The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies

by Erik Brynjolfsson

Moravec’s paradox,

In his book Why the West Rules—For Now, anthropologist Ian Morris starts tracking human societal progress in 14,000 BCE, when the world clearly started getting warmer.

social development (“a group’s ability to master its physical and intellectual environment to get things done”)

The Industrial Revolution, of course, is not only the story of steam power, but steam started it all. More than anything else, it allowed us to overcome the limitations of muscle power, human and animal, and generate massive amounts of useful energy at will.

Now comes the second machine age. Computers and other digital advances are doing for mental power—the ability to use our brains to understand and shape our environments—what the steam engine and its descendants did for muscle power.

This work led us to three broad conclusions. The first is that we’re living in a time of astonishing progress with digital technologies—those that have computer hardware, software, and networks at their core.

Our second conclusion is that the transformations brought about by digital technology will be profoundly beneficial ones.

Our third conclusion is less optimistic: digitization is going to bring with it some thorny challenges.

Rapid and accelerating digitization is likely to bring economic rather than environmental disruption, stemming from the fact that as computers get more powerful, companies have less need for some kinds of workers. Technological progress is going to leave behind some people, perhaps even a lot of people, as it races ahead.

However, there’s never been a worse time to be a worker with only ‘ordinary’ skills and abilities to offer, because computers, robots, and other digital technologies are acquiring these skills and abilities at an extraordinary rate.

In 2004 Frank Levy and Richard Murnane published their book The New Division of Labor.1 The division they focused on was between human and digital labor—in other words, between people and computers. In any sensible economic system, people should focus on the tasks and jobs where they have a comparative advantage over computers, leaving computers the work for which they are better suited.

The authors put information processing tasks—the foundation of all knowledge work—on a spectrum. At one end are tasks like arithmetic that require only the application of well-understood rules. Since computers are really good at following rules, it follows that they should do arithmetic and similar tasks.

When expressed in computer code, we call a mortgage rule like this an algorithm. Algorithms are simplifications; they can’t and don’t take everything into account

At the other end of Levy and Murnane’s spectrum, however, lie information processing tasks that cannot be boiled down to rules or algorithms. According to the authors, these are tasks that draw on the human capacity for pattern recognition.

As the philosopher Michael Polanyi famously observed, “We know more than we can tell.”

DARPA, the Defense Advanced Research Projects Agency, was founded in 1958 (in response to the Soviet Union’s launch of the Sputnik satellite) and tasked with spurring technological progress that might have military applications.

Improvement in autonomous vehicles reminds us of Hemingway’s quote about how a man goes broke: “Gradually and then suddenly.”

In addition to pattern recognition, Levy and Murnane highlight complex communication as a domain that would stay on the human side in the new division of labor. They write that, “Conversations critical to effective teaching, managing, selling, and many other occupations require the transfer and interpretation of a broad range of information. In these cases, the possibility of exchanging information with a computer, rather than another human, is a long way off.”6 In the fall of 2011, Apple introduced the iPhone 4S featuring “Siri,” an intelligent personal assistant that worked via a natural-language user interface.

As noted by Tom Mitchell, who heads the machine-learning department at Carnegie Mellon University: “We’re at the beginning of a ten-year period where we’re going to transition from computers that can’t understand language to a point where computers can understand quite a bit about language.”11

The word robot entered the English language via the 1921 Czech play, R.U.R. (Rossum’s “Universal” Robots) by Karel Capek, and automatons have been an object of human fascination ever since.

As the roboticist Hans Moravec has observed, “It is comparatively easy to make computers exhibit adult-level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.”27 This situation has come to be known as Moravec’s paradox, nicely summarized by Wikipedia as “the discovery by artificial intelligence and robotics researchers that, contrary to traditional assumptions, high-level reasoning requires very little computation, but low-level sensorimotor skills require enormous computational resources.”28* Moravec’s insight is broadly accurate, and important. As the cognitive scientist Steven Pinker puts it, “The main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard. . . . As the new generation of intelligent devices appears, it will be the stock analysts and petrochemical engineers and parole board members who are in danger of being replaced by machines. The gardeners, receptionists, and cooks are secure in their jobs for decades to come.”

Today’s factories, especially large ones in high-wage countries, are highly automated, but they’re not full of general-purpose robots. They’re full of dedicated, specialized machinery that’s expensive to buy, configure, and reconfigure.

In 2008 Brooks founded a new company, Rethink Robotics, to pursue and build untraditional industrial automation:

Boston Dynamics, yet another New England startup, has tackled Moravec’s paradox head-on. The company builds robots aimed at supporting American troops in the field by, among other things, carrying heavy loads over rough terrain.

