September 14, 2025
Me and my research partner Claude solving NP-hard problems this weekend.
[ed. — that's not what's happening here]
I've been working on building up enough scaffolding that Claude can identify poorly performing edge cases in our graph layout code and iterate on fixes by itself.
This is a nice, open-ended problem: you can try different parameters, you can brute force some things, you can add layout steps that handle special cases, you can use libraries or implement algorithms yourself.
We're not quite there yet. I'm still very much in the loop, even after quite a lot of poking at this off and on for a few days.
A lot of LLM code generation feel like this to me, right now:
1. Productive, but easy to slip into an unproductive round-and-round mode.
2. Long-running, self-directed iteration is not yet possible.
3. Tantalizingly close to producing what I want but not there yet without a lot of looking at the code, specifically directed refactoring, and hands-on "pair programming."
4. My lossily compressed computer science education is still better than Claude's lossily compressed computer science education.
Presumably all of these things will be largely solved in a couple of years.