AI 教程与实战

A sum is not anonymization. Whether a number over special-category personal data (Art. 9 GDPR) says...

A 1,145-star dev tool that claimed to run entirely locally fetched and executed code at import time. The five static checks that now gate every repo I install for my coding agents - and where they still fail.

AI Harness Engineering · Essay Four · derek wang (derekwang85) A century before anyone called...

Imagine a world where training a model with too many parameters or open-sourcing weight files lands...

Disclosure: I maintain FrameSprite, a browser-based AI game asset and sprite animation tool. This...

Hi, glad you found your way here. I'm Yash, a developer who contributes to open source, mostly p2p...

cheu·​gy ˈchü-gē - cheugier; cheugiest. slang: uncool, unfashionable It's happened enough times...

MCP — the Model Context Protocol — has gone from a niche Anthropic spec to something every AI-forward...

If you've added a CLAUDE.md or AGENTS.md file to your repo and felt like your coding agent still...

Building a security workbench for the tools agents call on the web The...

Part 2 of 2. Part 1 covered statelessness; here I tackle caching. Part 1 was about statelessness,...

Jev does not generate an answer string. It makes a typed decision in one forward pass. So,...

Your AI Agent Can Call APIs Now. Who Is Checking What It Does? The real problem with agentic AI...

This is part 5 of a series that started with the full reference diagram in part 0, covered identity...

The new model is out. The benchmarks look great. Your provider has already scheduled the sunset date...

Emad Mostaque 指出，Modal、Fireworks、Baseten 等美国公司凭借先进的 Nvidia 和 AMD 芯片，能以中国竞争对手十分之一的成本部署 Kimi K3。模型研发虽已大量转移至中国，但一旦针对 Nvidia Rubin 等下一代硬件优化，其运营成本有望再降 10 到 100 倍。

Beyond the Veto: Implementing 'Conditional Approval' and Dynamic Risk Tuning in AI Trading...

AI can write code faster than ever. But there's a problem. Writing code and solving problems are...

by shahbaz_ali. Built for the DEV × Sanity Challenge. #sanitychallenge When two of your company's...

TL;DR: I ran three Japanese classification tasks (250 rows each, same gold labels) through six...

DeepSeek's API charges two different prices for the same tokens. Which one you pay depends on what...

We gave Claude Code the same "never edit protected/" rule two ways, as one line in CLAUDE.md and as...

TL;DR: I shipped the second project in my AWS nanodegree, an AI support agent on Amazon Bedrock....

Built for Sanity Challenge, Path Two. What I Built Fliproom turns a room changeover into...

I have been thinking about a fairly ordinary failure mode in AI-enabled SaaS products. A support...

I don't lose terminals when they're running. I lose them when I come back. An hour ago I had five...

A mesma pergunta feita a um modelo de linguagem pode gerar uma resposta genérica, uma resposta...

TL;DR: I built MagicKit — a free AI toolbox for image generation, AI writing, and video making. No...

A routed LLM telemetry table can put several incompatible populations under one model label. The comparison becomes honest only after model epoch, role, analysis unit, and proxy eligibility are kept separate.

ImIR 提出用退化图像自身生成的图像指令替代文本提示，通过 VAE 保留结构、轻量 token mapper 将退化图像的视觉-语言嵌入向干净图像的嵌入偏移，从而驱动图像编辑模型完成修复。

V7 推出 V7 Go 智能体平台，通过 Context Graph 将企业文件转化为智能体可查询的记忆，其最难的图查询测试中 GPT-6 Astra 准确率达 89%。平台用 GPT-5.6 Luna 做结构化抽取、GPT-5.6 Terra 和 Sol 负责推理与工具调用，在 HERB 基准上检索系统比官方基线高 69%、不可答查询的幻觉减少 38%。

vLLM 在 GB300 NVL72 集群上对 Qwen3.8-2.4T 做 PD 分离部署，8K/1K 负载下高吞吐场景达到每 GPU 5000 total token 吞吐，低延迟场景每用户 180 生成 token，并给出完整 pareto 前沿。

How We Could Build a Civilization We No Longer Know How to Rebuild For centuries, anyone could look...

Disclosure: I'm the author of rebar, the open-source tool this post is about. It is alpha, and I'm...

Hey there, oji_ai_dev here. Working on AI agents and bots as a side hustle often brings a different...

The Invalid State of PR #48 The merge button was gray. Thirty cycles in, that was the only...

Most AI-in-product integrations fall into two camps, and both have a ceiling. A chatbot bolted onto...

When exploring large language models, you will often come across terms such as base model, chat model...

Iterations in one early set ate about 350 000 tokens each. The only reason anyone could say that out loud is that every task carried an estimate and a recorded fact — estimate, status, incidents and budget on one thread, or the cost of a r…

On a Thursday afternoon a traveller published 178 trip reports at once. Two years on the road, Canada...