Black Boxes

This is not a discussion of the ethics of AI. 

I am not fit to address the theft of others’ work for the training, the sublime laziness AI encourages, or its systematic devaluation of the creative arts. Nor is this a discussion on what a degree can be worth when students aren’t writing their own essays. Many others are better suited to discuss AI's impact on art, livelihoods, and education, and who have already done so.

But another question arises with this new technology: is it impacting more than just the creative world - the world of writer’s strikes and art theft? Is there a tangible effect beyond the replacement of in-house artists? And while we might just be seeing AI incorporated in software products like Instagram’s search or in ChatGPT, very much detached from our physical experiences, is it already affecting our lives right now?

This is a discussion about the power of AI. Not the power to swamp platforms with uncreative art - the kind that keeps the lights on. The kind that keeps homes at room temperature; the power that people rely on for things more essential than stealing actor’s voices or completing assignments.

How much? Too much?

Is AI just the latest in a line of trends? Like cryptocurrency, blockchains and NFTs, is it guzzling power on a scale comparable to small industrialised nations (Digiconomist, 2024)?

Think about it like this - we rarely consider the electricity we use for a simple Google search.

Probably because it’s difficult to wrap your head around electronics consuming power at the best of times. We see plugged-in computers as things that just work. Larger, more complex programs on a laptop naturally drain the battery faster, but it's easy to overlook the fact that, beyond keeping the laptop running, electricity is also being used for every letter I type.

People everywhere leave their computers running overnight because they don’t process that an electronic device draws electricity the way lights do. Asking someone to consider not just the power consumption on their own device, but also what's happening at the Google server they're accessing—something they can't even see—can seem excessive. After all, it's not their computer. That small disconnect is enough for most people to never give it a second thought

Because opening an encyclopaedia and finding an entry doesn’t register as an activity that requires energy when it does. In the same way, neither does Googling something.

Now consider the first wave of AI you encountered, back in 2022.

Chatgpt spits out answers to all your questions as well, but it takes longer than a Google search. Have you considered how much more power it takes to combine data, to create the response, as opposed to simply retrieving what it thinks you want and presenting it for you to read?

Have you considered the actual process Dall.E and other image generation AIs use? It’s not just filling in pixels; creating an image involves complex processes like cycles of generational genetic algorithms and "thinking" that require significant computational effort.

The evidence is quite straightforward, supporting the logical conclusion based on observable data.  While the average Google search uses 0.3 Wh to return results, the average ChatGPT query is an order of magnitude more intensive, consuming 2.9 Wh per request (IEA, 2024). That means that if Google were to fully implement current large language models tomorrow into the 8.5 billion requests it receives each day, the company’s energy consumption would exceed that of the country of Ireland on a yearly basis.

But on a case-by-case basis, this information isn’t game-changing. Google isn’t going to radically change to just providing plain-text responses overnight, and just comparing the two systems overlooks the multitude of differences between them. So, how much energy does it take to operate AI? It’s a new industry - so what’s the setup cost? Even the hardware is something that must be considered - the physical construct that is the brain of an AI.

This information is surprisingly vague. There are studies from 2021 on the energy requirements and carbon emissions produced for training AI - using ChatGPT’s predecessor, GPT-3, and others - and but in an industry that is expanding and growing like AI, 3 years is an aeon. There have been methods developed to make them far more efficient since then, but in The Verge article, ‘How much electricity does AI consume?’  the excellent point is made that AI has become more private as it has become more privatised. There is research underway, and it is a subject gaining more and more focus, but gone are the days of the companies behind the AI publishing the data themselves.

An approach taken by Alexander de Vries, a PhD student from the Netherlands, was to examine the hardware underpinning these AI models. NVIDIA hardware is the go-to, and they release both technical specifications and sales projections. de Vries used these to estimate that in general, due to the hardware and infrastructure required in the data centres that host AI, it is estimated that ChatGPT uses 564 MWh per day (de Vries, 2023).

AI is a new industry - whether it is an unnecessary or harmful one is not my position to argue - but any new ‘thing’ should be taken into consideration with due respect for what it is and does, and the impact it would have as an addition to our lives. Even if AI used only as much power per query as Google, it’s still another Google, kept up and running, kept cool with water, in massive massive data centres around the globe.

And as an industry that is expanding the way it is, it is absolutely mission-critical that people understand and are fully aware of what they are dealing with.

References

Vincent, J. (2024, February 16). How much electricity does AI consume? The Verge. https://www.theverge.com/24066646/ai-electricity-energy-watts-generative-consumption

Luccioni, A. S., Viguier, S., & Ligozat, A.-L. (2023). Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. Journal of Machine Learning Research, 24, 1–15.

Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D., Texier, M., & Dean, J. (2021). Carbon emissions and large neural network training (arXiv:2104.10350). arXiv. https://doi.org/10.48550/arXiv.2104.10350

Electricity 2024 – Analysis and forecast to 2026. (2024, January 24). IEA. https://www.iea.org/reports/electricity-2024

de Vries, A. (2023). The growing energy footprint of artificial intelligence. Joule, 7(10), 2191–2194. https://doi.org/10.1016/j.joule.2023.09.004 

Bitcoin energy consumption index. (n.d.). Digiconomist. Retrieved August 23, 2024, from https://digiconomist.net/bitcoin-energy-consumption/

Previous
Previous

Cryptocurrency – Innovation?

Next
Next

The age of social media: a new world?