Friday, 11 March 2016
Alphabet, Deep Minds, and the Story of Go
Alphabet, which you probably haven't heard of, is now the highest valued company in the world, last month eclipsing even Apple in its ability to generate billionaires through the magic of owning sufficient numbers of its shares. Its common stock is trading near the $750 mark. Why? It earned about $16 billion in profits in the last quarter because it is the proud owner of a company that you have certainly heard of --Google.
More than a year ago, Google reorganized itself. The result was Alphabet, a holding company sitting on top of Google and its various subsidiaries. Google had been buying startups and medium-size companies at a furious pace and shareholders were getting confused as to who was in charge of what. The value of Alphabet by market capitalization was reported in February, 2016 to be $546.8 billion which is bigger than the GDP of many countries. [A reader on a Google plus site objects to this comparison, so another way of putting this is a comparison between Alphabet's profit and a country's tax revenue. For many countries, $16 billion a quarter is huge.] Alphabet, or its subs, are thus able to buy many more of the companies and scientists doing leading edge work in robotics and artificial intelligence. This is yet another demonstration of the theory of the virtuous circle: the more successful a company is, the more opportunities it will get, and the faster it will grow. In the case of Alphabet, this is both a miraculous thing and an extremely dangerous prospect.
There is a plan afoot at Alphabet/Google to corner the markets in artificial general intelligence and autonomous robots. Thanks to Alphabet/Google we are speeding at an ever increasing rate toward what Ray Kurzweil calls the Singularity, the point (in 2029, he predicts) when the intelligence of machines will be far greater than the sum of all human smarts. Kurzweil's hope is that very soon -- before he dies -- various innovations now in pre-production at Alphabet/Google will permit him to upload his mind into an autonomous robot which will carry his Kurzweilness into the distant future though his body resides under the grass. His business model is based on the notion that many others will want to pay for this sort of machined immortality too. But immortality isn't his only story: Kurzweil says we'll soon be able to plug nanobots into our brains to augment our general intelligence and connect directly to an Internet of intelligent machines. Since Kurzweil is in charge of Google's project to back engineer the human brain, his musings will certainly turn into an Alphabet of commodities.
In case you think I'm making this up, just google Ray Kurzweil.
Artificial general intelligence means mimicking with algorithms the flexible ways in which living things adapt, adjust, and learn from interactions with the environment. As my last book SMARTS details, birds do it, bees do it, even brainless slime molds -- alone or in groups -- do it. In fact, everything alive has to teach itself: everything alive has to learn. For millennia, it was a given that only humans display this capacity. Charles Darwin was the first to study how other living things, specifically worms and even brainless plants, learn. For a long time his results were ignored, pushed to the back of Science's desk, but in the past twenty years, Darwin has been shown to have been more than right. Current science demonstrates that intelligence is shaped by particular embodiments (which makes me wonder what kind of intelligence we will get if the mind of Kurzweil is embodied in a robot like Atlas). And yet: in spite of truly fascinating work exploring intelligence in its many, many forms, mimicking human intelligence still drives most of the work on artificial general intelligence. This may be Alan Turing's fault. In a paper published in 1950, he set out a standard for a successful artificial intelligence--something that could fool a human into thinking it's human. Those leading the charge to create machines smarter than us are still running down Turing's road.
DeepMind is a UK based company started by a former junior chess master, Demis Hassabis, who went on to achieve double firsts in computer science at Cambridge, Turing's alma mater. He started a game company and then, years later, earned a doctorate in neuroscience. He had decided he wanted to figure out how human brains work so that machines may be imbued with human-like imagination as well as human-like memory. He started DeepMind with that aim and demonstrated that he could invent algorithms that learn by themselves, learn well enough to master complex games and beat humans playing them. DeepMind was purchased in 2014 for 400 million pounds by Google. Hassabis now leads Google DeepMind.
