If the dreams of companies making large sequence models do come to pass, and they really do scale up to where training and running large sequence models is a significant portion of world electricity usage, then of course it becomes a real problem, but the same would be true for any large electricity-intensive industry.
With AI there are strong incentives to make everything more efficient and use less energy. This is already happening: people are finding ways to run AI models on less powerful hardware (see LLaMA etc), and AI providers like OpenAI have managed to run things. Lee efficiently - that's presumably how they could drop the price of an API call by 1/10th for GPT3.5 compared to GPT3.
With Bitcoin efficiencies don't actually help, because the entire system is designed as a competition. The person who burns more energy wins more value. If the tech becomes more efficient the miners just increase their activity, so energy usage doesn't meaningfully decrease.
Then there's Ethereum, which just completed a major upgrade to convert its Proof of Work method to Proof of Stake and now uses a negligible amount of energy compared to other things, as shown in this article on their site[1].
And there are newer cryptocurrencies like Solana, that started as proof of stake to begin with and can handle at least 8500[2] transactions per second (compared to only 4 transactions per second with the much older bitcoin, and 30 transactions per second with Ethereum).
That's a hell of a lot more efficient as well, and also seems to be more efficient than ChatGPT is per query, as Solana's energy use per transaction is 3,290 Joules, or 0.00091 kWh[3], versus the conservative estimate of ChatGPT (3.5)'s estimated energy usage on this article [4] of 0.00198 kWh, or 7,128 Joules.
So I think that overall, crypto is still incentivized to become more efficient. Bitcoin is not, as more people care about the security provided by proof of work there than they do about energy efficiency, which is unfortunate.
[1]: https://ethereum.org/en/energy-consumption/
[2]: https://crypto.news/solana-transactions-per-second-hit-an-al...
[3]: https://solana.com/news/solanas-energy-use-report-september-...
[4]: https://towardsdatascience.com/chatgpts-electricity-consumpt...
It really hit home to me when I saw something (that I can't find right now for some reason) about ChatGPT-4 being much more complex and would need X times more resources than ChatGPT-3.5 for the same queries (the number I don't actually remember, but I want to say 2 or 3x). Gave me a bit of a pause from me asking it to generate a bunch of sitcom scripts on random silly concepts for my private, personal amusement (Seinfeld in particular seems to work pretty well).
I was able to find an article on Bloomberg[1] talking about how the increase in AI is also increasing its carbon footprint, so yeah, it's coming. Also I saw this article[2] that estimates ChatGPT's usage in January alone to be equivalent to he power used by 175,000 people. And that's before ChatGPT-4.
People who are fans of AI right now reading this, eventually you're going to have to come to terms with it being responsible for a ton of carbon emissions as well, as it takes up more and more resources and scales to billions of people as they use it more and more in their everyday lives and the models and computation get more complex.
[1]: https://www.bloomberg.com/news/articles/2023-03-09/how-much-...
[2]: https://towardsdatascience.com/chatgpts-electricity-consumpt...