AI series · Part 1
The 20-Watt Genius
What your brain still does better than a data center
On this page
An old dreamMeet the neuronThe journey of a spikeHow a brain learnsThe 20-watt trickWhat the machines keptClosing the gapWhere that leaves usRight now, a three-pound organ behind your eyes is recognizing these words, following the thread of an argument, and quietly keeping you upright in your chair. It is doing all of that on about 20 watts, roughly what a dim light bulb draws. Meanwhile, the machines that now write essays and code for us live in buildings that draw megawatts, enough for a small town.
How we ended up with the second kind of machine, and what it kept from the first, is a strange and rather lovely story. We will follow it from the first neuron-shaped idea to the chips being built today, with a close look at a real neuron along the way, how a signal moves through it, and how a brain rewires itself when you learn. Then comes the awkward comparison, and an argument researchers are still having: to make AI better, should we copy the brain more closely, or have we been right to ignore it?
An old dream, with a few false starts
The dream begins in 1943, when a neuroscientist, Warren McCulloch, and a young logician, Walter Pitts, noticed that a neuron behaves a little like a switch. It adds up what it hears and either fires or stays quiet. Connect enough switches together, they showed, and you can compute logic. A few years later Donald Hebb supplied the missing ingredient, learning: when one cell keeps helping another to fire, the link between them gets stronger.
In 1958 Frank Rosenblatt built the idea into a real machine, the perceptron, which learned to tell shapes apart from examples. Newspapers promised a thinking machine. Then a pair of critics proved that a single layer of such units could not handle even fairly simple problems, the money dried up, and the field entered the first of several “winters”. The idea survived anyway, kept alive by a few stubborn researchers who, in the 1980s, found a practical way to train networks with many layers. One of them, Yann LeCun, taught a network to read handwritten digits well enough to be used on bank checks.
But the computers and the data were not ready yet. The thaw came in 2012, when a deep network trained on gaming graphics chips won an image-recognition contest by a margin that startled the field. Five years later came the transformer, the design behind today’s chatbots, and a few years after that, millions of people were talking to one.
Now think about what happened at every step. Researchers borrowed one idea from biology (a switch, a strengthening link, a layered visual system), made it work in silicon, and then took it somewhere the brain had never been. Hold on to that thought, we will get back to it at the end.
Meet the neuron
Your brain holds about 86 billion neurons, and each is a small branching cell with a clear job: listen to many neighbors, decide whether enough of them are speaking up, and if so, pass the word along.
Start at the listening end. The cell’s dendrites spread out in a branching crown, and their surfaces are studded with tiny knobs called spines, which is where other neurons connect. A single neuron can take in thousands of connections this way. Everything flows into the cell body, where it is tallied. The nucleus sits there too, holding the DNA and handling upkeep, but it is not where the deciding happens.
That happens at the axon hillock, a small cone where the cell body narrows into one long cable, the axon. If the tally of “go” and “stop” signals crosses a threshold here, the neuron fires, and the axon carries the result away, sometimes more than a meter in a human leg. To make that journey fast, the axon is wrapped in segments of fatty insulation called myelin, made by helper cells. The short gaps between segments, the nodes of Ranvier, matter a great deal: the signal does not crawl along the cable, it leaps from gap to gap. At the far end the axon splits into terminals, little bulbs full of chemical messengers, each resting beside a dendrite of the next neuron. That meeting place is the synapse.
The journey of a spike
A resting neuron holds a stored-up voltage of about minus 70 millivolts across its membrane. Incoming signals nudge it up or down. If the total reaches roughly minus 55 at the hillock, channels in the membrane fling open and the voltage surges to about plus 30 and back, all in around a thousandth of a second. That is a spike, and it is all or nothing. There is no loud spike or quiet spike. What carries information is when spikes happen, how often, and which connections they travel through.
The spike races down the axon at up to about a hundred meters per second, then hands over to chemistry. Small bubbles called vesicles fuse with the membrane and release neurotransmitter into the synaptic cleft, a gap only about twenty nanometers across. The chemical drifts over, docks on receptors on the next neuron, and nudges its voltage up or down. Electricity, then chemistry, then electricity again.
