I think the argument is that we can already simulate a single neuron fairly accurately today, and so some decades from now we'll likely be able to simulate them all, plus their interconnects, and thus by pure brute force we'll get a human-level AI.