That seems to have nice properties, but 2019 was a while ago. Is the problem of top-k sampling still relevant with much better frontier models?