Faster, cheaper hardware

To understand the evolution of computer hardware – a system’s physical parts – join me in a thought experiment. Imagine getting into a VW Beetle and driving down the feeder road on to a motorway.⁹ For the first minute, you begin in the slow lane driving at only 5 miles per hour. Cars and trucks are streaming past you with lights flashing and horns blaring. There would probably be some curious hand signals coming your way too. Not wanting to seem unreasonable, after a minute you double your speed to a majestic 10 miles per hour and make a mental note to keep doubling your speed as each minute passes. In the early part of your strange journey, the rate of acceleration would be slow. By doubling your speed every sixty seconds it would take five minutes to creep past the UK speed limit and be travelling at 80 miles per hour. But, as the fifth minute passes to the sixth, you’d begin travelling at 160 mph. This is the power of what mathematicians call exponential growth. It’s only as it progresses that you see its true potency. If you continued this pattern of doubling your speed, in the 28th minute, unbelievably, you’d be hurtling along at 671 million miles an hour. During that time, your rattling VW Beetle would travel 11 million miles. That’s 275 times around the Earth (traffic permitting).

This slightly terrifying and, I admit, totally unrealistic metaphor describes how microchips (or integrated circuits, as they were once called) have increased in power since their invention in Silicon Valley in 1958. The exponential growth effect is known as Moore’s Law,¹⁰ after the Intel co-founder Gordon Moore who first described it back in 1965. Moore noticed that the number of transistors in an integrated circuit doubles every eighteen months or so. This regular exponential growth has occurred now for over sixty years. The result: the sort of computer speed changes we saw in your imaginary VW now occur every year in real life. What’s truly thought-provoking is Moore’s Law is not stopping any time soon. It will double again in the next two years, and again after that. And again, and again. But now, of course, the speed increases each time by far greater amounts, because we’re much later on in our ‘motorway’ journey.

Some argue that, as we reach the limits of computer chip miniaturisation, Moore’s Law will lose its vigour. However, most eminent computer scientists argue that there are plenty of new technologies that will keep it going, including stacking chips on top of each other, neuromorphic computing, which mimics the human brain, and quantum computing. Instead of using the binary digits of 1s and 0s to describe the world, quantum computing delves into quantum bits (qubits), which can be simultaneously on and off at the same time.¹¹ If you understand how that works, you’re smarter than I am. At this stage I just accept it’s possible.

The bottom line is that Moore’s Law explains the dominance of computing in our lives – and how it underpins AI. It has transformed our world and is likely to continue doing so for the foreseeable future. Its effects are staggering. It’s estimated that if a modern smartphone could have been built in 1958, it would have cost one-and-a-half times today’s global GDP, would have filled a 100-storey building three kilometres long and wide and would have used 30 times the world’s current power-generating capacity.¹² If Moore’s Law continues for another two decades, the total amount of computing power now available to Google will be available on an ordinary desktop computer.¹³ Just think what a super smart teenager might do with that.

Software that learns on its own

Computers work by following a set of rules or instructions called algorithms.¹⁴ An algorithm uses data from the world to build a model that it can then use to make predictions. It tests these predictions against more data to refine its simulation of the world.¹⁵ The reason for AI’s inexorable rise is this: humans are no longer writing all the algorithms. The machines are teaching themselves.

To understand this story, we need to time travel to New York City in 1997. The former world chess champion, Garry Kasparov, has just been unexpectedly beaten by IBM’s AI, Deep Blue. At the time, this was hailed around the world as a major leap forward. But, in terms of modern AI, it was just a baby step. Deep Blue relied on the sort of hardware speed improvements we just explored. It used ‘brute force’ computing: reviewing the chess board and then thinking through all the potential moves. Deep Blue didn’t really innovate the model of computing invented by Turing because it followed human rules. It beat the human champion using instructions written by IBM computer scientists who, in turn, relied on the advice of chess masters. It overcame Kasparov’s vast skill and experience through lightning speed alone. It was capable of examining 200 million moves per second, or 50 billion positions in the three minutes allocated for a single move.¹⁶ Reflecting on his unique position as the fallen standard bearer for the human race, Kasparov wryly noted: ‘Deep Blue was intelligent the way your programmable alarm clock is intelligent. Not that losing to a $10 million alarm clock made me feel any better.’¹⁷

