Part 1 ended on a claim. The brain, I said, is the cheapest thing in the building.
Time to try to prove myself wrong.
So let’s give the machine everything. Perfect memory. Instant recall. Arithmetic you couldn’t do in a lifetime. Then read both meters.
A petabyte on a dim bulb
Researchers at the Salk Institute measured the junctions between neurons and found about 26 distinguishable sizes of them. Roughly 4.7 bits of information each. That pushed estimates of brain memory capacity up tenfold, to at least a petabyte.
In an organ that draws about 20 watts. A dim light bulb.
And you’ll remember from part 1 that thinking as hard as you can adds one or two watts on top. The whole petabyte, searched, cross-referenced and put to work, runs on a bulb that barely flickers.
The tuition bill
Now count what it cost to fill. You spent perhaps 20 years learning to walk, talk, read faces, ride a bicycle and find the fridge in the dark, at 20 watts. That comes to about 3.5 megawatt hours, and some of it involved falling off the bicycle.
Training GPT-3, by the estimate in the paper that calculated it, took 1,287 megawatt hours. That was 10,000 graphics processors drawing about 330 watts each, for a little over two weeks.
Roughly 370 times your lifetime tuition. For a system that learned to write text and can’t walk to the end of the street.
And one of those processors, running flat out, drew 330 watts. You, running everything, including the plan and the shelter, draw 250.
What the machine does better
Fair is fair, so here’s where it wins.
It never forgets. It recalls exactly, where you recall approximately and with confidence, which is worse. Ask any husband about recall that never fails. Stored in perfect fidelity, indexed by date, retrieved on demand, usually mid argument. And when one robot learns something, a million of them can inherit it overnight, which no human skill has ever managed.
Those are real advantages, and they’re advantages of software. Take the software outside and the bill changes. The computer that goes inside a humanoid, Nvidia’s Jetson Thor, is rated at 40 to 130 watts. That’s two to six and a half times your entire brain.
Whether a model clever enough for survival fits on it is an open question. The biggest models live in data centres, and a machine in the wilderness doesn’t get to phone home.
Two systems
So here are both machines, read on the same meter.
One runs on 80 watts, thinks on 20, and carries its entire day as half a kilo of fat. It’s slow to learn, impossible to copy, and it forgets. The other thinks in copies, never forgets, and can’t carry its own day.
Give the robot a perfect mind. The meter doesn’t care. The brain was the cheapest thing in the building when I started this, and it still is. What costs a fortune, in watts and in kilograms, is everything the brain has to move.
There’s no winner on that scoreboard. Each system is strong exactly where the other is weak, and systems like that don’t stay indifferent to each other. They trade, or they compete, or they depend. In my experience, usually all three at once.
Until now I’ve kept them apart, one on each side of the bench. That was the easy part. The robot’s answer to where it plugs in is a grid that people built and people run. For now. And that’s the detail that turns this from a spreadsheet into a story, because we are the ones building them.
Next, I put both of them in the same world.
Next: Two Meters, One World.
Sources and the ballpark
Every figure above is rounded. The learning comparison is my arithmetic on two published numbers: 20 watts for 20 years is about 3.5 megawatt hours, set against 1,287.
- Brain memory capacity (about 4.7 bits per synapse, at least a petabyte) and 20 watts: Salk Institute news release.
- GPT-3 training (1,287 MWh, 10,000 V100 GPUs, about 330 W each, 14.8 days): Patterson et al., Carbon Emissions and Large Neural Network Training.
- Jetson Thor power range (40 to 130 W, developer kit specifications): Nvidia Jetson Thor page.