LLMs Learn Any World; Humans Are Built for This One

LLMs lack the domain-specific inductive biases humans have from evolving in a physical world, making them equally capable of modeling arbitrary or counterfactual physics.