Hacker News Reader

Comment by MrThoughtful | Hacker News Reader

Does AI really "adapt" a "work"? Or does it learn from looking at the world and then create its own works, just like a human does?

View on HN

That's the trillion dollar question.

My belief is that generative AI probably breaks both the letter and the spirit of the law, but it also renders the law pointless because its existence, at whatever quality level, means there's no longer a commercial justification for copyright to exist to protect creative output of that level or lower.

But there might be an evolutionary/signalling reason (think peacock tail) that overrides the commercial.

https://kitsunesoftware.wordpress.com/2022/10/09/an-end-to-c...

An AI learns by looking at the world the same way my camcorder does when I take it to a movie theater and point it at the screen.

It's actually the other way round.

The camcorder can reproduce every scene in detail. But it can not tell you the plot.

AI can tell you about the plot, but not reproduce detailed scenes.

So this is a prime example of how AI is humanlike and not copying anything.

The analogy falls apart when you consider that for instance stable diffusion was trained on 2.3 billion images, but the model itself is only around 10GB. That only works out to be 4 bytes per image. Clearly it's not the same as pointing a camcorder at a movie screen.

You could argue that training a model is essentially a lossy compression algorithm resulting a compressed artifact that can be non-deterministically decompressed.

So are humans. If I were to draw a Mickey mouse from memory, would it be amenable to copyright laws?

Well, yes.

That’s 10GB with everything included, including training weights. The models most people use are 2 or 4 GB.

Reply on news.ycombinator.com