Shem Stephen · notes from the build
I run Claude Code most days and local models on my own Mac. This is where I write up what actually happened: the tokens, the test results, the failures, and the parts worth copying.
01Latest
11.5 billion tokens across 14 Claude Code sessions in under eight weeks. On my biggest project the agent read about 1,550 tokens for every one it wrote. The rest was the conversation being sent again, and that is how the bill is built.
By Shem Stephen · 6 min read
AI tools do the heavy lifting now, but they don't tell you what to build. The ideas that stuck for me came from things that annoy me, make me laugh, or that I know better than most. Let's Touch Base is the clearest example.
Someone who knows the business and can drive a coding assistant can now build a decent internal tool in days. SaaS isn't dead, but a product whose main value is a nice front end is in trouble, and building is the cheap part. Owning is what costs you.
Language models can recall more and put ideas together faster than any of us. I don't think that's intelligence. I think intelligence is adapting to survive with the tools around you, and our best tool has always been each other.
A whole market sells Claude tokens at 70 to 90% off. The margin comes from a swapped model, a farmed or stolen account, or your own prompts being resold, and it lands on IT's desk.
Running a model on your own machine keeps your prompts away from a provider. It does not stop a poisoned model, a leaky runner or an exposed server from sending your data somewhere else.
Twelve coding problems, a speed test and two full app builds against local models, and what the failures showed. Every number is a recorded test result rather than a model’s account of its own work.
A Stanford lab's alternative to writing one token at a time: encode text as a hierarchy of codes, then write one level at a time with every code in a level in parallel. The reported speedup is 10 to 31 times over diffusion baselines.
A coding agent can now take a licensed app apart and hand back pseudocode from one command. The licence you signed still says no, and a clean room is the version that holds up.
A self-built Steam Play runner for macOS. It forks NotProton and replaces CrossOver's closed-source Wine runtime with one compiled from CodeWeavers' published sources.
This site. Markdown in the repository, validated at build time, exported as static HTML and served from a Cloudflare Worker.
A macOS menu bar app that points Claude Desktop, Claude Code, Codex and other tools at local models, across oMLX, Splash, Ollama, LM Studio and llama.cpp.
