“Misha Laskin believes the problem of creating a perfect reward model is "ASI complete," meaning a neural network that can accurately verify any outcome is likely a superintelligence itself.”