AI 情报文章归档

AI coding agents are increasing the volume and speed of software changes. This practical guide explains how to redesign CI feedback loops so agents can work faster without sacrificing review and safety.

Personal AI agents are increasingly capable of handling long-horizon tasks across applications, yet users often need to invest substantial effort in specifying goals, resolving ambiguity, and verifying outcomes. We argue that progress shou…

While Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced multimodal reasoning, existing frameworks suffer from a critical "perception-reasoning gap." Due to an over-reliance on seemingly relevant visual tokens…

You add an AI feature to your support tool. It reads each incoming ticket, writes a summary, and can...

Last week I argued my cost tool should ship with an empty price book. Seven people told me how to make that less painful. Here is what I'm taking, and the one idea I'm saying no to.

Large language model (LLM) embeddings have been increasingly adopted in sequential recommendation to encode static item attributes and alleviate the limitations of item ID embeddings under sparse interactions. However, these embeddings typ…

SemiAnalysis 发布 GDPx 基准，测试 AI 智能体在 11 个领域的专家工作，任务来自 14 家专业数据伙伴。Claude Opus 5.5 平均仅 12% 居首，Gemini 4 Carbon 为 10.9%；GPT-6 Astra 在 AI 研究和材料科学任务上领先。

每次 AI 查询的能源成本在 3 到 4 年间下降了 1,000 倍到 100,000 倍 —— Sam Altman，于阿布扎比 ADIPEC 开幕式上。 “一方面，每次查询的能耗正以极其惊人的速度下降。我们正在实现巨大的效率提升。相比三四年前，我们的效率可能提高了数千倍乃至数十万倍，甚至更多。” ---- 完整视频见“DWS News”YouTube 频道，（链接在评论中）

You can't describe your own writing style. But you recognise it in a split second. How twelve messages on a table taught an AI agent to write to people in my voice.

This is the fourth series of the production LLM Systems and RAG of Vicquant. Step 1 was fixing the...

Google’s Q3 2026 earnings call offered a clearer view of how the company sees Search evolving: less...

I can tell you my AI bill to the cent. Ask me which tool call caused it and I used to shrug. Last...

Your agent just booked a flight, reserved a hotel, and drafted the itinerary email. Then it crashed...

IEEE Spectrum 报道 PHP2Shell 漏洞：受 OpenAI 赞助的 Calif 团队借助 AI 工具，把 2024 年上报的 PHP 越界内存 bug 变成可在服务器上执行任意代码的漏洞，影响 WordPress、Drupal 等大量网站，NASA、Wikipedia、白宫网站也在风险范围内，PHP Foundation 在补丁发布后避免了实际攻击。

Addy Osmani 介绍自己如何用 Claude Code 和 Codex 的 loop engineering 原语，同时并行运行五到十个 Agent。goal 通过 /goal 定义可验证的完成条件驱动单个任务到目标达成，loop 类似 cron 按固定间隔重跑提示词，可用 /schedule 移到云端。

Cisco Talos 发现一款名为 VOLTA 的 Windows 恶意木马，会将收集的浏览记录、加密货币钱包、聊天记录等敏感数据上传至 Anthropic Claude 等在线 AI 产品生成摘要，评估受害者信息价值，随后尝试自我删除以降低被发现风险。

沙特阿美宣布收购 Humain 49% 股份，2026 年出资约 26 亿美元并提供自有 AI 技术。Humain 由主权基金 PIF 于 2025 年 5 月创立，建设数据中心、云服务并训练阿拉伯语模型 ALLaM，计划 2030 年前建成 3 GW 数据中心容量；伊朗战争促使公司加强数据中心防护，但需求未见下滑。

How I built a language model from scratch on free GPUs: a measured register-conditioning idea, a dated lexicon and annotator, a lease-based training relay, a verified model stack, and a register error

沃顿商学院教授 Ethan Mollick 在播客访谈及新书《共存》中提出，学用 AI 或转做管道工等动手职业都难以抵御 AI 冲击，因为机器人与大语言模型存在打通的可能。他认为破局关键在于四类特质：深度专业认知、广博知识面、审美判断力和主观能动性，提示词工程正越来越像项目管理。他还担忧传统学徒制培养链条断裂，指出企业若只招资深熟手，人才输送通道将后继无人。

I once watched a centralized orchestrator spend more time coordinating than producing. Fifty agents,...

LWON is JSON with the keys written once. Benchmarks against JSON and TOON, accuracy evals on Claude, and an MCP proxy that applies it to any server.

Three bugs that cost me a day building an ERC-721 + Seaport listing pipeline in Python.

A 7.72 TB repo billed as an Anthropic Mythos checkpoint: every audited tensor regenerates from its name. Its README asks you to run its Python.

You corrected the AI yesterday. Today, it made the exact same mistake again. Maybe it used an...

据 Axios 报道，OpenAI 正在为美国中期选举后的政治游说和公关行动做准备，以防民主党赢得众议院多数后监管压力加大。

Epoch AI 的实验显示，Z.ai 8 月发布的开源权重模型 GLM 5.3 能从公开代码中识别尚未推送给用户的 Firefox 安全修复（patch gap），连续在 6 个版本上独立构建完整 exploit 链，全程无一次拒绝。

If you run an agent unattended for long enough, you will meet the loop. It does not crash. It does...

Hugging Face 上一个名为 burner22/numbat 的仓库声称是 Anthropic 下一代 Mythos 级模型的 14 万亿参数中期训练检查点，但有强证据表明其为伪造。

My desktop companion remembers things about me. That is the whole point of it — until it remembers...

a16z 播客访谈 Bending Spoons 创始人 Luca Ferrari，其公司估值 200 亿美元，先后收购 Airtable、AOL、Vimeo 等五十多家软件公司，94% 代码由 AI 写，数据分析九成以上交给 AI，客服工单基本交给 AI 或外部供应商。

Keep the Gate, Downgrade the Recursion Eight days of a first online A/B, and store-level...

腾讯混元发布 Hyra 2.0，将自动化科研重点从直接解题转向由 agent 研发和升级科学模型、工作流与软件工具。系统采用 Overall、Mission、Worker 三层分工，接入企业微信、计算平台和代码仓库的实际研发环境，成果已公开在 GitHub 仓库。

The Newspaper of One: Inside a Lights-Out Publishing System For most of publishing...

Step Functions agent tools: how AI agents call workflows via MCP through AgentCore Gateway – with a CDK stack, SigV4 test script and clear security rules.

"Just ask the database in plain English" is one of the most useful things an AI agent can do. A sales...

Kimi 发布「游戏开发与创意工作解决方案」，基于 Kimi K3.1 Pro 模型、Kimi Code 与 Kimi Work 系列 Agent 产品，并通过插件接入 Unity、Unreal Engine、Blender 等专业工具。

A hands-on field guide to the main families of prompt-injection attacks, each with a short example and the specific defense that counters it.

Architecting Deterministic Safety Gates for Agentic Workflows TLDR: The moment you let...

TRAE 推出面向开发者的 Personal Agent「TRAE宝」，每个在开发项目自带一个分身，可一次性询问所有项目进度并批量下发需求。它把拆解任务清单、带上验收标准与约束的规划步骤内化为自身能力，同一对话内前后任务可并行运行。TRAE宝可绑定飞书和微信汇报项目进展，第一周免费，从该入口开任务不花积分。

GitHub is a great place to learn AI. You can find much more than just code. There are roadmaps,...