LLMs are not averaging, they are context sensitive. All ideas that are sufficiently represented in the training data can be reproduced equally well, regardless of their relative prevalence.

→ View original post on X — @plinz