I'm not convinced GPT can infer correct actions needed for novel scenarios for contexts that have not been directly experienced in training.

Consider:

I have never in my multiple decades of existence had a ball roll out in front of me while driving. But if I were driving by a playground and a ball rolled out in front of me from between parked cars, and the ball itself is not a hazard because it has already passed my lane, I can infer that I need to slow or stop because a child may rapidly follow that ball.

I'm not convinced that GPT or any AI can infer that context for handling novel edge cases at this point. It could just be a need for more training data to correlate different contexts, but I don't think it's there yet.