Week 0x27

I’m really back to the work I was doing before ripping out Letta: building state machines to make the loop a big more deterministic. The model gets nudged to stay on task, warned when it’s about to run out of “budget” for reading operations without making any changes, etc. It’s all very nuanced and I burn a lot of tokens with failed runs.

The funniest moment was when a poor model got stuck with Den telling it “no, finish the task!” which was to commit a file that I had deleted. It felt like I was rescuing it when I stopped the loop and talked about what options it should have in such a situation.

I spent most of the weekend really cleaning up the codebase. “Vibe coding” is supposed to mean forgetting the code even exists, but I don’t believe that’s sustainable. It’s still in the spirit, IMHO, to run evaluations and synthetic code reviews. With temporary access to Fable, I asked it for a thorough assessment of the architecture. I don’t agree with it completely, but it had enough really solid advice to impress me.

I will also admit to a touch of pride when Fable wrote:

This is an unusually disciplined architecture for an assistant platform.