Diary

David Runciman

It’s three weeks before Christmas and Los Angeles is in flames, though you wouldn’t know it from inside the bowels of the Long Beach Convention and Entertainment Centre, where all is cool and grey. I am here with eight thousand other attendees of the Neural Information Processing Systems (Nips) conference – the great annual get-together of people who work in machine learning. These are the men and women (mainly men, but we’ll come to that) who are building the artificial systems that may one day, perhaps quite soon, be able to perform many tasks that have traditionally been thought to require human intelligence. The prevailing mood of the conference is one of remorselessly practical problem-solving, mixed with occasional bursts of euphoria at how far machine learning has come in recent years. Meanwhile, about thirty miles to the north, wildfires have reached the Bel Air district, lapping at the edges of the UCLA campus and the Getty Museum. Smoke is drifting across the 405 highway, which carries 400,000 vehicles a day, the busiest stretch of road in the United States. If the people at Nips get their way, it won’t be long before most of these cars drive themselves through the haze, human cargo safely stowed. For now, though, it’s humans doing the driving and strung out firefighters doing battle with the elements. Plenty could still go wrong, though it will take a while for the news to reach Long Beach.

I’m here to take part in a symposium on different kinds of intelligence, natural and artificial. Yet this is not simply an AI conference. In fact, one of the reasons for the remarkably rapid recent progress in machine learning has been the deliberate detachment of algorithmic problem-solving from hoary questions about what counts as true intelligence. Trying to get machines to mimic how the human mind works – which has long been the holy grail of the AI community – turns out to be a distraction. Machines that are capable of learning do not have to be capable of thinking the way we do. What they require are vast amounts of data, which they can filter at incredible speed searching for patterns on the basis of which they make inferences, and then for further patterns in their own failures to draw the optimal inferences. In other words they learn from their mistakes, without stopping to wonder what it really means to learn or to make a mistake. It is amazing how much can be accomplished by machines that operate like this. It’s as though they were just waiting to be told that they should try doing it their way, not ours.

The prince of this brave new world – and the closest Nips has to a resident rock star – is Demis Hassabis, co-founder and CEO of Google’s DeepMind, best known for having taught an algorithm (AlphaGo) to play the fearsomely demanding game of Go better than any human being has ever managed.[*] Hassabis is here to talk about his latest breakthrough. AlphaZero is a machine learning system designed to teach itself games from scratch, without any human input outside the basic rules. It does this by playing against itself over and over again, working out its own lessons from its own mistakes, and then correcting for them. When AlphaZero was let loose on chess, it took just four hours before it could play well enough to defeat the best rival program, Stockfish, which could already see off any human grandmaster. Algorithms have been better than all human players at chess since 1997, when IBM’s Deep Blue beat Garry Kasparov, then world champion, in a six-game match. But twenty years ago computers had to be trained with a large number of historic games, so as to learn which human tactics tended to work, and which didn’t: it was taught openings and endgames and all the rest. AlphaZero, though, uses an entirely different approach: it isn’t given any information about what humans consider an advantageous position, and just calculates whether any given move will be more likely to make it win or not.When applied to Go – a far more difficult game – the AlphaZero algorithm, as Hassabis put it, acquired in the space of 72 hours the level of knowledge that human civilisation had painstakingly accumulated over three thousand years. Then they left it running. Within weeks it had taken Go to a level that was beyond anything that had been previously imagined. It did the same for chess, effectively turning it into a different kind of game. It made moves that made no sense, until much later they emerged as part of a winning strategy that only AlphaZero could have envisaged. After a month, the DeepMind team switched it off, not because it had stopped making progress, but because it wasn’t clear what else there was to gain by letting it gobble up so much energy. It hadn’t just proved its point. It was now making a point that it alone was able to reckon with.

