There are many cases where it does not work or does something insane. However, the number of hooks exposed are very low and I’m impressed with how often it works. When it goes off the rails, it is usually because it 1) tries to use functions that don’t exist (the distance() function is one that I had to relent and implement; but I’ve also seen euclideanDistance() occasionally), 2) defines functions of its own after invoking them in code, or 3) misunderstands the relationship between the user (whose position is exposed through “cameraPosition”) and planes. I’ve put very little effort into mitigating these but have found that simply adding functions for it to use in common scenarios helps a lot.

I think the window size of ChatGPT is quite limited so I can’t use prompts that are arbitrarily long but it’s not hard to see how increasing this window size and fine tuning on custom APIs could produce a model that would work more often than not.