so much slower

unless you're working on 10 features in parallel with smaller models in best-of-4 groups