porfiria posted:

Yeah but it also knows you can't build a real horse out of cabbage (but you can build a statue of one), that horses can't drive because they aren't smart enough and don't have hands, and so on. All this stuff may just be weighted values in a huge matrix or whatever, but it can build associations that are vastly more extensive and subtler than words just tending to appear near other words in the training data.

You edited your response a bit. So just to expand:

I'd say it does "know" what a horse is, but that's for some definition of "know." It doesn't have any kind of audio or visual model for a horse (although it probably will soon, so it's kind of a moot point). And of course it doesn't have any personal, subjective associations with a horse in the way that a human does.

But as a matter of language, I'd say yeah, it can deploy "horse" correctly, and "knows" just about all the facts about horses there are, and how those facts inter-relate to other facts about the world in a comprehensive way that, to my mind, meets a lot of the criteria for "knowing" something.