A lot of people are still referencing the environmental impacts of generative AI, so I feel like it’s worth reiterating some points on that which have already been made. Of course this is just one factor among others people have mentioned which I understand better, but I don’t get statements like “as we all know” about how severe the environmental impacts are. It seems like nobody really knows, and they’re probably less than what most people think.
From the other thread on this subject:
From earlier in this thread:
Back-and-forth from the comments under the blog post:
Obviously LLMs in general are resulting in new buildings being built and large amounts of energy being used etc. that weren’t formerly being used, but it seems like individual use of them is kind of negligible compared to everyday things like driving, watching TV, running other electronic devices in your house.
My impression is that probably the most intensive energy use period comes from the initial training of the model, when it has to learn from everything in the internat and scanned books and whatnot that can be found for it. If iNat uses an already trained LLM and just updates it with information from iNat, like how Discourse’s summary AI here does, then that skips a lot of the environmental impacts. If a brand new LLM is trained (I’m not sure exactly how they work but that seems unlikely to me?) then maybe more but that’s still far less than training on the corpus used by mainstream LLMs.