Just stumbled on an article submitted at the Neural Information Processing Systems Conference in Long Beach, CA – US, reporting on AlphaZero capability to learn to play games by being instructed on the rules (not surprising) and then playing with itself to learn the winning strategies (that is more interesting).
The result is that AlphaZero can become so proficient at playing the game(s) within 24 hours of self training that it can defeat the very best players in the field, humans and computers alike.
However, what surprised me is that according to the article’s writers AlphaZero ends up playing differently from a human, and differently from a computer. It looks like an alien form of intelligence is being created, a form that can outsmart both humans and “classical” computers!
AlphaZero is a creation of DeepMind (a British company acquired by Google in 2014, focussed on advanced artificial learning technologies based on neural networks). The article is reporting on AlphaZero learning to play Chess and Shogi after having been instructed on the rules of those games and left alone to develop playing strategies. To develop these strategies AlphaZero plays with itself and discover what works and what doesn’t. In just 24 hours of playing (which means millions and millions of games played) it gets sufficient experience to outsmart top level players and programs based on mimicking human approaches.
Observing AlphaZero playing one (an expert player…) would see the development of a playing strategy that is nothing like the one adopted by a human nor (obviously) one used by a normal program (not surprising since existing programs are attempting to mimicking humans strategy, thinking if you want).
Notice that this is not about being good at number crunching, being better than a chess player because you can examine the outcome of more options, of looking further down the game potential evolution. If it were so the strategy would still look a human one, just a better one.
What we are seeing is the emergence of a different sort of intelligence, it is not about being “more intelligent”, rather about being intelligent in a different way (that proves to be better than our intelligence).
This is tricky indeed, and this is basically what amazed me in reading the paper. AlphaZero by autonomously developing game strategies develops a new form of intelligence.
This is also strengthening the importance of the FDC IEE Symbiotic Autonomous Systems Initiative with an emphases on the symbiotic part. Establishing a “partnership” with machines like AlphaZero is not extending our human intelligence, it is complementing it with a different form of intelligence and this opens up new, unexplored, landscapes. It is not like using an electronic calculator that makes us faster (and more accurate) in doing computation. It is allowing us to partner with an alien intelligence to explore issues we would not be able to explore with the kind of reasoning that we use.
This reminds me of what Einstein once said: ” We cannot solve our problems with the same thinking we used when we created them”. In this ever more complex world we are facing a lot of problems, from climate to geopolitical struggle that we have created on our own. Getting a helping hand from an “alien” intelligence might be worth exploring.