Hassabis presented these results hot off the press to a packed auditorium at Nips. His sense of wonder at what he had accomplished was infectious. Hassabis is himself an exceptionally strong chess player – at 13 he was the second highest-rated player of his age in the world. He is also a five-time champion at the Mind Sports Olympiad. So when he showed some of the bewildering but ultimately successful moves made by AlphaZero playing chess – withdrawing its queen to a corner of the board during an intense middle game, sacrificing a rook for no apparent gain as though it were a worthless encumbrance – and described them as ‘chess from another dimension’, it was hard to disagree. Because he had built a machine that was free from human input, it didn’t know what it wasn’t allowed to do. No one had ever told it that a rook is worth five pawns and therefore inherently valuable. We have to assume it didn’t even know the meaning of sacrifice, which is why it found making sacrifices so easy. All it saw were threats and opportunities. As Hassabis put it, for AlphaZero ‘everything is contextual.’ And anything else is just prejudice.

The notion of a machine that rapidly surpasses human capabilities, and then surpasses the capabilities of other machines, and then keeps going, is vertiginous. Hearing its maker express his own astonishment at its prowess – Hassabis said he had no idea that it would turn out to be so good, so fast – was a little otherworldly, like a fleeting echo of the moment of divine creation. But the feeling soon passed. Really, is AlphaZero anything more than a toy? Hassabis’s critics, of whom there are plenty, even at Nips, point out that chess is a perfect information game, where everything you need to know is laid out on the board in front of you. Nothing is hidden and nothing is ambiguous. The ability of machines to learn in these circumstances is contingent on the fact that they can’t misread what has already happened; they can only misjudge what may happen next and then try not to make that mistake again. A computer won’t confuse a rook for a knight. But outside the world of perfect information games even the smartest machines regularly misidentify objects, misinterpret language and misunderstand nuance. They struggle with information whose source is unclear and whose character is open to question. As one of the speakers who followed Hassabis pointed out, deep learning machines are still capable of mistaking turtles for rifles (don’t ask me how). So no one should feel confident that this technology is ready to be weaponised. You might want AlphaZero fighting your chess battles for you. You wouldn’t want it fighting your wars.

That said, DeepMind’s summary mission statement is not lacking in ambition:

1. Solve intelligence.

2. Use that to solve everything else.

What solving intelligence means here is instrumentalising it: turning it into a tool that can be used by machines. Hassabis put this slide up at the start of his presentation and then he talked about chess and Go for forty minutes. Only at the end did he come back to the question of the wider social implications of his work, and expressed his hope that this technology would soon be delivering equally impressive results in healthcare, science and energy. He did not provide details. Is AlphaZero really a step on the road to solving intelligence? It all depends on what we mean by intelligence, which is the question that machine learning cannot defer for ever. When he showed us some of the astonishing chess moves made by his machine, Hassabis laughed with pleasure at the hilarity of it, as though it were a kind of joke. His audience laughed too because it was funny to see something so outlandish. But the computer wasn’t joking. We were only laughing at, and with, ourselves. Alpha-Zero may have overcome thousands of years of human civilisation in a few days, but those same thousands of years of civilisation have taught us to register in an instant forms of communication that no machine is close to being able to comprehend. Chess is a problem to be solved, but language is not and this kind of open-ended intelligence isn’t either. Nor is language simply a problem-solving mechanism. It is what enables us to model the world around us; it allows us to decide which problems are the ones worth solving. These are forms of intelligence that machines have yet to master.

Before Hassabis spoke, the child psychologist Alison Gopnik took to the stage to remind the audience that many human beings aren’t so hot at this kind of intelligence either. That included all of the humans in the room, notwithstanding their unusually high IQs. The most creative modellers of new information environments are children. The older we get, the less flexible our intelligence becomes. When we are very young, we are much more likely to be open to new ways of thinking and to countenance unlikely hypotheses. That doesn’t make children any good at getting things done, especially if it involves a lot of planning. Frankly, kids are terrible at concerted action. But they are excellent at imaginative reasoning. In Gopnik’s memorable phrase, children ‘think like theorists’, whereas adults think like decision-makers. Human development turns out to be a trade-off between plasticity and efficiency: as we become more efficient at making decisions, we gradually lose our capacity to absorb the things that don’t fit with the world we know. We become more intelligent with age, and we become less intelligent as well.

