A Google team has published a paper on Nature describing the new generation of AlphaGo. Well, the very name of the new generation is significant: AlphaGo Zero. One would have expected that the new generation would have been named “Old + 1”, yet they decided to start from scratch and name it Zero!
Indeed it is a generation Zero. It is the first system that has been designed to rely completely on itself to learn, rather then observing humans and learning from them. With AlphaGo we were surprised by seeing that it went beyond learning moving into the creation stage, becoming able to do things that have not been tried before. Now in a completely new approach the designers at Google have decided to let AlphaGo Zero to learn by itself.
The results are impressive. After just a few days of learning by competing with itself AlphaGoZero has been able to win against the previous version that competed with a Go master and after a month of self learning it won 100 to 0 against the latest AlphaGo, the one that defeated the Go world champion.
In the paper the team reached the conclusion that it is more effective to use a self learning approach rather than leveraging on human expertise. Humans have made redundant (at least in this field).
Have we reached an inflection point? Has AI become so smart that it can work on itself to become better and better without having to have humans in the loop?
AI is at the core of autonomous systems and so far it has been used to allow them an understanding of their environment so that they can apply the best (or at least a viable) approach to pursue their goal in a changing environment. Now we might have reached a thresholds where autonomous systems may start to design their own goal and take action not just to interact with an environment to play their game but also to act on the environment to change it to better serve their purposes. Are they becoming conscious? Would they make a further step and decide to change their goals? How can we control them if we are “out of the loop”?
These are all difficult questions that we are crafting in the FDC group addressing Symbiotic Autonomous Systems. Symbioses implies that “We” are in the loop, of course, but it is a loosely coupled loop. We might end up in a symbiotic relation where the smarter participant, which unfortunately is no “us”, takes advantage of the other to pursue its own agenda.
Trouble waters ahead….