Nice result, but their claims are a little misleading, as always. They have come full circle, having a very powerful language model they can attach it to other

Nice result, but their claims are a little misleading, as always. They have come full circle, having a very powerful language model they can attach it to other inputs(video, audio,...) creating an ensemble system that gives the illusion that it has "knowledge" about domains and can reason in context, as they ofter state in all their works. But we should know better that what they are doing is very different, eg. trowing a truckload of data into billion-parameter neural networks, spending millions, if not billions, of dollars, just to minimize a loss function in an adversarial environment (a GAN), in this case.