According to legend, the medieval philosopher and Franciscan friar Roger Bacon created an all-knowing artificial brain, which he encased in a bronze, human-like head. Bacon, so the story goes, wanted to use the insights gleaned from this “brazen head” to make sure Britain could never be conquered.
Following Bacon, a long-standing challenge for engineers and computer scientists has been to build a silicon-based replica of the brain that could match, and then exceed, human intelligence. This ambition pushes us to imagine what we might do if we succeed in creating the next generation of computer systems that can think, dream and reason for us and with us.
Today there is little talk of brazen heads, but artificial intelligence seems to be everywhere. Magazines and newspaper articles promote it endlessly, raising expectation and fear in roughly equal measure.
Certain forms of AI are indeed becoming ubiquitous. For example, algorithms execute huge volumes of trading on our financial markets, self-driving automobiles are beginning to navigate city streets, and our smartphones are translating from one language to another. These systems are sometimes faster and more perceptive than we humans are. But so far that is only true for the specific tasks for which the systems have been designed. That is something that some AI developers are now eager to change.
Some of today’s AI pioneers want to move on from today’s world of “weak” or “narrow” AI, to create “strong” or “full” AI, or what is often called artificial general intelligence (AGI). In some respects, today’s powerful computing machines already make our look brains look puny. They can store vast amounts of data, process it with exceptional speed and communicate instantaneously with other computers all around the planet. If these devices could be provided with AGI algorithms that work in more flexible ways, the opportunity would be huge.
AGI could, its proponents say, work for us diligently, around the clock, and drawing on all available data, could suggest solutions for many problems that have so far proved intractable. They could perhaps help provide effective preemptive health care, avoid stock market crashes or prevent geopolitical conflict. Google’s DeepMind, a company focused on the development of AGI, has an immodest ambition to “solve intelligence.” “If we’re successful,” their mission statement reads, “we believe this will be one of the most important and widely beneficial scientific advances ever made.”
Since the early days of AI, imagination has outpaced what is possible or even probable. In 1965 an imaginative mathematician called Irving Good, who had been a colleague of Alan Turing in the World War II code-breaking team at Bletchley Park, predicted the eventual creation of an “ultra-intelligent machine … that can far surpass all the intellectual activities of any man, however clever.” He predicted that such a machine would be able to turn its vast intellect to improving itself—each tweak would increase its ability to enhance its own powers, leading to a rapidly accelerating positive feedback loop. “There would then unquestionably be an ‘intelligence explosion,’” Good wrote, “and the intelligence of man would be left far behind.”
Good went on to suggest that “the first ultra-intelligent machine” could be “the last invention that man need ever make.” This led to the idea of the so-called “technological singularity” proposed by Ray Kurzweil, who argues that the arrival of ultraintelligent computers will be a critical turning point in our history, beyond which there will be an eruption of technological and intellectual prowess that will alter every facet of existence. Good added an important qualification to his “last invention” prediction: the idea that we would be able to harvest its benefits “provided that the machine is docile enough to tell us how to keep it under control.”
Fears about the advent of malign, powerful, man-made intelligent machines have been reinforced by many works of fiction—Mary Shelley’s Frankenstein and the Terminator film series, for example. But if AI does eventually prove to be our downfall, it is unlikely to be at the hands of human-shaped forms like these, with recognizably human motivations such as aggression or retribution.
Instead, I agree with Oxford University philosopher Nick Bostrom, who believes that the gravest risks from AGI do not come from a decision to turn against humankind but rather from a dogged pursuit of set objectives at the expense of everything else. Berkeley AI researcher Stuart Russell summarizes what he sees as the core of this problem: “If you, for example, say, ‘I want everything I touch to turn to gold,’ then that’s exactly what you’re going to get and then you’ll regret it.” If computers do become extremely intelligent, there is no reason to expect them to share any capability that people would recognize as justice or compassion.
The promise and peril of true AGI are immense. But all of todays excited discussion about these possibilities presupposes the fact that we will be able to build these systems. And, having spoken to many of the world’s foremost AI researchers, there is good reason to doubt that we will see AGI any time soon, if ever.
