Hacker News Reader

I thought LLMs had pretty big promise until I started trying to use ChatGPT to help me learn Godot. Literally every answer it confidently spit out was just dead wrong. Didn't matter if I specified that was using v4.22 a million times, it'd still give answers mixed in with v3, or in some cases even solutions from other frameworks like Unity. It also mixed up GDScript and its shader language.

Just absolutely awful and unusable. God help whoever has to review or modify a game made in part by a different dev using ChatGPT.

> 5. A negative view would be that AI coding assistants mean people don’t necessarily need “libraries” like this if an LLM can spit out the same results in a couple queries (rather than hunting around Github or the internet for pre-packaged code).

Why use a library when the LLM can spit of a unique reimplementation of the library logic every time you need it?

The future is terrifying.

People like having standards, that's why. They make processes more simply repeatable and experience more easily transferrable

A positive view would mean an uptick in library quality and sophistication, and a death-knell to the trend of shallow do-one-thing script wrapping libraries.

I've been doing similar mobile development with Godot, and have found the Phantom Camera library to be an example of a better developed and thought out library thats actually worth importing.

An LLM that just replies with an executable?

> Here’s a quick example prompt

You almost certainly should not be updating the character.position both every frame (in _process) and in response to a MouseMotion event. That makes no sense.

Yep, it is definitely worth clarifying that all the code has to be reviewed manually still. An aside is that the LLMs (both Claude and GPT) have common failure modes in using Godot 3 syntax instead of 4, etc.

I meant to make the point that, having worked on mobile game dev recently, if the behavior I'm thinking of seems simple, my workflow instincts have become to go to LLMs first rather than Github. Very possible this is a bad reflex. Just one I've observed directly from my experience implementing similar functionality as the parent library.

It looks like spywaregorilla's comment was downvoted at least once but I think it was making the valid point sarcastically that the idea of "LLM generated code replacing commonly used libraries is a bad one." My original post was positing on whether this might be a future trend though.

Also want to recognize the author of the library (prophesi) commented down below and definitely applaud their work in packaging and open sourcing their work plus making the tutorial video.