Sustainability of generative AI
We can start by agreeing that the energy requirements for developing and running large language models are significant.
However, I believe the problem is more complicated than most people appreciate, for the following reasons:
- Articles and blog posts focusing on the issue tend to have an outcome in mind, and then present information in a way that moves the reader in that direction i.e. they’re not neutral.
- AI is also being used to develop solutions to the climate crisis.
- Most of the energy used is during the training phase i.e. we’ve already ‘spent’ that energy. Limiting the use of generative AI isn’t going to bring it back.
- If generative AI leads to an abundance of human flourishing through improved education, creativity, and commercial opportunities, would it be worth the impact on the climate?
- What if the continued use of generative AI helped reduce the impact of climate change wrought by industrial age technologies?
I’m going to try and explore this topic in some depth in this section.
Energy consumption
There are sustainability challenges posed by the intensive energy and water costs involved in training and operating generative AI systems. The graphics process units (GPU) chips on which these models run require 10-15 times the energy of a traditional CPU with significant water use in cooling the servers. This has been described as the ‘dirty secret’ of the current hype wave with what Timnit Gebru has called the “corporate pissing contest” of ever larger models leading to spiralling ecological impacts. - Mark Carrigan
Stranded energy
I don’t know if this is true for AI but Bitcoin mining can be made to use ‘stranded’ energy i.e. energy that can’t easily put to productive use elsewhere.
Unlike other industries, Bitcoin mining is relatively mobile. In their quest for cheap and abundant energy sources, miners can set up new facilities fairly quickly all over the world, including the most remote areas. As a result, Bitcoin miners can tap into so-called ‘stranded’ energy assets that cannot easily be put to productive use by other industries. In those cases, Bitcoin miners are not competing with other industries or residential users for the same resources, but instead soaking up surplus energy that would otherwise have been lost or wasted. - https://ccaf.io/cbnsi/cbeci/comparisons
Water consumption
I’m going to go into this in more detail, but for now, I’ll just say that we use more water (a lot more water) in the production of almonds, than we do on generative AI. Now, I like almonds but I’d give up almonds any day of the week before I give up AI. If we’re going to complain about how much water we use in the creation of AI, I’d like to see proportionally more space given to complaints about almonds.
Material science
AI is increasingly helpful in the discovery of new materials, which are going to central in the fight against climate change. And Google DeepMind has an AI tool they used to create 700 new materials.
Nuclear energy
I’m not going to make an argument that nuclear energy is the best solution to the climate crisis. However, sustainable energy production through wind and solar isn’t currently close to what we need to keep society running. The only other way we know how to create abundant energy with low emissions is nuclear. And AI has a part to play.
Microsoft has started training AI to generate the paperwork necessary for regulatory approval of next-generation nuclear reactors1 that that will power data centres. The time taken to bring this technology online might be significantly reduced through the use of AI in the process, helping organisations move away from fossil fuel power.
Additional resources
Footnotes
-
Small modular reactors (SMRs) are ↩