The Future of the Mind (35 page)

—ALAN TURING

10
THE ARTIFICIAL MIND AND SILICON CONSCIOUSNESS

In February 2011, history was made.

An IBM computer called Watson did what many critics thought was impossible: it beat two contestants on a TV game show called
Jeopardy!
Millions of viewers were glued to the screen as Watson methodically annihilated its opponents on national TV, answering questions that stumped the rival contestants, and thereby claiming the $1 million prize money.

IBM pulled out all the stops in assembling a machine with a truly monumental amount of computational firepower. Watson can process data at the astonishing rate of five hundred gigabytes per second (or the equivalent of a million books per second) with sixteen trillion bytes of RAM memory. It also had access to two hundred million pages of material in its memory, including the entire storehouse of knowledge within Wikipedia. Watson could then analyze this mountain of information on live TV.

Watson is just the latest generation of “expert systems,” software programs that use formal logic to access vast amounts of specialized information. (When you talk on the phone to a machine that gives you a menu of choices, this is a primitive expert system.) Expert systems will continue to evolve, making our lives more convenient and efficient.

For example, engineers are currently working to create a “robo-doc,” which will appear on your wristwatch or wall screen and give you basic medical advice with 99 percent accuracy almost for free. You’d talk to it about your symptoms, and it would access the databanks of the world’s leading medical centers for the latest scientific information. This will reduce unnecessary visits to the doctor, eliminate costly false alarms, and make it effortless to have regular conversations with a doctor.

Eventually we might have robot lawyers that can answer all common legal questions, or a robo-secretary that can plan vacations, trips, and dinners. (Of course, for specialized services requiring professional advice, you would still need to see a real doctor, lawyer, etc., but for common, everyday advice, these programs would suffice.)

In addition, scientists have created “chat-bots” that can mimic ordinary conversations. The average person may know tens of thousands of words. Reading the newspaper may require about two thousand words or more, but a casual conversation usually involves only a few hundred. Robots can be programmed to converse with this limited vocabulary (as long as the conversation is limited to certain well-defined subjects).

MEDIA HYPE—THE ROBOTS ARE COMING

Soon after Watson won that contest, some pundits were wringing their hands, mourning the day when the machines will take over. Ken Jennings, one of the contestants defeated by Watson, remarked to the press, “I for one welcome our new computer overlords.” The pundits asked, If Watson could defeat seasoned game show contestants in a head-to-machine contest, then what chance do the rest of us mortals have to stand up to the machines? Half jokingly, Jennings said, “Brad [the other contestant] and I were the first knowledge-industry workers put out of work by the new generation of ‘thinking’ machines.”

The commentators, however, forgot to mention that you could not go up to Watson and congratulate it for winning. You could not slap it on its back, or share a champagne toast with it. It wouldn’t know what any of that meant, and in fact Watson was totally unaware that it had won at all. All the hype aside, the truth is that Watson is a highly sophisticated adding machine, able to add (or search data files) billions of times faster than the human brain, but it is totally lacking in self-awareness or common sense.

On one hand, progress in artificial intelligence has been astounding, especially in the area of raw computational power. Someone from the year 1900, viewing the calculations performed by computers today, would consider these machines to be miracles. But in another sense, progress has been painstakingly slow in building machines that can think for themselves (i.e., true automatons, without a puppet master, a controller with a joystick, or someone with a remote-control panel). Robots are totally unaware that they are robots.

Given the fact that computer power has been doubling every two years for the past fifty years under Moore’s law, some say it is only a matter of time before machines eventually acquire self-awareness that rivals human intelligence. No one knows when this will happen, but humanity should be prepared for the moment when machine consciousness leaves the laboratory and enters the real world. How we deal with robot consciousness could decide the future of the human race.

