“Tom McGrath reports that after applying their hallucination reduction techniques, the model showed essentially no degradation on benchmark capabilities.”