“Misha Laskin cites the Grok 4 release as an example where a model specifically trained with a tool for the "Humanities Last Exam" benchmark showed a significant performance improvement over the general model.”