Is this more general than LLMs? Is it possible to do something Chain-of-Thought-like in a transformer model that _isn't_ trained on language?