Example of how ChatGTP can assist modding... Fixing Bugs

Is it always right, no, but it is far smarter than me trial and error to find out what is wrong on my own. I would like the paid plan to not be limited in time to use.

 

I don’t know why anyone hasn’t posted on this. This is a game changer for beginner modders. But yes it won’t always be correct but that’s what you deal with when you use AI

 

Step zero to working with any AI assistant is to feed it as much context data as possible – I usually start with relevant wiki entries and similar files. Otherwise, you open it to a lot of guessing (it made a few mistakes in the video above).

Case in point: ChatGPT pretty much always treating jbeam files as json, which usually results in fixing "bugs" such as comments or extra commas, which are totally fine in jbeam.

Claude AI gives much better results, with proper prompting and safety checks you can get it to create correct node and beam structures or troubleshoot existing ones.

I can't agree with this though, even if Claude manages to get stuff right, it still does it once per a few tries. Because of this, a beginner modder may trap themselves trying to fix code that is fundamentally broken.

And with how much guidance it requires, at best it's a decent tool for experienced modders that want to automate some stuff they'd otherwise have to do by hand. I should mention here that usually you don't save that much time this way.

Spoiler: rant about ChatGPT

Current state of GPT is a joke compared to its competitors. The amount of gaslighting it forces upon you is insane. When it can't get some code right, it will add comments or reformat it just to show that it changed something, even if there was no functional change. Then you have to point out issues one by one, but at that point you may as well fix them yourself.

GPT4 lacked context to work on big files, 4o started being an dumb, o1 is an embarrassment when it comes to actual code work or reasoning (despite being a deep thinking model). I haven't checked o3 with Beam stuff yet, but it's already managed to do stupid things with basic html I threw at it, so my expectations are pretty low too.

No wonder DeepSeek disturbed the market, it gives you results as good as paid ChatGPT for free. And with some fine tuning, it can handle jbeams pretty fine.

Claude AI is much better at coding altogether, though. So far I've managed to get it to create things I didn't provide explicit examples for, i.e. adding extra parameters to engine sound tuning. You still need to know what you're doing, though the difference I've noticed in how it works is that it gets the context much better.