**The future of artificial intelligence begins with a game of Space Invaders.**From the start, the enemy aliens are making kills -- three times they destroy the defending laser cannon within seconds. Half an hour in, and the hesitant player starts to feel the game's rhythm, learning when to fire back or hide. Finally, after playing ceaselessly for an entire night, the player is not wasting a single bullet, casually shooting the high-score floating mothership in between demolishing each alien. No one in the world can play a better game at this moment.
This player, it should be mentioned, is not human, but an algorithm on a graphics processing unit programmed by a company called DeepMind. Instructed simply to maximise the score and fed only the data stream of 30,000 pixels per frame, the algorithm -- known as a deep Q-network – is then given a new challenge: an unfamiliar Pong-like game called Breakout, in which it needs to hit a ball through a rainbow-coloured brick wall. "After 30 minutes and 100 games, it's pretty terrible, but it's learning that it should move the bat towards the ball," explains DeepMind's cofounder and chief executive, a 38-year-old artificial-intelligence researcher named Demis Hassabis. "Here it is after an hour, quantitatively better but still not brilliant. But two hours in, it's more or less mastered the game, even when the ball's very fast. After four hours, it came up with an optimal strategy -- to dig a tunnel round the side of the wall, and send the ball round the back in a superhuman accurate way. The designers of the system didn't know that strategy."
In February, Hassabis and colleagues including Volodymyr Mnih, Koray Kavukcuoglu and David Silver published a Nature paper on the work. They showed that their artificial agent had learned to play 49 Atari 2600 video games when given only minimal background information. The deep Q-network had mastered everything from a martial-arts game to boxing and 3D car-racing games, often outscoring a professional (human) games tester. "This is just games, but it could be stockmarket data," Hassabis says. "DeepMind has been combining two promising areas of research -- a deep neural network and a reinforcement-learning algorithm – in a really fundamental way. We're interested in algorithms that can use their learning from one domain and apply that knowledge to a new domain."
DeepMind has not, admittedly, launched any products -- nor found a way to turn its machine gameplay into a revenue stream. Still, such details didn't stop Google buying the London company -- backed by investors such as Elon Musk, Peter Thiel and Li Ka-shing -- last January in its biggest European acquisition. It paid £400 million.
On DeepMind's website, the company's mission is explained simply as to "solve intelligence". As Hassabis describes it, it comes down to a multi-decade Apollo-style project to crack artificial general intelligence (AGI): rather than teach the machine to understand language, or recognise faces, or respond to voice commands, he wants machine learning and systems neuroscience to teach the network to make decisions -- as humans do – in any situation whatsoever. "The dream of AI is to make machines smart," he explains in the new six-storey King's Cross building that houses 150 DeepMind staff. "Most AI today is about preprogramming a machine. Our way is to program them with an ability to learn for themselves. That's much more powerful; that's the way biological systems learn. "We refer to it as an Apollo programme, a Manhattan project, in terms of the quality of the people involved -- getting 100 scientists, here from 40 countries, together to work on something visionary and trying to make as fast progress as possible. We've brought together the world's top computational neuro-scientists as well as machine-learning experts with huge engineering resources, to see how far can we push."
Artificial intelligence tends to get a bad rap in popular culture: as cyborg assassins in Terminator, or operating systems, like Samantha in Her, that lure us into unwitting love. So why do we need a general form of AI at all? "I think we're going to need artificial assistance to make the breakthroughs that society wants," Hassabis says. "Climate, economics, disease -- they're just tremendously complicated interacting systems. It's just hard for humans to analyse all that data and make sense of it. And we might have to confront the possibility that there's a limit to what human experts might understand. AI-assisted science will help the discovery process."
Still, AGI won't be here any time soon. "We're trying to build a single set of generic algorithms, like the human brain," he says. "We're trying to build things with generality in mind. The Nature paper was the first baby steps. You need to process vision; you need long-term memory; you need working memory so you can switch between tasks... Today you can create pretty good bespoke programs to solve specific tasks -- playing chess or driving a car. Our system could learn how to play chess, but it's not going to be better than Deep Blue. You give it all the knowledge it needs -- the moves, the openings, the endgames. But where does the intelligence reside in something like Deep Blue? It's not in theprogram, it's in the minds of the programming team. The program is pretty dumb; it doesn't learn anything."