Fast forward 19 years to 2016. Now we’re in Seoul, South Korea, watching another titanic battle between human and machine. This time Korean world champion and grandmaster Lee Sedol is taking on a ‘British’ opponent. The AI – called AlphaGo – was created by a team of techies from the Oxford University spin-off, DeepMind. The field of battle is a 3,000-year-old board game called ‘Go’. In Go, one player plays with small white pebbles and the other with black pebbles, on a square wooden board. The goal is deceptively straightforward: surround your opponent, cut them off to win territory and the game is yours. It looks simple, but the number of possible move outcomes dwarfs even that of chess – more than the number of atoms in the known universe, according to the DeepMind CEO and founder Demis Hassabis. Go was seen as the pinnacle of computer game playing because of its massive complexity and the perceived importance of intuition. This meant it was not possible for humans to programme AlphaGo with every possible scenario as they had with chess. Instead, AlphaGo used machine learning,¹⁸ in which the computer’s algorithms learn without needing to be explicitly programmed.¹⁹

AlphaGo was merely given a goal: to win the game. From trial and error, the algorithm built the most efficient path towards victory.²⁰ Underlying these processes is what’s called a ‘neural network’ – a type of algorithm that learns from observation. It gets its name from the fact that it processes information a little like a human brain, which is itself made up of billions of interconnected neurons.²¹

Prior to taking on Lee Sedol, AlphaGo warmed up by mastering the video games I played in my childhood in the 1980s: Asteroids, Space Invaders and Atari Breakout. In Breakout, the player tries to smash through a wall of bricks at the top of the screen by hitting a ball upwards with a paddle located at the bottom. Again, AlphaGo was given only the bare minimum of the sensory input (i.e. what you see on the screen) and a simple rule – to maximise the score. It didn’t even know what the controls were designed to do. The rate of learning was staggering. After ten minutes of training, the AI could barely hit the ball back. After two hours, it was playing like an expert. After four hours, it hit upon the best strategy. AlphaGo figured out that if it focused on one small point in the wall and dug a hole to allow the ball to break through to the other side, the ball would then bounce around on its own smashing bricks without the computer having to even move the paddle. This novel approach surpassed the best human players.²² If AlphaGo had been a toddler, you might have been tempted to put a lock on the outside of its bedroom door.

After seeing this demonstration, Google bought DeepMind for $500m. At the time, it had no profits. It didn’t even have revenues. The reason? The developers did not teach AlphaGo how to play video games, but how to learn how to play video games. A profound, and hugely valuable, difference. AlphaGo went on to win the globally televised Go tournament against Lee Sedol four games to one. As with Kasparov, artificial intelligence was emphatically the winner. But this time, the game had shifted significantly. AlphaGo didn’t need human instruction because it learned, and then wrote its own rules. The AI of today is the intellectual offspring of AlphaGo. Its ability to learn is why it is transforming our world. Systems are doing it for themselves.²³

Oceans of data

Metaphorically speaking, if AI is a car,²⁴ the silicon chips are the engine, algorithms are the engine’s control system – and data²⁵ is the fuel.²⁶ To accelerate the type of self-learning described above, you need lots and lots of data. Machine-learning AI uses feedback data from Atari Breakout, or anything else, to learn. The AI will try, fail, then try again millions of times, to work out the best way of achieving a goal. By sucking in vast data sets, AI can spot patterns and create insights humans would never see.

Fortunately for AI, data has never been so plentiful as it is now. Our smartphones are how many of us interact with AI. They are also how we proffer our data so AI can learn. The amount of information we all now create intentionally, and as ‘exhaust’ (by-product) data, is truly awe-inspiring. They say, in our digital society, data is the ‘new oil’. In this case, our devices leave a filmy smear of personal information as they record everything we do, say, or see in the form of emails, tweets, photos, videos and social media posts. But it doesn’t stop there. Just imagine, for a moment, the oily digital footprints you leave behind you every day: credit card payments, CCTV images in shops, offices, trains and buses, smart home devices, location sensors on your car, keystrokes and database entries at work. You are a one-person bobbing data oil derrick. And this allows AI to understand you so well it can predict your next move.

Other rivers feed into the data ocean, this time flowing from objects. Nearly everything you can see around you, and much you can’t see, is now connected, or soon will be. It’s dubbed the Internet of Things (IoT). This is the phenomenon that every conceivable item on earth – trains, planes, automobiles, washing machines, wind turbines, buildings, air conditioning units, ovens, buoys out in the ocean, clothes, underwear, shoes – will continually drip into the data sea. All products are now services thanks to the data trail, which reveals how they get used, or could be used better. All this is fed back, logged, analysed and interpreted by the unblinking eye of our machines.

