I found the conflation between training and running models very confusing or even disingenuous. The whole thing about evaluating a model frame by frame with low latency local on the device, that is a good point, but then having that as a follow-up to why they don’t train the models with lots of data on the server ... well, I don’t really know what to make of that.