it’d be deeply ironic if MLPs, sufficiently scaled, are enough for AGI