The Cloud of Poems and AI

A recent discussion reminded me of a short story I read a long time ago.

A technologically-minded alien god wanted to surpass human poets. So he did something very much in keeping with his nature: since a poem is nothing more than a combination of words, all he had to do was store every possible combination of words, and every poem that had ever been written — along with every poem that ever would be written — would, of course, already be in there.

To carry out this ultimate act of writing, the god's storage chips spread densely across an entire star system. At last, every possible poem had been stored.

But then a problem arose. The god had written "all poems," yet he could not build an algorithm capable of retrieving the good ones from the pile. When he needed to produce a good poem, what he had to do turned out to be no different from what an actual human poet does: pick, out of the countless possible combinations of words, the one that would move someone — and this was precisely what he could not, by any means, build.

He possessed all the poems, and still could not write a good one. After all, even without those storage chips, the space of possible word combinations already existed in the world.

I think AI is the same. The fact that a model's latent space "contains every possible image" is no more remarkable than the fact that every possible combination of words already exists in the world. The god and the poet are standing before the same sea of possibilities; the only difference is who does the choosing, and how. To pick, out of countless possibilities, the one that moves someone — this has always been where art happens, whether that sea is written on paper, stored on chips, or compressed into the weights of a model.