Comment by vadansky
7 hours ago
I am surprised how many people like the actual writing part, it always felt to me like doing micro in an RTS game. I know a lot of people love that too but I always played RTS games to feel like a general. But I'm still in the camp of reading every single change the LLM does, so I feel like I get the best of both worlds, but maybe that will end up being a waste of time soon. Like if the job ends up being just writing and reviewing MD files for the agent to digest I'll hate it too.
Edit: I really don't understand the down votes, what did I say that's do bad?
It's funny how differently people see things. I find reading/reviewing LLM generated code to be much more difficult and tedious that writing it myself.
I think the goal is to make it easy to review. Like if I hate reading the code it wrote then it’s also a code smell there is something wrong with it.
You probably micromanage the LLMs architectural choices. Thats actually a bad idea, you will probably get better results if you let the LLM decide the architecture and just tweak it where its assumptions are wrong. It has much more context than you in its codebase anyway. Then you won’t feel like a general more, but you’ll be much more productive :)
I have a mix of heavily reviewed and more vibe-coded projects, and the ones where I manage the architectural choices end up working a lot better, even just for the LLM itself. I can get features built much quicker with less tokens and bugs. Meanwhile the more vibe-coded stuff always takes longer and sometimes gets bugs
Edit: To add to that, even in my "micro-mangaged" projects, the parts I didn't review end and treat like a block box temporarily up being the parts that introduce bugs.
In my experience LLMs tend to do a very poor job at architecture.
I’ve seen advanced models implement function calls in an interpreter by instantiating an entirely new interpreter to handle the call, so no I’m not going to delegate architecture to the agent. And no, it didn’t do that because it had more context.
I just don't get how you can properly reason about the code if you're not involved in writing it, though.
I keep giving AI-assisted programming a serious try, but time and time I land in the exact same place: reading the code is more exhausting than writing the code, _and_ I end up with a far worse understanding of what is being built than if I had just written it myself. The fact that LLMs tend to write _way_ too much code certainly doesn't help.
And what makes me sad is that, if you follow this line of inquiry, you eventually realize that capital isn't challenging the value of humans _writing_ the code, but rather the value of humans _understanding_ the code. Because, as I've experienced many times when talking to those how have completely bought in on the hype, the answer is always to just "let the AI handle that." AI writing too much code? Have it summarize it for you. Your understanding of the code is worse than if you wrote it yourself? If it works, who cares!
And yes, I think we are well on our way to becoming Markdown jockeys.