AI 情报文章归档

Pydantic AI 现已作为受支持智能体接入 Scale AI 的开源框架 AgentEnv Framework，可在模拟邮箱、Slack、工单等真实环境中运行多步任务。

据时代财经报道，11 月 6 日起豆包工作个人版取消 5 小时短周期额度限制，个人用户可在周额度范围内灵活安排使用时间，周额度及其他规则仍以对应套餐为准，用户可在“设置-订阅与额度管理”查看套餐和剩余额度。豆包工作是豆包面向生产力场景的 AI 产品，可根据任务目标自主拆解任务、调用工具，完成调研分析、内容创作、数据处理、应用搭建等工作。

A cat crosses a room, jumps onto a windowsill, watches a bird, curls up and falls asleep. That became...

Quand un agent IA modifie des fichiers de votre projet et casse quelque chose, le réflexe est git...

Claude Code put your account email in the User-Agent header of its web requests, so every server log on the agent's path harvested a scam-ready lead, and the same email ships in the system prompt of every session, an issue open since July.…

Ai2 发布 ToyBlocks，一个围绕机械臂和桌面方块的开源仿真框架，作为其 FoundationWorld 框架的首个受控环境，用于研究机器人学习与评测。

OpenAI 宣布与安全公司 Hacktron 合作，双方工程师与安全研究员将共同挑选高优先级领域进行测试，重点检查 OpenAI 系统与第三方服务的交互方式。

Qwen2.5-0.5B's config declares its vocabulary-by-hidden matrix twice: once as the embedding table,...

A shopper asks ChatGPT for "a waterproof hiking jacket under $150 that ships to Dubai". The answer...

"Done." It's probably the most satisfying word in a coding-agent session. It can also conceal three...

Most agents are instrumented for the failures that announce themselves and blind to the ones that do not. Here is the pattern: the three streams...

Your request volume doubled this month. So you doubled your GPU spend. That is the most common...

清华大学教授、面壁智能联合创始人及首席科学家刘知远在《光明日报》发表署名文章，提出建立以实践为标准的人工智能评价体系。文章主张从评价具体应用、评估行业成效、衡量整体贡献三个层面完善评价体系，把效率和成本纳入评价，引导人工智能从规模竞争转向效率竞争、价值竞争。

Meta 宣布把基于 Muse Wonder 训练的专用模型推向 Muse 和 Meta AI 用户，Personal Superintelligence Index 得分从 38% 提升到 47%。

"We support MCP" used to mean something specific. This year it became ambiguous, because there are...

Perplexity 于当地时间本月 4 日宣布推出 Teammate 功能，一款作为用户“团队成员”、能执行持续性复杂工作的智能体，可简化使用 Perplexity Computer 时的规划工作。

小岛秀夫在接受日媒 Game*Spark 采访时表示，AI 如同 Photoshop 和文字处理软件一样能提升游戏开发效率，他并不反对。但他担忧效率提升会省略创作中最有价值的试错过程，就像拼积木的乐趣在于亲手组装而非成品。他认为若放任 AI 发展，最终可能在创意工作上取代人类，但未来 5 至 10 年应该不会有什么问题。

In April, Claude Code had /buddy for about a week: a tiny creature in your terminal that watched you...

11 月 6 日在上海进博会上，西门子医疗与阿里巴巴达摩院签署战略合作协议，计划将达摩院基于平扫 CT + AI 的一扫多筛技术与西门子医疗影像设备、临床应用结合，推广用一次 CT 扫描早期发现胰腺癌、胃癌、肠癌等高致死癌症及主动脉夹层、肺动脉栓塞等心血管危急重症。

荣耀 CEO 李健在 2026 移动网络大会提出，AI 入口之争并非存量零和博弈，行业应让入口主权回归用户。荣耀将携手中国移动及生态伙伴，由 YOYO 联动灵犀、千问、豆包、Gemini 等搭建智能体网络，并在荣耀 Magic9 系列定制机上首发 Token 套餐。

Cloudflare CEO 马修·普林斯在第三季度财报电话会议上预计，未来两年严重的网络安全威胁将持续出现，原因是 AI 正在挖掘传统软件中隐藏的安全漏洞。近期已有 6 个严重漏洞被公开披露，Cloudflare 已与多家主要 AI 研究机构和网络安全公司合作，提前获取未公开漏洞信息以保护客户。

Context: Dicamus is an AI self-help service, not therapy and not a crisis line. If you are in danger...

