The obvious way to try to get superhuman performance would be to optimize against a face discriminator, but current AI systems are not robust under such optimization (the “faciest image” according to the discriminator is not a face)
This argument is a form of strawman rather than steelman. The more obvious way to get robust superhuman performance today is to use a diffusion model, which optimizes a joint cooperative objective and doesn’t have such issues. These robustness advantages carry over to diffusion based planning. See also this comment for the extended version.
This argument is a form of strawman rather than steelman. The more obvious way to get robust superhuman performance today is to use a diffusion model, which optimizes a joint cooperative objective and doesn’t have such issues. These robustness advantages carry over to diffusion based planning. See also this comment for the extended version.
I agree with your point but you might find this funny: https://twitter.com/waxpancake/status/1583925788952657920