Current “thinking” AI models still can’t reason to a level that would be expected from humanlike artificial general intelligence, the researchers found.