The body is the missing link for truly intelligent machines

I feel that virtual or real embodiment is probably a good way to acquire sensory grounding of language and common sense. But I don't believe it's the only way to acquire it or necessary for the distribution or use of all of this knowledge.

What is necessary is a way to encode spatial-temporal information at different levels of abstraction. Maybe the best way we have to do that is neural networks. Or maybe something like networks of networks.

But it's a bit of a leap from the idea that training on text only is inadequate (which seems obviously true to me) to the conclusion that embodiment is necessary.

The next step might be to train on captioned video.