Github copilot can do some impressive things, but it also ignores my instructions to not try to run anything and leave testing to me that I’ve stopped bothering saying it and just block the attempt when it asks permission. Just yesterday, it confidently said it had figured out an issue I was debugging with it and made a bunch of code changes that literally only affected comments. If I leave it in agent mode (which allows it to edit code) when asking a question to clarify something and not intending any code changes but wanting to think about the answer (and telling it that), sometimes it still runs ahead and tries to make changes anyways.

When it does well, it’s uncanny how effective it can be these days, but it’s not reliable enough to be trusted to be in control of the whole system. Plus I don’t trust Microsoft enough to put my data on onedrive, and believe that access to data is the real reason behind their AI push, no matter how much usefulness and reliability improves.