Defending Digital Series, No. 12: Nearly seven years into the deep learning era, it’s clear that breakthrough benefits of AI are taking longer to develop than many predicted. This gives society plenty of time to address any risks and downsides, if and when they actually occur.
In March of 2016, the AlphaGo system developed by Google’s DeepMind unit defeated the highly ranked Go player, Lee Sedol, four games to one. Although IBM’s Deep Blue system had defeated the chess grandmaster Gary Kasparov two games to one (with three draws) a decade earlier, Google’s victory was seen as much more significant. This wasn’t because Go is considered an even more complex game than chess; it was because the Google system used previous Go games and deep learning technology to essentially train itself. In contrast, IBM took a traditional expert system approach, relying on functions and processes hard coded by humans.
After decades of disappointment, the artificial intelligence community had found a formula for broad-based innovation. There were three main components: 1) The Internet provided the vast amounts of data needed for self-training, whether it was for playing Go, recognizing images, translating languages, or countless other tasks; 2) the emergence of cloud computing meant that the required processing power was now readily and cheaply available on demand, and 3) the World Wide Web enabled new services to be deployed quickly and globally to businesses and consumers alike. In contrast, previous AI efforts lacked all three. There was insufficient data, expensive computing, and narrowly deployed applications.