By 2024 end, a model exhibits action recognition (video) equivalent to human level accuracy on Something Something V2?