I’m curious how they’re doing it. Obviously the standard bag of tricks (eg, speculative decoding, flash attention) won’t get you close. It seems like at a minimum you’d have to do multi-node inference and maybe some kind of sparse attention mechanism?