According to Russell, “We are several algorithmic breakthroughs away from having anything that you would recognize as general-purpose intelligence.” Tong Zhang, who was, until earlier this year, head of AI research at the Chinese technology firm Tencent, agrees: “If you want general AI, there certainly are a lot of obstacles you need to overcome.” “I just don’t see any practical drivers in the near future for a cross-sectional general superintelligence,” says MIT roboticist Cynthia Breazeal. Mark James of Beyond Limits also doubts that anyone is really on track to develop true AGI, saying that, “for the AI field to truly progress to the point of having a really human-like thinking machine, we need to rethink the problem from square one.”
I think James is right—after all, how can we engineer something that we cannot even define? We’ve never really managed to work out what natural human intelligence is, so it is not clear what engineers are trying to imitate in machines. Rather than intelligence being a single, physical parameter, there are many types of intelligence, including emotional, musical, sporting and mathematical intelligences. Zoubin Ghahramani, Cambridge professor and chief scientist at Uber, agrees: “I actually don’t think there is such a thing as general intelligence,” he told me. And if there is no such thing as a general intelligence, there is no hope of building one, from either synthetic or biological parts. Ghahramani goes further still, arguing that “our view of intelligence is “pre-Copernican.” Just as the Earth is not at the center of our solar system, the human brain does not represent the pinnacle of intelligence.
What all this means is, even if we could emulate the intelligence of the human brain, that might not necessarily be the best powerful route to towards powerful forms of AGI. As leading AI researcher Michael Jordan, from the University of California, Berkeley, has pointed out, civil engineering did not develop by attempting to create artificial bricklayers or carpenters, and chemical engineering did not stem from the creation of an artificial chemist, so why should anyone believe that most progress in the engineering of information should come from attempting to build an artificial brain?
Instead, I think engineers should direct their imaginations towards building computer systems that think in ways that we cannot: that grapple with uncertainty, calculate risk by considering thousands or millions of different variables and integrate vast quantities of poorly structured data from many different sources.
None of this is to take away from the power of increasingly adaptable AI algorithms, or to ignore the risks that they could one day pose through unanticipated side effects or malign applications. But if we have reason to believe that a machine with generalized human-like intelligence is impossible, many concerns about AI evaporate; there is no need to write any rigidly defined moral code or value system into the workings of AI systems. Instead, our aim should be to make them controllable and highly responsive to our needs. Many first-rate researchers and thinkers are devoting a great deal of time and energy to preempting problems associated with AI before they arise.
Russell thinks the key to making AI systems both safer and more powerful is in making their aims inherently unclear or, in computer science terminology, introducing uncertainty into their objectives. As he says, “I think we actually have to rebuild AI from its foundation upwards. The foundation that has been established is that of the rational [human-like] agent in optimization of objectives. That’s just a special case.” Russell and his team are developing algorithms that will actively seek to learn from people about what they want to achieve and which values are important to them. He describes how such a system can provide some protection, “because you can show that a machine that is uncertain about its objective is willing, for example, to be switched off.”
Work like this is important, particularly because Russell and his collaborators are not simply flagging up ill-defined risks, but also proposing concrete solutions and safeguards. This is what Stanford AI professor and former head of Google Cloud Fei-Fei Li meant when she said to me, “It’s not healthy to just preach a kind of dystopia [about AI]. It’s much more responsible to preach a thoughtful message.”
If the only message presented about the great leaps in physics during the early 20th century had been dire warnings of imminent nuclear Armageddon, we would not now have all the amazing discoveries that have stemmed from our understanding of atomic structures and quantum mechanics. The risks associated with AI must be kept in perspective and responded to with constructive action and regulation, rather than hand-wringing and alarmism.
Between the extraordinarily optimistic and the terrifyingly pessimistic, lies a more realistic future for AI. Long before they achieve anything even remotely resembling ultraintelligence, computers will continue to change how we live and think in ways that are both far-reaching and hard to predict. As our computers become smarter, people will also get smarter and more capable. We will need the processing power and increasingly intelligent insights generated by machines to take on our most pressing global challenges—from tackling climate change to curing cancer—and to seek answers to the deepest questions about ourselves and our place in the wider universe.
Attempts by medieval alchemist Roger Bacon notwithstanding, engineers have so far failed in their attempts to emulate the human brain in machine form. It is quite possible that they will never succeed in that ambition. But that failure is irrelevant. Regardless of the fact that today’s advanced AI systems think in distinctly non-human-like ways, they are among the most powerful tools we have built. If we wield them wisely and responsibly, they can help us build a better future for all humanity.