Raw LLM Responses

I've used AI to learn k8s faster than I normally would, and to kinda co-architech with me since I'm the only backend dev at my job. There's no reasonable way it can replace us anytime soon, but I do think companies will try. I have to be very verbose and direct when I need it to generate code. "app/handler/story_handler.py needs to do this". And even then it struggles. In my experience using AI heavily for the past 6 months: It's decent at architecture. It's nice to get feedback of any kind honestly, but it often just agrees with you. So you have to let it make the first decision then refine it and have it defend its position. It's decent at boilerplate. I have a style guide and new service guide. It adheres to them maybe 75% of the time. But even then, when scaffolding new applications with a lot of boilerplate (Dockerfile, Jenkins file, deployment files for k8s, folder structure for the app, entry points, etc), even half the time is very nice. It's terrible if you do any kind of SOLID, DRY, KISS, etc. I have it mentioned in 10 different places to keep my interfaces generic (RedisInterface vs CacheInterface).If it didn't specifically generate the function name of something, it either tries to change it on you for no reason or calls a non-existent function. Copilot specifically when editing a file will occasionally remove my code and out # ...existing code in place of it. Which would be fine in chat mode but in edit mode you just replaced my whole file with that. People keep making these "one shot" videos of AI making like a very basic game. I think when everything is in one file it works pretty well. More complex programs need a ton of direction though. If I didn't already know what I'm doing to an extent, it wouldn't get anywhere. Overall, I think it has helped me be more productive. But check its code a lot and frequently. And think of it like a feedback system, not something to do your job, because it will fail.