Same's true for the LLMs. Seventy bajillion parameters from every written word they could scrape up, with millions of hours of human tagging assistance and trillions of CPU hours, and they're still not quite there. Most humans get to this point of literacy with a few hundred books (at most) and maybe a couple thousand hours of actual reading time.
So yeah. Machine learning has done some great things, but it's still really pathetically inefficient.