As long as there have been computers, there have been programmers with dreams of building an app to play chess. Indeed, Chess playing programs have been one of the major testbeds for both programming skills and AI development since the 1950s. And they’ve received various levels of amusement and condemnation from the global chess-playing community along the way. During that time, something of a religious debate arose: 

On the one side were the scientists and programmers, who saw chess as a purely logical problem that, due to its inherent limited possibilities (64 squares, 32 pieces) could be “solved” by brute force. In other words, for every single possible board position, a best move could be determined. But a simple analysis finds that that path leads to madness: indeed, even with such a limited playspace, the combinatorial possibilities are unfathomable and near infinite. And so, the computers, instead of “solving” the position, had to be trained in how to “think” like a grandmaster: to plan strategies, to analyse trade-offs, to attempt to determine what the opponent’s plan was, and to counter it. This did lead to some interesting developments and curiosities. However, no chess program was ever taken seriously at the world-class competitive level. At the top of the field, human grandmasters beat chess programs, every time.

And so, on the other side, were the Humanists. Those who said that, at its highest levels, Chess was a game of genius, creativity, and intuition… qualities that, no matter what happened, could never be modelled by mere machines, functioning on logic. These were human qualities… these were the essence of the indomitable human spirit.

But everything changed in 1997. When IBM launched something of a moonshot, and determined to pour in as many resources (monetary, compute power, and manpower) as necessary to, once and for all, build a computer to play the game of chess, and to beat the best human players in the world.

And we might ask ourselves:

What is the actual Game that we are playing?!?

A Brief Timeline of AI Chess Engines

The year was 1992…

• In 1992, IBM launches their “moonshot” AI Chess Project
  • codename: “Deep Blue”
  • (technical note: in 1988, IBM launched “Deep Thought,” which did reasonably well in grandmaster-level competition, until its weaknesses were assessed and then relentlessly exploited)
  • Deep Blue given an ~$200 million budget
  • assigns ~100+ engineers to work for 10+ years
  • IBM retains the services of ~20 international chess grandmasters — expert trainers to identify weaknesses in the program and to help articulate and codify deep strategies, tactics and gambits
  • The effort ended up building one of the world’s fastest custom built supercomputers of its day: Deep Blue
  • RESULT:
  • Feb 1996 :
  • Kasparov(human) wins, 4-2
  • IBM invests heavily in upgrades
  • May 1997 :
  • Deep Blue wins, 3-2-1 (3 W, 2 L, 1 draw)

  • …this victory — a computer beating the reigning World Chess Champion — essentially is the Harbinger for the coming Age of AI — which doesn’t really hit its stride for another 2 decades…

Fast forward 18 years… 2015

• By 2015,
  • Any $500 smartphone running a free chess program could beat *any* human, including Kasparov, and including Deep Blue
  • side note: comparing computing speeds : how your smartphone represents the pinnacle of human technological achievement, and how you don’t even use it.
• At this time international championship chess was divided into two classes:  human vs. human tournaments, and computer vs. computer tournaments. It no longer made any sense for humans to attempt to seriously challenge computers. They lost 100% of the time.
• In fact, international chess grandmasters stopped training the engines, and instead started taking lessons from the computer engines. In less than 2 decades, the script had been flipped.
• …but this was not yet AI.
• These engines were not self-taught.
• These were merely the evolutionary apogee of a long line of meticulously trained and hand-coded expert systems.

Fast forward another 24 months…

2017: The year of the AI Chess Apocalypse.

On December 5, 2017, Google’s AI division, DeepMind, told its general purpose AI, dubbed AlphaZero, to learn the game of chess. The AI was trained on a mesh network of 5,000 TPUs (Tensor Processing Units, specialized chips designed to meet the specific demands of AI and neural nets), along with 64 master TPUs that served as trainers.

The DeepMind AI began without any knowledge whatsoever of the rules or even the representations of the game, nor any access to any existing gameplay / championship databases that were the lifeblood of all chess computing engines up until that time. The sole method of learning was self-play; there was zero interaction with any other human player or machine engine. The only measurement of success was win or lose. The goal was to have the AI train itself to a high level of competency starting from ground zero. To achieve this goal, it played 44 million sequential games, taking a total of 9 human hours to do so. 

After that brief training, DeepMind went up against then world champion Stockfish 8 in a 100-game tournament (the standard contest in computer vs. computer matches). It is profound to note that while Stockfish at the time was a “brute-force” engine that evaluated over 70 million positions per second, AlphaZero is an “intuitive” engine that runs on a simple machine with just 4 TPUs, and analyzes less then 60,000 positions per second (in other words, AlphaZero uses an analog of “intuitive guidance” to look at only the most promising lines, in the process ignoring 999 out of 1,000 positions that Stockfish is looking at) .

——-

The result? AlphaZero utterly annihilated its opponent, with a final tournament score of 28-0-72 — (28 wins, 0 losses, 72 draws).

