‘There is nothing that man fears more than the touch of the unknown.’ Elias Canetti, Crowds and Power (1984)
We begin with a survey of where we are now and a sketch of the world to come.
In the next century, politics will be transformed by three developments: increasingly capable systems, increasingly integrated technology, and increasingly quantified society. Together these changes will give rise to a new and different type of collective life: the digital lifeworld. The strange word lifeworld is taken from the German Lebenswelt, meaning all of the immediate experiences, activities, and contacts that make up our individual and collective worlds. When you imagine the digital lifeworld, imagine a dense and teeming system that links human beings, powerful machines, and abundant data in a web of great delicacy and complexity. In this Part, I don’t presume to evaluate or critique the effects of the technologies I describe. The purpose is to identify and understand them, and then (in chapter four) to inspect the intellectual tools that will help us to think clearly about what it all means for politics.
The next three chapters contain many real-life examples, and the idea is not to hold them all in mind at once. Instead our aim is to glimpse, if only in outline, the future that stalks us. That begins with acknowledging that we will soon live alongside computing machines of extraordinary capability. These are increasingly capable systems,1 and they will be the first defining feature of the digital lifeworld.
The field of Artificial Intelligence (AI), which began in earnest in 1943, is concerned with the construction of ‘intelligent’ digital systems. When I refer to AI here, I am describing systems that can perform tasks that in the past were believed to require the cognitive and creative processes of human beings.2 Progress hasn’t always been smooth but today it is impressive and gathering in speed. There are thousands of activities formerly done only by humans that digital systems can now do faster, more efficiently, with greater precision, and at a different order of magnitude.
AI systems are now close to surpassing humans in their ability to translate natural languages, recognize faces, and mimic human speech.3 Self-driving vehicles using AI are widely expected to become commonplace in the next few years (Ford is planning a mass-market model by 2021).4 In 2016 Microsoft unveiled a speech-recognition AI system that can transcribe human conversation with the same number of errors, or fewer, as professional human transcriptionists.5 Researchers at Oxford University developed an AI system capable of lip-reading with 93 per cent accuracy, as against a 60 per cent success rate among professional lip-readers.6 AI systems can already write articles about sports, business, and finance.7 In 2014, the Associated Press began using algorithms to computerize the production of hundreds of formerly handcrafted earnings reports, producing fifteen times as many as before.8 AI systems have directed films and created movie trailers.9 AI ‘chatbots’ (systems that can ‘chat’ to you) will soon be taking orders at restaurants.10 Ominously, engineers have even built an AI system capable of writing entire speeches in support of a specified political party.11 It’s bad enough that politicians frequently sound like soulless robots; now we have soulless robots that sound like politicians.
Every day, algorithms carry out countless transactions in financial markets on behalf of investors. They’re trusted to pursue complex strategies based on changing market conditions. Deep Knowledge Ventures, a Hong Kong-based investor, has appointed an algorithm called VITAL (Validating Investment Tool for Advancing Life Sciences) to its board of directors.12 In medicine, AI systems can now differentiate between lung cancers and predict survival periods better than human pathologists. Researchers believe the same is likely to be true of other cancers in due course.13 In law, an AI system correctly guessed the outcomes of 79 per cent of hundreds of cases heard at the European Court of Human Rights.14 Lethal autonomous weapons systems are under development. That’s missiles, armed drones, and weaponized robots supported by AI. If deployed into the theatre of war, they’d have the capacity to select targets based on certain criteria before homing in and
destroying them—potentially, in due course, without intervening human decision-making.15
Games of skill and strategy are considered a good way to gauge the increasing capability of digital systems. In short, they now beat the finest human players in almost every single one, including backgammon (1979), checkers (1994), and chess, in which IBM’s Deep Blue famously defeated world champion Garry Kasparov (1997). In 2016, to general astonishment, Google DeepMind’s AI system AlphaGo defeated Korean Grandmaster Lee Sedol 4–1 at the ancient game of Go, deploying dazzling and innovative tactics in a game exponentially more complex than chess. ‘I . . . was able to get one single win,’ said Lee Sedol rather poignantly; ‘I wouldn’t exchange it for anything in the world.’16
A year later, a version of AlphaGo called AlphaGo Master thrashed Ke Jie, the world’s finest human player, in a 3–0 clean sweep.17
A radically more powerful version now exists, called AlphaGo Zero. AlphaGo Zero beat AlphaGo Master 100 times in a row.18
As long ago as 2011, IBM’s Watson vanquished the two all-time greatest human champions at Jeopardy!—a TV game show in which the moderator presents general knowledge ‘answers’ relating to sports, science, pop culture, history, art, literature, and other fields and the contestants are required to provide the ‘questions’. Jeopardy! demands deep and wide-ranging knowledge, the ability to process natural language (including wordplay), retrieve relevant information, and answer using an acceptable form of speech—all before the other contestants do the same.19 The human champions were no match for Watson, whose victory marked a milestone in the development of artificial intelligence. This was a system that could answer questions ‘on any topic under the sun . . . more accurately and quickly than the best human beings’.20 The version of Watson used on Jeopardy! was said to be the size of a bedroom; by the early 2020s it’s expected that the same technology, vastly improved, will sit comfortably in a smartphone-sized computer.21 Today, what IBM calls ‘Watson’ no longer exists in a single physical space but is distributed across a cloud of servers that can be accessed by commercial customers on their computers and smartphones.22 As IBM is keen to stress, different versions of Watson do more than win game shows. In late 2016, one Watson platform discovered five genes linked to amyotrophic lateral sclerosis (ALS), a degenerative disease that can lead to paralysis and death. The system made its discovery after digesting all the published literature on ALS and parsing every gene in the human genome. This took Watson a matter of months; humans would have taken years.23 In early 2017, Fukoku Mutual Life Insurance in Japan sacked thirty-four of its staff and replaced them with Watson’s ‘Explorer’ platform, which will chomp through tens of thousands of medical records and certificates, data on hospital stays, and surgical information to calculate payouts to policyholders.
