Comment by proc0

There's probably an analog to it, but I really doubt it's the same or even close to it. Just think about how we read, it's not tokens it's visual representations. I think LLMs are tapping into some larger knowledge structure that we also use.