[mplanchard]

It writes code well enough in simple contexts, sometimes. But that code is also easy to write, indeed often easier to write than to review. It struggles in more complex contexts and with more complex constraints. Unfortunately, it’s the latter case where I most often have any desire to reach for an aid, and it has failed so consistently and so often there that I have largely stopped trying.

It’s nice when you need to do something simple in an unfamiliar but simple context, though.

It seems though that a lot of the narrative here from its proponents is that we’re just not trying hard enough to get it to solve our problems. It’s like vimmers who won’t shut up about how it’s worth the weeks of cratered productivity in order to reach editing nirvana (I say this as one of them).

Like with any tool, the learning curve has to be justified by the results, but the calculation is further complicated by the fact that the AI tooling landscape changes completely every 3-6 months. Do I want to spend all that time getting good at it now? No. I’ll probably spend more time learning to use it when it’s either easier to get results that actually feel useful or when it stops changing so often.

Until then I’ll keep firing it up every once in a while to have it write some bash or try to get it to write a unit test.