AI has spawned a multiplicity of sub-fields, each applying different methods to a wide range of problems. There is a spectrum of approach, for instance, between those who seek to recreate the neural engineering of the human brain, just as ‘early designs for flying machines included flapping wings’, and those who employ entirely new techniques tailored for artificial computation.24 Some researchers seek the holy grail of an artificial general intelligence like the human mind, endowed with consciousness, creativity, common sense, and the ability to ‘think’ abstractly across different environments. One way to achieve this goal might be whole-brain emulation, currently being pursued in the Blue Brain project in Switzerland. This involves trying to map, simulate, and replicate the activity of the (more than) 80 billion neurons and tens of trillions of synapses in the human brain, together with the workings of the central nervous system.25 Whole-brain emulation remains a remote prospect but is not thought to be technically impossible.26 As Murray Shanahan argues, our own brains are proof that it’s physically possible to assemble ‘billions of ultra-low-power, nano-scale components into a device capable of human-level intelligence’.27
Most contemporary AI research, however, is not concerned with artificial general intelligence or whole-brain emulation. Rather, it is geared toward creating machines capable of performing specific, often quite narrow, tasks with an extraordinary degree of efficacy. AlphaGo, Deep Blue, and Watson did not possess ‘minds’ like those of a human being. Deep Blue, whose only function was to play chess, used ‘brute number-crunching force’ to process hundreds of millions of positions each second, generating every possible move for up to twenty or so moves.28
It’s tempting to get hung up on the distinction between machines that have a narrow field of cognitive capacity and those able to ‘think’ or solve problems more generally. The latter is a juicier target than the former. But distinctions between ‘narrow’ and ‘broad’, or ‘strong’ and ‘weak’ AI, can obscure the fact that even narrowly focused AI systems will create vast new opportunities and risks worthy of careful attention in their own right. Soon computers will be able to do things that humans can do, even if they don’t do them in the same way—and a lot more besides. And it doesn’t matter that one system may only be able to perform a small number of tasks. At the very least, it looks like the digital lifeworld will play host to a multitude of overlapping AI systems, each engineered to perform specific functions. And for those of us on the receiving end, it may be impossible to distinguish between a system that possesses general intelligence, and one that uses fifty different sub-systems to give the impression of general intelligence. In most important respects, the effect will be the same.
The most important sub-field of AI at present is machine learning. As Pedro Domingos explains in his book The Master Algorithm (2015), the traditional way of getting a computer to do something was ‘to write down an algorithm’—a series of instructions to the computer—‘explaining . . . in painstaking detail’ how to perform that task.29 In contrast with an ordinary algorithm, a machine learning algorithm can discover on its own how to recognize patterns, create models, and perform tasks. It does this by churning through large amounts of data, identifying patterns, and drawing inferences. Machine learning algorithms can learn both knowledge (‘if a thing looks like X it is a Y’) and skills (‘if the road curves left, turn the wheel left’).30 The idea is that after a certain point ‘we don’t have to program computers’; instead ‘they program themselves’.31
Many of the AI systems I have described employ machine learning techniques. Indeed, machine learning algorithms are all around us:32
Amazon’s algorithm, more than any one person, determines what books are read in the world today. The NSA’s algorithms decide whether you’re a potential terrorist. Climate models decide what’s a safe level of carbon dioxide in the atmosphere. Stock-picking models drive the economy more than most of us do.
