@matty @Jonny @mjdigspigs because that's the trick - they don't really *know* anything, it's just the ( as far as I can seen ) the probability of that being the correct word to add on the end.

It's like a super elaborate mad-libs trick. I don't think you can even glean things like "what are the words most associate with X" from them.

That's not to say you could *make* a model that did this, but it would involve a much different way of storing information.