BOOM AND BUST CYCLES IN AI

It is difficult to foretell the fate of AI, since it has gone through three cycles of boom and bust. Back in the 1950s, it seemed as if mechanical maids and butlers were just around the corner. Machines were being built that could play checkers and solve algebra problems. Robot arms were developed that could recognize and pick up blocks. At Stanford University, a robot was built called Shakey—basically a computer sitting on top of wheels with a camera—which could wander around a room by itself, avoiding obstacles.

Breathless articles were soon published in science magazines heralding the coming of the robot companion. Some predictions were too conservative. In 1949,
Popular Mechanics
stated that “in the future, computers will weigh no more than 1.5 tons.” But others were wildly optimistic in proclaiming that the day of the robots was near. Shakey would one day become a mechanical maid or butler that would vacuum our carpets and open our doors. Movies like
2001: A Space Odyssey
convinced us that robots would soon be piloting our rocket ships to Jupiter and chatting with our astronauts. In 1965, Dr. Herbert Simon, one of the founders of AI, said flatly, “
Machines will be capable, within 20 years, of doing any work a man can do.” Two years later, another founding father of AI, Dr. Marvin Minsky, said that “within a
generation … the problem of creating ‘artificial intelligence’ will substantially be solved.”

But all this unbounded optimism collapsed in the 1970s. Checker-playing machines could only play checkers, nothing more. Mechanical arms could pick up blocks, but nothing else. They were like one-trick ponies. The most advanced robots took hours just to walk across a room. Shakey, placed in an unfamiliar environment, would easily get lost. And scientists were nowhere near understanding consciousness. In 1974, AI suffered a huge blow when both the U.S. and British governments substantially curtailed funding in the field.

But as computer power steadily increased in the 1980s, a new gold rush occurred in AI, fueled mainly by Pentagon planners hoping to put robot soldiers on the battlefield. Funding for AI hit a billion dollars by 1985, with hundreds of millions of dollars spent on projects like the Smart Truck, which was supposed to be an intelligent, autonomous truck that could enter enemy lines, do reconnaissance by itself, perform missions (such as rescuing prisoners), and then return to friendly territory. Unfortunately, the only thing that the Smart Truck did was get lost. The visible failures of these costly projects created yet another AI winter in the 1990s.

Paul Abrahams, commenting about the years he spent at MIT as a graduate student, has said, “It’s
as though a group of people had proposed to build a tower to the moon. Each year, they point with pride at how much higher the tower is than it was the previous year. The only trouble is that the moon isn’t getting much closer.”

But now, with the relentless march of computer power, a new AI renaissance has begun, and slow but substantial progress has been made. In 1997, IBM’s Deep Blue computer beat world chess champion Garry Kasparov. In 2005, a robot car from Stanford won the DARPA Grand Challenge for a driverless car. Milestones continue to be reached.

This question remains: Is the third try the charm?

Scientists now realize that they vastly underestimated the problem, because most human thought is actually subconscious. The conscious part of our thoughts, in fact, represents only the tiniest portion of our computations.

Dr. Steve Pinker says, “
I would pay a lot for a robot that would put away the dishes or run simple errands, but I can’t, because all of the little problems that you’d need to solve to build a robot do to that, like recognizing objects,
reasoning about the world, and controlling hands and feet, are unsolved engineering problems.”

Although Hollywood movies tell us that terrifying Terminator robots may be just around the corner, the task of creating an artificial mind has been much more difficult than previously thought. I once asked Dr. Minsky when machines would equal and perhaps even surpass human intelligence. He said that he was confident this would happen but that he doesn’t make predictions about dates anymore. Given the roller-coaster history of AI, perhaps this is the wisest approach, to map out the future of AI without setting a specific timetable.

PATTERN RECOGNITION AND COMMON SENSE

There are at least two basic problems confronting AI: pattern recognition and common sense.

Our best robots can barely recognize simple objects like a cup or a ball. The robot’s eye may see details better than a natural eye, but the robot brain cannot recognize what it is seeing. If you place a robot on a strange, busy street, it quickly becomes disoriented and gets lost. Pattern recognition (e.g., identifying objects) has progressed much more slowly than previously estimated because of this problem.

