Commentary on X
@nabla_theta @diatkinson @So8res @AIImpacts Yes, agree that there are lots of ML people who can't even understand questions talk about GPU versus TPU topology and why model training batch sizes and update regimes should depend on speed of interconnects - *and they were the ones are being asked if LLMs can scale further*!
@davidmanheim @diatkinson @So8res @AIImpacts the ML-systems gap is probably bigger (ML here is labspeak for algorithms); lots of ML people often dream up stuff that will never be systems-efficient because they don't understand or appreciate the importance of systems friendliness.
Posts from X
2 posts@nabla_theta @diatkinson @So8res @AIImpacts Yes, agree that there are lots of ML people who can't even understand questions talk about GPU versus TPU topology and why model training batch sizes and update regimes should depend on speed of interconnects - *and they were the ones are being asked if LLMs can scale further*!
@davidmanheim @diatkinson @So8res @AIImpacts the ML-systems gap is probably bigger (ML here is labspeak for algorithms); lots of ML people often dream up stuff that will never be systems-efficient because they don't understand or appreciate the importance of systems friendliness.