A single neuron is not clever. The cleverness comes from the crowd, each cell wired to thousands of others, with a thought being a shifting pattern of spikes sweeping through the web. Figure 2 shows the smallest crowd we can manage, three neurons in a row. Watch Bo. It does not fire on the first spike. Charge builds up in its cell body, and only when enough has piled up does it pass the message on. And notice what happens each time a signal gets through: the connection thickens. That is the seed of learning.
How a brain learns
A brain does not learn by filing information away. It learns by rewiring itself. Picture a lawn with no paths, and people who need to get from one side to the other. The first few people walk wherever they like. Before long the most convenient routes are worn into the grass since people walk these more often, while the grass on the trails nobody uses grow over again. That is the picture in Figure 3, with neurons as little creatures and connections as footpaths.
At birth, most neurons are asleep, not yet wired into anything. Then comes an explosion. In early childhood the brain overproduces synapses, laying down far more paths than it will ever need, and experience does the sorting. Connections that are used together get stronger, which is Hebb’s old idea, and its cellular form is called long-term potentiation. Rehearse a skill, a route or a chord or a word in a new language, and its path turns into a highway. Paths nobody walks get pruned away. That is not damage but tidying, because a leaner network is faster and cheaper to run. And while you sleep, the brain seems to replay the day and lock in the parts that mattered.
The lawn never stops changing. Spines keep growing and vanishing throughout life, which is why an adult can still learn a new language: slower than a child, but for real.
The 20-watt trick
So how efficient is all this wiring? The brain is about two percent of your body weight and uses about twenty percent of your energy, at roughly twenty watts. A single rack of NVIDIA’s latest accelerators, the GB200 NVL72, is specified at around 120 to 132 kilowatts (HPE, Supermicro). That is about six thousand brains’ worth of power in one cabinet weighing more than a ton, and a big AI data hall holds many such racks.
The gap in learning is even starker. Twenty years of running a brain costs about 3.5 megawatt-hours, my own back-of-envelope figure for twenty watts over twenty years. Outside estimates for training a GPT-4-class model run from roughly thirty to sixty gigawatt-hours, about ten thousand times more (one such estimate). The labs do not publish these numbers, so treat that as a ballpark. And a child hears something like a hundred million words by early adolescence, while a large language model reads on the order of ten trillion, about a hundred thousand times as much.
Two caveats keep this honest. A person arrives with a genome that evolution spent hundreds of millions of years tuning, and learns from sight, sound and touch, not just text. Still, even with generous allowances, the ratios are enormous. The brain learns far more from far less.
So where does the difference come from? Not from better parts. Neurons are slow, noisy and leaky when sitting next to transistors. The savings come from how the work is organized. Most of your neurons are silent at any given moment, and a cell spends energy mainly when it spikes. Each synapse stores its strength right where the signal passes through, so nothing has to be fetched from somewhere else. And learning happens in place, through local changes at each connection, rather than by sending error signals back through the whole system. A GPU does close to the opposite on all three counts, and much of its power goes not on arithmetic but on moving numbers back and forth between memory and processor. To see why the machines ended up this way, it helps to look at what they borrowed from the brain and what they threw away.
What the machines kept, and what they left behind
Artificial neural networks kept the skeleton of the idea. They are made of many simple units joined in layers. Each connection carries a weight, the digital cousin of a synapse’s strength. Each unit sums what it receives, applies a threshold-like function and passes the result on. Learning means adjusting the weights, which is the same idea as strengthening and pruning synapses. Even the layering echoes the brain: early layers of an image network pick out edges and later ones pick out objects, much as the visual system does.
Almost everything else was dropped. Artificial units pass smooth numbers on every step, while neurons send sparse, timed spikes and stay silent most of the time. A real neuron’s dendrites do computing of their own; one study found that imitating a single cortical neuron took a small deep network, not a single unit. Most artificial networks learn by backpropagation, which sends exact error signals backward through the whole network, and nobody has found a clear equivalent in the brain, which seems to rely on local rules and chemical signals. A GPU keeps its weights in memory and hauls them to the processor again and again, while a synapse stores its weight exactly where the computing happens. And once a language model is deployed, its weights are frozen. It adapts through the conversation in front of it, through notes and tools, or through later retraining, never by quietly rewiring itself as you chat. A brain never stops.
