I mean, a bunch of rocks falling on a keyboard can also generate tokens. So can an autocomplete software running on a 10W phone. But the quality is so far apart to the point you would find most people do not consider them on the same categories of "inference" as an LLM, not to mention a real human's mental capacity.