Hassabis runs DeepMind according to a "20-year roadmap". "General AI isn't the sort of thing you can wake up one morning and say, that would be a cool thing to do a startup on. I'm trying to fuse together a deep understanding of computer science and neuroscience, to understand systems neuroscience -- the algorithms the brain uses, the representations it has for knowledge, its architecture. As opposed to the EU Human Brain Project, which is trying to reverse engineer at the cortical-column level. That's too low-level for us."
So how will his project affect our lives within 20 years? "Science for sure will benefit -- within drug discovery, protein folding, anything where there's a huge amount of exploration," he says. "Of course we'll have self-driving cars -- but that's narrow AI. We'll have things that can start being creative in 20 years. A lot of things that look very complex, when you break them down it becomes clear how the apparatus works. I studied imagination. We did brain scans, found areas of the brain involved, built models. That made me think that most processes can be understood, including creativity." An AI making an entertaining movie? "I'm thinking more on a basic level -- putting disparate things together to make a new hypothesis. A novel or film is many decades away, though with music, a more limited domain, there are already passable projects that hint at what's possible."
In the near term -- say, five years -- he sees DeepMind's work "making our everyday tools more smart and adaptive". "Better search, so it understands your context and intent better -- you could be more ambiguous and it will understand what you're trying to do. Smartphone assistants are pretty limited in how they work as they're programmed. Wouldn't it be great if they could learn to adapt? You'd say, 'I want an amazing trip round Europe, book me all the hotels and restaurants and flights,' and it would know the archaeological sites, would take you via a vineyard if wine's your thing. Or, 'I'm moving to a new city, I've got small kids, what area should I look for with decent Ofsted reports?' That's the sort of tech I'd hope to see embedded in lots of places in five years. Reinforcement learning will be as big as deep learning is now -- the decision-making part of our algorithms, working out the best action to take, which works initially by trial and error. Trading would be big -- anywhere you have sequential decision-making tasks. We were talking to a weather company about the data they're collecting -- trying to detect the pattern for climate modelling. Predictive analytics will be huge -- you'll see this learning adaptable intelligence seeping into different products."
And after? "Algorithms will be as good as radiographers at looking at scans -- some aspects of those tasks will be augmented by AI. Ten years-plus, it's the AI scientist. And maybe there'll be an AI listed among the authors of a Nature paper. That will be pretty cool."
Demis Hassabis was four when he became curious about the chess game his father and uncle were playing. His father, a Greek Cypriot singer-songwriter who once ran a Finchley Central toy shop with Demis's Chinese Singaporean mother, humoured him and let him play; within two weeks the boy was beating the adults. At five he was competing nationally; at six he won the London under-eight championships; at nine he was captaining England's under-11 team -- when England was second in the world to the Soviet Union. "I was an introspective thoughtful child, I guess, always trying to work things out," he recalls. "You can't help asking, how's my brain coming up with these moves? You start thinking about thinking."
Aged eight, he bought his first computer -- a ZX Spectrum -- with £200 prize money from beating his American opponent, Alex Chang, 3:1 in a four-game match. "The amazing thing about computers in those days is you could just start programming them. I'd go with my dad to Foyles, and sit in the computer-programming department to learn how to give myself infinite lives in games. I intuitively understood that this was a magical device which you could unleash your creativity on."
His father began home-educating him while his mother worked in John Lewis. Then, aged 11, at the local comprehensive, Christ's College Finchley, he discovered AI after buying himself a Commodore Amiga to program games. "I wrote AI opponents for Othello, as chess was too complicated for the machine to run, and it beat my younger brother."
By 13, he was the world's second-highest-rated chess player for his age after Hungarian grandmaster Judit Polgár. Within a year he'd decided that computers were more interesting than chess. He completed his GCSEs by 14, took maths A level at 15, and further maths, physics and chemistry at 16, and applied to Cambridge after seeing a Jeff Goldblum movie, The Race for the Double Helix, about the discovery of DNA. "I thought, is this what goes on at Cambridge? You go there and you invent DNA in the pub? Wow."
His school hadn't sent anyone to Oxbridge for years, and Hassabis, 15, was not prepared for the Queens' College interview. "It was really cold. The professor was asking a question about computer science. If I wanted to visit all 30 colleges, he asked, what's the standard path? I said 30 factorial. Then he said, 'So what is that, then?' I was like, what? How am I supposed to calculate 30 factorial? I came out with the answer immediately: 10 to the 25th -- not that far from the actual answer. He was totally taken aback. He'd just wanted me to say, 'It's a very big number.' I didn't tell him how I did it. I was bored in an A-level maths class, and was playing to work out the shortest number that would break my scientific calculator. Turned out to be 60 factorial. I used that to estimate what 30 factorial was."