On Star Trek, tricorders and person-to-person communicators were separate devices, but in the real world the two have merged in the smartphone. They enable their users to simultaneously access and generate huge amounts of information as they move around. This opens up the opportunity for innovations that venture capitalist John Doerr calls “SoLoMo”—social, local, and mobile.

3D printing, also sometimes called “additive manufacturing,” takes advantage of the way computer printers work: they deposit a very thin layer of material (ink, traditionally) on a base (paper) in a pattern determined by the computer.

The digital progress we’ve seen recently is certainly impressive, but it’s just a small indication of what’s to come. It’s the dawn of the second machine age. To understand why it’s unfolding now, we need to understand the nature of technological progress in the era of digital hardware, software, and networks. In particular, we need to understand its three key characteristics: that it is exponential, digital, and combinatorial.

To highlight it better, we’ll change to logarithmic spacing, where each segment of the vertical axis represents a tenfold increase in tribbles: an increase first from 1 to 10, then from 10 to 100, then from 100 to 1,000, and so on. In other words, we scale the axis by powers of 10 or orders of magnitude. Logarithmic graphs have a wonderful property: they show exponential growth as a perfectly straight line.

This view emphasizes the steadiness of the doubling over time rather than the large numbers at the end. Because of this, we often use logarithmic scales for graphing doublings and other exponential growth series. They show up as straight lines and their speed is easier to evaluate; the bigger the exponent, the faster they grow, and the steeper the line.

Kurzweil’s distinction between the first and second halves of the chessboard inspired a quick calculation. Among many other things, the U.S. Bureau of Economic Analysis (BEA) tracks American companies’ expenditures. The BEA first noted “information technology” as a distinct corporate investment category in 1958. We took that year as the starting point for when Moore’s Law entered the business world, and used eighteen months as the doubling period. After thirty-two of these doublings, U.S. businesses entered the second half of the chessboard when it comes to the use of digital gear. That was in 2006.

Researchers in artificial intelligence have long been fascinated (some would say obsessed) with the problem of simultaneous localization and mapping, which they refer to as SLAM. SLAM is the process of building up a map of an unfamiliar building as you’re navigating through it—where are the doors? where are stairs? what are all the things I might trip over?—and also keeping track of where you are within it (so you can find your way back downstairs and out the front door). For the great majority of humans, SLAM happens with minimal conscious thought. But teaching machines to do it has been a huge challenge.

A Google autonomous car incorporates several sensing technologies, but its most important ‘eye’ is a Cyclopean LIDAR (a combination of “LIght” and “raDAR”) assembly mounted on the roof. This rig, manufactured by Velodyne, contains sixty-four separate laser beams and an equal number of detectors, all mounted in a housing that rotates ten times a second. It generates about 1.3 million data points per second, which can be assembled by onboard computers into a real-time 3D picture extending one hundred meters in all directions. Some early commercial LIDAR systems available around the year 2000 cost up to $35 million, but in mid-2013 Velodyne’s assembly for self-navigating vehicles was priced at approximately $80,000, a figure that will fall much further in the future. David Hall, the company’s founder and CEO, estimates that mass production would allow his product’s price to “drop to the level of a camera, a few hundred dollars.”

Multiplying 62.34 by 24358.9274 is an example of a floating point operation. The decimal point in such operations is allowed to ‘float’ instead of being fixed in the same place for both numbers.

That Waze gets more useful to all of its members as it gets more members is a classic example of what economists call a network effect—a situation where the value of a resource for each of its users increases with each additional user.

Digitization, in other words, is the work of turning all kinds of information and media—text, sounds, photos, video, data from instruments and sensors, and so on—into the ones and zeroes that are the native language of computers and their kin.

Like so many other modern online services, Waze exploits two of the well-understood and unique economic properties of digital information: such information is non-rival, and it has close to zero marginal cost of reproduction. In everyday language, we might say that digital information is not “used up” when it gets used, and it is extremely cheap to make another copy of a digitized resource. Let’s look at each of these properties in a bit more detail.

While the very first copy of a book or movie might cost a lot to create, making additional copies cost almost nothing. This is what is meant by “zero marginal cost of reproduction.”

This huge body of information was not cheap to generate, but once it’s digitized it’s very cheap to replicate, chop up, and share widely and repeatedly. This is exactly what a service like Google Translate does. When it gets an English sentence and a request for its German equivalent, it essentially scans all the documents it knows about in both English and German, looking for a close match (or a few fragments that add up to a close match), then returns the corresponding German text. Today’s most advanced automatic translation services, then, are not the result of any recent insight about how to teach computers all the rules of human languages and how to apply them. Instead, they’re applications that do statistical pattern matching over huge pools of digital content that was costly to produce, but cheap to reproduce.