In one sense, children are the AlphaZeros of the natural world: they start a lot of their thinking from scratch and try to work things out for themselves. Context trumps prejudice, until we learn better. But in another sense, children are nothing like AlphaZero, because they are doing their learning in an open information environment, packed with human input and ambiguity. They are also learning how to play, which is something else kids are exceptionally good at. Despite being a game, chess is not play. It is a practical business. Some of AlphaZero’s moves may look playful, but only humans would be capable of seeing them in those terms. To the machine, everything is about getting the job done. The danger of conceiving intelligence in these terms is that it makes everything else about getting the job done too. What we gain in efficiency we lose in plasticity.

Undoubtedly, there are plenty of jobs that could be done better than humans manage at present. That is the great promise of machine learning. Healthcare, energy and transportation systems are all riddled with inefficiencies that these machines could help to eradicate. Until I arrived in Long Beach, I had never used Uber before. After a few days I couldn’t imagine living without the service it provides, in a city where I had to walk half a mile just to cross the freeway outside my motel. There was something almost magical about the time, money and peace of mind it saved me. I know Uber has a very mixed reputation and some people told me I should be using Lyft, which treats its drivers better. Maybe I should have thought harder about that. But my Uber experience was so spectacularly efficient that I didn’t think about it at all.

Is Uber intelligent? Ironically, that is what I was at Nips to talk about. Not Uber specifically, but corporations in general were my remit at the symposium. Just as adult human beings are not the only model for natural intelligence – along with children, we heard about the intelligence of plants and animals – computers are not the only model for intelligence of the artificial kind. Corporations are another form of artificial thinking machine, in that they are designed to be capable of taking decisions for themselves. Information goes in and decisions come out that cannot be reduced to the input of individual human beings. The corporation speaks and acts for itself. Many of the fears that people now have about the coming age of intelligent robots are the same ones they have had about corporations for hundreds of years. If these artificial creatures are taking decisions for us, how can we hold them to account for what they do? In the words of the 18th-century jurist Edward Thurlow, ‘corporations have neither bodies to be punished nor souls to be condemned; they may therefore do as they like.’ We have always been fearful of mechanisms that ape the mechanical side of human intelligence without the natural side. We fear that they lack a conscience. They can think for themselves, but they don’t really understand what it is that they are doing.

When corporations misbehave, we look for human beings to blame. If we don’t like Uber, we try to pin it on Travis Kalanick and punish him where we can, which turns out to be harder than it looks (as I write he is poised to sell $1.4 billion of shares in the company). The problem is that Kalanick is not Uber – he goes, but the business goes on. Of course, corporations are not inherently bad, any more than any other kind of machine. But they are not neutral either. They provide many benefits for their human creators, which is why we can’t live without them. Corporations have shown themselves to be remarkable devices for pooling resources, finding efficiencies, driving innovation and increasing value. But there is still something ghostly about corporate identity. We inhabit a world dominated by these entities and yet we have little understanding of how they think. We don’t really know whether they think at all. Three hundred years of worrying about corporate responsibility has produced an enormous patchwork of laws and regulations trying to define it – and a vast army of lawyers and accountants trying to redefine it – but there has been little progress in understanding what goes on inside the black box of the corporate mind. What does Uber want? We could still be asking that question long after the company has taken over the world’s transportation systems.

As well as being a showcase for the latest research, Nips is also a job fair, and all the big tech companies are represented at glossy booths staffed by fresh-faced young recruiters. The most famous names are present, of course, including Google, Facebook, Amazon and Microsoft. So too, at even bigger and glossier booths, are the giant Chinese companies, such as Alibaba, Baidu and Didi (the Chinese Uber). The Chinese have increasingly been driving research in machine learning, particularly as a tool for targeting consumers. There is also a host of newer entrants to the market, with wonderfully indeterminate names: Recursion, Quantum Black, XPRIZE, CrowdFlower, Criteo Research, Voleon Group. Some of these are the offshoots of far more established players: Quantum Black is the machine learning wing of McKinsey, the ubiquitous management consulting firm. All seem to be engaged in an unspoken contest to see who can come up with the most aspirational mission statement. CrowdFlower: ‘The essential human-in-the-loop AI platform.’ Criteo: ‘Let’s connect more shoppers to the things they need and love.’ XPRIZE: ‘An innovation engine. A facilitator of exponential change. A catalyst for the benefit of humanity.’ That one probably wins.

