This argument only works if you assume all output of an LLM comes merely from its training data, and that it receives no alignment via RFHL, no outside data ground truth via RAG and so on. The engine might be a probabilistic token predictor, but the car is the sum of its part and those parts are not just the engine.
No it still works, because in that case all the LLM is doing is making an API call based on the same token prediction, and passing on the result. It still doesn't "know" anything about anything.
