The problem with potemkins in AI models is that they invalidate benchmarks, the researchers argue. The purpose of benchmark ...