[gpugreg]

I think I was using GitHub Copilot when I made the experience that led me to this statement. I guess the experience of using LLMs can be quite different depending on model version and harness.