Are the questions as simple as a = F/m = mg / m = g?

Anyways. If humans put effort into optimizing a tool for accurate simulation, we can treat it more like an alignment problem rather than pure scientific judgment.

You can update the knowledge in the physics engine if you want.

This is what I got: GPT-3 Answer.

Apparently, the model arrives at the wrong answer without mentioning the air resistance. I have tried many times the results are consistent.

Considering the free fall rules should be encoded in some text books (which should have been included in the pre-training datasets), these results are even more striking to me.