Stephen Hawkings. Bill Gates. Elon Musk. They are some of the greatest minds of our era, and for years they have been united in calling for regulation of Artificial Intelligence (AI) development, or what Musk calls “our greatest existential threat”.
There is a constant fear that AI will progress to the point where it collides with human civilisation to ruinous results. It sounds like the stuff of science-fiction, but it is a reality we may face in the not too distant future.
While not quite as concerning as some of the predictions made by the men above, a milestone for AI was nevertheless established last week when AlphaGo – a system devised by the team from Google’s DeepMind – defeated Go champion Lee Sedol in a five-game match.
This was important for two reasons:
Firstly, though computer programs like Deep Blue have defeated humans at board games in the past, AlphaGo’s Deep Learning system sets it apart. Deep Blue used a network of computers to determine what was the best move possible. AlphaGo’s system, however, was created to learn the difference between a good move and a bad move. It learnt by playing against itself, and evolving its game through thousands of simulated matches.
Secondly, the game of Go is considered one of the most intricate in existence. So intricate, in fact, that in 1997 experts claimed any system that could beat a champion Go player was at least a century away from being realised. This is due to the staggeringly high of possible ways a single game can be played: approximate 2.082 × 10^170 (2 followed by 170 zeroes).
AlphaGo became so experienced at the game, that at one point it made such an ‘outside the box’ move that Sedol left the room for 15 minutes to regain his composure. Ironic, really.
The result of the match lead to a range of articles speculating on how the system’s victory proves that, before long, all human jobs will be made redundant as workers are replaced by machines. Some of these, such as Scott Santen’s piece for Medium.com, make a good point. At a risk of mimicking the experts who claimed no computer could beat a Go professional, I’d have to say we’re still some time away from that being a reality.
The one round in which Sedol beat AlphaGo hinged on a single move, known as a ‘wedge’. Sedol played the wedge due to the vast amount of moves that could result from it, essentially ‘confusing’ the AI. It proved that when faced with a non-routine situation, AlphaGo could not react instinctively enough to adapt like a human might. Until an AI develops intuition, there are many roles it would be unable to fill, and it is this intuition that prominent critics of AI fear the most.
Another issue is energy use. The human brain uses an average of 20W to operate. AlphaGo, on the other hand, uses an estimated 1MW, or 50,000 times as much as Sedol. Until the power levels are brought closer to parity, a broader comparison of human/AI efficiency can not truly be made.
The line between AI being an aid and a threat is only starting to be drawn, but wherever it lies, few can disagree that this latest advancement in intelligent systems is particularly exciting.
I’ll leave you with a Maclean’s interview with the ‘Godfather of Deep Learning’, Geoffrey Hinton. If you are interested in what the future holds, Hinton knows as well as anyone.
What do you think of AI? Exciting, or terrifying? Let us know in the comments.