Notes for March 23–29

Work ate the week again. I’m exhausted, running on fumes, and daylight saving time stole an hour of sleep I could not afford–the biannual clock shuffle is one of those vestigial absurdities that nobody can be bothered to abolish, and I’m starting to take it personally.

I did manage to get my AI minions to do a proper refactor of the piclaw codebase (which was desperately needed), spent a bit of time cursing at the SK1 (and concluding that I need a new 3D printer), and that’s about it. Meetings, deadlines, the usual corporate grind.

This was the week of the great piclaw codebase reckoning–299 commits, all of them aimed at breaking apart the monolithic mess it was becoming before it got completely out of hand.

Agentic development works, but, again, you need to have taste and force a feedback loop to get good results. But you can automate away the boring parts, mostly.

The whole thing was driven by an autoresearch-style loop–I basically adapted Karpathy’s approach of having an LLM do research on a codebase, generate a plan, execute it, and then verify the results, except in my case the “research” phase also involves running the test suite and feeding failures back as context. It works very well for mechanical refactoring like this, where the risk of hallucination is low and the reward for not doing it by hand is immense:

I still haven’t written up those SBC benchmarks I keep promising, but I have been using the SBC extensively–in fact, I dusted off my macemu fork (BasiliskII/SheepShaver with Raspberry Pi-optimized SDL2 builds, Docker images, and .deb packages) and got that running on it.

It’s been sitting at v1.0.5-patch-3 since February, but Marchintosh guilt finally got me to fire it up and poke at it. No new Mac Classic replica yet, but at least the emulator is working, and I am back trying to get an ARM64 JIT working in it, which is a fun challenge:

I expect to have something to show on that front… this year?