Comment by microtonal

It’s not been a cakewalk. I’ve needed to toss out large swaths of LLM-generated code, and rewrite by hand, but, for the most part, it’s been a huge help.

But your anecdote is much more balanced and more in line with my personal findings. Not like all the AI astroturfing that happens here. I like to use LLMs as well, but in a very targeted way in places where LLMs shine.

Yes, LLMs can make you much more productive. But so could assembly -> C -> Python or Rust, or switching an IDE with code completion and support for refactoring. Each step makes you more productive.

Sure, LLMs can spit out greenfields projects. But on large projects with complex requirements, you still need senior engineers to guide them, carefully review the output, and as you say throw out code and write it from scratch in a better way.

I had some friends who ended up in bits of AI psychosis. They exclaim that a swarm of agents was writing all their code, but every time I ask them to show the end-result, all they have is a pile of code that they don't understand, nor doesn't really work either. At the same time, they stopped getting any actual work done.

At any rate, somebody had a great analogy on HN recently: think of it is a vector, LLMs can significantly increase the magnitude of the vector, but you still have to make sure that the orientation of the vector is correct.