General-purpose large language models outperform specialized clinical AI