Source context: This field note was first shared through Eric's Advisory Hour Substack notes stream. View the original note.
The note
Accidentally gave up my GPU for the calendar day to an OpenClaw instance so it could run a new set of experiments in packing more intelligence per megabyte in a nanochat model using the ‘push your own code into the weights’ ideas that are floating around.
On the plus side, my office is VERY warm.
Why this belongs in the SLM field notes
This is the practical shape of small-model work: not abstract model talk, but a workstation tied up for a day because the experiment is trying to move capability from runtime scaffolding into the model itself.
The interesting question is intelligence per megabyte. If a nanochat-sized model can absorb more of the behavior normally carried by code, prompts, or orchestration glue, then the local agent becomes cheaper, smaller, warmer-running, and more portable.
It sits next to the other Custom Model Training notes on pico models, nanochat, local inference, and ability training: a tiny operational receipt from the moment the GPU became the scarce resource.
Original source
This local copy preserves the public note text and source link. The public page did not expose meaningful inline media during import, so there are no localized assets for this note. Canonical Substack note URL: https://substack.com/@advisoryhour/note/c-232281577.