Because many of these enterprises have only been around for a decade or less, they are the children of the corporate world. It’s tempting to think they might have some of that childlike curiosity and imaginative capacity that characterises the early stages of human development. Certainly that’s how they would like to present themselves: all of their iconography is intended to signal that these are places where out-of-the-box thinking is encouraged and human intelligence is given free rein. But it’s a false prospectus, because corporations aren’t really children: they are still just money-making machines. Too often, the story we hear about the coming of AI focuses exclusively on the interaction between super-smart individuals, like Hassabis, and super-smart algorithms, like AlphaZero. Yet behind both lie super-powerful corporate actors, which are neither. One day, perhaps, humans will build the intelligent robots capable of supplanting us. But for now, the biggest threat to our collective survival comes from the corporate machines we first started building hundreds of years ago to make our lives easier and have never really learned how to control. If we do end up manufacturing killer robots, it won’t be because individual humans made it happen. It will be because the corporations for which they worked didn’t know how to stop it.

Something else that made the sponsors’ hall feel a little strange was the gender distribution on display. More than 85 per cent of the attendees at Nips are male. In a crowd of this size, that still leaves plenty of room for diversity, but taken together the impression is one of uniformity: lots of men, dressed in jeans and T-shirts, talking about cool new stuff in a language it was hard for an outsider like me to understand. By contrast, a disproportionate number of the recruiters staffing the corporate booths were women. These companies were clearly keen to put their best face forward, to indicate to would-be employees that they were diverse and progressive places to work. On the basis of what was going on elsewhere at the conference, this was mainly just for show.

Looking for a different perspective, I went in search of an organisation called Women in Machine Learning (WiML) to ask if I could sit in on their event. I assured them I was an academic writing for a respectable publication, but they suspected I was just another greedy hack hoping for lurid tales of the horrors of being a woman in a male-dominated industry. (After I left Nips I saw this notice posted on the conference website, under the heading ‘Statement on Inappropriate Behaviour’: ‘Nips has a responsibility to provide an inclusive and welcoming environment for everyone involved in the fields of AI and machine learning. Unfortunately several events held at – or in conjunction with – the 2017 conference fell short of these standards. We are determined to do better in 2018 and beyond.’ I didn’t notice any of this while I was there. Perhaps I should have been looking harder.) The nice people at WiML welcomed me in anyway. The atmosphere wasn’t so different from the rest of the conference: practical, serious-minded, problem-oriented. The president of WiML reminded me that her organisation did not exist to provide a haven for women within machine learning but a platform for more women to do machine learning. It was a support group and a professional network. Despite the overall balance of numbers at Nips, it seemed to be doing its job well. Most of the presentations at WiML were as technically oriented, and for me as daunting, as the presentations happening outside.

But not all. The opening talk was given by Raia Hadsell, a research scientist on the Deep Learning team at DeepMind. She didn’t discuss her own work, but rather her route into the machine learning business, which had been a circuitous one. Her undergraduate degree was in philosophy and religious studies. At the same time, she had always been a lover of games and puzzles, and after graduating she took a further degree in maths, before retraining as a computer scientist. The title of her PhD thesis was ‘Learning Long-Range Vision for Off-Road Robots’. Now here she was, still open to new ideas. She also talked about her mother, who had worked as an artist for thirty years before learning in her sixties how to code and becoming a successful computer game designer. Hadsell’s message was to stay curious, but also to stay committed. Her talk didn’t have the evangelical fervour of the presentation given by Hassabis, who is now her boss. In some ways it was a dose of solid good sense that could have been delivered at any academic or professional conference. But that made it particularly resonant at this one. It was very human and it was still with me later as I got an Uber back to my motel through the clear Los Angeles night, trusting that someone, somewhere had put out the fires.

[*] Paul Taylor wrote about DeepMind and AlphaGo in the LRB of 11 August 2016.