Testing the cognitive limits of large language models Hoover Institution Testing