I have books on AI that are thirty years old. I think I can say they cover somewhere between 80% and 90% (if not more) of what AI is today. The difference is computing that is thousands, millions, of times faster, massive amounts of storage, etc. In other words, one could very well argue we haven't done much in 30 years other than build faster computers.
If we take the example of a puppy, it seems to generalize pretty well using something like one-shot learning, but is it? I cannot confirm for sure how much data a puppy has already digested before being able to do what we could call "one shot learning". So maybe, the exposure to data is already there, waiting for a specialization toward a particular task.
Giving the ability to a network to be probabilistic enable it to do inference using uncertainty, which is clearly a neat feature when you are gravitating toward AGI for scene understanding.
In the case of video compression, scene understanding may introduce more artifacts IMO: Even if the scene is captured with high end cameras, on a pixel level basis, the edges will never be perfectly neat. I think this will decrease the ability of any network to "understand" which object is at the edges, this results in low classification rates on them, resulting in bad compression/decompression quality (?) for features that are important to the human eye.
All in all, I'm not sure that NN are the right tool for this kind of problems. But we are diverging from the main subject VVC, Thanks for the very interesting comments :)