I do find ChatGPT-4 does a semi-reasonable job of guessing why I might be asking something basic and trying to anticipate major "head-desk" errors. We've come a long way from Codex.
It's still not good, but it saves me a lot of time typing boilerplate for simple but long functions. It saves me TONS of time debugging Java errors. Anything complex I have to do by hand, and it's bad at application-wide refactoring of class/method hierarchies, and it's bad at general architecture. But it's 80% great for "wtf caused this compiler/runtime error" and spewing out boilerplate that I still have to double-check but I can double-check it super fast and almost always ask ChatGPT to fix things that I notice.