And yet it’s still a good indicator of general performance. Any model that scores under GPT-4 on that benchmark, but above it in other, tends to be worse overall.