GANs are different. They may "learn" something plausibly enough, even if they don't know what they are learning. A GAN model could, in theory, generate "plausible" private keys, by learning what private keys are, without knowing anything about their corresponding public keys. Now, from this space of plausible keys that the model generates after we train it, we would just use the results on a subset of the entire Bitcoin network address space, until we find a fit. That's the "brute force" part of it. GANs are coming somehow before that, and they may, because of how they work, shorten the path to a working private key.

I'm not sure I understand this part: "Isn't the whole point of encryption that the fastest way to generate a private key from a public key is to use brute force."