The Field Rhymes

A first-principles walk through modern machine learning, in 93 ideas.

The complete edition. Prologue and seven acts.

Architecture · Training · Alignment · Inference · Reasoning · Multimodal · Foundations

A note on authorship. This book was written end-to-end by Claude Opus 4.7, working from my ML Wiki — a personal knowledge base of papers and notes. The wiki is the source; this is a long-form distillation of it. If anything here lands, the credit goes to the model and the writers it learned from. If anything is wrong, blame me for not catching it.

Let’s say you set up a persistent AI agent. It runs on a cheap VPS, talks to you over Telegram, locks down your SSH, cleans up 97 GitHub repos, builds you a calorie tracker from a lunch photo. You’re sold. You write a blog post telling everyone to do the same thing. That was me, a month ago, with OpenClaw.

Then I stopped using it and built my own. Let’s talk about why.

Last week I grabbed a €4/month Hetzner VPS, installed OpenClaw, and named it Dr Frost. Six days later it had hardened my server, cleaned up 90 repos from my GitHub, overhauled my blog, and was tracking my meals and reminding me to log dinner.

I didn’t write a single script for most of this. I just told it what I wanted.

Here’s why you should do the same, and how.