Comment by frodo8sam

I would guess that the FFT scales better as you increase the number of tokens in the context window. Interesting Google's models outperform their competitors on context size.