I've been using chat gpt plus quite a bit for supporting me in programming tasks. It's too flaky to trust blindly and you need to narrowly scope what it does. I

I've been using chat gpt plus quite a bit for supporting me in programming tasks. It's too flaky to trust blindly and you need to narrowly scope what it does. It can also be a bit hand wavy with some things and I've had issues getting it to generate working code for some more complicated things. But where it really shines is in documenting things, explaining things, and suggesting problem areas in code and possible solutions. And you can do fun things like point it at an openapi schema and asking it some questions about that API. I've given it screenshots of websites and asked it to criticize the design or document what is visible. It's amazingly good at supporting localization work. I was working with some geospatial code for an algorithm that generates UTM coordinates from GPS coordinates a few weeks ago. I needed a Kotlin implementation that I could use on multiple of its platforms (js and jvm, i.e. no java dependencies allowed). It was kind of useless writing the code for me for this (that's the first thing I tried obviously) but it unblocked me a couple of times and helped me figure out some details. I ultimately found several old Java implementations that contained a lot of undocumented magical numbers and I just asked it the meaning of those numbers (about a dozen) and it came back with good explanations (basically things like wgs84 ellipsoid parameters, radius of the earth, etc.). The code wasn't quite working (it failed a test I wrote) so I asked it to identify possible causes for that and it came up with some uncovered edge cases that I was able to cross confirm with other implementations. In the end I was able to piece together a working implementation by combining different elements from several implementations. Each of them individually had issues. A lot of this code is ancient and there is apparently a lot of copy paste reuse in this space.