No, I said the best way to learn to predict human text is to capture human knowledge. I did not say how best to capture human knowledge. Obviously doing gradien

No, I said the best way to learn to predict human text is to capture human knowledge. I did not say how best to capture human knowledge. Obviously doing gradient descent over a huge text corpus is massively inefficient for a human. Whether its the best way for a machine to learn it is another matter. After all, the amount of data that humans experience through their senses during formative years, as well as the "computational work" that went into evolving human brain architectures surely dwarfs GPT-3's training regime.