This week, Google DeepMind announced one more milestone on the road to achieving what Hassabis describes as artificial general intelligence via an algorithm. If you want to see him interviewed, check Nature Magazine's truly awful venture in online science reporting. Hassabis and his colleagues create game playing algorithms which embody the way the human hippocampus remembers and imagines. When IBM's super computer defeated chess grand master Gary Kasparov, chess was described as an almost impossibly difficult game for a machine to master, involving as the calculation of at least 20 possibilities for each move made [thank you dear reader for the correction about the supercomputer too]. But DeepMind pursued a much bigger challenge-- machine mastery of the game of Go.
I first heard of Go many years ago when I was still a student. My husband's colleague, Bonnie Kreps, a leading feminist and documentary filmmaker working at CTV's W5, was married to a high energy physicist at University of Toronto where I studied politics and English. Bonnie's husband Rodney kept trying to get me to study physics no matter how many times I confessed to being hopeless in higher math. One day I gave in: I sat in on one of his seminars. After it was over, we went for coffee in the faculty lounge. There were game boards of Go littered about. I'd never heard of it before, though I did recognize it from a photograph I once saw in a National Geographic magazine. The photo was of two Japanese monks sitting on a stone wall in front of a temple playing a board game. The game looked exotic and oddly beautiful in that photo, with its round, white and black stones massed together on a wooden board. Rodney said Go was much more interesting, much more difficult than chess. When I went home, I asked my husband if he'd ever played it. Not really, he said, but it was the game of choice in a lot of mathematics disciplines. He'd first come across it as a grad student at McGill. One of his economics professors had a Go board set up in his office. A sign beside it said: anyone who touches this board will not graduate.
Go is a simulation of a war, a game of territory, a game requiring future planning, sacrifice in order to advance, group maneuvers, and most of all intuition. For every move in Go there are at least 200 possible variations. The difficulty of calculating the impact of each move grows astoundingly huge in no time flat.
Hassabis's first try at generating an algorithm that could teach itself to play Go was able to beat a human on a smaller version of the standard board (thus reducing the complexity). But this week, his enhanced algorithm beat the world grand master of Go in Seoul, a man named Lee Sedol. (Oops, revise that: the algorithm just won game two! Oh no, revise again, the algorithm just won game three!! And one of the commentators described this win as, in a sense, so leisurely on the part of the algorithm it made him feel physically ill.) The algorithm is not programmed to play a specific sequence of moves. It is programmed to efficiently adapt and learn, which permits it to imagine and predict the results of its own and its opponent's behavior. It is programmed to intuit but most of all, to win as efficiently as possible.
So where does that take us?
Nowhere we should worry about we are told -- constantly -- by the corporate leaders of Alphabet/Google. We should not worry about the loss of jobs to intelligent and autonomous robots in spite of predictions that half of the jobs we now rely upon will be taken over by smart machines in the near future. We should not worry about lawyers, doctors, writers, even artists being replaced by algorithms (did you see the recent stories about the algorithm that can predict a winning Broadway musical?). Why don't we have to worry? As one of the commentators on this site has pointed out: a guaranteed annual income will solve the problem.You may have noticed that our political leaders are beginning to sing this tune. I don't know whether to be grateful that they can see where this revolution is going before the brown stuff hits the wall, or, to throw up my hands in despair because by these musings they make it obvious that they don't intend to try and shape that future in any way.
Eric Schmidt, the executive chairman of Alphabet, who is worth $11 billion, has spoken on this point repeatedly. “I think that this technology will ultimately be one of the greatest forces for good in mankind’s history simply because it makes people smarter,” he told SXSW last year. However, other folks who are not sitting on a mountain of Alphabet shares, such as Stephen Hawking, Bill Gates, and Elon Musk, beg to differ. They're telling us these innovations could lead to the abrupt end of our position as the leading species on earth due to machines who have their own ideas about how things should be. Musk was so shocked by the Google DeepMind Go win that he declared AI had just made a ten year advance in one jump.
So, with regard to danger: who do you believe, Musk or Schmidt?