Yeah as models get better, valid benchmarks become more "trust me bro".