He won his place to read computer science, but Cambridge wouldn't let him start at 16. He had come second in a games competition in Amiga Power magazine and won a job at Bullfrog Productions, Peter Molyneux's development house, where he spent his year off. Aged 17, he wrote a multi-million-selling game called Theme Park, "cementing my view that AI would be this incredible advance" and making him enough money to finance university. On graduating with a double first, he joined Molyneux's new company, Lionhead Studios, where he was lead AI programmer for Black & White. A year later, he started his own studio, Elixir, which grew to 60 people. "I wanted to create a political simulator, Republic, where you had to overthrow the dictator by any means. We simulated a whole country, a million people. It took five years -- we were too ahead of our time."
Hassabis decided to take a PhD in cognitive neuroscience at University College London, focusing on memory and imagination. "I thought that would be a good thing to study because computers do episodic memory badly. My work was investigating imagination as a process -- how do we visualise the future?" He tested the imagination of amnesiac patients with a damaged hippocampus and found that their descriptions of, say, being on a beach were impoverished, suggesting that the hippocampus incorporated a visualisation engine. His published paper was listed by Science as one of the top ten breakthroughs of 2007. He then studied computational neuroscience at UCL's Gatsby Computational Neuroscience Unit, with stints as a visiting researcher at MIT and Harvard. By now, he also held a record as five-times winner of the Mind Sports Olympiad.
When in 2011 Hassabis was ready to launch DeepMind, he knew he wanted funding from Peter Thiel, Facebook's initial lead investor. Trouble is, he didn't know how to reach Thiel. "It took me a year to work out that I'd need to give a talk at one of the AI conferences he was sponsoring, at which there would be a speaker event where I'd probably have a one-minute chance to pitch him," he recalls. He researched Thiel and found that he too played chess. "So I thought that would be a more interesting 'in' than being the hundredth person pitching to him. I not so subtly weaved into the conversation a question about why chess had survived so successfully. He was intrigued. The reason, I said, was that the knight and the bishop are perfectly balanced, causing all the creative asymmetric tension. He said, 'Why don't you come back the next day and do a proper pitch?'"
Thiel invested; so did Elon Musk, Skype cofounder Jaan Tallinn, and Li Ka-shing's Horizons Ventures. Antoine Blondeau, cofounder of Sentient, an AI startup which has raised $143 million, was asked by Horizons to investigate DeepMind. "They impressed us with their laser focus on solving a complex scientific problem in an agile, pragmatic way," he says. "And we liked the quality of the group." It was Musk who told Larry Page about this company trying to crack AGI. A couple of months later, Alan Eustace, Google's senior VP of knowledge, emailed to suggest a meeting with Page. "It's not the sort of invitation you turn down," Hassabis says. "It took a year [of negotiations]. One reason we chose Google was culturally we were a good fit, but also AI is something Larry passionately cares about."
But why did he sell? "We weren't planning to, but three years in, focused on fundraising, I had only ten per cent of my time for research," he says. "I realised that there's maybe not enough time in one lifetime to both build a Google-sized company and solve AI. Would I be happier looking back on building a multi-billion business or helping solve intelligence? It was an easy choice. And there was something Larry said to me: 'I spent 15 years building Google -- why don't you just come and take advantage of the opportunity we've built here?' I didn't have a good answer to that."
There were questions, even among its investors, over whether yet another London startup was selling out too soon. Hassabis dismisses this notion as "silly". "People should be caring about where is the work being done. We have full control of what we work on. Google is investing in the UK -- that's brilliant for UK science. Who owns the company is neither here nor there. And we're a research company. "I also want to show that this is a more efficient way of doing science. There have been Bell Labs and Xerox PARC and Microsoft Research -- but they've been run more like academic departments. This is hybrid. All the great advances will come when two worlds merge, whether neuroscience and machine learning, or academic thinking and startup thinking, combined within a big company."
It was non-negotiable that DeepMind would stay in London. Partly because it offers a talent advantage: "If you've got a PhD in physics from Cambridge and want to do some world-changing technology, there aren't that many options here -- in Silicon Valley there are thousands. And if you're focusing on a long-term goal, the Valley can be a bubble -- people trying to create the next Snapchat every five minutes. There can be a lot of noise in the system."
