2 ms·

I always appreciate reading a take from someone who's well versed in the domains they have opinions about. I think longer-term we'll eat up any slack in effici

I always appreciate reading a take from someone who's well versed in the domains they have opinions about. I think longer-term we'll eat up any slack in efficiency by throwing more inference demands at it -- but the shift is tectonic. It's a cultural thing. People got acclimated to shlepping around morbidly obese node packages and stringing together enormous python libraries - meanwhile the deepseek guys out here carving bits and bytes into bare metal. Back to FP!

vonneumannstan 2y ago
This is a bizarre take. First Deepseek no doubt is still using the same bloated Python ML packages as everyone else. Second since this is "open source" it's pretty clear that the big labs are just going to replicate this basically immediately and with their already massive compute advantages put models out that are extra OOM larger/better/etc/ than what Deepseek can possibly put out. Theres just no reason to think that e.g. a 10x increase in training efficiency does anything but increase the size of the next model generation by 10x.