Comment by cubefox

LLMs are trained with a form of imitation learning, they imitate human (and other) text. It seems indeed not likely that pure LLMs will advance far beyond human ability, since even a perfect LLM could only imitate human text perfectly. But other approaches will follow.