@aeva yeah, typically what I end up doing when I go that route is that I go into something vision related and end up specializing in classical techniques that are applied downstream of the ML magic (sometimes this involves meshes, sometimes image processing, once I got to do some tracking/controls adjacent stuff). Which was fun work back before the AI coding agents got good enough that it no longer made sense to ignore them. Honestly it's probably still fun, I also made the mistake of taking on a project that puts me inside the ML bits because I wanted to learn more about the ML bits. And *that* has turned into a few hard lessons about how knowing one set of things really well doesn't translate into knowing other things really well, especially if I don't already have decades of experience in the other things.
Also also, the roles I land in these domains aren't nearly as lucrative as the ones my spouse reads about in the news and wants me to get.
But yeah, the coding agents feel like they're too helpful to not use, but also too dangerous to use heavily, and I haven't settled on a middle ground that works for me. If you think that the right middle ground is "don't use them at all", I'm open to hearing you out.