17 FIVE DECADES (OR MORE) FROM NOW
THE FITS AND STARTS OF AI DEVELOPMENT
Now that we’ve created digital automata that can outperform humans at tasks like recognizing objects in images, transcribing recordings of human speech, and playing games like Go, what can we expect to see them do in the next 50 years? Plenty. But before we speculate on where we’re headed next, let’s briefly look at how far we’ve come.
A lot of the ideas from the past 20 years that have created excitement in the AI community were the same things that drummed up excitement around AI half a century ago, in the late 1960s. Back then, the field of AI felt like it was roaring ahead, with improvements in neural networks, development of algorithms to play games like chess and Go, excitement at conferences (where AI felt like it was bursting at the seams) and hardware that was growing exponentially with the advent of microprocessors—all just before the field of AI went into a dark period known as the AI Winter. Funding for AI research dried up for several decades. AI even became “a term of derision” among some researchers.1 The funk lasted through much of the 1980s and 1990s, before the field’s rebirth over the past two decades.
In other words, the progress we’ve seen since just before the turn of the century is noteworthy, but it’s not an isolated burst of technological advancement, even in the field of AI. It’s part of a longer, sustained series of developments in AI—a series of developments that comes in fits and starts.
The automata our ancestors created in the 18th century were also part of a sustained development of technology spanning many decades. In Europe, mechanics created automata during the 18th and 19th centuries, but the trend had been going on globally for a much longer period. A trio of Persian brothers created a programmable flute-playing device as early as the ninth century, and the Greeks had developed primitive steam engines by the first century AD.2 We should expect our modern, digital automata to follow a similarly long arc of progress, interrupted by periods of slow progress.
HOW TO REPLICATE THE SUCCESSES IN THIS BOOK
Many of the machines we’ve seen might look superficially different, but they have an enormous amount in common. Classifiers enabled these intelligent machines to perceive the world. Finite state machines and recurrent neural networks enabled them to keep track of what was going on around them—what they had done, what they are doing, and what they still need to do—and to focus on only the most salient parts of their environment. Search algorithms enabled them to brute-force their way to the best among millions of options. And reinforcement learning gave them the ability to learn from their experience. These “statistical elements” were then combined via remarkably similar architectures into the machines we saw, which could drive autonomously, predict humans’ preferences for movies, answer Jeopardy! questions, and play games of strategy with stunning precision.
But the design of these statistical machines was only part of the story. All of these machines required prolonged and well-organized human effort. The smallest team in this book to “succeed” was the one that created IBM’s Deep Blue; made up of only a few people, it gained and lost members here or there like a rock band over its dozen years of work. But Deep Blue attacked the problem for an entire decade. Many of the other teams we saw took less time to develop their products but were much larger—often dozens of researchers and engineers working on a project for a year or more, typically reaching tens or hundreds of person-years of research and development. And this required careful management of these teams’ efforts.
Sebastian Thrun’s experience in organizing the effort around Stanley the self-driving car set an excellent standard for such a high-functioning team. Sometimes he needed to make the tough but necessary decision to tell someone their part of the project, on which they might have been working for months, would not make the cut to be in the final robot. But his team members, whom he had carefully selected, recognized that this was for the good of the project.3 For them, winning was a group effort, and everyone—including the leaders—made sacrifices. Sebastian explained:
During this phase of the project, everyone on the core team fully understood what it meant to play with the team. Getting lunch for the team was as noble a deed as writing cutting-edge software. To the present day, I continue to be amazed by the willingness of every single team member to do whatever I asked him or her to do. And I tried to lead by example. My personal highlight was the day I spent building a tank trap out of PVC pipes. After bolting together three pipes, my team noted that the surface wasn’t sufficiently similar to rusty metal. So I went back to the store to buy spray paint, and then spent hours applying a combination of paint and dirt to give the trap the look of a World War Two tank trap. This was not exactly the type [of] job for which I had come to Stanford. But it was magically gratifying to keep my hands dirty and to spend my time on mundane things of no scientific value whatsoever.4
These teams also couldn’t have succeeded if they hadn’t been embedded within larger communities of engineers and research scientists that shared their knowledge broadly. This was by design in competitions like the DARPA Grand Challenge and the Netflix Prize, but it was also true for projects like AlphaGo. Although AlphaGo was created by about 20 people in a private company, many of the ideas in AlphaGo—such as Monte Carlo Tree Search, evaluation functions, reinforcement learning, and deep neural networks—had been developed in the decades before DeepMind worked on the problem. Most of these projects succeeded not just because they were driven by a large engineering team with a clear goal and funding—but also because the ideas from which they came had been incubated by a publicly funded research community that offered the collective wisdom of decades of supporting research and experimentation. This was true even for privately funded projects: some of the core researchers on AlphaGo, for example, cut their teeth at the University of Alberta, and IBM Watson drew heavily from talent and ideas in the academic community.
Walter Isaacson came to a similar conclusion in his book The Innovators. He noted the difficulty in attacking an ambitious problem in a vacuum. Virtually none of the major advances in the history of the computer were the result of a lone tinkerer in his garage. The same is true of advances in AI and machine learning.
