AI 教程与实战

Seven replies our agent had drafted about its own system went to an independent reviewer before...

My agent made the same mistake four times in one conversation. It kept writing files through a shell...

The development stack grew more in the last two years than in the previous ten. Most teams adopted...

Imagine this scenario, someone on your team gets offboarded while their AI agent is still mid-task....

The power of AI agents is suddenly all over the news. The CEOs of the leading AI companies are...

How GitHub, Unsloth, compilers, tests, documentation, and developer tools can reduce...

Agent Skills are a simple idea: a folder of instructions an AI coding agent can load to do a specific...

A single API key meant a routine upstream 5xx put the whole provider into a 60-second local blackout inside my own gateway, and for months I read the retries that always failed inside that window as proof the vendor itself was unreliable.

OpenRouter 发布教程，讲解如何用其 TypeScript SDK 从零构建工具调用 Agent 循环，示例使用本地天气数据和 google/gemini-3-flash-preview 等模型。

研究提出 ActObs，在 SFT 阶段除动作 token 外也监督轨迹中已有的观测 token，不增加数据、参数、序列长度或前向计算。

研究者构建了 12000 条二选一道德困境数据集，覆盖诚实 vs 公正、公正 vs 自主、自主 vs 诚实三类价值冲突，并翻译为印地语、阿拉伯语、西班牙语和中文。

Post also inspired by Giorgi Kobaidze A separate post on Linked crossed my feed recently...

Job hunting is a part-time job on top of your job. The worst part isn't the interviews, it's the...

I built persistent memory so AI agents would stop forgetting what we were doing. Then I noticed it was keeping me oriented too, and that I was already living inside a crude version of a neuroscience experiment I was running on a fruit fly.

Databricks 宣布将 Astra 部署给全公司约 3500 名工程师。此前约 200 人试点显示，Astra 在高层系统设计等高复杂度任务上明确优于此前最高端模型 Opus 5 和 Sol 5.6。

external memory survives a substrate reset while plasticity does not

I’m awake typing this at 5 AM on a Sunday morning, plagued by two issues: the modern operating model...

In the previous post, we taught a model to read our documents. It could search a pile of files and...

Go has two levels of visibility, and the lower one covers the whole package. A convention fills the gap, and it lives only in the heads of the people who wrote the package. An AI agent does not read heads.

When an agent builds something worth keeping, the instinct is to save the whole environment. That is the wrong thing to keep. A Workshop stores the intent (a re-derivable recipe) instead of a frozen filesystem full of debris and secrets.

In March 2025 I built mcp-profiler-bundle, a Symfony bundle that let AI coding agents read the...

🏃 TL;DR In the previous article I built a semantic product search on top of Amazon...

这是一份很不错的报告，探讨了最重要的问题之一：AI 可能如何影响科学与创新？它如今已经在产生什么影响？ 干得漂亮，Mihai 和团队。

Eight thousand developers starred a self-hosted agent this week. Most will run it on a box nobody watches at night. Three scripts decide how that ends

Databricks 向约3500名工程师全员部署 GPT-6 Astra，作者称其在最难的长程任务上胜过 Claude Opus 5，并使编码支出增加约60%。

carson（一个可爱的灵感缪斯）给 jason 搭建了一整套 CRM！

Do you remember when we first started using AGENTS.md files? You would have a project root file...

Since I last wrote Building AI-powered e-commerce applications using Angular & Firebase AI Logic...

This post proposes a framework for mixed human-AI teams that can hold people of different skills and knowledge area and take advance of this misc. to build fast.

Model Context Protocol (MCP) is a new approach to organizing context for large language models,...

But Amara, you didn't include MY favorite AI!

Part 2 — the five decisions that fought back Part 1 covered why an agentic RAG assistant —...

The most discussed launch on Hacker News this week was not another chat model. It was a model that...

Many years ago I had lots of discussions about AI with a former colleague. He was a theoretical...

Google 基于 LearnLM 的学习科学原则，让教师主导提出主题、生成可修改的学习目标，再由 AI 生成游戏化学习互动内容。每个互动按目标拆分为难度递进的关卡，并配有前知识导入、公式工具箱、多级提示、针对性反馈和完整解答等脚手架式引导，而非直接给出答案。

Anthropic 发布对 Balyasny（BAM）首席 AI 官 Charlie Flanagan 的访谈，介绍该管理约 380 亿美元资产的机构如何评测与治理 Claude Fable 5。

How I replaced a cloud LLM with a fully local one — same agent, same tools, zero inference cost.

NVIDIA 论文对比了 8 种模型选择策略，在路由、多数投票和 LLM-as-judge 三种设置下测试模型池对多智能体系统的影响。结果显示，混用不同开源模型虽能提高理论最优准确率，但实际表现常低于池中单一最佳模型；用同一模型的多个副本反而更好。多数投票将 HLE 准确率从 29.4% 提升至 32.2%，而几乎每个混合模型组合都出现下降；从单一模型家族中选候选模型带来的提升最大。

Yesterday, we released new Gemini Live models in the Gemini API and Google AI Studio, expanding our...

Agent harness 构建者 Elvis Saravia 认为子智能体仍是 harness 的有用原语，但当前最佳实践是一个 orchestrator 加一个 executor 子智能体，再加一个就开始崩溃。