This is almost 2 years old, but I’ve stumbled upon it just now.
In October 2017 DeepMind published an article explaining how it set up a Go match between two of their AI programs:
- AlphaGo, the famous AI that defeated the human world champion Lee Sedol, and
- AlphaGo Zero, their newest Go AI software.
The difference
The difference between the two AIs was that AlphaGo learned from a huge database of the past Go games (played by humans) and then played many more games on its own in order to train.
On the other hand, AlphaGo Zero started from scratch, without having a database of the past Go games. It had to learn how to play the game on its own (only playing against itself) and master it.
The result?
AlphaGo Zero beat AlphaGo in all 100 matches.
As Garry Kasparov put it in an interview with Wondery: This result showed that the human Go knowledge was not an advantage, but a liability.
