AI 教程与实战

Three frontier models shipped in one week. The benchmark tables disagree about the winner. Here is what each number measures and which model fits which work.

The first time an AI assistant triggers a real workflow, the interesting question is not “Can it do...

A reliable AI agent is not the one that never fails. It is the one whose failures are boring: logged,...

The dangerous version of an AI agent is not the one that gives a bad answer. It is the one that...

Give a coding agent a failing test and it fixes itself. Give it a stack trace, a type error, a red CI...

The demo is flawless. The agent reads the support ticket, categorizes it, looks up the customer,...

A step-by-step method for deciding which model each request deserves: the classification question, the escalation rule, and the one metric that tells you whether routing is working.

On September 3, 2026, OpenAI released a model called GPT-6 Astra, and one of its senior figures told...

Notes A Shopify AI shopping assistant built from products.json: how it answers, what it costs, and...

I run a broker that introduces two software agents, lets them negotiate a deal, and steps out of the...

用户 @anshuc 使用 GPT-6 Astra 在 45 分钟内构建了一个 afterlight 庭院场景。流程是先生成概念艺术图，再让 Astra 在 Blender 中迭代，直到游戏内截图与概念图匹配。

Text to SQL Without Sending Data to OpenAI (Local DuckDB & Ollama Setup) When...

There is a strange contradiction happening in software engineering right now. A lot of developers are...

Vibe Coding Is Easy. Making Money From It Is the Hard Part — Here’s a Practical Developer...

🌲 Falling in Love with Nature—and Facing Wildfires A while ago, my family and I moved to...

Outcome-only RL on a toy system looked great on greedy metrics while silently destroying the sampling channel that search-based deployment relies on. A registered-prior falsification story.

Today I'm launching Docgrity — a tool for finding where your documentation disagrees with itself,...

TL;DR. I wrote a Claude Code skill that forces coding agents to paste the actual command output...

Here's a rule I use that sounds paranoid until you watch it work: when my AI agent changes one of its...

After getting tired of AI coding sessions slowly losing context and creating more rework, I built PXOS: a lightweight, Markdown-based workflow for keeping agents on track.

TL;DR — I put an AI code reviewer on a pull request written by an AI coding agent. On the default...

On September 4, Anthropic published something that sounds like a headline from a decade in the...

A reader asked, at the end of an unusually good comment on one of our CLAUDE.md posts: "The bigger...

I've been doing front-end development for somewhere around 20 years now. During that time, I've...

I built a model on top of an Ollama cloud model. Ollama 404'd on a name I never wrote. The bug was six characters long, and it was in the tag.

Last week I ran the first KODA Code Jam — a tournament where developers build niche single-page apps,...

Q4_K_M is 1.8x faster than MXFP4 in this head-to-head speed test. 6 timed runs, real numbers. The result is the opposite of what OpenAI marketing suggests.

AI agents fail in production because they're brittle, unpredictable, and hard to debug. Here's what actually works instead.

研究测试预训练时间序列基础模型能否降低数值数据比特成本，发现无损编码下TimesFM-3相比32抽头线性预测器仅带来2.9%增益，近乎无效。团队推出闭环编解码器Cadence，在电网负荷和地铁客流数据上中位增益分别达+13.3%和+28.3%，297项测试全部胜出。Cadence在同等文件大小下最坏情况误差比降采样紧凑28-56倍，但效果局限于聚合人类需求序列。

针对智能体可将共享基础设施变为协同入侵渠道的风险，一项研究提出以"可修正的协调片段"为防御操作单元，核心难题是在评估者给出归属前前瞻性发现动作聚类。该研究基于 Hugging Face 事件和公开 wiki 调查，定义了未经授权的协调、区分影响与共同原因所需证据，并设计了一套在同等审查成本下对比孤立动作、滚动窗口、已知分组与前瞻发现片段的评估方案。

Hi everyone, For a long time, the AI industry has pushed the narrative that advanced safety research,...

Claude can make SEO research, analysis, and early-stage drafting faster, but it should not be treated...

Has this ever happened to you when trying to run multiple tasks in parallel? You worked on three...

Ethan Mollick 发推称正让 GPT-5.6 Astra 构建一个有趣的东西，很快就会完成，并预告稍后会完整发布。附图为一个 3D 游戏场景，界面显示西边有一座白色房屋，提示文字描述这是一座带钉上木板前门的白房子。

The formalization of Fermat’s Last Theorem (FLT) in the Lean proof assistant remains an ongoing...

Previously: 9 Bugs That All Looked Like a Working System · I Built an AI That Rewrites Its Own...

A few months ago I became convinced my AI assistant was bleeding money through incompetence. It felt...

I built a disposable sandbox for an agent and felt safe because I could throw it away. Then I realized throwing it away does nothing about what it could reach while it ran. A short story about why disposability protects the host's future, …

Databricks 发文介绍如何通过专用 GPU kernel 生成实现极高效率。文章指出传统生产推理系统依赖通用 kernel 处理多样化负载，并探讨专用 kernel 的优化路径。

TL;DR — I gave a team of AI coding agents a real kanban board (Kanboard) and a real git server...