I want to preface this article by admitting that I am not an expert on Artificial Intelligence and I am definitely not a chess expert. But one of the youtube chess analysts that I follow, agamator, has recently analyzed some games between two programs. On one side we have stockfish, a renowned chess game and on the other we have Alpha zero, an all purpose AI based on re-enforced learning with deep neural networks.
Their game was very impressive and a link with good commentary can be found here: https://www.youtube.com/watch?v=lb3_eRNoH_w&t=719s
Some background can be found here:
So, why is this impressive? Well lets talk numbers first. I am rated around 1800+ (unofficial) and in general amateur chess players are rated in the 1000-1400 range. Magnus Carlsen, the best human chess player nowadays, is close to 2800. Stockfish’s rating is a whopping 3400. Yeap! Chess programs have surpassed human players that much. I don’t believe there will ever be a natural born human being that can compete with chess engines in the future.
And then you have Alpha Zero. A “general purpose AI” meaning a program that can learn how to do various things with the same “coding”. Alpha Zero’s game, at least the 10 or so it played vs Stockfish, exhibits some sort of extremely high level human intuition. My impression of those games was really, an amazing human player (beyond belief amazing) versus a chess program. True, stockfish shows elements of intuition and has introduced novelties in the game. But it is always using a calculation of best “position valuations” basically using some metrics to find the best move forward. Theoretically, this approach should lead to victory all the time. Think of an error free method that always chooses the best foreseeable future and incrementally reaches victory.
Alas, Alpha zero shows true “intuition”. It employs amazing sacrifices that pay off way later in the game, including setting up traps, and exhibiting an almost human like ability to “intimidate” the opponent. Combine that with the “clinical” execution of the endgame (where it totally becomes machine like with always the best move and the fastest way to victory) and you have one of the best chess players I have ever seen. My mind is blown really. And the fact that this is “general purpose” AI is even more mind boggling.
Alpha zero is said to have picked up the game in 4 hours and then played against itself improving as it went. A technique known as reinforced learning. All sorts of questions arise from that. For example, the original games it followed were played by humans. Could it be that Alpha zero took their ideas, human intuition etc, and brought them to another level? Is this why we see elements of trickery and intimidation in its game? Is that really intuition on the philosophical level?
You see, as human players we would like to think, that there is something more than position valuation in chess. An ability to cut through incalculable computations to find the right move by human intuition that makes the game truly magical. So far, computers showed us that by maximizing the gains of position valuation they can outperform any human player. But … this is a shinning beacon of hope for what I call “true intelligence”. What is kind of scary though is we have no clue how the choices arise for Alpha Zero ( a thing that I face in my research too, where I can’t really interpret what my code does nor how it does it) since neural networks are notoriously hard to explain using simple techniques. The “black box” approach is useful when it comes to results, but impossible to work with when it comes to explanations.
Finally, I believe that computers will surpass us, and I am actually hoping it will happen sooner than later since I don’t mind being surpassed, as long as progress is maintained and humanity benefits from it. If intelligence 2.0 is computer based, I think it is prudent to pursue it and nurture it, guide it and teach it to be ethical. I want to close this article with a nice book that i recently read by Mag Tegmark called Life 3.0. It is perfect for those of us that worry a bit about the rise of AI but also see the great potential.