It is probably a double whammy kind of thing. Just the cost of running these things is high and if cutting a few functions or accuracy can shave some dollars of

It is probably a double whammy kind of thing. Just the cost of running these things is high and if cutting a few functions or accuracy can shave some dollars off, they might take it. The other side is that with things like language and voice models is they sort of have a Red Queen problem. They need a surprising large amount of consistent new data that is vetted just to stay in place. If there are cuts made and the quality diminishes even slightly, it become a feed back loop of lower quality. I remember Jaron Lanier saying that about language translation. They need a consistent flow of new data otherwise the models would drift surprisingly quick as our language changed without us really noticing.