Nvidia already has a product that does this. It’s no more a violation of Shannon theory than neural audio codecs. All of these optimal compression models produc

Nvidia already has a product that does this. It’s no more a violation of Shannon theory than neural audio codecs. All of these optimal compression models produce a string of length -sum(p log p). If you can come up with a better representation of likely events, that implies stronger assumptions on what the state space can actually be occupied by, and so you reduce the surprisal of individual observations, thereby reducing code length.