miniblog.

I've been impressed with code written by Fable in my testing: Difftastic: found small optimisations in a hot loop I'd already profiled extensively. Helped me prototype Dijkstra to A* too (hard to find a good heuristic). Garden: Found some real bugs in my simplistic typechecker.

I find that I'm choosing AI tools based on the quality of the harness rather than the model. For example, I'm using Claude Code Web because its model of remote VMs is extremely convenient. I'm using Perplexity because it's tuned really well for web searches.

One fun way of testing new AI models: take an existing codebase you have and just ask them to "review it and fix bugs". In principle this should find more issues over time as models get smarter. I've found a few bugs this way at least.