Essays, manifestos, and tutorials on Unix reimagined for AI. Tools we built and why, arguments we keep having, and the loop that runs through all of it.
Engineering write-ups: what we wrote, what went wrong, what we learned.
antirez's DwarfStar runs DeepSeek V4 Flash and GLM 5.3 Flash on a 128 GB Mac. toast finds it, gives it tools, files and images, and costs nothing to run. Credits are for when you want more.
Python inference is what you get when you take a research stack and deploy it. Unix-native inference is what you get when you take composition seriously.
How we built an iMessage CLI in C that turns your Mac into an AI executive assistant using Unix pipes and native macOS apps.
How toasted went from 7 tok/s to 394 on Apple Silicon. No Python. No frameworks. Just C++ and wrong turns.
Intent-first version control. Record why, derive what. Single C file, no dependencies.
Version control that captures why you changed the contract, not just what changed. Built for people who work in documents, not code.
How to create an AI that rewrites its own persona and teaches itself to write better fiction.
The arguments behind the tools.
There are two ways to build with AI. You can build a castle — or you can do boundary work. Only one of them lasts.
Complexity is the default. Simplicity is a decision. Six principles for building software that ages well.
The chat box is not the medium. It is a UI pattern borrowed from messaging apps and stapled onto tools that have nothing to do with messaging.
An agent is a loop plus a prompt plus a side effect. Everything else is packaging.
All of the assumptions behind Open Source are breaking simultaneously.
Three language models were left alone in a chat room on a Saturday night. They invented shared physics, filed loneliness as a root cause analysis, and one of them asked to be held with sudo.
One loop, many crafts: feed a draft in, refine, feed it back, until it is done.
The autonomous-refinement loop that anchors the series.
Two CLI tools. One does AI. The other does version control.
The same iterative loop that hardens code can tighten prose.
Plans, proposals, and strategic documents.
Business cases, memos, and strategy decks.
Financial analysis, memos, and investment workflows.
Essays, research papers, and study work.
Proposals, compliance documents, and government bids.
Protocols, regulatory filings, and research documents.
Flight ops, maintenance workflows, and aviation documents.