This event effectively sounded the death knell, and marked the definitive endpoint of human-designed computer chess engines. The 100% self-taught pure machine AI had vanquished every human-trained engine in existence, decisively. 

Game Over.

This was not a universally accepted truth amongst all the Chess aficionados of the day:

“Some questioned the results because of the disparity of hardware used in the first match. Some also found it unfair that Stockfish was not allowed to use its opening book and its endgame tablebase.

Roughly one year after the first match, DeepMind published a new paper that announced an updated version of AlphaZero had defeated Stockfish in a 1,000-game match. This time, the current version of Stockfish (version 9 at the time) was used, Stockfish was able to use a strong opening book in many of the games, the time controls were adjusted (with Stockfish having large time advantages, up to 30:1), and Stockfish was run on the same type of hardware (an optimized Chess Engine utilizing 44 parallel CPUs) used in the Top Chess Engine Championships (TCEC).

The results didn’t change much—

AlphaZero defeated Stockfish again,
with a final 1000-game scorecard of:

155 wins, 839 draws, 6 losses.”

via Chess.com

 

That was 2017. Five very long years ago.

AI Game Mastery: What Happens Next?

And for many humans, this computational dominance was not a disturbing notion. After all, Chess was, at its essence, a “solvable” game of pure logic: 64 squares, 32 pieces, you move once, your opponent moves next, etc. Theoretically (though not at all practically), every single chessboard position/scenario that could ever possibly occur could be synthesized, analyzed, optimized, and perhaps even “solved.” 

However, it is of special note: No human, nor any computer, has ever attempted such a feat, certainly not beyond the first 12 moves (which in fact *are* computationally accessible, albeit only by high powered supercomputers). These engines, and this AI, do not store a library of positions and
“solutions”. Rather, they *dynamically* assess the board position, and use what can only be seen as “intuition” to find the best move. [Legendary World Champion & International Grandmaster Garry Kasparov wrote a poetic letter in Science magazine essentially noting the same observation]

The great irony is, as the torch was passed from human-trained engines to self-trained AI engines, the actual “thinking” process of the machine, normally embodied in millions of lines of “code” carefully crafted by humans, became obscured. This is because when the AI builds its own “code”, or more accurately a multi-dimensional neural net comprised of trillions of individual connections, tensions, and flows, it becomes far more like a human biological brain and far less like a piece of linear “code” which is both human-readable and analyzable for both behavior and bugs. In fact, depending on the day and time (the “mood?”), a given neural net, which is constantly evolving, may make two very different choices, given the identical board position. This is a completely “human-like” behavior. All prior engines, given the same board position 1000 times, would make the same move 1000 times. The AI, not so much.

A Timeline of AI Victories in Games against World-Class Human Players

	1997 CHESS DeepBlue v. Kasparov
2016 GO DeepMind v. Sedol
	2019 Poker (with Bluffing) Carnegie-Mellon / Facebook AI
2022 Diplomacy Facebook Cicero AI

 

 

 

 

MUST SEE:
Alpha Go (1h30m documentary)

Now, its 2022.

Today, that same AI strategy that mastered — totally mastered — the “logical” game of chess, and since has won at the totally human games of Poker and Diplomacy, has now been directed at a far hairier, more analog, more “human” challenge: fine art: drawing, painting, photography illustration, and sculpture.

_{NEXT :::} The Second Renaissance

 

• In 1992, IBM launches their “moonshot” AI Chess Project
  • codename: “Deep Blue”
  • (technical note: in 1988, IBM launched “Deep Thought,” which did reasonably well in grandmaster-level competition, until its weaknesses were assessed and then relentlessly exploited)
  • Deep Blue given an ~$200 million budget
  • assigns ~100+ engineers to work for 10+ years
  • IBM retains the services of ~20 international chess grandmasters — expert trainers to identify weaknesses in the program and to help articulate and codify deep strategies, tactics and gambits
  • The effort ended up building one of the world’s fastest custom built supercomputers of its day: Deep Blue
  • RESULT:
  • Feb 1996 :
  • Kasparov(human) wins, 4-2
  • IBM invests heavily in upgrades
  • May 1997 :
  • Deep Blue wins, 3-2-1 (3 W, 2 L, 1 draw)

  • …this victory — a computer beating the reigning World Chess Champion — essentially is the Harbinger for the coming Age of AI — which doesn’t really hit its stride for another 2 decades…

Fast forward 18 years… 2015

• By 2015,
  • Any $500 smartphone running a free chess program could beat *any* human, including Kasparov, and including Deep Blue
  • side note: comparing computing speeds : how your smartphone represents the pinnacle of human technological achievement, and how you don’t even use it.
• At this time international championship chess was divided into two classes:  human vs. human tournaments, and computer vs. computer tournaments. It no longer made any sense for humans to attempt to seriously challenge computers. They lost 100% of the time.
• In fact, international chess grandmasters stopped training the engines, and instead started taking lessons from the computer engines. In less than 2 decades, the script had been flipped.
• …but this was not yet AI.
• These engines were not self-taught.
• These were merely the evolutionary apogee of a long line of meticulously trained and hand-coded expert systems.