To test this hypothesis, Erik asked Google if he could access data about its search terms. He was told that he didn’t have to ask; the company made these data freely available over the Web.

Digitization can also help us better understand the past. As of March 2012 Google had scanned more than twenty million books published over several centuries.18 This huge pool of digital words and phrases forms a base for what’s being called culturomics, or “the application of high-throughput data collection and analysis to the study of human culture.”

“If you want to have good ideas you must have many ideas.” —Linus Pauling

With their typical verbal flair, economists call innovations like steam power and electricity general purpose technologies (GPTs). Economic historian Gavin Wright offers a concise definition: “deep new ideas or techniques that have the potential for important impacts on many sectors of the economy.”7 “Impacts” here mean significant boosts to output due to large productivity gains. GPTs are important because they are economically significant—they interrupt and accelerate the normal march of economic progress. In addition to agreeing on their importance, scholars have also come to a consensus on how to recognize GPTs: they should be pervasive, improving over time, and able to spawn new innovations.

When multiple GPTs appear at the same time, or in a steady sequence, we sustain high rates of growth over a long period. But if there’s a big gap between major innovations, economic growth will eventually peter out. We’ll call this the ‘innovation-as-fruit’ view of things, in honor of Tyler Cowen’s imagery of all the low-hanging fruit being picked. In this perspective, coming up with an innovation is like growing fruit, and exploiting an innovation is like eating the fruit over time. Another school of thought, though, holds that the true work of innovation is not coming up with something big and new, but instead recombining things that already exist. And the more closely we look at how major steps forward in our knowledge and ability to accomplish things have actually occurred, the more this recombinant view makes sense.

As he summarizes in his book The Nature of Technology, “To invent something is to find it in what previously exists.”14 Economist Paul Romer has argued forcefully in favor of this view, the so-called ‘new growth theory’ within economics, in order to distinguish it from perspectives like Gordon’s. Romer’s inherently optimistic theory stresses the importance of recombinant innovation.

Romer also makes a vital point about a particularly important category of idea, which he calls “meta-ideas”: Perhaps the most important ideas of all are meta-ideas—ideas about how to support the production and transmission of other ideas. . . . There are . . . two safe predictions. First, the country that takes the lead in the twenty-first century will be the one that implements an innovation that more effectively supports the production of new ideas in the private sector. Second, new meta-ideas of this kind will be found.16

The Web itself is a pretty straightforward combination of the Internet’s much older TCP/IP data transmission network; a markup language called HTML that specified how text, pictures, and so on should be laid out; and a simple PC application called a ‘browser’ to display the results. None of these elements was particularly novel. Their combination was revolutionary.

If this recombinant view of innovation is correct, then a problem looms: as the number of building blocks explodes, the main difficulty is knowing which combinations of them will be valuable. In his paper “Recombinant Growth,” the economist Martin Weitzman developed a mathematical model of new growth theory in which the ‘fixed factors’ in an economy—machine tools, trucks, laboratories, and so on—are augmented over time by pieces of knowledge that he calls ‘seed ideas,’ and knowledge itself increases over time as previous seed ideas are recombined into new ones.18 This is an innovation-as-building-block view of the world, where both the knowledge pieces and the seed ideas can be combined and recombined over time. This model has a fascinating result: because combinatorial possibilities explode so quickly there is soon a virtually infinite number of potentially valuable recombinations of the existing knowledge pieces.* The constraint on the economy’s growth then becomes its ability to go through all these potential recombinations to find the truly valuable ones.

The advances we’ve seen in the past few years, and in the early sections of this book—cars that drive themselves, useful humanoid robots, speech recognition and synthesis systems, 3D printers, Jeopardy!-champion computers—are not the crowning achievements of the computer era. They’re the warm-up acts. As we move deeper into the second machine age we’ll see more and more such wonders, and they’ll become more and more impressive. How can we be so sure? Because the exponential, digital, and recombinant powers of the second machine age have made it possible for humanity to create two of the most important one-time events in our history: the emergence of real, useful artificial intelligence (AI) and the connection of most of the people on the planet via a common digital network.

We’ve also recently seen great progress in natural language processing, machine learning (the ability of a computer to automatically refine its methods and improve its results as it gets more data), computer vision, simultaneous localization and mapping, and many of the other fundamental challenges of the discipline.

The economist Julian Simon was one of the first to make this optimistic argument, and he advanced it repeatedly and forcefully throughout his career. He wrote, “It is your mind that matters economically, as much or more than your mouth or hands. In the long run, the most important economic effect of population size and growth is the contribution of additional people to our stock of useful knowledge. And this contribution is large enough in the long run to overcome all the costs of population growth.”