At 11, Mustafa Suleyman was buying Irn Bru bars and Refreshers wholesale at 7½p a pack and selling them at Queen Elizabeth's School, Barnet, for 25p. The venture was profitable -- until teachers closed it down. Suleyman won a Young Enterprise award for borrowing hospital wheelchairs to create a guide to London for young disabled people. He'd grown up best friends with Demis Hassabis's younger brother, but was more motivated by social impact than business. "Demis and I had conversations about how to impact the world, and he'd argue that we need to build these grand simulations that one day will model all the complex dynamics of our financial systems and solve our toughest social problems. I'd say we have to engage with the real world today."
Suleyman, the son of a Syrian-born taxi-driver father and English nurse, won a place at Oxford to read philosophy and theology, but dropped out in the second year, and instead helped start the Muslim Youth Helpline. At 22, he went to work for London mayor Ken Livingstone, advising on human rights policy, but found that government wasn't the vehicle to promote radical systemic change. He had always been the "well-spoken interlocutor" at home, helping parse his father's broken English. As DeepMind's 30-year-old co-founder and head of applied AI, he's responsible for integrating the company's technology across Google's products -- and ensuring clear communication among the top engineers.
So how is Google starting to apply DeepMind's research? "I've got five teams, working on YouTube, search, health, natural-language understanding and some Google X projects," Suleyman says. "We're working at applying the core engine that sits behind the Atari games player across the company. One is in YouTube-recommendation personalisation. We try to learn from the types of videos that lots of users are watching in aggregate to better recommend at the right time, in the right place, what they'd like. Then search: think of search as a process of querying an engine, browsing through links generated, then refining your query in this iterative feedback cycle. Over time we can improve results."
DeepMind is under no pressure to help Google boost advertising revenues. "There was never any expectation we'd do that -- we're a long-term research effort," Suleyman says. "Think of deep learning as one of the first steps on the way to building general-purpose learning systems. If we just search for solutions involving products we can imagine today, we're constraining the limits of our imagination."
Projects include natural language understanding, boosted by DeepMind's recent acquisition of Oxford University spin-off companies Dark Blue Labs and Vision Factory: "We're using neural methods to see if we can create very large machine reading systems without any hand coding," Suleyman says.
He is also focusing on health. "Preventative medicine is the area I'm most excited about. There's huge potential for our methods to improve the way we make sense of data." Suleyman has a final role: overseeing "ethics and safety" at DeepMind.
In November 2014, Elon Musk submitted a (subsequently deleted) comment to Edge.org: "The pace of progress in artificial intelligence (I'm not referring to narrow AI) is incredibly fast. Unless you have direct exposure to groups like DeepMind, you have no idea how fast -- it is growing at a pace close to exponential. The risk of something seriously dangerous happening is in the five-year time frame. Ten years at most. This is not a case of crying wolf about something I don't understand. I am not alone in thinking we should be worried. The leading AI companies have taken great steps to ensure safety. They recognise the danger, but believe that they can shape and control the digital superintelligences and prevent bad ones from escaping into the internet. That remains to be seen..."
Musk had previously issued a warning that AI was "potentially more dangerous than nukes" and of "summoning the demon". He had also explained that his DeepMind investment was not to make money but purely "to keep an eye on what's going on with artificial intelligence".
Last May, Stephen Hawking, too, referenced the movie Transcendence to warn that dismissing highly intelligent machines as mere science fiction "would be a mistake, and potentially our worst mistake in history. One can imagine such technology outsmarting financial markets, out-inventing human researchers, out-manipulating human leaders, and developing weapons we cannot even understand." Apple cofounder Steve Wozniak has also voiced concerns recently.
Jaan Tallinn, the Skype cofounder who was an early DeepMind investor, was instrumental in establishing an independent advisory committee on ethics. "The main reason I backed DeepMind was strategic: I see my role as bridging the AI research and AI safety communities," Tallinn says. "Google and DeepMind agreed to set up an ethics and safety board as part of the acquisition. An important conference earlier this year demonstrated remarkable consensus about long-term issues in AI impacts, culminating in an open letter plus an associated research programme through the Future of Life Institute that Elon Musk backed with a $10 million [£6.6m] donation." Musk, Hawking and DeepMind's founders signed the letter.
