Can we talk about the week GPT-style models wrote half my code and I shipped 3 features in 5 days
I gave the assistant full access to my repo on Monday and by Friday I had three features merged that normally take me two weeks, mostly because it caught a null bug in my auth flow I would have stared at for hours. I still read every line, but at this rate hand typing boilerplate feels like washing dishes by hand when there's a machine sitting right there. Has anyone else noticed the gain is way bigger on boring glue code than on the actual tricky logic?
Wait, did it really catch a null bug on its own or did you point it at the auth flow first? The part I'd push back on is the idea that it "caught" anything, since in my experience the model mostly spots stuff when you aim it at a file or ask it to review a diff. Left alone it'll happily write new buggy code right next to the old buggy code. So the win isn't some magic bug radar, it's that feeding it a diff is way faster than you re-reading the same block for the tenth time. Either way the glue code point holds up, that stuff is where it saves the most time.