AI 情报文章归档

IDC 发布横跨 25 个研究领域的 250 项 FutureScape 2027 年度预测，称企业到 2028 年将在开源模型上运行 60% 的 AI 工作负载，并借助模型路由策略降低成本、对冲供应商锁定风险。

Your coding agent says the code is faster. Put both commits in worktrees, time them with hyperfine and benchstat, and read the p-value before you land.

A tested setup guide for spin-that-dice, the Spotify party dice: roll from the 305-track starter crate first, then add your own app or music server.

Pydantic AI 团队以模型可迁移性、持久执行、人工审批、前端协议、追踪和商业支持等标准，对比 Pydantic AI、LangGraph/LangChain、OpenAI Agents SDK、Google ADK、Microsoft Agent Framework、Strands Agents、CrewAI、Claude Agent SDK 和 Agno。

OpenAI 构建了 LifeVal 评测，覆盖照护、健康、教育、家庭事务、个人理财、购物和政府服务等日常任务，用于追踪模型过去一年在日常任务上的表现。该评测基于美国时间使用调查等两项大型研究，并新增 3995 名受访者的代表性调查来确定用户最需要帮助的领域。OpenAI 称最新模型在多项 LifeVal 任务上得分很高，GPT‑6 Astra 在部分任务上达到 100%。

清华与微软论文提出 HarnessMM，让模型预先写出带校验条件的后续命令，工具结果通过校验就直接执行、跳过 LLM 调用。在 98 个 HumanEval 任务上，DeepSeek Harness 成功率从 71.43% 升至 85.71%，token 用量比原版减少 87.93%；Codex 成功率保持 90.82%，token 用量减少 29.54%。

土耳其竞争管理局宣布对谷歌搜索中的“AI 概览”“AI 模式”功能启动调查，以确定其是否违反该国《竞争保护法》。调查涵盖这些功能如何改变用户获取信息方式、是否损害新闻出版商和内容创作者流量、内容使用协议是否含不公平条款、是否产生排他性影响以及内容筛选或许可中是否存在区别对待。

Illustrated Chief workflow from the asklog film. I kept opening another coding-agent terminal and...

"All tests pass." If you use an AI coding agent, you've probably read that sentence several times...

AI made building cheap, so nothing forces developers to choose anymore. Three rules that stop fast output from becoming a longer to-do list.

Y Combinator 与 Olam Labs 的研究人员在《我的世界》中测试上百个智能体协作与对抗，依托 Mindcraft 和 Mineflail 等工具开展实验。

Hi! I am Ricardo Sueiras, leading SDLC transformation at Indicium AI. That means I get to work with...

Why Senior Engineers Are the True Targets of 'The AI Layoff Trap' There is an ironic...

I wanted to know why AI coding agents get expensive on real codebases, so I stopped guessing and...

agent_summary asks the agent's own model to label its last action in 3-5 words. How the timer works, how to measure its share, and why the numbers disagree.

An agent with write tools is coordinating distributed transactions over an unreliable network, and...

Nextcloud 将 Olmo 3 纳入其本地 LLM 后端 llm2，并给出公司最高伦理 AI 评级“绿色”，该评级要求开源训练与推理软件、可自托管模型及公开训练数据。Nextcloud 工程师 Marcel Klehr 称 7B 版本可在消费级硬件上运行，无需 16 或 32 GB VRAM，用户可在 Nextcloud Assistant 等应用内本地调用模型。

巴西一家劳工法院对两名律师处以约 1.6 万美元罚款，因其诉状中用白底白字嵌入提示词注入指令，要求 AI 对诉状仅做表面处理、不质疑内容。该注入未实际起效，法院认定行为有损司法尊严，按涉案赔偿金额 10% 罚款并通报律师协会。判决书由巴西利亚大学教授 Vladimir Aras 在 X 平台分享，此前当地还发生过当事人嵌入隐藏提示词被要求线下提交纸质材料等类似事件。

From DevOps to AI Platform Engineering — Part 1 I'm a DevOps engineer. I've spent years running...

Many RAG failures that look like model problems are ingestion problems. If a parser flattens a PDF...

Deterministic contract tests for structured LLM output on the JVM, and what 120 real documents taught...

阿里云宣布 Memory 服务升级，面向 Personal Agent 场景增强对用户偏好、计划和历史约定的记忆能力。Memory 从对话中提取关键信息保存为记忆与用户画像，供后续会话检索使用，减少重复询问。

A model writes the change. A model reviews the change and says it looks good. The pull request now...

阿里云发布面向专业机构投资者的 AI 投研工作台 Qovest，用户一句话提出研究需求即可自动查数据、收集资料、分析并生成报告，还能按设定频率持续跟踪研究对象。Qovest 内置权威金融数据库和自研金融场景 Agent 工程框架 Harness，研究结果可沉淀到投研空间继续推进。该产品已对个人用户开放邀测，后续订阅套餐为每月 99 元或 199 元。

“右”边请看，近距离了解 Apollo Go！🚗 从我们在伦敦的首次媒体试乘，到在香港的首次大规模公众试乘活动，最新一期 AI Pulse 将追踪 Apollo Go 在这两个右舵市场的进展。此外，还有 11 月 12 日百度世界 2026 的预告。 了解更多 ↓

NVIDIA 以 Apache 2.0 许可证开源 Auto Ontology，一个从企业元数据构建语义层的框架，供智能体按认证业务定义生成 SQL。

阿布扎比王储哈立德会见 OpenAI 首席执行官 Sam Altman 后，双方相关机构签署四项协议，覆盖政府、能源和投资领域。

开发者 Vixhal 在 GitHub 以 MIT 许可发布实验模型 qwenGPU，把一个 4 通道 SIMT GPU 编译为 45 层未训练权重，放在 Qwen3.5-2B 自身 24 层之前，使模型在前向传播中运行真实的 CUDA（编译为 PTX）。

Why an agent's freshly written helper often duplicates an existing utility, and how silent forks drift into real bugs.

git commit, npx, apt, ssh, psql, terraform apply, gh auth login: each can stall an AI coding agent on a prompt. The flag or env var that stops each one.

Claude Looked Blocked. My Policy Had Never Fired. From the agentpolicy v0.1.0 release: a...

10 Prompts That Turn ChatGPT Into a Senior Data Analyst I once spent an entire weekend...

A 6B design model peaked at 64.4 GiB on an A100 80GB. Its transparent PNGs depend on a line break, so verify the alpha channel yourself.

When you ask Claude Code to "make a 60-second horror reel", it will happily start writing components....

We replayed 45 findings from a careful human code review against automated rules. They caught 9. Here is what that taught us about which problems tools can own, and the toolkit we built around it.

Developers discover that Claude Code ignores custom instructions because they set them in the wrong layer. Learn the three ways to shape Claude Code behavior, what the system prompt actually controls, and which parts remain locked no matte…

Getting a language model to answer WhatsApp messages takes an afternoon. Getting it to stop answering...

The problem nobody talks about AI agents are getting persistent memory. mem0, Zep, Letta,...

MIT 的 Regina Barzilay 与 James Collins 获得 2026 年 Pritzker AI 科学研究卓越奖，获奖成果为融合深度学习、生物工程与实验验证的 AI 抗生素发现工作。

A tampered Claude Code extension sat in Cursor's marketplace for about 11 hours. Here is the delivery chain that made it look like a normal update.