home > Technology >

Open Outcry

Date: 2001-12-11 Author: Chris Wenham Artist: Amara Telgemeier Best permanent link

Summary: When you give a computer multiple senses and let the universe become your co-programmer, expect bizarre emergent behavior. Third in a series exploring machine intelligence.

If you believe the theory of Evolution, then you should also believe the universe is perverted enough to make a duck.

The price of a hamburger tends to stay the same because McDonald's makes a promise to its farmers that it'll buy their beef at X-dollars a pound 6 months from now. The farmer then takes the contract he and McDonalds forged—called a future—and shows it to his bank to secure any loan he might need in case the months inbetween are not so good. The farmer wins because he knows how much money he'll get 6 months from now, and McDonalds wins because they know exactly how much beef will cost in advance. Of course, suited executives from McDonalds don't spend any time haggling with farmers on porches, nor do they care who the farmer turns out to be, because they buy their futures on The Market, and the biggest market for futures in the whole world is the Chicago Mercantile Exchange—a place that physically resembles the mosh-pit of a rock concert at a hundred and fifty decibels a second.

A trader in the midst of the Exchange's perpetual human earthquake can direct resources measured in billions of dollars just by pointing and waving, for the din is so loud that a system of hand signals had to be developed so orders wouldn't go misrecorded. Tiling all four walls of the pit are electronic scoreboards that log the prices and sales volume of futures and options on everything from beef, to pork bellies, to frozen concentrated orange juice, all displayed large enough to see with a squint from one side of the floor to another. Even the traders themselves embody information, wearing jackets color coded to their job (red is a trader, yellow is a runner, light blue is a market reporter) and placards etched with their trader letters and clearing firm number. Concentrate hard enough on listening, and you can use one of the hundreds of telephones that line the arena. The effect of so much human activity is that, from 7am in the morning until 4pm at night, a non-corporeal intelligence emerges; one that discovers prices, reacts to shortages, and exhibits an unpredictable consciousness all of its own.

Dad, when snow melts, where does all the white go?

Emergent behavior is one that arises from the activity of a network, but isn't spelled out by the design of it. That means you can't point to a dohickey in a car and say “that's where the speed comes from”, because speed is a result of all the parts working together properly, but not the explicit function of any given one of them. The same is true of your mind: there isn't a big master brain-cell in the center of your head that does all the thinking, instead it's the result of over 100 billion little ones each doing a very simple and unintelligent job.

Furthermore, your thoughts don't have purely internal origins either, they all began from somewhere outside your head and made their way in via your senses. And thanks to the invention of language, some of these thoughts have already made a trip through somebody else's mind and come to yours in a purified form, stripped of the background noise. Both types undergo processing in your own mind and, as the result of a mechanism we're only beginning to understand, you have your own original ideas and conclusions. In a way, the universe was your co-programmer along with your mom and dad.

So sensory input must be the name of the game for producing a human-like machine intelligence, because just as the traders of the pit would accomplish no price discoveryD without some buyers and sellers on the outside phoning in their orders, a mind isn't going to do any thinking unless it has something to think about. Senses are where it'll gather material, and the design of the machine is how those sensations will be crunched until behaviors emerge. It can also be compared to Evolution, where a mechanism is established somehow and works on the environment it's in. In Evolution's case the mechanism is a way to pass on variations of a design (mutated genes), combined with an environment that eliminates whatever doesn't fit (natural selection), and the emergent behavior is the diversification of species.

You ever notice ducks always look like they can't take life seriously? They have this silly smile on their bills all the time. Just how can you look at such a funny animal and not crack a smile in response?

Artificial intelligence research wouldn't be nearly so fascinating if it wasn't for the whole idea of building a machine that you cannot predict the behavior of, and comparing that with decades of utilitarian computer programming that has been exactly about writing software which is completely predictable. When a program does something the programmer doesn't expect, it's because the programmer alone screwed up. More profoundly, a programmer who designs and implements a piece of software gets to understand it in such a mundane way that the final product doesn't hold any mystery for him, and the unexpected behavior—called “bugs”—are an annoyance. Simulating human-like thought would put desirable surprises back into his work.

