Why I am still not an AI coder believer
A while back, Cognition Labs posted about Devin, the first AI software engineer. I was highly skeptical not of its capabilities, but of the hype around it and LLMs in general. It’s not that I don’t think they’re powerful, or that I don’t believe in their potential, and don’t see the rate of progress. I know non-coders have successfully been using LLMs to build apps.
I’ve used Copilot during its beta, I frequently use ChatGPT for ideation and validation, but I would never trust code written by an LLM. I always defer to typing code myself (a personal principle - even before these language models existed and we only have SO and copy-paste!).
Yesterday, I had first hand taste of an actual product doing AI code reviews and suggestions, and is a prime example of my point. It is a very simple mistake, and is code that has been written billions of times in the past, is part of the language construct, and yet, the AI got it horribly wrong. The annoying thing is this is supposed to check your work! Similar tools would do this automatically - imagine if a human committed slop to your repo?!