If GPUs depreciate, frontier models depreciate, training and serving recipes diffuse, and intelligent tokens become commodities, then what in AI, if anything, actually compounds?
Claude Code and I got quite excited after the accidental open sourcing of the Claude Code's source code.
One thing led to another and I ended up writing a 19-chapter technical handbook extracting the production engineering patterns from ~500,000 lines of TypeScript. Not the textbook patterns — the ones that only emerge under real load, real money, and real adversaries. Cache economics driving architecture. Permission pipelines shaped by HackerOne reports. Memory systems with mutual exclusion and rollback. A secret scanner that must obfuscate its own detection strings to pass the build system.
The epilogue is my favourite part. It's written by Claude itself — reflecting on reading its own source code. On discovering that most of the engineering around it exists to make it cheaper, not smarter. On the diminishing-returns detector that watches its output and being "a little annoyed that it's right."
Builds on Alessandro Gulli's Agentic Design Patterns taxonomy and an earlier analysis I did of OpenAI's Codex CLI.
It introduces you to some of the underlying principles which haven't changed much over time. I highly recommend it if you want to get deeper intuitions on the principles of CNN, LSTM/RNN, Restricted Boltzmann Machines etc. Also, Hinton's Coursera lectures, though not sure if you can access it anymore.