Post

My own spectacular LLM failure was also with sim/geometry/rendering C++ code. It simply couldn't stay consistent about our right-handed coordinate system, kept sneaking in left-handed stuff. But that was forever ago (i.e. about three months) and I'm told the models are much better now.

If you're going through the chat interface that's very different than what most people are talking about. I don't know how often the agents one shot correct solutions because the power is in their ability to run code, test, and iterate.

I believe you are indeed using it differently. Try an environment designed for coding: Cursor or Antigravity are state-of-the-art and have free trials. ChatGPT won't be great.

Agree with others here: the results you get from a tool that can actually compile and then run the code are a world apart from if you copy and paste things back and forth with ChatGPT I really like Claude Code, Codex CLI and Gemini CLI for this

Have you tried using a IDE with an agent integrated like Cursor or VsCode? VsCode+copilot has a free tier, I recomend you to try that instead of copy-pasting to Gemini, that way the model doesn't have the right context and can't modify your code, that is how we used to do it 2 years ago! :P