> My impression from a couple years ago was that it was fairly decent at coding, it was just slow to go from question -> code, and the tooling around that has improved significantly so that it's all pretty quick

Your experience is very different from mine. Early GPT/LLM tech was hilariously wrong. It famously hallucinated code out of nowhere, made breaking changes all the time, failed to follow very simple instructions. I remember when it couldn't play Tic Tac Toe! It hallucinated board positions and rules. I used to break it all the time, for fun (and it didn't take much, it mostly fell down the stairs on its own). Now it can play far more complex games.

Was I right to be skeptical? Well, based on what I saw, I was right. GPT was impressive and fun but also hilariously wrong most of the time. Until they weren't!