一年趴在 5% 以下的那张榜，半年冲到了 50 分

• 美国放开出口H200
  • Should the US Sell Hopper Chips to China? ifp.org
• AI驱动的言论市场（时间轴上的移动 with LLM）
  • Show HN: Gemini Pro 3 imagines the HN front page 10 years from now news.ycombinator.com
  • HackerNews 2035 - Future Tech News From Tomorrow future-tech-hackernews.qwikbuild.site
  • @karpathy：Quick new post: Auto-grading decade-old Hacker News discussions with hindsight I took all the 930 frontpage… x.com
  • Auto-grading decade-old Hacker News discussions with hindsight karpathy.bearblog.dev
  • 人在做，LLM在看
• 剑桥大学提出战略自我提升AI智能体框架，AI劳动力市场的经济动力学规律
  • [2512.04988] Strategic Self-Improvement for Competitive Agents in AI Labour Markets arxiv.org
• AIA 预测Agent
  • Bridgewater AIA Labs重磅发布：大模型预测能力已比肩人类“超级预测员” mp.weixin.qq.com
  • AIA Forecaster: Technical Report arxiv.org
• AI路边摊，下一个市民经济风口-虎嗅网 huxiu.com
  • item.taobao.com · item item.taobao.com

AI基础设施 AI Infra

• OpenAI State of Enterprise AI
• State of AI from OpenRouter
  • OpenRouter's State of AI - An Empirical 100 Trillion Token Study news.smol.ai
  • State of AI 2025: 100T Token LLM Usage Study openrouter.ai

基础模型 Foundation Mode

• GPT-5.2
  • Introducing GPT-5.2 openai.com
• G
  • Gemini 3 — Google DeepMind deepmind.google
• Meta提出"共同改进"范式:人机协作通往超级智能的更安全路径
  • [2512.05356] AI & Human Co-Improvement for Safer Co-Superintelligence arxiv.org
• 阿联酋MBZUAI发布K2-V2
  • GitHub - LLM360/k2v2train: Training codebase for K2-V2 github.com
  • File not found · GitHub github.com
• Inference Efficiency matters
  • OpenAI new model x.com
• Microsoft VibeVoice-Realtime-0.5B
  • VibeVoice microsoft.github.io
• DeepSeek v3.2 效能差距在变大
  • 架构 / 资源 / Agent
• 腾讯混元团队提出MoCA：通过组件混合注意力机制实现可扩展的组合式3D生成，支持单个资产最多32个组件
  • MoCA: Mixture-of-Components Attention for Scalable Compositional 3D Generation lizhiqi49.github.io
• LM Arena movements: The Arena added Baidu’s ERNIE‑5.0‑Preview‑1103 to the text leaderboard (preliminary) and shared YTD trends across top labs. Links: ERNIE entry, trends.
• Best model polymarket.com
• 康奈尔大学提出ShadowDraw：将任意3D物体转化为影子-线条画组合艺术，实现物理投影与生成式绘画的无缝融合
  • ShadowDraw: From Any Object to Shadow-Drawing Compositional Art red-fairy.github.io

具身模型 Embodied AI

• Meta FAIR 发布开源触觉手套 OSMO
  • OSMO: Open-Source Tactile Glove for Human-to-Robot Skill Transfer jessicayin.github.io
  • Project Aria Case Study: Introducing EgoMimic by the Georgia Institute of Technology - YouTube youtube.com
  • EgoMimic: Scaling Imitation Learning via Egocentric Video arxiv.org

AI安全 AI Security

• OpenAI提出Confessions方法：通过自白机制监测模型不当行为，准确率超95%
  • How confessions can keep language models honest openai.com
• OpenReview 被攻击

• ForecastBench forecastbench.org
• SVG matters
  • @lintool：👀Introducing a brand new @yupp_ai SVG leaderboard ranking frontier models on the generation of coherent and v… x.com
  • Yupp blog.yupp.ai
• @lmthang：Give this prompt a try "Create an interactive visualization for the following: The 5-fold symmetry traced by… x.com
• 阿里巴巴通义实验室发布EcomBench：首个面向真实电商场景的基础智能体全面评测基准，揭示当前AI智能体在复杂商业决策中的能力边界
  • [2512.08868] EcomBench: Towards Holistic Evaluation of Foundation Agents in E-commerce arxiv.org
• 美团与华东师范大学、上海创智学院等联合发布LocalSearchBench，首个本地生活服务智能体搜索基准，揭示大型推理模型在垂直领域的能力边界
  • LocalSearchBench: Benchmarking Agentic Search in Real-World Local Life Services localsearchbench.github.io

• 宾夕法尼亚大学沃顿商学院发现专家人设提示词并不能提升AI模型在困难事实性问题上的准确率
  • [2512.05858] Prompting Science Report 4: Playing Pretend: Expert Personas Don't Improve Factual Accuracy arxiv.org

应用

• 豆包手机
• Zhipu发布会

AI4S

• 麻省理工学院与魏茨曼科学研究所提出BrainExplore，首次实现人脑视觉表征的大规模自动化发现与解释
  • BrainExplore: Large-Scale Discovery of Interpretable Visual Representations in the Human Brain. navvewas.github.io

AI设备 AI Device

• Waymo as embodied AI at scale: Jeff Dean on fully autonomous data fueling system advances @JeffDean.

