I don't know if this is right because the models rank all of each other's work fully pairwise (and anonymized) and seem to agree on the ranking.

→ View original post on X — @karpathy