When a robot walks into a room, it has to perform trillions of calculations, breaking down the objects it sees into pixels, lines, circles, squares, and triangles, and then trying to make a match with the thousands of images stored in its memory. For instance, robots see a chair as a hodgepodge of lines and dots, but they cannot easily identify the essence of “chairness.” Even if a robot is able to successfully match an object to an image in its database, a slight rotation (like a chair that’s been knocked over on the floor) or change in perspective (viewing the chair from a different angle) will mystify the robot. Our brains, however, automatically take different perspectives and variations into account. Our brains are subconsciously performing trillions of calculations, but the process seems effortless to us.

Robots also have a problem with common sense. They do not understand simple facts about the physical and biological world. There isn’t an equation that can confirm something as self-evident (to us humans) as “muggy weather is uncomfortable” or “mothers are older than their daughters.”
There has been some progress made in translating this sort of information into mathematical logic, but to catalog the common sense of a four-year-old child would require hundreds of millions of lines of computer code. As Voltaire once said, “Common sense is not so common.”

For example, one of our most advanced robots is called ASIMO, built in Japan (where 30 percent of all industrial robots are made) by the Honda Corporation. This marvelous robot, about the size of a young boy, can walk, run, climb stairs, speak different languages, and dance (much better than I do, in fact). I have interacted with ASIMO on TV several times, and was very impressed by its abilities.

However, I met privately with the creators of ASIMO and asked them this key question: How smart is ASIMO, if we compare it to an animal? They admitted to me that it has the intelligence of a bug. All the walking and talking is mostly for the press. The problem is that ASIMO is, by and large, a big tape recorder. It has only a modest list of truly autonomous functions, so almost every speech or motion has to be carefully scripted ahead of time. For example, it took about three hours to film a short sequence of me interacting with ASIMO, because the hand gesture and other movement had to be programmed by a team of handlers.

If we consider this in relation to our definition of human consciousness, it seems that our current robots are stuck at a very primitive level, simply trying to make sense of the physical and social world by learning basic facts. As a consequence, robots are not even at the stage where they can plot realistic simulations of the future. Asking a robot to craft a plan to rob a bank, for instance, assumes that the robot knows all the fundamentals about banks, such as where the money is stored, what sort of security system is in place, and how the police and bystanders will react to the situation. Some of this can be programmed, but there are hundreds of nuances that the human mind naturally understands but robots do not.

Where robots excel is in simulating the future in just one precise field, such as playing chess, modeling the weather, tracing the collision of galaxies, etc. Since the laws of chess and gravity have been well known for centuries, it is only a matter of raw computer power to simulate the future of a chess game or a solar system.

Attempts to move beyond this level using brute force have also floundered. One ambitious program, called CYC, was designed to solve the commonsense
problem. CYC would include millions of lines of computer code containing all the information of common sense and knowledge necessary to understand its physical and social environment. Although CYC can process hundreds of thousands of facts and millions of statements, it still cannot reproduce the level of thought of a four-year-old human. Unfortunately, after some optimistic press releases, the effort has stagnated. Many of its programmers left, deadlines have come and gone, and yet the project still continues.

IS THE BRAIN A COMPUTER?

Where did we go wrong? For the past fifty years, scientists working in AI have tried to model the brain by following the analogy with digital computers. But perhaps this was too simplistic. As Joseph Campbell once said, “Computers are like Old Testament gods; lots of rules and no mercy.” If you remove a single transistor from a Pentium chip, the computer will crash immediately. But the human brain can perform quite well even if half of it is missing.

This is because the brain is not a digital computer at all, but a highly sophisticated neural network of some sort. Unlike a digital computer, which has a fixed architecture (input, output, and processor), neural networks are collections of neurons that constantly rewire and reinforce themselves after learning a new task. The brain has no programming, no operating system, no Windows, no central processor. Instead, its neural networks are massively parallel, with one hundred billion neurons firing at the same time in order to accomplish a single goal: to learn.