The issue with efficient summarization is the necessity of context. For a message to be effectively summarized you require a bit of "theory of mind"[0], you need to have at least a decent idea of what the receiver already knows. This is something that, especially when done for a single message at a time (with no other info), likely doesn't have a global solution.

Edit: I should be clearer, I believe that the solution shown here does not have this capability, and the effort of using ML to do message summarization is flawed in the general case.

[0] https://en.wikipedia.org/wiki/Theory_of_mind