据知情人士，Palo Alto Networks 正洽谈以 5 亿美元现金加里程碑付款收购专注 AI 智能体系统访问安全的 Clutch Security。

World Labs 在两个月前推出自回归世界模型 Atlas 后，发布 Frameboy 演示，让用户以精确相机控制在云端 GPU 上以最高 15 fps 的流式传输实时探索 AI 生成的世界。fal 表示其 serverless 推理引擎为 Frameboy 提供算力支持。

GitHub: https://github.com/ESeungJun/fix-proposal I do most of my work with Claude Code now, and...

A support bot that says "I don't know" is annoying. A support bot that confidently says "No, we don't...

webAI 开源了基于 IBM Granite-4.2-8B 的 TwIL-8B 推理模型，针对数学、代码和指令跟踪优化，采用非商业许可。该模型在 MATH-500、AIME 2025 两项数学基准上取得参比模型最佳成绩，强于尺寸数倍的 Gemma 4 31B，LiveCodeBench 同为同组最佳，IFBench 与 Gemma 4 12B 持平。

1. The argument is being filed under the wrong heading Most debates about agent...

An SF Tech Week experiment: we gave an AI a phone and had it call San Francisco businesses. Only 1 in 3 kept talking - most businesses are not ready for the AI assistants that are about to call them.

Aravind Srinivas 认为智能体将越来越多地被其他智能体调用，因为编排更简单，且可集中到一个以 git 为核心、不断复利的系统底座；当智能体以 code-mode 运行并对 trace 做版本控制，就形成可用减少人工介入来度量的递归循环。配图显示人工输入的队友操作占全部队友轮次的比例从第一周的 31% 降至最近的 22%，其余来自子智能体、跟进和自动化。

We asked six AI chat models what is at 41,526 points of San Francisco, from memory. The best three tie at 86% on the city map, against 61% for a trivial answer, and none of them can draw the Marina's streets.

Citation Is Not Truth: A Three-Metric Evaluation of an Extractive Research...

Agent Memory Architecture – A Practical Guide 📋 Concise Diagnosis...

A stateless Streamable HTTP MCP server in Next.js: keyless discovery, Bearer keys, readOnlyHint on 26 tools, preview then confirm on 18 write tools.

Looping is a powerful inductive bias for reasoning. By repeatedly iterating a recurrent block, looped models can scale their effective depth and enable multi-step reasoning in latent space. However, existing looped models are not inherentl…

Instruction-tuned large language models (LLMs) answer at greater length than a request needs. We ask whether a training-free controller inside the decoding loop can remove that surplus while preserving content, judged against the simplest …

记者 Kate Clark 援引 WSJ 报道称，Anthropic 的年化收入（run rate）在 10 月底突破 800 亿美元。链接报道同时提及部分大型投资者对 Anthropic IPO 能否达到 1.5 万亿美元估值持怀疑态度。

Do this first: go down the list and tick what your harness already does. Each unticked item is a...

Multi-agent LLM systems increasingly aggregate judgments across model instances, but reasoning about other agents can itself introduce failure. We study a setting in which LLMs shift their public decisions after observing ambiguous social …

We present PhysioWave-T, a token-efficient wavelet foundation model family for EEG, ECG, and sEMG. Dynamic scale folding recombines J+1 wavelet subbands prior to tokenization, compressing sequence length from (J+1)CS to CS tokens for J lev…