LLMs don't have access to reality, though. They're trained on the symbolic output of other models.