The limit could also be in the training data. If we could train LLMs with a corpus generated by some super intelligent entity, maybe we'd consider them super intelligent even with the same model size/architecture we have today.