Comment by pixl97

I think a lot of it is the massive amount of compute we've got in the last decade. While inference may have been possible on the hardware the training would have taken lifetimes.