The exponential curves of different technologies are combining and making each other steeper. As well as silicon chips, a number of other technological developments have been observed to be growing exponentially. Memory capacity, for one. In 1980, IBM created a cutting-edge hard drive that could store a princely 2.5 gigabytes. I have a Seagate external drive on my office desk to back-up my laptop that is about the size of a pack of cards. It can hold four terabytes of data. To put that into perspective, it stores roughly 1,600 times more data than IBM’s supercomputer, which was the size of a refrigerator and weighed 250 kg. The IBM hard drive cost around £200,000 in today’s money; my little Seagate drive set me back £82.99 from Amazon.²⁷ Not surprisingly, sensors, LEDs and the number of pixels in digital cameras are also all growing at an exponential rate. This means the cost of gathering richer information is falling as more is collected. And so it continues to snowball.²⁸

This means the world is now capturing and recording more data than ever before. It’s been estimated that at the start of the twentieth century the sum of human knowledge was doubling every century, and that by the end of the Second World War this was taking place every twenty-five years. Now it takes months. IBM has estimated that as IoT becomes a reality, this doubling may be down to days and even hours.²⁹ It means around 90 per cent of the data on the Internet has been created since 2016.³⁰ We’re back on that motorway in the hurtling VW. But this time the foot on the accelerator is information, not hardware.

Head for the high ground

Think of these human superpowers – dexterity, social understanding, emotional skill, deriving meaning, common sense, creativity, critical thinking, humour, human contact and collaboration – as snow-capped mountain peaks. Elaborating on his paradox, Hans Moravec describes a flood, with the valleys below the peaks being submerged by AI. The deepest canyons of this skills geography contain capabilities such as ‘rote memorisation’. We are all aware how the use of smartphones means we no longer need to remember numbers, directions and addresses that we used to habitually keep in our head. The story of NASA’s doomed human computers illustrates the submersion of another valley: ‘arithmetic’. One of the smaller foothills we can still see through the translucent waters might be labelled ‘chess playing’. Fifty years ago, the waters submerged most filing jobs such as record clerks. Now the flood has reached the ridges where tax return preparers, office administrators, personal assistants and taxi drivers perch. What ledge do you currently inhabit?

Wide, not narrow

At the time of writing, investigators are looking into the demise of a Boeing 737 MAX 8 which crashed just after take-off near Addis Ababa in Ethiopia. It’s believed the disaster may have been caused by a confused AI.³⁵ The plane’s computer might have been fed faulty sensor data that showed the aircraft was level, when in fact it was plummeting to the ground. The pilots could not override the AI, and all 157 people onboard were killed. The plane hit the ground so fast the engines were buried in a 10-metre-deep crater. The tragedy is it was obvious the plane was heading for the ground. It wasn’t just the pilots that could see this. Any child looking out of the window could have diagnosed the situation. But the AI was only concerned with its narrow understanding of the data it received.

ANI is efficient, but it has zero common sense. Humans are far better at seeing the big picture. We intuitively ‘get’ situational context. We can also skilfully integrate different stages in a process, and different domains of knowledge. Put simply, we do wide, computers do narrow. Max Tegmark explains it this way: ‘We humans win hands-down on breadth, while machines outperform us in a small but growing number of narrow domains.’³⁶

AI has a long way to go to recreate our human superpowers. It lacks our ability to think creatively about ‘everything’ and to link it together. Instead, it focuses only on the problem we’ve asked it to crack. AI is very good at responding to specific questions, and providing options and solutions; good at gathering data and finding hidden patterns within the data. It works when we ask it to oversee predictable processes. In these areas it delivers exponentially faster, cheaper consistency – and often quality too.

The goal of each AI system is totally focused on a single, very specific goal. IBM’s Deep Blue beat Kasparov in the narrow domain of chess. AlphaGo did the same against Lee Sedol in the similarly cramped field of Go. Neither AI could have then gone on to make a cup of tea, empathise with their fallen opponent or offer support and sympathy. Nor could they then tactfully change the subject and discuss a treasured memory, or an alternative line of work. Neither machine even knew it had won the game. AI is currently a one-trick pony. You, on the other hand, are a general intelligence genius with a cleverness that’s remarkably adaptable, broad and therefore creative.