Does this mean that a lone researcher shouldn’t bother to start with a project if he doesn’t have a big budget and a team of researchers? Not at all, but it can still help to join or organize a larger effort down the road. Remember, for example, that the team called Pragmatic Theory started out as “two guys, absolutely no clue.” But they carefully studied what the best teams did, which enabled them to rise quickly within the community and to eventually join what became the winning team. The team that created the chess-playing program Deep Blue also started out small, but eventually its members joined IBM, where they continued to develop Deep Blue over the following eight years before beating Garry Kasparov. And ultimately all of these projects started with one person who had an idea.
Sometimes the people with the ideas don’t even have to solve the problem to have an impact: as we saw, they can organize a competition to encourage researchers to coalesce around a common cause. Is it possible that these competitions don’t always foster advances, and instead just provide more transparency into progress that’s already happening? This probably happens sometimes, but the Netflix Prize is a shining example of a competition that clearly added impetus to a field.
When Netflix planned their competition they made several important decisions that can serve as an example for future competition organizers. First, the dataset they released to the community was large enough to be valuable—it was 100 times the size of other public datasets of the same type—yet it was small-enough, and Netflix had cleaned it up well enough, that it was easy to work with. Second, they offered a large cash prize to the winners. Netflix also chose a good target for the Grand Prize: 10 percent was a difficult but not impossible target for teams to achieve.5 They created a lively community around the project, offering an online forum where participants could share ideas and where a leaderboard could foster excitement. And finally, Netflix helped the researchers to move along by requiring winners to write reports before they could claim either a Progress Prize or the Grand Prize; these reports were widely read by members of the community.6
Competitions have the benefit that they can change the way a research community invests its time. One way they do this is by standardizing research. We’ve seen the same thing in financial markets: that publicly traded securities are fungible—that is, exchangeable with one another—means that they can be objectively evaluated, priced, and, ultimately, compared with one another. This helped with the ImageNet Challenge in 2012, where a neural network was the undisputed winner. Since all of the entrants to the competition were evaluated on the same criteria, it was clear that the network was the fair winner. Other teams immediately jumped aboard the deep-learning bandwagon, and in subsequent years the top contestants all used deep convolutional neural networks in their submissions.7 While the 2012 team won by a large margin, nine teams in 2013 beat the best 2012 team, and progress was rapid in the ensuing years.
PERVASIVE USE OF DATA
Another recurring theme in the development of the statistical machines we’ve seen was their pervasive use of experiments and data. In some cases, large quantities of data were available because that data had been collected and organized by passionate gamer geeks. We saw this with games like Go (for which online games had been recorded) and Jeopardy (for which fans had collected questions from televised episodes). In other cases, academic researchers and companies put together comprehensive, well-labeled datasets.
In yet other cases, researchers found ways to create their own data. Sebastian Thrun and his Stanford team drove around in a car covered with sensors to collect training data for their terrain-detecting classifier. The Atari-playing neural network played millions of games in the Arcade Learning Environment to collect the data that it needed to improve its play. And the creators behind AlphaGo, the DOTA 2 bot, and the backgammon-playing program turned their programs against themselves so they could create their own training data. The only bottleneck to how much data these game-playing programs could train on was the time it took the computers to play through their games.
WHERE WE GO NEXT
I’ve intentionally avoided much speculation about the future of AI in this book because I’m an engineer, not a philosopher, economist, or historian. But I do believe we’ve seen enough evidence in the development of these intelligent machines that I can say a couple of things about the future with fairly high confidence (although many of these things may take centuries, not decades, to occur).
First, the automata we create in the future will invariably still follow programs. This is a constraint of the media we’ve used to create these automata and the physical laws of the world we live in. These machines will follow programs that will grow more and more complex, and it will become more and more difficult to discern what they’re doing, but it will always be possible to trace every action they perform back to a deterministic set of instructions.8 Some philosophers have argued that this suggests that machines will never think.9 My own belief is that humans are machines as well—we’re analog machines—and if we believe that humans can think, then there’s nothing to preclude us from designing digital computers that will also someday think. Rather, it’s inevitable that our machines will someday think, and that they will develop emotions, opinions, and the desire for self-preservation—which will someday conflict with our own.
Second, we will continue to build machines that can replicate our intelligence and behavior more and more accurately, until there is no discernable difference between their abilities to perceive and reason and our own abilities to do these things—except that the machines will be better than us in many ways. We’ve been trying to do this since long before Vaucanson and his contemporaries tried their damnedest to create automata that looked and acted human.
As we continue to build better automata, these efforts will inevitably feed the perception that these machines are a threat to humanity—that they will steal our jobs and destroy our livelihoods. At the very least, these machines will make us uncomfortable in their uncanny resemblance to us. Remember: Vaucanson himself was forced to close of one of his workshops because a religious official considered it “profane.”10 And to some extent it will be true that these machines will be a threat to us: machines will take peoples’ jobs precisely because they will do them more cheaply. Robots will be the “immigrants” blamed by future politicians, and their creators will market them carefully, just as IBM carefully positioned Watson. This will require our leaders to make thoughtful decisions to ensure that the benefits of improving technology are fairly distributed, and we should expect no less of them.
But however well our society can absorb these agents, we will continue to build them to meet and exceed our abilities as long as our technology—our hardware, our theory, and the software architectures behind them—continues to improve. Some of this will be driven by economics and business, but the drive to build such machines will continue long after any economic motivation has disappeared. Building machines in our image is a human endeavor, and certain qualities of human nature—curiosity, aesthetics, hubris, and vanity, but mostly curiosity and aesthetics—will compel us to continue.