>As Smith writes, “Questions like this are simple for humans living in the real world but difficult for algorithms residing in MathWorld because they literally

>As Smith writes, “Questions like this are simple for humans living in the real world but difficult for algorithms residing in MathWorld because they literally do not know what any of the words in the question mean.” Eh, that's mathematically false. Text descriptions of the world can be thought of as <n dimensional projection of a n dimensional real world, at least as seen by humans. With enough projections from different points the real world can be reconstructed, although it's of course likely that the gpt-3 architecture isn't capable of it regardless of how big the model is, that all text data in the world doesn't contain enough information, or both at the same time. That's a different argument however. I think from a theoretical pov, only the former is true - as even without being able to analyze videos and images it's possible to build an internal world model and humans after reading a book about physics and genetic human code, but computational requirements as opposed to being trained directly on videos, images, and simulated worlds make this approach absurd in practice.