Hacker News Reader

Comment by CuriouslyC | Hacker News Reader

AI writes code according to the instructions given. You can instruct it to architect and organize the code any way you want. You got the fruit of your inarticulation.

View on HN

This is what I heard from many people online.

I have tried beign specific, I have even gone as far as to feed its prompt with a full requirement document for a feature (1000+ words), and it did not seem to make any significant difference.

Need to use the Wittgenstein approach: “What can be shown cannot be said.” LLMs need you to show it what you want. Its model has seen nothing. Built up a library of examples. Humans also work better this way. Just good practice.

I think that's why cursor works well for me. I can say, write this thing that does this in a similar way to other places in the codebase and give it files for context.

You are experiencing the Dunning-Kruger effect of using AI. You used it enough to think you understand it, but not enough to really know how to use it well. That's okay, since even if you try and ignore and avoid it for now, eventually you'll have enough experience to understand how to use it well. Like any tool, the better you understand it and the better you understand the problems you're trying to solve, the better job you will do. Give an AI to a product manager and their code will be shit. Give it to a good programmer, and they're likely to ask the right questions and verify the code a little bit more so they get better results.

> Give an AI to a product manager and their code will be shit. Give it to a good programmer, and they're likely to ask the right questions and verify the code a little bit more so they get better results.

I'm finding the inverse correlation: programmers who are bullish on AI are actually just bad programmers. AI use is revealing their own lack of skill and taste.

You can literally stub out exactly the structure you want, describe exactly the algorithms you want, the coding style you want, etc, and get exactly what you asked for with modern frontier models like o3/gemini/claude4 (at least for well represented languages/libraries/algorithms). The fact that you haven't observed this is an indicator of the shallowness of your experience.

> modern frontier models like o3/gemini/claude4 (at least for well represented languages/libraries/algorithms). The fact that you haven't observed this is an indicator of the shallowness of your experience.

I'm not chasing the AI train to be on the bleeding edge because I have better things to do with my time

Also I'm trying to build novel things, not replicate well represented libraries and algorithms

So... Maybe I'm just holding it wrong, or maybe it's not good at doing things that you can't copy and paste from github or stackoverflow

It always feels like the height of HN when some pseudo-genius attempts a snobbish reply but instead just confidently confirms their lack of literacy on the subject.

LLMs write code according to thousands of hidden system prompts, weights, training data, and a multitude of other factors.

Additionally, they’re devoid of understanding.

I truly hope you do better in the future.

I have had great success at guiding LLMs to produce my desired output. Your baseless snark is a sign of your incompetence.

Or that your use case is well-documented across the internet and simple.

That would be more inline with how the technology works and less a sign of your inherent genius though.

> AI writes code according to the instructions given

Nope. The factors are your prompt, the thousand of lines of system prompts, whatever bugs may exist inside the generator system, and the weights born from the examples in the data which can be truly atrocious.

The user is only partially in control (a minimal part of that). With standard programming workflow, the output are deterministic so you can reason out the system's behavior and truly be in control.

Reply on news.ycombinator.com