OK, what is it that I am missing when people go on about environmental impact of ML? Strubell et al estimate of 284 tons CO2equiv for big transformer training is about what you get for 300 tons of steel: short railway bridge. Industrial countries produce megatons steel annually.
My guess: original point "your computation actually has environmental impact and cost" got transmuted - by not comparing to other *industrial* things - into "your computation is a serious environmental issue".
Energy cost of computation/comms *as a whole* does matter (Koomey's law needs to be speeded up), but the ML focus seems more be to knock something with currently high prestige compared to corporate database management or webservers.
Ah, think I found one decent argument: ML models are growing fast (esp. since there are scaling benefits), hence they will in a while become environmental issue. But main conclusion is to urge fix the later clause.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
