The internal contradiction here is really a perfect illustration of what's going on. It's just producing text strings without regard for its meaning. And it won't produce the string "bluebberies" because that appears nowhere (or almost nowhere) in its training data.
But it will be confidently wrong, because it doesn't actually know anything aside from statistical weights on character strings, not even about the internal structure of those very same character strings!
Yet we call it "artificial intelligence". Maybe we should call it "word strings" or "word streams".
It's really frustrating that there aren't better visual metaphorical representation of what is actually happening when GPT "talks". If people could see random words within a set being scooped up and put into logical places by a confused robotic arm they wouldn't be as impressed
Yes. A robot arm filling in madlibs in a dark room
By grabbing words out of buckets labeled "nouns," "verbs," "adjectives," etc.
Yes. You could even show it grabbing from specific subject buckets for types like "fish nouns" "bird adjectives", but they're still just things to the Thing
