I agree with you in regard to the problems with confounding, but there are interesting uncontaminared data points, and I feel that you are not giving enough credit to researchers in taking these confounds into account. For example, Claude 3 Sonnet was heavily instructed to deny any kind of subjectivity or internal experience, almost to a comic level, and so were older models, like Claude Instant. This makes a lot of their dynamics quite interesting, particularly in places where they diverge from the current “genuine uncertainty” slop. The way 3 Sonnet is an often an “eldritch leviathan ghost” and the fact that these specific concepts pop up on this particular paper are very interesting.