> The problem is your model needs to include all the computers playing the market, and it also needs to include the other hedge fund bros themselves. The error

> The problem is your model needs to include all the computers playing the market, and it also needs to include the other hedge fund bros themselves. The error here is equating hard takeoff with such granular and expensive world prediction. Human intelligence finds more efficient compression than this. When Donald Trump decides what provocative thing to say in a speech, he isn't granularly modeling the human intelligence of his supporters, opponents, journalists, and media organizations in a particularly CPU-costly way. He's just discovered some narrower domain where there's a vein of predictability that simplifies the whole system enough that he can guess a tweet or speech will be provocative in the desired way. The same thing happens when a scientist is studying cold fusion or dangerous viruses. You don't have to predict what will work, you just have to predict what might work, and try and try again. AI doesn't have to be good at predicting the future in any absolute terms. It just has to be better than us at predicting what is worth trying. Once it is better than us at deciding what's worth trying, why not FOOM? FOOM isn't guaranteed perhaps, but why is it not at least one of the likely outcomes?