Embeddings aren't objective, their point of view is some kind of complicated aggregate of the training data, which is itself (as of now) mostly human written te

Embeddings aren't objective, their point of view is some kind of complicated aggregate of the training data, which is itself (as of now) mostly human written text. I'd honestly love to see AI fending off the eternal winter on message boards. Is question redundant and basic? Direct user to a specialized AI that can explain the topic well and save the regulars from having to have the same discussion as two days ago. But it is certainly the case that stronger machine text understanding can be used for censorship and oppression. As pretty much all powerful, general tools can be used for nefarious purposes. But it can also be used for a wide range of great purposes as well.