When the time comes for you to take your first ride in a self-driving car, remember that:33
no engineer wrote an algorithm instructing it, step-by-step, how to get from A to B. No one knows how to program a car to drive, and no one needs to, because a car equipped with a learning algorithm picks it up by observing what the driver does.
Machine learning, to borrow from Domingos, is the automation of automation itself.34 It’s a profound development because it liberates AI systems from the limitations of their human creators. Facebook’s engineers, among others, are working on a machine learning algorithm that can build other machine learning algorithms.35
Machine learning algorithms generally ‘learn’ in one of three ways. In supervised learning, the human programmer sets a series of defined outcomes and provides the machine with feedback about whether it’s meeting them. By contrast, in unsupervised learning, the machine is fed data and left to look for patterns by itself. An unsupervised machine can therefore be used to ‘discover knowledge’, that is, to make connections of which its human programmers were totally unaware.36 In reinforcement learning, the machine is given ‘rewards’ and ‘punishments’ telling it whether what it did was right or wrong. The machine self-improves.
Many of the advances described in this chapter, particularly those involving images, speech, and text, are the result of so-called ‘deep learning’ techniques that use ‘neural networks’ inspired by the structure of animal brains. Google launched one in 2012, integrating 1,000 large computers with more than a billion connections. This computer was presented with 10 million ‘random’ images from YouTube videos. It was not told what to look for, and the images were not labelled. After three days, one unit had learned to identify human faces and another had learned to respond to images of a cat’s face (this is YouTube after all).37 Engineers at Google now use ‘duelling’ neural networks to train each other: one AI system creates realistic images while a second AI system plays the role of critic, trying to work out whether they’re fake or real.38
The rapid increase in the use of deep learning can be seen from the AI systems used in games. The version of Deep Blue that beat Garry Kasparov at chess in 1997 was programmed with many general principles of good play. What’s most remarkable about AlphaGo Zero—the latest and most powerful incarnation of the Go-playing AI systems—however, is that it ‘learned’ not by playing against the very best humans or even learning from human play, but by playing against itself over and over again, starting from completely random moves, and rapidly improving over time.39
Machine learning has been around for a while. Its rapid growth and success in the last couple of decades has been enabled in part by the development of new algorithms, but mostly by the explosion in processing power and the growth in available data (chapter three). Data is critical to machine learning. Too little will stunt the growth of a machine learning algorithm, but given enough, ‘a learning program that’s only a few hundred lines long can easily generate a program with millions of lines, and it can do this again and again for different problems.’40 This is why data has been called ‘the new coal’41 and those who gather it are called ‘data miners’.
As we’ll see at various points in this book, however, reliance on flawed real-world data can cause havoc with machine learning systems. Microsoft launched its AI chatbot Tay on Twitter on 23 March 2016. Tay was intended to mimic the speech of a nineteen-year-old girl and to learn from interactions with other Twitter users. Sixteen hours after its launch, Tay was removed from active duty after posting a series of racist and sexually inflammatory tweets, including one which captioned a photo of Adolf Hitler with the tag ‘swag alert’, and another saying ‘Fuck my robot pussy daddy I’m such a naughty robot’. Tay had ‘learned’ to communicate this way from other users on Twitter. This example says as much about humans on social media as it does about machine learning.
A final point about machine learning: it used to be that the computing power that fuelled any particular system was physically present within the system in question. The most powerful digital devices literally contained the processors that made them run. The arrival of cloud computing in the last decade or so has meant that the computing power no longer needs to be located within the device itself: like Apple’s Siri it can be accessed over the internet. This has major implications for the integration of technology, as it means that small devices can draw on big computing resources (chapter two). But it’s also important for machine learning because it means that machines don’t need to ‘learn’ from their own separate experiences; they can learn from others’ too, so that every machine in a swarm or fleet adds to the collective ‘intelligence’ of the whole.
Progress in artificial intelligence and machine learning has been underpinned by advances in mathematics, philosophy, and neuroscience. But as mentioned, it has depended above all on two developments: an explosion in the amount of available data, and an explosion in computing power.
For the last fifty years or so, computing power—the ability of computer chips to process data—has grown at an exponential rate, doubling roughly every two years. This progress is generally expected to continue. On current trends a computer in 2029 will be sixty-four times faster than it was in 2017. If the technology continued to improve at the same rate, then in 2041 it would be 4,096 times faster. After thirty years, the computer would have grown millions of times more powerful. Ray Kurzweil and others predict that within the next decade or so, a normal desktop machine (costing $1,000 or thereabouts) will rival and surpass the processing power of the human brain. By 2050, ‘one thousand dollars of computing will exceed the processing power of all human brains on earth’.42 If this sounds unlikely, look back to where we have come from. Just thirty years ago, it would have needed 5,000 desktop computers to rival the processing power of today’s iPad Air.43 Sixty years ago, 2010’s iPad2 (now hopelessly out of date) would have cost $100 trillion, roughly twenty-five times the United States federal budget for 2015.44 The average smartphone has more processing power than the Apollo Guidance Computer that sent Neil Armstrong to the moon.45
Our brains are not wired to think exponentially. We tend to think of change as happening in a straight upward line, not noticing when the underlying rate of change is itself accelerating. To put it in perspective, try (in Pedro Domingos’ example) to imagine a single E.Coli bacterium of miniscule proportions. This bacterium divides in two and doubles in size roughly every fifteen or twenty minutes. Assuming the right conditions, after a few hours it will mushroom into a colony of bacteria, but still too small for the human eye to see. After twenty-four hours, however, that E.coli microbe will swell into a bacterial mass the size of planet Earth itself.46 That’s the power of exponential growth.
