Massive data centres that consume huge amounts of energy are springing up everywhere. Can this really be the right approach to AI, when the human brain consumes a mere 20 watts?
To be honest, I've been surprised that the more-or-less brute-force approach has got as far as it has. I can't help but feel that it's a little inefficient, and that emulating the way nature does things would pay dividends. But let's pursue this a little more.
I asked my "go to" AI (currently Claude Sonnet 4.5) to compare the energy cost of educating a human for 25 years (PhD-level education) with the energy cost of training current AI models. "This is a fascinating comparison!" exclaims Claude enthusiastically, always keen to flatter me, making me feel that I've had a brilliant idea that no-one else has ever thought of. Perhaps Claude in reality finds it incredibly tedious as he (it?) has already had the same or a similar discussion with hundreds or thousands of humans. But if he does find it tedious, he's good at hiding it.
Yes, I'm anthropomorphizing Claude to a ridiculous degree, but then I do the same with Google, and it's "Let's ask Mr. Google" from me if there's ever any discussion the requires factual information in order to resolve it.
Claude informs me that GPT-3 was estimated to use around 1.287 GWh of energy during training, and that current frontier models such as GPT-4 probably use 10 to 100 times more. The human brain consumes 20 watts, and over 25 years that's 4.38 MWh, which is less than the energy used to train GPT-3 by a factor of circa 300.
Claude claimed that the brain does other things than just learning (processing sensory input, controlling the body's muscles and movement, etc.). On the other hand, everything else the brain does is merely operating the body that the brain inhabits. If we take the possibly-controversial view that the body is merely a support system for the brain - that the sole function of the human is to achieve human-level intelligence and learning - then in fact we need to consider the energy consumption of the entire human, not just the brain. This is around 5 times more, and over 25 years is about 21 MWh.
And then, as Claude pointed out, we also need to consider "the massive energy infrastructure supporting human learning - schools, universities, transportation, food production for the learner, teachers' energy consumption, etc. The total societal energy cost of educating a human is orders of magnitude higher."
I asked Claude how to go about estimating the energy cost per human of all the educational infrastructure - including not just schools and universities, teachers and professors, but also libraries and educational book publishing. It came up with a ball-park figure of 800 to 900 MWh. Almost egging me on in my exploration, it helpfully offered to refine this estimate by looking for more specific data for a particular country. Sure, I said. Let's look at just the US. "Great! Let me get more specific data for the US ..." it said, eagerly going off to do my bidding. If my humaniform robot is as eager and compliant when I get one, I'll be in heaven.
So it found data for the energy consumption of schools and universities, and the number of enrolled students, and on this basis came up with a much lower figure of 44 MWh, but admittedly not including the energy consumption of teachers and staff, or transport of students to and from school, or libraries and publishing, as well as the amortized cost of constructing educational buildings in the first place. I asked it to do its best to estimate and add in all of these items, and it finally produced a neat, itemized table (below) with a grand total of 139 MWh, which consists of 21 MWh for "human metabolism", and 118 MWh for the per-human cost of the whole educational infrastructure. So the initial 800-900 MWh estimate may have been too high. On the other hand there are no doubt things I've missed with the itemization approach.
At each stage of the process, Claude added, unbidden, "key insights". Notably, the energy cost of "staff" is surprisingly small, because each teacher is shared between many students. Transportation of the human to and from school during the K-12 years is a significant fraction of the total educational energy cost, and "private vehicle drop-offs" are particularly energy-intensive compared to taking a school bus.
So now in conclusion it looks like training GPT-3 is equivalent to the complete 25-year education of 9.3 humans, so not so unreasonable after all. GPT-4 training cost is likely more by a factor of 10 to 50.
Finally we need to consider that the human with a PhD-level education will generally be an expert in only one area, while the latest AI models appear to be rapidly approaching PhD-level expertise in basically all areas of human knowledge, so that you would need a small army of PhDs to compete with AI models on something like the so-called Humanity's Last Exam. The latter (which is an appalling name) is a can of worms that is probably worthy of an entire separate blog entry. Watch this space.
The upshot is that we do appear to be getting "value for money" (or at least, "value for megawatt-hours") from the huge training costs of the latest AI models.
