Sam Altman told an audience at The Indian Express this week that it’s “unfair” to compare the energy costs of AI to the energy costs of a single human inference. His reasoning: humans take twenty years and all that food before they get smart. A hundred billion people lived and died learning not to get eaten by predators so that you could exist. Measured that way, he said, AI has probably caught up on an energy efficiency basis.
The problem is not that this is wrong, though I feel it is; and is quite silly for that matter. The problem is that it only makes sense if you have already decided that the purpose of a human life is to produce outputs, and that the food and years consumed along the way are overhead against production.
A child eats breakfast because she is a person and people need to eat. That is true whether she grows up to be a software engineer or whether she never enters the labor force at all. The resources spent raising a human to adulthood are the baseline obligation of a society that has decided its members deserve to exist, and once you start treating those resources as a line item to be compared against the electricity bill or water usage for a GPU cluster, you have wandered into a very specific kind of accounting. One where the logical terminus is always the question of whether certain people are worth the expense of keeping around. Altman probably does not think he is making that argument, but the structure of his comparison makes it for him, whether he intends it or not, because the moment you talk about human sustenance as “the cost of training” you have implicitly accepted that humans who do not produce sufficient output are a bad investment.
Let’s be candid, the comparison does not even survive contact with its own premises. A sixteen year old can learn to drive in hours of practice and study and experiential learning. To teach a machine to drive, you must feed it every video of someone driving ever recorded, outfit it with LIDAR and infrared sensors that no human requires, and even then it needs a human operator to bail it out with some regularity. The entire corpus of recorded chess can produce a grandmaster in a human. Feed that same corpus to a large language model and it will lose track of its own pieces in a dozen moves. Human neurons, when arranged in biological neural networks, learn faster than silicon ones. A person generalizes from sparse data in ways that remain, after decades of research, distant to the people building these systems, and they do it on a caloric budget that would not cover the cooling costs of a single rack in a modern data center.
This way of thinking about people has consequences that extend well beyond energy policy. The reason a donor makes a transformational gift to a university or a hospital is not that a predictive model identified them as high propensity. To that I say: post hoc, ergo propter hoc. It is that over months or years, through conversations and meals and campus visits and the slow accumulation of trust, a human relationship developed in which the donor came to see their own values reflected in the institution’s mission. That process is expensive and inefficient by the metrics a technologist would apply to it, and it is also the mechanism by which charitable giving in this country works. There is no version of it that gets better by removing the human from the middle.
This is why i’ve grown somewhat disillusioned with the push for AI engagement in the name of scale. I have no doubt that it produces revenue in the narrow sense, but when you strip the human element from an enterprise whose entire value proposition is human connection, you do not get a more efficient version of the same enterprise. You get a different enterprise, one that is worse at the thing it was supposed to do.
Civilizations are expensive because they are made of people, and people eat food and tell stories and take decades to grow up and spend their working lives in ways that cannot be reduced to just their outputs. Every organization I have worked with/for exists because someone, at some point, decided that bearing those costs for one another was the point of organizing a society. The technology Altman is selling only exists because enough people made that decision for long enough to shape the world in which his company could be built.