Epilogue

Trustworthy AI, grounded in reasoning, commonsense values, and sound engineering practice, will be transformational when it finally arrives, whether that is a decade or a century hence.

Already, over the past two decades, we have seen major technological advances, mostly in the form of “blank slate” machine learning applied to big data sets, in applications such as speech recognition, machine translation, and image labeling. We don’t expect that to stop. The state of the art in image and video labeling will continue to progress; chatbots will get better; and the ability of robots to maneuver and grasp things will continue to improve. We will see more and more clever and socially beneficial applications, such as using deep learning to track wildlife and predict aftershocks of earthquakes. And of course there will also be advances in less benign domains, such as advertising, propaganda, and fake news, along with surveillance and military applications, long before the reboot that we have called for arrives.

But eventually all that will seem like an appetizer. In hindsight, the turning point will be seen not as the 2012 rebirth of deep learning, but as the moment at which a solution to the challenges of common sense and reasoning yields deeper understanding.

What will that mean? Nobody knows for sure; nobody can pretend to predict what the future will look like in all its ramifications. In the 1982 film Blade Runner, the world is filled with advanced AI-powered replicants that seem all but indistinguishable from humans—and yet at a crucial moment Rick Deckard (Harrison Ford) stops at a pay phone to place a call. In the real world, it’s a lot easier to replace pay phones with cell phones than to build human-level AI, but nobody on the film crew considered that part of the chronology. In any set of predictions about fast-moving technologies, ours or anyone else’s, there are bound to be some pretty glaring errors.

Still, we can at least make some educated guesses. To begin with, AI that is powered by deep understanding will be the first AI that can learn the way a child does, easily, powerfully, constantly expanding its knowledge of the world, and often requiring no more than one or two examples of any new concept or situation in order to create a valid model of it. It will also be the first to truly comprehend novels, films, newspaper stories, and videos. Robots embedded with deep understanding would be able to move safely around the world and physically manipulate all kinds of objects and substances, identifying what they are useful for, and to interact comfortably and freely with people.

Once computers can understand the world and what we’re saying, the possibilities are endless. To begin with, search will get better. Many of the questions that confuse current technology—“Who is currently on the Supreme Court?,” “Who was the oldest Supreme Court justice in 1980?,” “What are the Horcruxes in Harry Potter?”—would suddenly be easy for machines. And so too will many questions we wouldn’t dream of asking now: A screenwriter could ask some future search engine, “Find a short story involving the leader of one country being an agent of another, for screen adaptation.” An amateur astronomer could ask, “When’s the next time that Jupiter’s Great Red Spot will be visible?,” taking into account the local weather forecasts as well as the astronomy. You could tell your video game program that you want your avatar to be a rhinoceros wearing a tie-dyed shirt, instead of choosing from a dozen preprogrammed options. Or ask your e-reader to track every book you read, and sort how much time you spend reading literature from each continent, broken down by genre.

Digital assistants, meanwhile, will be able to do pretty much anything human assistants can do, and they will become democratized, available to all rather than just the wealthy. Want to plan a corporate retreat for a thousand employees? Your digital assistant—equipped with deep understanding—will do most of the work, from figuring out what needs to be ordered to who needs to be called and reminded, encompassing what Google Duplex hopes to do (dialing and interacting with people on the other end) but not just for a haircut or a dinner reservation with a preordained script, but for a vast, customized operation that might involve dozens of staff and subcontractors, from chefs to videographers. Your digital assistant will be both liaison and project manager, managing parts of a hundred people’s calendars, and not just your own.

Computers will also become vastly easier to use. No more poring through help menus and memorizing keyboard shortcuts. If you suddenly want to italicize all foreign words, you could just ask, instead of going through the whole document yourself, word by word. Want to copy forty different recipes from forty different web pages, automatically converting all imperial measurements into metric, and scaling everything for four portions? Instead of hunting for an app that has just the right set of features, all you will have to do is ask, in English, or whatever language you prefer. Essentially anything tedious that we currently do with our computers might be done automatically, and with much less fuss. Web pages full of “Chrome annoyances” and “PowerPoint annoyances” that detail minor but important contingencies that corporate software developers forgot to anticipate, much to the aggravation of their users, will disappear. In time, that new freedom will seem as life-changing as web search, and probably more so.

The Star Trek holodeck will become a reality, too. Want to fly over Kilauea while it’s erupting? Or to accompany Frodo to Mount Doom? Just ask. The rich virtual reality worlds imagined in the book and film Ready Player One will become something anyone can try. We already know how to make graphics that good; AI powered with deep understanding will make rich, intricate human-like characters possible, too. (And for that matter, psychologically rich aliens, too, who make plausible choices given bodies and minds very different from our own.)

Meanwhile, domestic robots will finally become practical—and trustworthy—enough to work in our homes, cooking, cleaning, tidying, buying groceries, and even changing lightbulbs and washing windows. Deep understanding might also be just the ticket for making driverless cars genuinely safe, too.

Over time, the techniques that allowed machines to have an ordinary person’s understanding of the world could be extended, to match the even richer understanding of human experts, extending beyond basic common sense and into the sort of technical expertise that a scientist or doctor possesses.

