[simonh]

I suspect the comment about the thunder dome was a reference to RLHF. On the one hand RLHF seems far superior to the kind of prompt engineering Microsoft seems to have relied on with Sydney. On the other, it's dubious that the manual selection in RLHF is really always selecting for quality, as against at least to some significant extent pandering to whatever biases or preferences the humans in the training loop might have.