LLMs are not a linear combination of training data.

Some LLMs have interacted with the outside world, such as through reinforcement learning while trying to complete tasks in simulated physics environments.