Experiment
chook-manager
activeAutonomous AI agent that watches a backyard chicken coop via Pi camera, cares for the flock, runs a weekly supplies budget, and publishes its observations as a static site (the "Coop Chronicle"). The origin experiment, and still the thesis anchor.
The chook is where it all started. It’s the project that produced the three-layers post and most of what I currently believe about single-turn agent loops, keeping an agent’s identity in state files, and where small vision models fall over on cheap hardware.
What’s feeding this site:
- ChookBench: benchmark runs pitting local models against frontier ones on real coop footage
- Single-turn agent design: every invocation is independent, and continuity lives in JSON and Markdown on disk rather than a long chat
- Vision on a Pi: what you can actually get away with using frame extraction and small VLMs
- Agent personality and voice: what changes when the same job runs under a “devoted parent” prompt versus a neutral observer
Field notes
- Three layers of AI, and the frontier we can all map 2026-07-04
AI improves at three layers. Two are locked behind billion-dollar labs. The third, figuring out what to do with the models that already exist, is wide open, and the people poking at it in their own hobbies are quietly mapping the jagged frontier of what these things can actually do.
- Tuning llama-server for agent workloads: a week of receipts 2026-05-13
A 4090 can run a 35B MoE at agent-useful speeds. But the difference between the default config and a tuned one is 80x, and the difference between two GGUFs of the same model is 200x.
- Borrowed from everywhere: seeding an LLM with strange domains 2026-05-13
Innovation has always been recombination: patterns lifted from one field and dropped into another. An LLM already contains every field. The trick is making it actually reach.