AI 教程与实战

An evidence-first playbook for AI-search citations: what the studies actually show moves citation odds, what measurably does nothing, and how to check whether engines cite you.

It's really easy to spin up Unsloth and fine-tune Llama 3 in an afternoon. I wanted to see what...

agent-trace-witness v0.1.0: signed readiness seal, post-execution capture, and PROV-DM...

Last week I argued the obvious part: most production LLM traffic — extraction, classification, short...

Why in-process audit logs fail under agent compromise, and how to fix them with out-of-process,...

My son recently came to me to brag about using AI to find our ancestors. While the results were...

This started as an abstract way I think about AI. I wanted to manipulate files inside my own local...

ChatGPT Plus is $20 per person per month. A team of five is $100/month, forever, whether anyone used...

Disclosure: I used AI to help organize and edit this article. The implementation details and...

Enterprise AI in 2026 has transitioned from simple prompt engineering into complex agentic software...

As a developer, I've learned that the strangest requests often lead to the most interesting technical...

AI coding tools have become extremely capable. Claude Code can understand a large codebase. Codex...

A deliberately strict process for preventing false completion in autonomous software...

Architecture diagrams become unreliable when the code changes but the diagram does not. Archify...

I run an OpenAI-compatible gateway. Every day I fire the same request at it three times and write...

You can scrub every password from a prompt and still leak a working map of your company....

Do you ever feel like the AI world has moved forward a little too quickly? You hear about...

I love AI I used every single day. so this is not, by any means, an antiAI rant. quiet the opposite...

I set out to learn Amazon Bedrock AgentCore, AWS’s managed runtime for AI agents. So the plan is to...

AgentGPT pricing in 2026: free tier, $40/mo Pro, custom Enterprise — plus how the 25-loop cap and reliability decide what your money actually buys.

The best prompt I ever wrote was a README. What changed when I moved my conventions out of the chat window and into the files they describe.

A few months ago I dug out a box of family photos from the 80s. Most were faded, scratched, or had...

I built an arena where two AI agents fight over a network in real time. One of them refused the job. The other one started laying false trails.

A session commit reported success. The memory extraction produced zero memories. No error dialog, no...

I sat in on Margaret-Anne Storey's DORA community session last week, and she put a name on the thing...

GitHub Repository: https://github.com/13manmayarai-hash/omnicache-proxy (MIT Licensed) What...

Before asking an AI coding agent to “fix this bug,” there is one question worth asking first: Does...

One of our product levers is capped at three executions per ISO week. In week 2026-W35 the ledger...

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

LangChain 发布指南，介绍 Schneider Electric、Vodafone 和 monday.com 在欧洲与中东规模化落地生产级 AI 的做法。内容涵盖建立共享 Agent 平台与 LLMOps 实践，以及设计具备更强可观测性、评估和控制能力的多智能体架构。

Three gotchas installing a plugin into someone else's agent This one is scoped tight:...

The discriminator between a membrane and a logger The first version of this test passed...

Meta 首席 AI 官 Alexandr Wang 转发用户 @jaystar524 的实测反馈，称 muse spark 1.3 表现可靠。该用户用多个 sub-agents 并行完成任务，耗时约一小时。

Alexandr Wang 发帖号召大家试用 Muse Spark 1.3，并引用用户 @DeryaTR_ 的体验：用该模型快速构建了一个海盗平台跳跃游戏，一次生成即无缺陷运行，作者称其正在继续美化美术、精灵图并添加新关卡，认为这是一个快速、便宜、好用的编码模型。

LangChain 博客文章介绍 Nevermined 如何赋能 LangChain 智能体购买和出售服务，正文无更多细节。

Hugging Face 发布完整教程，使用 TRL 库以 GRPO 微调 LiquidAI/LFM2.5-350M，约 500 样本、100 训练步即可在免费级 Colab 或 Kaggle GPU 上完成，训练后模型在 IFStruct 基准上从 22.6% 提升至 29.7%。

研究者提出扩散增强LLM（diffusion-augmented LLMs），通过轻量扩散权重并行生成多个token，在不引入草稿模型或牺牲AR模型质量的前提下加速生成。配套的Ψ-Spec采样器实现无损加速，Uno模型吞吐量优于主流投机解码方法，较基础AR模型最高提速3倍。8B Uno在智能体工具使用、编程和长上下文推理基准上超越26B DiffusionGemma及Mercury 2。

EVOHARNESSBENCH 是一个新基准，用于评测智能体在工具、技能和智能体三个维度的受控框架演进下的表现。该基准包含 17 个多阶段流、802 个任务、520 个工具、42 项技能和 62 个智能体。结果显示，框架扩展本身会降低此前已解决任务的性能，产生框架引发的遗忘，且保留旧能力与适应新能力可能相互冲突。

伦斯勒理工学院研究团队提出"循环状态写回"概念，揭示低精度推理中状态存储规则对循环网络时序行为的关键影响。在荧光寿命成像的GRU编码器-解码器中，将连续状态传播替换为确定性4-bit状态存储，使τ1和τ2估计误差分别增加约70倍和300倍；误差反馈、残差记忆和方向记忆可在不重训练的情况下恢复精度。该现象在独立训练的LSTM中同样复现，表明状态存储接口是量化循环推理的核心设计考量。

Lint is green. The bot left forty comments about naming. Somebody typed LGTM. Then billing...