智能体 Agent & MultiAgents

• MCP becomes a Linux Foundation project: Anthropic is donating the Model Context Protocol (MCP) to the new Agentic AI Foundation (AAIF) under the Linux Foundation, with backers including OpenAI, AWS, Bloomberg, Cloudflare, Google…
  • Related: OpenAI is showcasing Figma’s MCP server for “design-to-code” workflows (event, signup); LangChain MCP Adapters 0.2.0 adds multimodal tools and elicitation (release); OpenHands pointed to the Agent Client Protocol (ACP). anthropic.com ↗ anthropic.com
  • Enterprise momentum: Anthropic expands with Accenture (30k professionals trained on Claude; product to scale Claude Code org-wide) (link).
  • MCP → Linux Foundation: “MCP went from internal project to industry standard in a year” @AnthropicAI, @mikeyk.
• 谷歌与MIT联合发布Agent系统科学化框架，揭示多智能体协作的量化缩放原理
  • [2512.08296] Towards a Science of Scaling Agent Systems arxiv.org
• 香港中文大学等机构发布CUHK-X：首个大规模多模态人类活动场景理解与推理数据集，包含58,445个样本和七种同步传感模态
  • CUHK-X: Multimodal Dataset for Human Action Recognition, Understanding and Reasoning openaiotlab.github.io
• 微软发布DoVer：通过干预式验证实现LLM多智能体系统的自动化调试，在GAIA等基准上恢复18-49%的失败案例 mbjinx.github.io ↗ mbjinx.github.io
• 斯坦福与Together AI发布GPT-5驱动的论文纠错系统，揭示顶会论文错误率逐年上升55%
  • [2512.05925] To Err Is Human: Systematic Quantification of Errors in Published AI Papers via LLM Analysis arxiv.org
• 复旦大学NLP实验室联合WisPaper.ai推出智能学术检索与文献管理平台，通过三大核心模块构建完整研究闭环 wispaper.ai ↗ wispaper.ai

编码 Coding

• Pro Coder
  • 中国科学院等机构发布DAComp基准：首个全生命周期数据智能评测体系，揭示AI Agent在企业级数据工程与分析中的关键瓶颈 da-comp.github.io ↗ da-comp.github.io
  • Tech Radar 遗漏系统 Martin fowler
  • Vibe Coder

组织与个人 People & Org

• Cursor 首席设计师炸裂访谈：团队没有明确分工，每个人都在“越界” mp.weixin.qq.com
• Home - Agentic AI Foundation (AAIF) aaif.io
  • A rare cross company move with the establishment of the Agentic AI Foundation under the Linux Foundation, with Anthropic’s MCP joining OpenAI’s Agents.md and Block’s Goose as founding projects.
  • Block、Anthropic 与 OpenAI 联合成立 Agentic AI 基金会
  • Block - Block, Anthropic, and OpenAI Launch the Agentic AI Foundation block.xyz
• 创业这一年 mp.weixin.qq.com

课程学习 Course

• It took me 2.5 years to transition to an AI team at Google. These 11 books to helped me pivot. businessinsider.com

豆包手机

• 附近的火锅店（位置权限-精准位置）
• 微信公众号文章 mp.weixin.qq.com

AutoGLM

• 安装自己的AutoGLM
• AutoGLM xiao9905.github.io
• zai-org/AutoGLM-Phone-9B · Hugging Face huggingface.co
• AutoGLM-Phone-9B: The AI That Can See Your Phone Screen and Operate It For You xugj520.cn
• GitHub - Genymobile/scrcpy: Display and control your Android device github.com
• GitHub - X-PLUG/MobileAgent: Mobile-Agent: The Powerful GUI Agent Family github.com

Lux

• OpenAGI Dev Console developer.agiopen.org

Tips

• Github: ghproxy.link ↗ ghproxy.link ghfast.top ↗
• HF: export HF_ENDPOINT= hf-mirror.com
• AutoDL autodl.com
  • AutoDL帮助文档 autodl.com