Where AlphaGo and Boston Dynamics are still falling short (putting Lee Sedol’s loss in a different context)

Posted on April 4, 2016

(Last Updated On: April 6, 2016)

Two big things in the world of artificial intelligence have happened recently. This:

And AlphaGo, a computer program, beat Lee Sedol in the game of Go.

AlphaGo lee sedol

The accomplishment of the robot is likely obvious. With Go, it’s like when IBM had a computer beat a top rated human chess player. Except Go has way, way, more moves than chess. Furthermore, AlphaGo beat a top ranked player by actually learning how to play the game, rather than explicitly programming “If this, do that, else do this.” It came up with its own strategy after observing (many) other games.


In this lecture six years ago, Marc Raibert, founder of Boston Dynamics, said their current innovation of the time -a dog like robot rather than a human like one- had an energy efficiency of something like 2 miles per gallon, or 1 mile per 2 gallons. (He wasn’t sure on the exact number.) Then it would poop out, needing a gasoline refill. Obviously, we have ourselves a bottleneck here.

In what is called cost of transport -(power needed) / (mass * velocity)- it had a cost of transport of 15. The lower this number, the better.

For the new Atlas robot, cost of transport is actually worse. At 20.

For context, a human comes in at like 0.15. Or 133 times better.

Even a robot built to minimize cost of transport only hopes to come in at 1. About 10 times worse than a human. That robot currently can walk for a couple hours, or a little over 2 kilometers. The makers hope to get it able to walk 10 kilometers. About six miles.

For many humans, going for a 10k run is called “warming up.” Considering water is not a source of chemical energy for people -it has no calories- we can go what, three weeks or more, without needing to be “recharged.” Even with water, or not going too long without sleep, we’re talking a few days.

Think about a gallon of gasoline. That equates to about 31,000 calories. Humans, walking, burn about 100 calories per mile. That’s 310 miles we could go on a human equivalent gallon of gasoline. At 2,000 calories per day, that’s 15 days worth of food.

Artificial Intelligence

The notion the Go game of DeepMind versus Lee Sedol was man vs machine isn’t quite accurate. It was man vs machineS.

This is the founder of DeepMind:

Demis Hassabis power of alphago

Fan Hui was a lesser competitor than Lee Sedol.

According to their January paper, in the Fan Hui match they used-

  • 1202 CPUs
  • 176 GPUs

DeepMind is affiliated with Google. DeepMind = DeepPockets.

Now this gets to be a pain in the ass to figure out, so if some computer scientist comes across this and finds it’s off (despite me asking my brother who is finishing up his CS degree), forgive me (and let me know), but we’ll say DeepMind is using the most powerful stuff we have. Something like the Intel Core i7 Extreme Edition microprocessor. According to Wikipedia, this uses ~145 watts.

Pinning down GPU power supply recommendations is tougher. This chart has GPUs which get up to 1700 watts. We’ll go with that.

  • (1202 CPU * 145 watts per CPU) + (176 GPU * 1700 watts per GPU) = 473,490 watts

This matches up with a pioneer of AlphaGo’s technology and Google employee’s take, which is that AlphaGo was likely using “hundreds of kilowatts.” 473,490 watts => 473 kilowatts. We’re right there.

Alright, and what does the human brain get up to?

  • 20 watts

That’s not a typo!

These numbers we’re saying are the amount of power at any given moment. Let’s say each respective player used this amount of power throughout the entire match. Energy is measured in terms of kilowatt hours. We’ll then say AlphaGo was using 443 kilowatt-hours of energy. In California, where Google is based, it’s something like 20 cents per kilowatt-hour.

  • 443 kilowatt-hours * 20 cents = $88.6

That would be the hourly rate. That translates to a salary of $184,288. It’s worth examining this number because for artificial intelligence and robots to supplant human workers, the financial incentive is critical.

-> One of the legitimate concerns about raising the minimum wage is you decrease this threshold. Does $15 an hour suddenly make automated McDonald’s orders make more sense?

With our robot above, while humans are way more efficient at using calories, for a human to supply itself cheaply with 2,000 calories is something like the cost of a gallon of milk. Say $3. Give or take the same price as a gallon of gasoline, which can supply 31,000 calories! Humans can way more efficiently use calories; we’re way worse at getting them cost effectively.

Say over the course of a workday, 8 hours, the robot needs eight gallons of gasoline for four miles of work. That’s still only $24 in cost. Way less than any developed country worker is going to do labor for. In other words, if a job was to strictly walk a certain amount of miles each day, no state in the U.S. would even allow a human to be paid the robot’s rate.

At the other extreme, with AlphaGo, we’re talking $184,288 just to supply the program’s electricity. We’re not considering maintaining the hardware (need to pay someone(s)), maintaining the software (need to pay someone(s)), the 20 or so people it took to create the program over the course of a couple years, all the kilowatt-hours it took to train the program- basically 443 kilowatt hours for every hour for 18 months. The thing ran non-stop, which is likely a big reason this happened now. Someone like Google 1) had the hardware 2) was able to pony up the money 3) doesn’t need an immediate financial return. All opposed to some gigantic breakthrough in AI, as deep learning and neural nets have been around decades. (Note much of the above applies to the robot as well.)

Now how many jobs out there require something analogous to walking a few miles each day? How many jobs require the computation of Go? How many jobs require more computation than Go? This is after all only a board game. It doesn’t matter there are more potential moves in this game than atoms in the observable universe. Machines were trained to conquer this game in about 18 months. We are still nowhere near a computer having fluent conversation despite decades of trying. In fact, the same company, Google, funding AlphaGo has significant funding in speech recognition, as does a similar company, Apple, yet here we are regularly calling Siri a bitch. How many kilowatt-hours will understanding sarcasm require?

A lot of concern has been placed towards lower level jobs. Perhaps factory workers really need to worry about things like what Boston Dynamics are working on, and now have some more insight as to why we’ve already seen places like Amazon, who are cost ruthless, be able to replace so many humans with robots. On the other end, if something customer service oriented -good deal of speech recognition- ends up requiring anywhere near an AlphaGo type of hourly wage, then it’s not the lower level workers we should be worried about. It’s the higher paid ones. All this advancement and someone still listens to my order and hands me my money at McDonald’s, and we all breathe a sigh of relief when a person picks up the phone rather than an automated response. However, if you make $200,000 a year but there is a program which can cut you off by 15 grand…

But fear not higher paid worker? There’s even more to add to that $184,288. AlphaGo didn’t physically move the pieces on the board itself, but Lee Sedol did. Sedol only took up the space in the chair he sat, while who knows how much space DeepMind needed for all that hardware. They even used an internet connection. (A special one. They brought their own fiber to the event.)

Internet cables take up a wee bit of room:

undersea internet cables around world

They’re underwater.

Wired mentions AlphaGo connecting into Google’s data centers around the world. I actually just went in for a tour of a data center a month ago. (My dad has been working in them for years.) A single data center is roughly the size of a one floor Target. They are massive, extremely costly buildings, which increases the program’s hourly rate.

-> The places are run damn near black, to save money on lighting.

And Lee Sedol’s brain was doing other work whether he wanted it to or not. It was holding his posture, unconsciously listening for a potential threat, and god knows how many other unconscious activities it has a role in. Where maybe only 10 watts are able to be used for computation.

The two big advances of 2016 so far, in the above context,

  • The Atlas robot is 133 times worse at walking than us
  • AlphaGo needed enough cable to connect the entire world, 22,000 times more watts of power, and still lost a game!

We humans still have some tricks up our sleeves. For now at least.

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.