I hadn't seen anything three years ago produced by an LLM that looked better than the most mediocre humans. The argument here isn't about today's output, it's about the potential of LLMs to outperform humans. And it's far from clear that the architecture is bound this way, or that it only repeats stuff it's already heard.