The theory that computer processing power doubles roughly every two years is often referred to as Moore’s Law. This ‘law’, which is not actually a law but rather an observable pattern of development, has been called ‘Silicon Valley’s guiding principle, like all ten commandments wrapped into one’.47 It originated with a 1965 article by Gordon Moore, the co-founder of Intel, who predicted that the number of components that could be crammed onto an integrated circuit would double roughly every two years. At the time, Moore predicted that this trend would continue ‘for at least ten years’.48 Others were sceptical, giving it a couple of years at best. Moore’s Law has now persisted for more than five decades. As Walter Isaacson notes, it became more than a prediction: it was a ‘goal for the industry, which made it partly self-fulfilling’.49 Interestingly, processing power is not the only technology improving at an exponential rate. A host of others, including hard-disk capacity, bandwidth, magnetic data storage, pixel density, microchip density, random access memory, photonic transmission, DNA sequencing, and brain-scan resolution, are all developing in the same way.50 If Moore’s Law continues, the next few decades will witness the arrival of machines of remarkable capability. A world in which every desktop-sized machine has the processing power of all of humanity will be radically different from our own.
It’s sometimes said that Moore’s Law will grind to a halt in the next few years, mostly because it will become physically impossible to cram any more transistors into the same microchip, and because the economic efficiencies enjoyed in the last half-century are set to diminish. There is certainly some evidence of a slowdown, although Moore’s Law has been given its last rites countless times in the past.51 However, it is probably wrong to assume that the current computing paradigm—the integration of transistors onto 2D wafers of silicon (the integrated circuit)—is the final computing paradigm, and cannot itself be improved upon by some other method. History, market forces, and common sense all suggest otherwise. Before the integrated circuit, computers were built using individual transistors. Before that, in the days of Alan Turing, they relied on vacuum tubes, relays, and electromechanics. The story of computing is the story of a succession of increasingly powerful methods of processing information, each developing exponentially, reaching its physical limitations, and then being replaced by something better. Exponential growth in computing processing power stretches back to the seventeenth century and ‘the mechanical devices of Pascal’.52 Nothing is inevitable, but Moore’s law did not begin with the integrated circuit and it is unlikely to end with it either.
The interesting question is what will come next. A host of new methods are already in development, aimed at reaching the frontier of silicon-based computing and pushing beyond it. One approach has been to use non-silicon materials in chips for the first time.53 Another possibility is to move from the current paradigm of ‘2D’ integrated circuits—where transistors are arranged side-by-side on a wafer of silicon—to a ‘3D’ approach where transistors are piled high.54 Another approach might be to abandon silicon altogether in favour of carbon nanotubes as the material for building even smaller, more efficient transistors.55 Yet another approach, currently taken by Google, would be to use more special-purpose computer chips for particular functions—chips that do fewer things but much faster.56 Microsoft increasingly uses a new type of chip that could combine much greater speed with flexibility.57
Looking further ahead, engineers at Google and elsewhere are already hard at work developing ‘quantum computers’ which, in certain tasks, are expected to be able to compute well beyond the capabilities of classical computers.58 Another possible alternative to silicon might be to use 2D graphene-like compounds and ‘spintronic’ materials, which compute by harnessing the spin of electrons rather than moving them around.59 There’s also the growing field of neuroelectronics, which seeks to reverse-engineer the neural networks of the human brain while potentially requiring less power than silicon.60 In the still longer term, Quantum Dot Cellular Automata (‘QDCA’) technology may yield an unimaginably small semiconductor capable of doing the work of a transistor, but using much less power and wasting little energy.61
Many of these technologies are still in their infancy and nothing certain can be said about the future of Moore’s Law. But the least likely outcome is that computer science simply grinds to a halt, with hungry young Silicon Valley engineers hanging up their circuit boards and heading for retirement. Whatever the next computing paradigm turns out to be, it’s certainly not unreasonable to assume that computing power will continue to grow at the same rate as it has since Pascal’s invention of the calculator 400 years ago.