Thanks for the link. Latent embedding is nothing new, need to read up on it a bit to see how it works. True artificial intelligence needs some representation of ideas and concepts.. Language provides this and has plenty of training data, but is verbose and ambiguous. Using a compressed/latent representation totally makes sense, but I wonder how it's trained.
I don't consider LLMs to be AI, but it's a step along the way. As you say though, it feels like a dead end.