Disagree. The input/output structure (tokens) is the interface for both inference and for training. There is an emergent intellect embedded in the weights of the model. However, it is only accessible through the autoregressive token interface.

This is a fundamental limitation, much more fundamental than appears at first. It means that the only way to touch the model, and for the model to touch the world, is through the tokenizer (also, btw, why tokenizer is so essential to model performance). Touching the world through a tokenizer is actually quite limited.

So there is an intelligence in there for sure, but it is locked in an ontology that is tied to its interface. This is even more of a limitation than e.g. weights being frozen.