Suleyman warns that "getting caught up in the language of existential risk" is mere speculation when DeepMind's focus is on designing systems which are narrow and applied. "'The AI' is a sci-fi-inspired meme which is not helpful," he says. "We're not building AIs that will be equal to us -- they'll help us speak naturally and ask questions directly to a Google search engine. We would certainly be against any kind of autonomous tool of war that could wander round taking kill decisions. That's why we've kickstarted this conversation. Constraining the capacities of our systems is really central."
Hassabis is more direct. "There's a lot of unsubstantiated hype from people who are smart in their own domains but don't work on AI," he says. "Elon tends to shoot from the hip sometimes. I don't think getting hysterical is a good way of stimulating healthy debate -- you end up with unnecessary fearmongering." As for Hawking and Wozniak, "These are people who are not actually building something, so they're talking from philosophical and science-fiction worries, with almost no knowledge about what these capabilities can do."
So there's no risk of agents having moral autonomy? "Of course we can stop it – we're designing these things," Hassabis sighs impatiently. "Obviously we don't want those things to happen. We've ruled out using our technology in any military or intelligence applications. What's the alternative scenario? A moratorium on all AI work? What else can I say except extremely well-intentioned, intelligent, thoughtful people want to create something that could have incredible powers of good for society, and you're telling me there's people who don't work on these things and don't fully understand. I wouldn't purport to lecture Stephen Hawking on black holes -- I've watched Interstellar, but I don't know about black body radiation to the extent that I should be pontificating to the press about it."
Other AI practitioners also see the debate as misjudged. Andrew Ng, who founded Google's first deep-learning team and now runs research at Baidu, sees talk of an AI "superintelligence" as a distraction. "There's a real risk of technology creating unemployment. And there's a big difference between intelligence and sentience. Our machines are becoming more intelligent. But this does not mean that they are becoming sentient. Most AI researchers don't see it as realistic for machines to become sentient any time soon."
Shane Legg was failing at primary school and his teachers wanted to keep him back a year. Aged nine, although he had begun to program his Dick Smith VZ200 computer, he could barely read or spell. His concerned parents in the small New Zealand city of Rotorua sent him to an educational psychologist. "I distinctly remember getting to the end of the intelligence test, and the psychologist was quite annoyed, telling my mother, 'What the hell's going on here? This is a ridiculous waste of time.' "The psychologist took a deep breath, and said, 'He doesn't have a limited intelligence. In fact, I can't measure his intelligence -- it's off the chart. I thought, oh really? Because I was used to being the dumb child."
Legg was diagnosed with dyslexia and trained to use a keyboard; soon he was in his school's top one per cent. "I started developing games that would beat friends at chess, teaching myself at Rotorua public library. In Encyclopedia Britannica, there was an article on alpha-beta search. I thought, I could use that to program chess. I was 12."
Legg studied complexity theory at university in New Zealand, and took a PhD at IDSIA in Switzerland, focusing on how to measure machine intelligence. He then wanted to learn neuroscience, which brought him to the Gatsby Computational Neuroscience Unit at UCL -- where he first met Demis Hassabis. Over lunch they decided it was the right time, in 2011, to start a business. Legg would be chief scientist. "We'd turn up at conferences, say we're a small company trying to solve AI and people laughed at us," Legg recalls. "But one in ten would say, that's really cool. And as we'd get top researchers joining, credibility grew."
Today, behind King's Cross station in rooms named after Alan Turing, Leonardo da Vinci and Nikola Tesla, the DeepMind team quietly computes and prioritises its way along a 20-year roadmap. And where will we be in 20 years? A scenario like that of the movie Her, in which the lead character falls in love with his computer's Siri-like operating system? "That's science fiction," says Legg, 41. "Will we get there one day? I hope so. But most AIs may not be so human-like. Language is quite a sophisticated thing. Think of it as we're trying to build insects. Years from now, you might get to a mouse. Our systems are very good at Space Invaders, they can play Breakout, but they're struggling with Pac-Man. There's a long gap from here to having a system where you can sit and debate philosophy."
Hassabis agrees. "We're decades away from anything that's nearing human-level general intelligence. But we're going to have systems doing useful things in five or ten years. We're on the first rung of the ladder. As to how many rungs there are -- there could be ten or 20 more breakthroughs before we've solved what intelligence is."
David Rowan is editor of WIRED. He wrote about Tony Fadell in 07.14
This article was originally published by WIRED UK
