The golden rule of AI training, and the difference between a work-saving tool that transforms your business and a productivity and PR disaster, is the data you put in and the systems you use to train your algorithms.
This has become easier thanks to the development of general-purpose AI, which can adapt to a wider range of tasks with far less tuning necessary than was required in years past.
This evolution can be seen through the development of AI systems designed to solve what was believed to be an unsolvable game.
Following Deep Blue’s defeat of Garry Kasparov in 1997 and the milestone achievement this was for AI, one caveat was that chess was a game that could be solved by simply programming every possible move into the game and using complex computing to narrow down moves.
It was very effective, but did not necessarily reflect artificial intelligence as much as machine intelligence. By contrast, Go was a far older and far more complex game that would require an artificial intelligence that worked far closer to that of a human player in order to succeed.
When AI had solved chess, Go computers still only played at an amateur level, because the approach to playing the game was so fundamentally different.
This began to change with the development of AlphaGo at DeepMind, which almost from its inception in 2014, showed how much of an advancement deep learning could be.
An extremely powerful distributed version of AlphaGo running on nearly 2000 CPUs and nearly 300 GPUs defeated Lee Sedol, one of the best Go players alive, four games to one, with the fourth game highlighting a logic gap in the software. He would retire, saying that he could never beat an AI player.
AlphaGo was superseded by AlphaZero and later MuZero, which were not trained to play Go but were designed to learn the game themselves efficiently, something that is highly applicable to modern AI technology.