So who wins? It depends on what you ask for. The brain wins on energy and on how little it needs to learn. The machine wins on nearly everything else you would want from a tool. A spike takes about a millisecond, while a transistor switches in a fraction of a nanosecond. A trained model can be copied exactly, run in a million places at once, never gets tired, and shares what it learned instantly. A brain forgets, tires and cannot be backed up. The trade makes sense. Evolution built under a strict energy limit, so it used slow parts, sparse activity and memory stored in place. Engineers had abundant power and used it, brute-forcing the problem with speed and data.
Closing the gap
Researchers disagree about how much of the brain to bring back. Picture a spectrum running from “copy the brain as literally as you can” to “ignore it completely”. At one end is whole-brain emulation, scanning and simulating a real brain in detail, which today is a research dream. Next comes brain-style hardware, then brain-inspired ideas in ordinary software, then networks that are only loosely inspired, which describes most of today’s top models, and finally methods with nothing biological about them, such as search and planning, which show up inside larger systems.
Brain-style hardware, called neuromorphic computing, goes after the two big sources of waste. First, work happens only when an event arrives, so units sleep until a spike taps them on the shoulder. Second, memory sits next to compute, so nothing makes long trips. Figure 5 draws this as sleepy workers, each with a pantry beside them, next to a conventional chip whose workers are always awake and whose delivery trucks haul data back and forth from a distant warehouse.
Real versions exist. Intel’s Hala Point links 1,152 Loihi 2 chips into a system with 1.15 billion neurons that draws at most 2,600 watts. That is about an owl’s brain, and roughly one percent of a human’s, and it remains a research system (Intel). IBM’s NorthPole puts memory in with the compute; it is brain-inspired though not spiking. IBM reports that this research prototype is about 25 times more energy-efficient than a GPU made on a comparable process, and about 73 times more efficient per token than an H100 when running a three-billion-parameter language model at low latency (IBM). Smaller chips from companies like SynSense and BrainChip aim at always-on sensing, such as a camera that wakes only when something changes. The big obstacle is software: spikes are hard to train with standard methods, and the tools are young.
Meanwhile, the quieter route has already paid off. Convolutions came from the visual cortex, attention is a kind of selective focus, training replays old experience, and mixture-of-experts models wake only part of the network for each input, much as only a fraction of neurons fire at once. Live research asks whether models can keep learning without forgetting, and whether something like sleep could help them consolidate.
Then there is the opposing view, usually put as planes versus birds. Planes do not flap. Modern language models reached remarkable ability with almost no biology in their design, through a recipe of enormous pretraining on text, fine-tuning on human feedback, reinforcement learning on problems with checkable answers, and tools and agent loops that let them plan and act. They are not spiking, not efficient, and do not learn after shipping, yet they write code and pass hard exams. Whether that amounts to general intelligence is contested; “jagged” is a fair description, superhuman at some things and brittle at others. Nobody outside the labs can say how far the recipe goes.
Where that leaves us
The brain and the data center solved the same problem under opposite constraints. Evolution had a twenty-watt budget and no deadline, and produced slow, sparse, self-rewiring tissue that learns from very little. Engineers had megawatts and a deadline, and produced fast, exact, copyable machines that learn from almost everything ever written.
The interesting work now sits in between: chips that wake only when needed, models that keep learning after they ship, systems that take the brain’s best tricks without its limits. Whether the next big step comes from copying biology or ignoring it, the benchmark is still sitting in your skull, running on a light bulb’s worth of power, and nobody has matched it yet.
A note on sources. The history, the neuron anatomy and the voltages are standard background written from general knowledge. Rack power comes from HPE and Supermicro specifications, the GPT-4 training estimate from this analysis, and the chip results from Intel and IBM. Figures described as rough are approximate, and the figures’ illustrations are simplified.
Comments