alt Hacker NewsIt’s worthwhile to me under specific conditions, mainly that the codebase has existing patterns and abstractions that the LLM utilizes, and they’re actually good. It’s not a common case but I’m lucky that one of the codebases I’m in is one I built myself mostly by hand, and LLMs are really effective in it at the moment. My main concern is making sure the generated code maintains the high standard, but it’s definitely saving me time.
I’ve kind of changed my mind on prompting, it’s definitely a skill. It’s a skill based on your own skills in the domain. I’m at the point where I can get the LLM to generate the same code (roughly speaking) that I would have written. So it’s basically generating the same thing I would write, just faster. So it’s like reading your own code. Using it as a crutch to do things you aren’t capable of is where people run into trouble. That’s where the massive amounts of code review come into play. For me, I’m only ever reviewing 100-300 loc changes at a time. Often less. Because I know what I’m doing and can break things down into manageable diffs.
Can’t see myself going back, but also can’t see myself doing it without the experience I have without LLMs. Which is a bit of an issue for new developers. Not sure what the solution is for that.
Replies
That plus if you’ve ever stared at your code and then searched StackOverflow to see if you could find a better way of doing it, it’s like having that running continuously.
I’ve reached a similar conclusion, but there’s a part that worries me: the expertise that allows us to judge AI’s output was itself built by doing the work we’re now delegating. So there’s a risk that our judgement will decay over time. I’ve been thinking about the problem as choosing where we can afford to “borrow” comprehension, versus where we need to keep exercising it, and how to “claim back” the critical comprehension we lost.