The Brainpower Paradox

Thinking about: Biology

Your brain is an energy hog. Although it represents only 2 percent of your body’s mass, it consumes 20-25 percent of your body’s energy, which works out to about 20 watts. In biological terms, that’s crazy—unless having a brain dramatically increased your evolutionary ancestors’ chances of surviving and reproducing, which it did.

Primitive nervous systems, like those found in flatworms, allow an organism to reflexively respond to environmental changes. More complex nervous systems allow animals to observe and associate events and phenomena. They might, for example, come to associate a rustling noise in the grass with the presence of a snake. With even more brainpower, they can predict the future on the basis of what has happened and is happening, and can therefore act in a way that improves their future circumstances. This ability dramatically increases their chances of surviving and reproducing. In evolutionary terms, then, the 20 watts of power consumed by human brains is a real bargain.

This energy consumption, by the way, is fairly constant, regardless of whether you are thinking hard or sleeping. This suggests that your brain is still at work when you are fast asleep—which is only to be expected since, during that time, your brain’s “dream studio” is active. Furthermore, if you have spent a significant portion of your waking hours trying, unsuccessfully, to solve a problem, your muse might also be hard at work.

In the 1990s, computer scientists assumed that human thought processes could be simulated by brute-force computing. Consider, for example, the thought processes of a chess player. When deciding what move to make, they might think of all the moves they could make, how their opponent could respond to those moves, how they could respond to those responses, and so on. Three levels into this process, however, most players’ brains boggle. There are just too many options to keep track of. IBM’s Deep Blue computer was programmed to look six or so moves ahead, which gave it an edge on human players. By 1997, Deep Blue was savvy enough to beat chess champion Garry Kasparov.

Besides deflating the egos of human chess players, Deep Blue exposed the shortcomings of the brute-force approach. With each additional move it considered, the number of possible next moves exploded in an exponential manner. Computer scientists at DeepMind therefore turned to a different approach. Rather than relying on brute force computation, their AlphaZero computer played games against itself, and along the way, gained the ability to spot patterns, on the basis of which it could develop playing strategies. Inspired by the way human brains use neural networks—networks, that is, of neurons—these same scientists decided to rely on artificial neural networks embedded in silicon chips.

We can argue about whether AI systems are capable of genuine thought. One thing that is indisputable, though, is that they are prodigious consumers of power. In my research, I asked Claude to quantify the energy needs of a single large AI training facility. Its response: “Amazon’s Project Rainier—specifically intended for training and running Anthropic’s models, meaning models like me—will use 2.2 gigawatts of electricity, equivalent to the power consumption of 1 million households.”

Meanwhile, human brains are clearly capable of genuine thought—despite, as we have seen, drawing only 20 watts of power. This gives rise to the brainpower paradox: How is it possible for brains to outthink AI systems at a fraction of the energy cost? This in turn gives rise to a new question: Can we have the benefits of AI without having to pay this energy bill? It turns out that these questions are connected.

AI researchers made a major breakthrough when they started imitating, in the architecture of their circuits, the neural networks of human brains. Those engaged in neuromorphic computing are also turning to the human brain for inspiration. Their goal is to continue the development of AI but at a fraction of the energy cost, and they are exploring a number of avenues. Some attribute the brain’s thinking ability not only to its reliance on networks of neurons, but also to its unique way of processing information. As a result, they are experimenting with “spiking neural networks” that communicate the way neurons do—not in continuous mathematical signals but in timed electrical spikes.

Even more radically, researchers are experimenting with biocomputing, in which they grow actual human neurons on silicon chips. And where, you might be wondering, do they get their neurons? It turns out that online, you can buy neurons derived from induced pluripotent stem cells that in turn started out as skin cells.

Your brain is an impressive organ, both for what it can do and for how energy-efficient it is. You therefore have every reason to be proud of it.

Realize that our brains are the result not of intelligent design but of hundreds of millions of years of mindless evolutionary trial and error. Am I alone in thinking that it is ironic that some of the smartest people on the planet, assisted by AI, would turn to a mindless process for key insights?

Need more food for thought? Click here for my past essays, listed by title.