Sebastian Grebe
实战笔记
项目
实验室
自测
履历
联系
订阅
订阅
实战笔记
把
AI
送上生产环境,全程公开。
研究、开源工作和诚实的复盘。每月一篇长文。
下一篇实战笔记
AI 试点为什么卡住
每月深度文章:真实数字、哪里出了错,以及一份你可以直接照抄的模板。先发给邮件订阅者。
发到我的收件箱
免费就绪度自测
基线
人工复核
负责人
示例答案。
开始自测 →
全部
Announcement
Research
Engineering
Open Source
AI
Books
2026年8月29日
Winding Down Tandemu — the One I Said I Was Keeping
Twelve days ago I shut down Substaff and pointed at Tandemu as the project I was most excited about. Now I'm shutting Tandemu down too. Here's what changed, what I got wrong about building a measurement product, and why the code stays open.
Announcement
→
2026年8月17日
Winding Down Substaff — What I Learned Trying to Run a Company of Agents
I'm shutting down Substaff, my platform for hiring and managing AI employees. Low activity, real infrastructure bills, and a thesis that turned out to be early rather than wrong. Here's the honest retrospective.
Announcement
→
2026年6月16日
Winding Down Kreye — and Open-Sourcing It on the Way Out
After a long run, I'm shutting down Kreye, my AI-powered dynamic canvas workspace. Low activity, real infrastructure bills, and a stronger pull toward other projects made the call. Here's the honest retrospective — and the code, now open source.
Announcement
→
2026年5月18日
State Runtime — Can a Hybrid SSM Be the Application?
I built a web app where the only state is the hidden state of a Mamba/Transformer hybrid. No backend routes, no React useState — just a model, a Postgres, and a WebSocket. What worked, what didn't, and the two things that almost killed it: model intelligence and decode speed.
Research
→
2026年5月10日
DeepVector — How Much Retrieval Signal Lives in Mamba's Latent State?
I spent two weeks asking whether Codestral-Mamba-7B's hidden states are good enough for code retrieval without training. Mean-pooling is a disaster, MaxSim recovers a 35× lift, and four different frozen architectures all fail to beat it. The thing that finally worked wasn't more architecture — it was composing the latents with tree-sitter and an LLM.
Research
→
2026年5月6日
AI Code Ratio Benchmarks: What Percentage of AI-Generated Code Is Healthy?
0-20% barely used, 20-60% healthy, 60-90% heavy, 90%+ rigorous review. Benchmark ranges for AI-to-manual code ratio — and the drift metric that matters more than the absolute number.
Engineering
→
2026年5月2日
autoportfolio — A Stateful AI Portfolio Manager for Claude Code
I open-sourced an autonomous portfolio manager that runs as a Claude Code plugin. Interactive trading sessions, overnight monitor mode, multi-currency P&L, all driven by a single instruction file. Inspired by Karpathy's autoresearch.
Open Source
→
2026年4月17日
nextjs-indexnow — Ping Bing and Yandex Every Time You Deploy a Next.js Site
I published an open-source Next.js plugin that submits new and changed URLs to IndexNow after every build. Zero sitemap dependency, Turbopack-safe, diff-based. Here's why it exists and how it works.
Open Source
→
2026年4月8日
Launching the Lab — Free AI Templates, Prompts & Agent Blueprints
I rebuilt my personal website from a static portfolio into a creator hub. Today I'm opening the Lab — a free resource library with AI agent templates, trading prompts, and starter kits.
Announcement
→
2026年4月4日
Friction Detection: How to Identify When AI Coding Assistants Waste Developer Time
AI tools fail silently. Your DevEx dashboard can't tell a productive session from a prompt loop. Friction detection is the missing metric.
AI
→
2026年4月2日
How to Detect AI-Generated vs. Human-Written Code at the Commit Level
Legacy SEI tools guess whether code is AI-generated from IDE telemetry. First-party attribution gives you deterministic, per-file ground truth without monitoring developers.
Engineering
→
2026年3月31日
Beyond the Prompt: Why Your AI Teammate Needs a Memory
Most AI coding tools operate in a vacuum. Tandemu introduces persistent memory that learns your style, adapts to your personality, and gives leads real metrics.
AI
→
2022年2月1日
Full-Stack Web Development with GraphQL and React — 2nd Edition
Second and overhauled guide to building a scalable, production-ready full-stack application using React, Apollo, Node.js, Express, GraphQL, Sequelize, and AWS-based CI/CD.
Books
→
2019年1月1日
Full-Stack Web Development with GraphQL and React
Practical guide to building a scalable, production-ready application using React, Apollo, Node.js, Express, and GraphQL.
Books
→
Sebastian 的实验室
每月一封邮件,讲把 AI 送上生产环境。
新的实战笔记,加上就绪度自测和实验室里的新东西。随时可退订。
邮箱地址
订阅