Training is a continuous expenditure. We’re nearly ten years into this craze and we’re still continuously pumping out new models. Whether they’re trained from scratch or not is immaterial. Both processes still consume energy. If you want to justify the claim that training cost is negligible, you would have to show that this cost is actually going down over time and that it’s going down sufficiently quickly.
Whether they’re trained from scratch or not is very much material because it takes far more energy to do that. Meanwhile, we consume energy as a civilization in general. And frankly, a lot of energy is consumed on far dumber things like advertisements. If you count all the energy that goes into producing and displaying ads, that dwarfs AI energy use. So, it’s kind of weird t0 single AI energy use out here as some form of exceptional evil.
You know what else takes far less energy than training a single model? One query. Yet, you argue that it’s the main contributor to the energy consumption. Why is that? It’s because there’s a very high volume of them, thus bringing up the total energy consumption. At the end of the day, it’s this total energy consumption that matters, not the cost of doing it once. Look at the total energy expenditure of training, not just the cost of doing it once.
So, it’s kind of weird t0 single AI energy use out here as some form of exceptional evil.
We’re talking about AI here because that’s the topic of this thread. I’ve never seen anyone say that it’s the only problem worth addressing. Plus, if you want to compare energy usage of ads (or anything else) compared to AI, you would first need to know how much energy AI is actually using.
Yes, and my point is that operational cycle of the model dominates total energy consumption. And turns out that it’s not actually that high in the grand scheme of things, and continues to improve all the time.
Meanwhile, it’s absolutely necessary to contextualize AI energy use in relation to the other ways we use energy to understand whether there’s something exceptional happening here or not. All the information for figuring out how much energy AI is using is available. We know how much energy models use, and rough numbers of people using them. So, that’s not a big mystery.
Training is a continuous expenditure. We’re nearly ten years into this craze and we’re still continuously pumping out new models. Whether they’re trained from scratch or not is immaterial. Both processes still consume energy. If you want to justify the claim that training cost is negligible, you would have to show that this cost is actually going down over time and that it’s going down sufficiently quickly.
Whether they’re trained from scratch or not is very much material because it takes far more energy to do that. Meanwhile, we consume energy as a civilization in general. And frankly, a lot of energy is consumed on far dumber things like advertisements. If you count all the energy that goes into producing and displaying ads, that dwarfs AI energy use. So, it’s kind of weird t0 single AI energy use out here as some form of exceptional evil.
You know what else takes far less energy than training a single model? One query. Yet, you argue that it’s the main contributor to the energy consumption. Why is that? It’s because there’s a very high volume of them, thus bringing up the total energy consumption. At the end of the day, it’s this total energy consumption that matters, not the cost of doing it once. Look at the total energy expenditure of training, not just the cost of doing it once.
We’re talking about AI here because that’s the topic of this thread. I’ve never seen anyone say that it’s the only problem worth addressing. Plus, if you want to compare energy usage of ads (or anything else) compared to AI, you would first need to know how much energy AI is actually using.
Yes, and my point is that operational cycle of the model dominates total energy consumption. And turns out that it’s not actually that high in the grand scheme of things, and continues to improve all the time.
Meanwhile, it’s absolutely necessary to contextualize AI energy use in relation to the other ways we use energy to understand whether there’s something exceptional happening here or not. All the information for figuring out how much energy AI is using is available. We know how much energy models use, and rough numbers of people using them. So, that’s not a big mystery.