I think it's the different feeling you get from using an end-to-end streaming service (compute, not videos) versus the one that does a lot of intermittent buffering. It's quite subtle actually. Using a vanilla language model can feel like that if it's also sufficiently small but they are going towards the opposite direction very rapidly now because cloud.