I think we will learn more about the biased sample that transcribed speech has vs private speech. I have just started thinking about this but there are huge area of speech that I use with my family and wife, and coworkers around the coffee room, that I would never consider putting into writing or a YouTube or whatever. And more types of things that I might say when with some younger folks while, trying to debug s production emergency but would not make it into the post mortem. The training I guess is not going to be able to benefit from these more private speeches, or maybe we will have to have people become convinced they are sentient and fall in love with them and share their whole humanity with the LLM.