> environmental impact of a single average prompt (generating 400 tokens' worth of text, or about a page's worth) was relatively minimal: just 1.14 grams of CO2 emitted and 45 milliliters of water consumed
While it’s non-zero, it doesn’t strike me to be “hurting the planet” as some people would want me to believe I’m doing when I decide to use LLMs.
Yes, the training has a much bigger impact but the benefits of training are shared will all users and it’s a one-time cost per model.
I did the math and if I’m right the environmental footprint of a single LLM training, emitting 13,600 metric tons of CO2 and consuming 187,333 cubic meters of water annually, represents 0.000026% of global greenhouse gas emissions and 0.0000047% of freshwater use.
How many trainings and retraining are happening as we speak? How many more when the transition from millions of jobs replaced by AI do you expect? What about inference across all of that?
> How many trainings and retraining are happening as we speak?
Quite a lot. Let's assume a hundred different LLMs of this scale are being trained at the same time. If you multiply the global use percentages by a hundred you'll get: 0.0026% of global greenhouse gas emissions and 0.00047% of freshwater use. Still a literal drop in the bucket.
> How many more when the transition from millions of jobs replaced by AI do you expect?
Dunno, but the argument is that I should feel bad about my current impact on the environment as I use my LLM to autocomplete my code or answer my questions. We have no idea what the future will hold. We can and of course should do everything to minimize the environmental impact of everything we do, but that's a different discussion. For example switching to clean energy sources will make a big positive impact on these numbers.
> What about inference across all of that?
The report speaks about that, the inference cost in marginal when compared to the training cost (~15% for CO2 and ~9% for water consumption).
> Quite a lot. Let's assume a hundred different LLMs of this scale are being trained at the same time.
It won’t be 100, you are underestimating it to make the number be small, ignoring the fact that people are talking about GW worth of continuous power, not counting the refresh rate of GPUs (every 3-5 years the whole infrastructure is renewed).
> Dunno, but the argument is that I should feel bad about my current impact on the environment as I use my LLM to autocomplete my code or answer my questions.
That’s not the argument. The argument is that you should be aware of your consumption and therefore the impact it has. Right now people use everything as a ‘’dumb’’ magical API that just spits things out from nowhere with no impacts.
> The report speaks about that, the inference cost in marginal when compared…
Don’t ignore how many of these are happening as we speak. ChatGPT went from 0 to 100mi users within months, all submitting hundreds of queries.
> It won’t be 100, you are underestimating it to make the number be small
Make it a 1000 (I seriously doubt there are one thousand simultaneous training runs of Mistral Large 2 scale models going on every second) and it's still a drop in the bucket.
> not counting the refresh rate of GPUs (every 3-5 years the whole infrastructure is renewed
I am accounting for this by citing annual usage instead of one-time cost.
Not sure what you think demand is, but operators are building 10 GW AI datacenters. Assuming a GPU consumes ~1 kW, the number is potentially (upper bound) 10 GW / 1 kW, way larger than ‘1000’. For one company.
It is not a drop in the bucket when we are talking about a factor of 1000000 (and not 100 as your initially calculated), on par with buildings and transportation, only behind agriculture.
I’ll stop using ChatGPT when private jets are banned, families drop down to one car and stop having more than two kids. Seriously there are probably more emissions from a few supertankers/cargo carriers than all LLMs combined.
No one is asking people to stop using the Internet. However, becoming aware of one’s consumption is important, and as of now, people are generally oblivious to their digital footprint beyond power consumption, which is only one of the aspects.
Of course they are. We continuously hear that we have to stop having cars, kids, and consuming in a way that the elites have described as irresponsible, all whilst they travel the world in private jets where they definitely do not eat tofu.
it made me look. a private jet emits 4,900 g of c02 per mile (how the author's mind didn't explode mixing the measurement systems is beyond me) vs 1 prompt emitting 1.14 g.
I fail to see a point in this comment. We are comparing 100 million users to a single jet flight. If you believe that a flight is more important that a few million conversations for anyone, sure I can see how you can be concerned.
Most private jet flights are more than one mile in length and there are many thousands of them per day. There’s no where you can go (effectively) that you can’t get to with a major airline.
reply