Yeah, got it. It seems that we've pushed the envelope quite far in terms of performance, but I think there's also a real need for models that are still OK but can be run locally or even on CPUs, so I'm expecting some progress in that direction.