You need to expend some energy (1000/2000/5000/6 billion years of energy!) to organize your knowledge well enough to get that bit of entropy that tells you that you can shape the leaves this way to make a fan, or you can put copper and iron together to make a battery.
In a similar vein: we could have come up with StableDiffusion with just parchment, if ten of us had just written down the weights one at a time, 5 kilobytes per day, for several decades. Of course, what do we write down? That takes energy.
You'd still need calculators to run it on with sub-year response times. 100 teraflops times three minutes would 18 quadrillion multiplications of 8-digit numbers. At one minute per multiply you'd need 30 billion person-years of calculation to generate one image.
The training process was indeed slower still, but maybe that is because we are doing it inefficiently, as we did our biological evolution.
At a fundamental level, it's entropy and energy.
You need to expend some energy (1000/2000/5000/6 billion years of energy!) to organize your knowledge well enough to get that bit of entropy that tells you that you can shape the leaves this way to make a fan, or you can put copper and iron together to make a battery.
In a similar vein: we could have come up with StableDiffusion with just parchment, if ten of us had just written down the weights one at a time, 5 kilobytes per day, for several decades. Of course, what do we write down? That takes energy.