I think a more interesting experiment would be to train an LLM on a dataset that had zero references to "first person POV" sentences. And then see if the final model somehow develops that concept.
Seems like a way to justify stock valuations of OpenAI while picking up some good talent to go after vertical stack as the foundation models become commodified. Smart play.
What is the "crypto crapware" specifically you are referring to in the BlueSky project? There is no blockchain or token associated, though they are using DIDs to sign each post as well as standard asymmetric key encryption with the browsers WebCrypto API. https://github.com/bluesky-social/atproto/blob/main/packages...