To be honest, I feel a bit of relief every time AI fails to do something. Like, okay, we've got a few more years...
It looks more like refining a search pattern(like one might do with an LDAP query) on the part of the operator than it does an algorithm "fixing", "changing", and "being guided". It's interesting how we anthropomorphism the output of this algorithm compared to other APIs, even though the algorithm is closer to oher APIs than it is a human as far as we understand.
What is really interesting with ChatGPT compared to other interactive software is that you can give instructions the way you would do it with a human. You can literally copy paste a compilation error, with no more context, and it will fix the previous program it generated. Even just pointing vaguely to something like “that does not look correct, you forgot some edge cases” will result in an improved version.
Yeah those are impressive but basically toys. Although at the moment it's clearly still a research prototype. For some things it works really well and beats Google by saving dozens of clicks and repeated searches, for others it's just plain wrong.