It can’t, it needs hundreds of thousands (if not multiple millions) of human hours spent making content for it.

Training data? Luckily we have all of human history and culture to digitize to solve that problem.

Which is fine if you like human culture to never change for the next ten thousand years. LLMs can only ever interpolate or extrapolate things they’ve already been trained on.

I don’t accept the premise, it’s hard to take anyone seriously who is parroting the “stochastic parrot” argument in 02026.

Your assertion that we’ll somehow get brand new human culture from something that isn’t human might be a little shaky too.