4 ms·

I'd assume its a totally different architecture that isn't based on storing a compressed dataset of all digital human text.