When this understanding is achieved, perhaps decades from now, after a great deal of hard work, machines will be able to start doing expert-level medical diagnosis, digest legal cases and documents, teach complex subjects, and so forth and so on. Of course, political problems will remain—hospitals will still have to be persuaded that the economics of better AI makes sense, and better sources of energy, even if they are invented by machines, still have to be adopted. But once AI is good enough, many technical challenges will be surmountable for the first time.

Computer programming might finally be automated too, and the power of any one individual to do something new, like build a business or invent an art form, will be vastly greater than it is now. The construction industry will change too, as robots start to be able to do the skilled work of carpenters and electricians; the time required to build a new house will be reduced, and the cost will decrease, too. Nearly anything that is dirty and dangerous, even if it requires expertise, will become automated. Rescue robots and robotic firefighters will be widespread, delivered from the factory with skills ranging from CPR to underwater rescue.

Artists, musicians, and hobbyists of all stripes will be empowered by AI-powered assistants that can radically enhance the scope of any project. Want to practice your Beatles tunes with a robot band, or conduct a live symphony orchestra, consisting of robots that play in perfect time? Play doubles tennis with your spouse against licensed replicas of the Williams sisters? No problem. Build a life-size castle made out of Legos, with robots that joust? A formation of flying drones that replicates Stonehenge for your next visit to Burning Man? AI will be there to assist with every calculation you require; robots will do most of the labor. Individuals in virtually every field will be able to do things they never could have imagined; each one will be able to serve as the creative director for a whole team of robot helpers. (People will also have more free time, with AI and robots doing much of the tedious work of everyday life.)

Here again, of course, things may not proceed in lockstep. We may well achieve expert-level deep understanding in some domains before others; particular areas of quantitative science, say, might see advances, even as AI still struggles to reach child-level abilities in others. The perfect musical helper might well come years before the perfect AI paralegal.

The ultimate, of course, is a machine that can teach itself to be an expert, in any domain. That too will come, in time.

Eventually, the pace of scientific discovery itself might vastly accelerate, once we combine the sheer computational power of machines with software that can match the flexibility and powerful intuitions of human experts.

At this point, a single advanced computer might replicate what a team of carefully trained humans can do, or even do things that humans simply can’t, because we can’t, for example, track in our heads the interrelations of thousands of molecules, particularly not with the kind of mathematical precision that comes naturally to machines. With such advanced forms of AI, it might finally be possible to use sophisticated causal inferencing in conjunction with vast amounts of neural data to figure out how the brain works (far too little is known now) and how to make drugs that cure mental disorders (where hardly any progress has been made in three decades). It’s not unreasonable to think that AI with genuine scientific prowess could also help devise more efficient technologies for agriculture and clean energy. None of this will come soon or easily, because AI is so hard, but it will come eventually.

Which is not to say that all will be good. On the plus side, if Peter Diamandis is right, with all that automation will come abundance, and the price of many things, ranging from groceries to electricity, may come down. In the best case, humanity might well attain the vision of Oscar Wilde: “amusing itself, or enjoying cultivated leisure…making beautiful things, or reading beautiful things, or simply contemplating the world with admiration and delight, [with] machinery…doing all the necessary and unpleasant work.”

Realistically, though, employment is likely to become scarcer, and debates about guaranteed basic income and income redistribution are likely to become even more pressing than they are now. Even if the economic issues can be resolved, many people may have to change how they derive a sense of self-worth, from a model that is based in large part on work, to a model that finds fulfillment through personal projects like art and creative writing, once a vast amount of nonskilled labor is automated. Of course there will be new jobs (like robot repair, which at least initially may be difficult to automate), but it seems unwise to assume that the new professions will fully take the place of older ones. The fabric of society may very well change, with increased leisure time, lower prices, and decreased drudgery but also fewer employment opportunities, and possibly greater income inequality. It’s fair to expect that a cognitive revolution in AI will spark as many changes in society as the Industrial Revolution did, some positive, some not, many quite dramatic. Solving AI won’t be a panacea, but we do think it will generally be positive, with all that it can do for science, medicine, the environment, and technology, provided that we are smart and cautious about the way we pursue it.

Does this mean our descendants will live in a world of abundance in which machines do almost all the hard work, and we live lives of cultivated leisure, à la Wilde and Diamandis? Or a world in which we upload copies of ourselves to the cloud, as Ray Kurzweil has suggested? Or, bolstered by yet-to-be-imagined advances in medicine, achieve genuine immortality in Woody Allen’s more old-fashioned way, which is to say by not dying? Or by merging our brains with silicon? Nerd Rapture may be near, it may be far, or it may never come. We have no idea.

When Thales investigated electricity in 600 BC, he knew he was onto something, but it would have been impossible to anticipate exactly what; we doubt he dreamt of how electricity would lead to social networking, smart watches, or Wikipedia. It would be arrogant to suppose that we could forecast where AI will be, or the impact it will have in a thousand or even five hundred years.

What we do know is that AI is on its way—and we are best off trying to make sure that whatever comes next is safe, trustworthy, and reliable, directed as much as possible toward helping humanity.

And the best way to make progress toward that goal is to move beyond big data and deep learning alone, and toward a robust new form of AI—carefully engineered, and equipped from the factory with values, common sense, and a deep understanding of the world.