The Brain Runs on Twenty Watts — Will Copying It Cut the Power?

So just copy the brain

The previous piece showed why AI eats electricity: memory and calculation live apart, and most of the power goes into carrying numbers between them. The brain keeps the two in one place, so it never pays that cost. Put that way the answer looks obvious. Build a brain.

But the brain is cheap for three reasons, not one, and each has a price. Once you see the price, what should be copied and what should not begin to separate.

First reason — it works in one place

Each nerve cell both remembers and computes. The junction between two cells is called a synapse; how firmly it is joined is the memory, and a signal crossing it is the calculation. There is nothing to carry.

That nerves run on electricity was first seen in a frog's leg. Galvani touched metal to a dead frog and the leg twitched, and from that came the idea that there is electricity inside animals. Imitating that signal in semiconductors, two hundred and forty years later, is what is being attempted now.

Second reason — it fires only when needed

A computer chip runs to a clock. It ticks billions of times a second, and at every tick the whole chip wakes up, whether or not there is anything to do. That is the electricity.

A nerve cell has no clock. It sits quietly until enough signal has gathered, fires once, and goes quiet again. Of the brain's eighty-six billion cells, only a small fraction are firing at any moment. The rest are asleep, and sleeping costs almost nothing.

Third reason — it need not be exact

A computer must not err. One plus one must be two a million times running. So it sends its signals crisply, either on or off, and a crisp signal must be a strong one, and a strong one costs power.

The brain is approximate. A weak signal will do; a single cell can be wrong, because the thousands beside it give roughly the same answer and the whole comes out right. Recognising a face does not need ten decimal places. It does approximate work approximately, and that is cheap.

But there is a price

Turn those three habits over and you find what the brain cannot do.

1. It cannot multiply. Built for approximation, it cannot do a twelve-digit product exactly. What a calculator does in a second takes the brain minutes, and it gets it wrong.
2. It is slow. A nerve signal travels at most about a hundred metres a second; a signal in a wire travels at the speed of light. A difference of more than a million times.
3. It cannot be copied. Twenty years of one person's learning cannot be handed to another without twenty more years. An AI is copied by copying a file.
4. Its memory is unreliable. It cannot tell you what you had for lunch yesterday.

So copy the brain wholesale and you get a computer that cannot multiply, is slow, and cannot be copied. Nobody wants that. What is worth copying is not the brain but its three habits: work in one place, fire only when needed, and do approximate work approximately. Take those, and leave the multiplication to today's computers.

Those three habits, built in semiconductor, are what is called a neuromorphic chip. That is what is being made at KAIST in Daejeon.

A little further in

So what is real and what is inflated?

Where the idea came from

That the brain is made of separate cells was itself a discovery. In 1888 Cajal saw under the microscope that nerve cells stand apart from one another; before that the brain was thought to be one continuous net. Only once they were separate could anyone think of imitating them one by one.

In 1943 McCulloch and Pitts wrote a nerve cell down as an on-off switch, and from then on brains and computers were discussed in the same words. In 1949 Hebb proposed that cells that fire together wire together, which is the root of how AI learns today.

That idea went through two winters and came back to life in the 2010s when it met the graphics chip. But a graphics chip is not built the brain's way; it is warehouse and calculator kept apart. A brain-inspired AI was run on a chip of the opposite design. That is why it eats power. The work now is to move that AI back onto the brain's design.

What you would study to do this

1. Someone who measures how nerve cells actually signal — neuroscience. The person who knows precisely what is being copied.
2. Someone who builds those habits into semiconductor devices — materials and electronics. The work at KAIST in Daejeon.
3. Someone who writes the AI that fits the chip — computer science. The emptiest seat right now: the chip exists and this person is scarce.
4. Someone who joins brain to machine — biomedical engineering. Building devices like Neuralink and putting them safely into a body.
5. Someone who grows nerve cells in a dish and records them with electrodes — laboratory technique in the life sciences.

The fifth line is the unexpected seat. To copy the brain you must first measure how its cells behave, and keeping those cells alive in a dish, placing electrodes and recording the signals is done not by the professor but by the laboratory technician. It needs steady, exact hands, and a college laboratory-science course leads straight to it. Behind every neuroscience paper stands this person.

So

The brain runs on twenty watts, but those twenty watts were bought at the price of not multiplying, being slow, and not being copyable. The way to cut AI's electricity is not to build a brain but to take its three habits and leave the rest to the computers we have.

Of the previous piece's three answers, this is the second: build machines that eat less. The chip is in the laboratory; the AI that fits it is not yet written. In the making.

The question that remainsIf a machine could run on twenty watts at the cost of never multiplying, what work would be worth giving it?