But if the route is to invite the chaotic and mindless universe to be a co-programmer by attaching senses to the machine, then you'd better be ready for bizarre and wacky results to accompany the sensible ones. After all, if not the work of a deity with a sense of humor, then the universe was perverse enough to give us the duck.

There's more to this than simple unexpected results, too, because it will help us answer the inevitable philosophical question AI research prompts: if a mind can be successfully simulated on a machine that adheres to concrete mathematical principles, then does it not have free will?

That, at least, is implied if a mind can indeed be simulated. For assuming you have a second computer with a copy of the same mind, but is twice as fast, wouldn't you be able to predict what the first will think before it thinks it?

My last washroom break cost me $20,000.

Back to the trading floor of the Chicago Mercantile Exchange, where for every successful trader there are at least five who are now driving cabs. Every transaction follows strict rules laid down by the Exchange itself. Every option has an expiration date, every future must be settled on the day it's made out for, every share sold short paid back to the lender. Mathematically rigorous? Yes. Predictable? Not according to the guy who loses half a college fund while he goes to take a leak.

In Santa Fe, New Mexico is a firm called The Prediction Company who use computers programmed to look for trends in the market and automatically take advantage of them by trading electronically. They're bankrolled by UBS Warburg, and make an undisclosed sum of money every month from their financial and technological experiment. What's running on those computers isn't AI (it's just very sophisticated statistical analysis), and yet while an information beseiged trader on the noisy floor of the CME can lose twenty grand in five minutes, The Prediction Company boast of making a million dollars while taking a lunch break.

Is one of these fellows in the wrong career?

While throwing a word processor at every secretary has done precious little for worker productivity, networking bank and exchange computers together is single-handedly responsible for an explosion in the volume of money that gets exchanged every day. It begins with the credit card networks and ends at NASDAQ, where there is no mosh-pit, there are no hand signals, and nobody has to wear a bright colored jacket just to know who to give an order to. The button pressed by the celebrity to start the market every morning in NASDAQ's televised Times Square show-room isn't even wired up to anything, because it is entirely automatic, entirely digitized.

With computer networks, money siphoning techniques which almost sound illegal are possible and practiced every day. The system of hand-signals used by traders and runners in the CME shares a name—ArbitrageD—with a technique that takes advantage of minute differences between the price of a security on one exchange and another (like currency exchanges, where one that isn't updated as frequently as another could sell you a hundred German Marks at a low price, which you then sell on a different exchange for a higher price), but requires the speed of a computer to do effectively, least the prices swing the wrong way to soon and put you in the poor-house.

It's getting to the point where the human element is turning into a disadvantage that banks and exchanges are now trying to remove, but paradoxically is essential for the system to make money at all. Once when I watched a trader making money by short-selling currency—borrowing German Marks to sell, waiting for the price to drop, then buying them back to repay the lender while pocketing the difference—I asked him, "so if the amount of currency at the end of each day is always the same, where does all your money come from?"

And that's when he wheeled around in his chair, jabbed a finger at my chest, and said

YOU!

If every pound of beef was bought without first locking-in a price by negotiating a future, then the price of a hamburger at a McDonald's would swing wildly up and down from one day to the next. Today it's a dollar, tomorrow it's two dollars, the day after that it's dropped to 80 cents. And that's because farming conditions change all the time, subject to weather and disease, which affects the supply which in turn affects the price. The mass computerized networks of New York and Chicago may run algorithms with behaviors that—alone—can be predicted to a decimal point, but the world it measures and serves does not, so the emergent behavior of those networks can't be predictable either.

And so our hypothetical machine intelligence retains the semblance of free will after all, because to predict thoughts before they're made it wouldn't be enough to construct a computer that can simulate a mind, you'd have to build a computer that could simulate the whole universe it works on, too.

The complete series

A Network intelligence is one that exists on no single computer, but emerges as the function of many that are networked together. Our brains are like this, divided into connected lobes and hemispheres that are in turn divided into billions of networked cells. And like our intelligence only emerges because our brains must do something with sensory input, the same applies for the network of computers. Whether the senses are cameras, microphones, or conversations with other people, the machine must have input, and the programmer must let the disorderly real-world universe have a say in the ultimate design of its mind, just like it had a say in the design of the duck.

And that brings us to the subject of the fourth article in the series; because if something looks like a duck, walks like a duck, and quacks like a duck

then does that make it a duck?