Since an AI Defeated the Go World Champion, Humans Also Play Better

My book on artificial intelligence, When Monkeys Learn from Monkeys, argues that, taken under the wing of a higher intelligence, we could increase our own intelligence. Just as Francine Patterson took the female gorilla Koko under her wing, taught her sign language, communicated with her, and opened up a new world for her and for us, we could use a sophisticated AI to act as a lifelong personal assistant and teacher, allowing our intellectual and cognitive potential to unfold.

What was still considered a steep thesis in the book published at the beginning of 2020 now seems to be coming true in the first simple ways. It seems to be possible that humans can learn new procedures and thought models from an AI.

A group of researchers published an analysis titled How Does AI Improve Human Decision-Making? Evidence from the AI-Powered Go Program, in which they explore the question of whether AlphaGo’s victory over Korean Go world champion Lee Sedol in March 2016, with some surprising moves, led to a change in the quality of play of other human Go players.

The study examined 750,990 moves in 25,033 games played by 1,242 professional Go players since the AlphaGo victory. The results are clear. Especially in the early stages of the game, the quality of moves improved, with younger players showing the most improvement, increasing their chances of winning.

Certain moves that were previously used mainly for defensive purposes had been used as offensive moves by AlphaGo. The level of improvement is noteworthy in that Go has been played in this form for more than 2,500 years, and it was reasonable to assume that the game had been exhausted to some extent and that only minor improvements were possible. But the analysis of the games of a Go-playing AI like AlphaGo, and the even stronger playing successor systems of AlphaGo, showed new successful paths and strategies, which an AI could explore and find better and faster than a human.

Finding new ways of doing things and thinking seems to work similarly, as another experiment showed. Researchers at Switzerland’s Spiez Laboratory trained an AI specialized in identifying new drugs to do the exact opposite. It was to locate as many harmful molecules as possible – read: compounds suitable for chemical weapons. The results startled the researchers. In just six hours, the AI had suggested 40,000 new compounds, some of which were more toxic than the most potent nerve agent known today, VX.

While the Swiss researchers intended to show how quickly disarmament and non-proliferation treaties can be rendered absurd and new approaches to banning chemical weapons (or drugs and other pollutants) are needed, the AI algorithms used suggest how they can be used for good.

Appropriately set up machine learning and algorithms used allow much faster to recognize new patterns, follow all possible paths and classify the results. Where usually a human first has to dig into the shallows and unearth and explore the paths step by step, an AI can be sent off with instructions and bring back the result. We then no longer have to laboriously and slowly discover the best approaches and thought models; the AI presents them to us on a silver platter.

If we can apply that across domains other than Go, and create an AI assistant and teacher, then it opens up new possibilities for us that could benefit us all. My book is available here, by the way:

Leave a Reply