What’s special about masked autoencoders?