I’m pretty sure that LLMs can generalize a deep model and can create novel assemblages and are not just search engines.

The burden of proof would be on you to find the existing mathematical, erhm, proof.