Game Masters
Vol. 1, No. 3 — It had become common practice. A slightly degrading but ever-so-necessary ritual before the start of any in-person competition. A quick wave of an oval wand across the ears, once up the back of the neck, and then a neon pink beam of light into each eye. Done. No leaks or bugs. From chutes and ladders to blackjack to chess, each participant has to be sure they are playing against another human. Well, against another human mind at least.
Computers have been really good at games for quite a while. Already in 1952, a tic-tac-toe simulator was programmed which could play perfectly against a human opponent. The computer it ran on filled an entire room and required 11 kW of power to operate.
To get an artificial intelligence to “play a game perfectly” is also called “solving” the game. It can play a game at the highest possible human level and will never make a mistake because it already knows every possible game. There are 255,168 possible games in tic-tac-toe. If you go first, you have 131,184 ways to win. If you go second, there are only 77,904 ways to win. A draw occurs in 46,080 additional cases.
Those are already pretty big numbers for such a small game, right? A quarter-million different games from nine positions with two different pieces? It’s surprising. But computers are great at this kind of task. They know every possible game and look up a path to the next best move which leads to a win.
What about Connect Four? It also involves two different pieces but the grid is 6x7, not 3x3. So we’ve increased from nine to 42 positions. Maybe five times more complex? Not quite. Connect Four has 4,531,985,219,092 possible games. It’s nearly 18 million times more complex than tic-tac-toe.
Computers can still easily play a perfect game of Connect Four. Even at 4.5 trillion combinations, it’s well within their reach. And checkers? Checkers has 500 quintillion possible games. Still solvable. If you play against a computer, it will just look up which game it’s in and know the next best move.
But what happens when we cross the line into games so complex that we cannot know all possible outcomes? Chess is one such game. It is estimated to have between 100 quindecillion and 100 quinvigintillion combinations. Those are unimaginably large numbers that flirt with the number of atoms in the known universe. The computer can no longer know beforehand which move is best. It needs to decide independently. It needs to learn to play the game.
In 2010, a little research lab in England named Deepmind showed up to help do just that.
Previous attempts to “teach” a computer a certain skill, such as playing a game, required the programmers to explain all the rules of the game, known strategies, etc. Deepmind built a new system that enabled the computer to learn by doing. Their program would play a game over and over and eventually got better after thousands and thousands of rounds. And here’s the kicker, it wasn’t built for a single specific game. It was a general game playing intelligence.
By 2013, their gaming AI could play a handful of classic Atari games (think Space Invaders, Pong, Breakout) at superhuman levels. In 2020 they announced that it had now surpassed human performance on 57 Atari games. It had learned how to play games in general and done so simply by playing them.
Now the world has seen successful chess computers at this point. Already in the late 90’s the famous matches between Garry Kasparov and IBM’s Deep Blue computer proved that AI could topple even human grandmasters. However, Deepmind set their sights on unconquered territory: the game of Go. They would use their self-learning technology to build an artificial intelligence capable of beating the Kasparovs of Go.
The game of Go is infinitely more complex than chess. Add a couple hundred zeroes to the number of possible moves in chess and you’re getting close to the complexity of Go. It’s wild.
There were several versions and improvements to Deepmind’s technology to accomplish this but in 2016, their AlphaGo system beat world champion Lee Sedol in a series of five games. Lee Sedol is a 9-dan ranked Go player, the highest possible rank in the game. There are less than 200 in the world. This was an incredible outcome. Until this point, the best Go computers in the world could be beaten by amateur Go players. The sudden advancement in computer capability sent shockwaves through the community. It was not expected to happen for another decade. Something truly new had been unlocked.
Self-learning is the natural way to learn. We try something, we see a result, and we learn. We try again. However, once a computer is capable of self-learning, the playing field is tipped very steeply in its favor. It comes down to speed.
A human can get good at chess by playing chess. One estimate to achieve the top rank of Grandmaster in chess places it at an investment of 12,480 hours. So if it were your job, you would play chess for eight hours per day, five days per week, 52 weeks per year for six years to reach that level. Now how long would that take a computer?
The Deepmind computer began with no information about chess at all, only being told if it had won or lost a game. In nine hours it had played 44 million games against itself and learned enough to play at a superhuman level. Time and experience have a different relationship when it’s a silicon brain doing the thinking.
So far these “general” intelligence agents have still been limited to related tasks. But what happens when they become truly generalized? What happens when we have a single AI that can play chess, write a song, file your taxes, be your therapist, and tell you how to cook a mean omelet? That’s not AI, it’s AGI: Artificial General Intelligence. Once we crack AGI, the singularity may be peaking around the corner.
The ability for a computer to learn anything nearly instantaneously walks the thin line between exciting and terrifying. Good thing they’re tied to our desks and power outlets, right? As long as no one loads this stuff onto a robot or a drone we’ll be fine. Wait…what’s that? You did what now? Oh boy…