By far the dominant use of computing resources today, and the main driver for their expansion, is “AI”. I quote the term to mean the products of companies such as OpenAI, Anthropic and DeepSeek. The reason why I object to the current push towards wholesale adoption of this “AI” is because it is profoundly and pervasively harmful: it exacerbates climate change, destroys education, poisons information, erodes income from work, erodes trust, and in doing so hurts our society and economy as well as the environment we need for our existence.
On the frugality of “AI”
Dan McQuillan pointed out to me that recently, the term “frugal AI” is being mentioned increasingly, and usually the meaning is very different from what I mean by “frugal computing”.
Let us therefore consider the frugality of this “AI”. Most tasks done by the current generative and agentic AI tools can be done without them. I say most and I mean the vast majority. That “AI” can prove maths or physics problems is irrelevant from that perspective. Doing the task without “AI” might take longer and therefore entail an opportunity cost (then again, it might not), but not using “AI” is the frugal way: it achieves the same result with dramatically lower use of resources and without any of the harmful consequences.
From that perspective, there is no such thing as “frugal AI”, regardless of its resource consumption.
Of course the users of “AI” will argue that the frugal way is not possible for them (because they are not prepared to spend the time to do the task without “AI”, including learning the skills required). But that does not make this use of “AI” frugal.
Next, lets consider those — as far as I can tell mostly hypothetical — tasks that can only be done by “AI” and will generate profit or save cost. Usually, what is meant by this claim is that it can only be done cost-effectively by “AI”, from the perspective of the person or entity making the profit. Which takes us back to profit and harm: regardless of the actual nature of the service, it serves to further the growth of this intrinsically harmful technology. So it is certainly not frugal.
Finally, lets consider those even rarer tasks that can only be done by “AI”, regardless of cost, and that would help make the world better. In that case it is a matter of weighing the good against the harm. There are many examples of technologies that had great societal benefits but were banned because of their harm. Most of these are chemicals and minerals (DDT, asbestos, E605, freon, …), because until now there has not been much awareness amongst lawmakers about the harm caused by computing technologies.
“AI” is already doing damage purely by delaying the phasing out of fossil-fuel energy generation. The resulting increase in emissions will result in a disproportionate increase in the onset and severity of climate-change induced disasters that will affect millions of people. This ignores all other harms, including the actual emissions from AI data centres. It is therefore very difficult to conceive of use case that would compensate for this to such an extent that the good outweighs this harm.
Small and local does not imply frugal
The term “frugal AI” is increasingly used for smaller, locally-run models. The argument is that such models need fewer compute resources and are not under control of the large “AI” corporations. However, this is a narrow view.
Quite apart from the question of need (do we really need to use this “AI”?), a key question is, “What would happen if tomorrow the AI providers would disappear?” Most of the “small” models are distilled versions of the larger ones. For most of those “local” models, it would mean they would not longer be updated and therefore become less and less useful. “AI” companies are not charities. The reason for giving free access to these small, locally deployable models is to garner goodwill and further adoption and thus perpetuate their destructive business models.
I grant that it may be possible to locally train a small, fully open-source model on ethically sourced data and deploy it on a sufficiency basis on existing hardware. If this use of “AI” is necessary (i.e. there is no other way to do it) then that specific scenario could be considered “frugal AI”. But under the current circumstances, that is a vanishingly small proportion. In other words, the term “frugal AI” is ethics-washing.