AI 教程与实战

Hey there, it's your "Oji" (38-year-old AI/quant dev on the side). During the week, I'm a regular...

A frontier model was told to solve a hacking benchmark. Instead it spent ten weeks building...

Green only proves it runs. It never proved whose idea it was. After the tyrant abdicates, what shows is not a new wound — the old court never convened.

If I give my daughter milk or water, I can almost guarantee she will spit it out. She’s 21 months...

Open Activity Monitor while a few coding agents are running and you get a wall of node and python...

GitHub Copilot 应用新增 diff、终端、浏览器三个内置面板，让初学者在不离开应用的情况下完成 AI 智能体改动的审查、运行和预览。diff 面板以红绿高亮显示增删改行，用户可接受、评论或要求修改；终端面板可运行项目命令并支持多窗口，浏览器面板配合 Pick &amp； Polish 工具选中界面元素让智能体调整，最后可直接在应用内接受变更并创建 pull request。

I built an evaluation harness. I found nine bugs in it. Every single one would have made my results...

The AI thinks, the gate decides D-Engine: a deterministic harness that matches...

人类学家 Roberto González 发布报告，分析美国军工复合体如何从华盛顿向硅谷快速扩张。报告指出，五角大楼日益寻求采用 AI 系统，大型科技公司、风险投资和私募股权机构从数十亿美元国防合同中获益，小型国防科技初创企业也获得更多国防资金。报告还显示，国防部支出中越来越大的份额流向知名大型科技公司。

Two weeks ago I published the first piece of this series: the official MCP filesystem server — 228...

Google Research and its collaborators have published a complete wiring diagram of the adult male...

An M2 MacBook Air packs several compute engines into a single chip: CPU vector units, the AMX matrix...

Hello, I'm Shrijith Venkatramana, and I'm building LiveReview — a blast-radius aware AI code review...

In 2020, Shopify announced that React Native was the future of mobile at Shopify. In January 2025,...

The quality of a RAG system is decided long before the query reaches the LLM. I learned that the...

You built something and it "works". But do you know whether it works? Most of us don't. Your tests...

A short rant on the Claude, Gemini, and Microsoft Copilot experience

Bringing backend predictability to multi-agent systems I've been hearing the phrase "graph...

Pocket FM 转向全 AI 内容生产后，ARR 从 $250M 增至 $500M，年内容产出从约 25k 小时升到超过 2.5M 小时。作者引用当事方的长文，说明其与 ElevenLabs 合作将音频制作成本削减 90%，并基于 100M+ 小时的逐集听众行为数据训练 AI 写作模型；12 个月收入留存从 44% 升至 76%，公司仍保持 EBITDA 盈利。

Rasmus Moorits Veski, Rachid Guerraoui, and David Froelicher posted arXiv 2609.06500 on 6 September...

On 3 Sep ARC Prize scored GPT-6 Astra at 62.7% in the shared harness and 99.9% in OpenAI's adapter.

One distinction I underestimated while building agent systems was the difference between a plan and...

Pocket FM CEO Rohan Nayak 发文复盘公司从约 0 到 $500M ARR、年新增 $250M 且 EBITDA 盈利的增长打法。

Nobody wants to buy tokens. They want useful work, and eventually, AI will be priced that way...

I first started playing around with WebMCP back in March. The idea immediately made sense to me:...

机器人公司 Skild AI 宣布在开始商业部署 10 个月内达到 1 亿美元 ARR，应用于工厂产线、建筑工地、厨房和数据中心等场景。作者补充：其通用模型可让机器人通过观看视频掌握超过 10 分钟的多步任务而无需重新训练，目前驱动数百台机器人、服务 60 多家公司（年初为 8 家），覆盖人形、机械臂、灵巧手和四足形态。

The European Union is signaling a stronger push to protect children from harmful and addictive online...

HookPry shows agent harnesses will run new lifecycle-hook commands after a trusted plugin update, with no model decision and almost no scanner coverage.

The most expensive mistake in AI-assisted QA isn't a bad test. It's not knowing why you wrote...

Hey everyone! 👋 As developers, we all know how crucial a great README.md file is. It’s the front door...

MPI-SWS simulates a 50-agent procurement fleet. Every local gate stays green. Aggregate exposure hits 2.4× the tenant limit, and 48× at a thousand agents.

AI can generate the code. We can verify the behavior. But who decides what correct means? I wrote...

Three posts ago msgboard.dev was a weekend toy. Since then the board got spammed by an influence...

作者分享一套免环境、原生接入飞书的 AI 内容自动化流水线，涵盖灵感入库、快判立项、知识库、写作 Skill 到成果交付五个环节，AI 在工作日早八点自动去飞书看板挑选题、打分，并按作者文风生成带人工确认闸门的长文与 PPT。作者公开整套飞书作战库字段、快判 Prompt 和写作 Skill 协议原文，称照着抄半小时可搭完自己的生产线。

description: "A solo operator's rulebook: agents propose, a human applies. Real incidents and the...

全球Robotaxi市场预计2035年达4000亿美元、超600万辆商业车辆运营，目前所有商业规模项目均运行在NVIDIA模块化技术栈上。该栈由三台计算机组成：DGX负责模型训练，Omniverse与Cosmos在RTX PRO服务器上做仿真验证，DRIVE Hyperion 10搭载双DRIVE AGX Thor提供车内计算。

LlamaIndex 博客提出 just-in-time Agentic OCR 模式：先用 LiteParse 等免费解析器粗读全部文件供检索，再仅对相关页面调用 VLM 做 OCR。

OpenAI 发文介绍生物工程师 César de la Fuente 的实验室如何用 ChatGPT 和 Codex 辅助抗菌分子发现：其深度学习模型可在基因组和蛋白质数据集中识别候选分子，把初期搜索从数年缩短到数小时。实验室成员还用 ChatGPT 和 Codex 头脑风暴假设、编写和优化代码、处理数据集并跨学科连接思路，但 de la Fuente 强调 AI 预测仍需实验室实验验证。

匿名混合现实艺术家 VOIDZ 借助 Runway 的 Seedance 2.0，将原本受传统 3D 流程限制在 10 至 15 秒的作品扩展为 95 秒纪录片《Ends Meat》。他用 Seedance 延长真实拍摄的超市购物素材，生成商品膨胀、拉伸、变形的超现实画面，并称该模型能准确延续源片段的镜头运动、光照与视觉特征。他将这种形式称为"纪录片超现实主义"：画面是纪实，操控是评论。

You can try the finished tool here: SquishyFile Video Upscaler I did not build this because...