DeepMind, the British based, Google owned artificial intelligence company made headlines this past week by defeating highest “9-dan ranked” GO master Lee Sedol in 4 out of 5 games. Go is a very complicated game to explain and even more so to master; there are more possible board positions than atoms in the universe. AlphaGo, the program by Google Deepmind, uses advanced Monte Carlo tree search (ask a computer scientist) and a neural network to select the best possible move for victory.
Experts in AI have always cited GO as a great challenge for computers, saying that mastering this field could be decades away. The rate of how AlphaGo, conceived in 2014, managed to decipher one of the most ancient, challenging games is absolutely astonishing.
So how did it succeed?
According to DeepMind, the company secret is in the neural network and so called “deep learning”. AlphaGo got good in a process of, simply put, “playing against itself”. It played against older versions of itself over millions of iterations and improved with each game.
A human would require two decades to learn and master GO; AlphaGo managed to achieve such success in 18 months. Our species have come a long way, but are limited by the effects of evolution. Although technology has not yet reached our level of intelligence, AI has only progressed over the last 50 years. Its greatest and most frightening strength is the ability to continuously improve. Until very recently this has been a result of direct human intervention; however, AlphaGo was able to improve mostly by itself. If such a machine managed to improve completely autonomously, this would become very, very dangerous.
Imagine an AI device designed to complete a specific task as efficiently as possible: a side effect of this goal would be to avoid anything that may prevent it from succeeding at its task. Humans suddenly realize that the device has become too powerful, and wish to make changes or deactivate this device; meanwhile the device, focused solely on its objective, will make every effort to avoid these changes and as a result, we have just lost control of the machine.
The above example is taken from an open letter, signed in early 2015 by Stephen Hawking, Elon Musk and many others warning of the dangers of artificial intelligence. Elon Musk has even described AI as “summoning the demon.” Of course, the letter mentions the great potential of artificial intelligence, with the possibility of eradicating disease and poverty, but more importantly it states the critical long term research required to maintain control of these devices. According to the letter and research from Stanford, losing control of these AI systems is a serious issue, and research to avoid these catastrophes must be done in congruence with current AI developments. So while we are busy creating amazing machines like AlphaGo, we must also create a way to disable them or a way for us to intervene.
Leave a Reply