I can invent a grammar, explain it to an LLM, and it can apply it correctly in ambiguous situations. If its just generating output based on statistical probabil

I can invent a grammar, explain it to an LLM, and it can apply it correctly in ambiguous situations. If its just generating output based on statistical probability based on its training corpus, how would that work? It feels like there's something emergent there that isn't just a very good probabilistic calculator.