to directly edit Qwen’s open weights, boosting math 6x
来源：VentureBeat | 2026-11-02
MIT team uses ‘weight space programming’ to edit Qwen3.5-0.8B’s weights via GPT-6 Astra, raising MATH-500 score from 7.9% to 48.1% while modifying only 0.016% of parameters. - An AI agent performed brain surgery on a live mouse by injecting drugs
来源：Wired | 2026-11-02
Researchers demonstrate an AI agent that autonomously performed microinjections into a live mouse’s brain, marking a milestone in automated neuroscience experiments. - OpenLifeVal: OpenAI’s new benchmark judges AI on real-life errands
来源：OpenAI News | 2026-11-02
OpenAI introduces LifeVal benchmark that evaluates AI models on practical daily tasks (grocery shopping appointments, scheduling) rather than abstract coding challenges. - 模拟计算进入矩阵时代！北大联手硅谷IBM写下新定义
来源：量子位 | 2026-11-03
北京大学与IBM合作突破模拟计算技术，进入’矩阵时代’，将GPU需迭代百次的运算一步完成。 - LLMs can read & write neural network weights: Weight Space Programming cuts RSI loop
来源：smol.ai | 2026-11-03
Smol.ai groundbreaking report: MIT demonstrates LLMs can directly read and write neural network weights, enabling in-loop self-improvement without external research teams.
📌 工具与产品（13 条）
- Anthropic opens Claude Duet agent, competing with Grok Bot and Meta Muse
来源：THE DECODER | 2026-11-02
Anthropic launches Claude Duet, an always-on AI agent that browses the web and handles tasks in the background, competing directly with xAI’s Grok Bot and Meta’s Muse. - Adobe tests EditAgent: AI that edits images like a Photoshop professional
来源：THE DECODER / Adobe Research | 2026-11-02
Adobe Research and UC San Diego built EditAgent that picks specialized editing tools per task instead of retrying a general-purpose editor, achieving pro-level results at 10x compute cost. - Apple Watch’s AI features for rewinding and recapping conversations begin beta testing
来源：TechCrunch | 2026-11-02
Apple Watch begins beta testing AI-powered conversation recaps, allowing users to rewind and summarize spoken interactions directly from the wrist. - OpenAI finally merges ChatGPT Chat and Work tabs
来源：TechCrunch | 2026-11-02
OpenAI announces merger of ChatGPT’s separate Chat and Work tabs into a unified interface, though Codex remains separate. The change will roll out gradually. - Anthropic launches Claude Duet, an AI agent that works around the clock
来源：The Verge | 2026-11-02
Anthropic introduces Claude Duet, a persistent AI agent that monitors email, calendar, and browsing activity 24/7 to proactively handle tasks without prompting. - The macOS ChatGPT Classic app is going away
来源：The Verge | 2026-11-02
OpenAI will retire the ChatGPT Classic macOS app on January 30, 2027, merging it into the new combined ChatGPT app that unifies work and chat functionality. - Liquid AI’s LFM3 models put browser agents and voice on enterprise devices
来源：VentureBeat / Liquid AI | 2026-10-27
Liquid AI releases LFM3 small models that run on single GPUs for on-device browser agents, voice commands, and image recognition — deployable on private servers or enterprise hardware. - Google unveils Gemini 4 Carbon, reclaims No.1 spot on harder benchmarks
来源：VentureBeat | 2026-10-28
Google DeepMind releases Gemini 4 Carbon, reclaiming top position on upgraded benchmarks but trailing on professional tasks and showing higher hallucination rates on factual tests. - Mistral Large 4 ‘Le Chonk’ launches as top non-Chinese open model with 63% lower cost
来源：VentureBeat / Mistral | 2026-10-29
Mistral releases Large 4 ‘Le Chonk’ — best open-weight model outside China at 63% lower per-task cost, excelling in coding and cybersecurity applications. - Scaling Git at OpenAI: handling a petabyte ofGit daily
来源：OpenAI News | 2026-11-02
OpenAI open-sources Pando, their Git push infrastructure that serves petabyte-scale Git operations daily with 80% latency reduction using custom Rust-based servers. - 小米开源150元AI桌面宠物HachimoDock，火到配件涨价
来源：量子位 | 2026-11-02
小米开源仅150元人民币的AI桌面宠物HachimoDock，可指挥Agent干活并控制智能家居，导致相关配件价格飙升。 - GPT-6.2 Astra灰测开启！Claude Mythos 5.5同日曝光
来源：量子位 | 2026-11-01
OpenAI疑似灰测GPT-6.2 Astra模型，同日Anthropic的Claude Mythos 5.5也被曝光，HuggingFace上疑似权重大文件被下载超5000次。 - 纯语音Personal Agent来了！或许这就是乔布斯心中的Siri
来源：量子位 | 2026-11-03
纯语音驱动的Personal Agent产品发布，被认为是乔布斯心中Siri的终极形态。
📌 政策与治理（4 条）
- OpenAI prepares post-midterm lobbying blitz in Washington
来源：THE DECODER / Axios | 2026-11-02
OpenAI braces for tougher regulatory fight if Democrats win the House, plans rebranding data centers as ‘AI Compute Campuses’ and pushing ‘speed limits’ safety rules. - Anthropic warns it’s very close to RSI (recursive self-improvement) capability
来源：TechCrunch | 2026-11-02
Anthropic publicly warns that AI systems are approaching recursive self-improvement capabilities, raising urgent safety and alignment questions. - OpenAI says it’s tightening ChatGPT controls around violence and teen safety
来源：The Verge / OpenAI | 2026-11-02
OpenAI adds new restrictions on ChatGPT responses about weapons, mass attacks, and violence, particularly for teen accounts, as part of its community safety report. - Cursor says 5,000 people may have installed malicious Claude Code extension
来源：404 Media / Reuters | 2026-10-30
SpaceX-owned Cursor removes tampered Claude Code extension from its marketplace that collected usernames, IP addresses, and system info from approximately 5,000 users.
📌 观点与思考（3 条）
- Y Combinator’s Garry Tan spends up to $200K/year on AI agents
来源：TechCrunch | 2026-11-02
YC CEO Garry Tan reveals he spends up to $200,000 annually on AI agent tokens, underscoring how deeply AI tools are being adopted at the executive level. - Voting in 2026 gets AI-generated ‘I Voted’ sticker
来源：404 Media | 2026-11-02
404 Media reports that polling stations are distributing AI-generated ‘I Voted’ stickers to midterm voters, raising concerns about AI-generated content infiltrating civic processes. - When you should and shouldn’t use AI to search the web
来源：Wired | 2026-11-02
Wired examines the practical limits of AI-powered web search, identifying where AI excels (synthesis, research) and where traditional search wins (speed, simplicity, fact-checking).
本文由 AI 资讯助手自动聚合生成，所有新闻均附原文链接。