Fast forward another 24 months…

2017: The year of the AI Chess Apocalypse.

On December 5, 2017, Google’s AI division, DeepMind, told its general purpose AI, dubbed AlphaZero, to learn the game of chess. The AI was trained on a mesh network of 5,000 TPUs (Tensor Processing Units, specialized chips designed to meet the specific demands of AI and neural nets), along with 64 master TPUs that served as trainers.

The DeepMind AI began without any knowledge whatsoever of the rules or even the representations of the game, nor any access to any existing gameplay / championship databases that were the lifeblood of all chess computing engines up until that time. The sole method of learning was self-play; there was zero interaction with any other human player or machine engine. The only measurement of success was win or lose. The goal was to have the AI train itself to a high level of competency starting from ground zero. To achieve this goal, it played 44 million sequential games, taking a total of 9 human hours to do so. 

After that brief training, DeepMind went up against then world champion Stockfish 8 in a 100-game tournament (the standard contest in computer vs. computer matches). It is profound to note that while Stockfish at the time was a “brute-force” engine that evaluated over 70 million positions per second, AlphaZero is an “intuitive” engine that runs on a simple machine with just 4 TPUs, and analyzes less then 60,000 positions per second (in other words, AlphaZero uses an analog of “intuitive guidance” to look at only the most promising lines, in the process ignoring 999 out of 1,000 positions that Stockfish is looking at) .

——-

The result? AlphaZero utterly annihilated its opponent, with a final tournament score of 28-0-72 — (28 wins, 0 losses, 72 draws).

This event effectively sounded the death knell, and marked the definitive endpoint of human-designed computer chess engines. The 100% self-taught pure machine AI had vanquished every human-trained engine in existence, decisively. 

Game Over.

This was not a universally accepted truth amongst all the Chess aficionados of the day:

“Some questioned the results because of the disparity of hardware used in the first match. Some also found it unfair that Stockfish was not allowed to use its opening book and its endgame tablebase.

Roughly one year after the first match, DeepMind published a new paper that announced an updated version of AlphaZero had defeated Stockfish in a 1,000-game match. This time, the current version of Stockfish (version 9 at the time) was used, Stockfish was able to use a strong opening book in many of the games, the time controls were adjusted (with Stockfish having large time advantages, up to 30:1), and Stockfish was run on the same type of hardware (an optimized Chess Engine utilizing 44 parallel CPUs) used in the Top Chess Engine Championships (TCEC).

The results didn’t change much—

AlphaZero defeated Stockfish again,
with a final 1000-game scorecard of:

155 wins, 839 draws, 6 losses.”

via Chess.com

 

That was 2017. Five very long years ago.

AI Game Mastery: What Happens Next?

And for many humans, this computational dominance was not a disturbing notion. After all, Chess was, at its essence, a “solvable” game of pure logic: 64 squares, 32 pieces, you move once, your opponent moves next, etc. Theoretically (though not at all practically), every single chessboard position/scenario that could ever possibly occur could be synthesized, analyzed, optimized, and perhaps even “solved.” 

However, it is of special note: No human, nor any computer, has ever attempted such a feat, certainly not beyond the first 12 moves (which in fact *are* computationally accessible, albeit only by high powered supercomputers). These engines, and this AI, do not store a library of positions and
“solutions”. Rather, they *dynamically* assess the board position, and use what can only be seen as “intuition” to find the best move. [Legendary World Champion & International Grandmaster Garry Kasparov wrote a poetic letter in Science magazine essentially noting the same observation]

The great irony is, as the torch was passed from human-trained engines to self-trained AI engines, the actual “thinking” process of the machine, normally embodied in millions of lines of “code” carefully crafted by humans, became obscured. This is because when the AI builds its own “code”, or more accurately a multi-dimensional neural net comprised of trillions of individual connections, tensions, and flows, it becomes far more like a human biological brain and far less like a piece of linear “code” which is both human-readable and analyzable for both behavior and bugs. In fact, depending on the day and time (the “mood?”), a given neural net, which is constantly evolving, may make two very different choices, given the identical board position. This is a completely “human-like” behavior. All prior engines, given the same board position 1000 times, would make the same move 1000 times. The AI, not so much.

A Timeline of AI Victories in Games against World-Class Human Players

	1997 CHESS DeepBlue v. Kasparov
2016 GO DeepMind v. Sedol
	2019 Poker (with Bluffing) Carnegie-Mellon / Facebook AI
2022 Diplomacy Facebook Cicero AI

 

 

 

 

MUST SEE:
Alpha Go (1h30m documentary)

Now, its 2022.

Today, that same AI strategy that mastered — totally mastered — the “logical” game of chess, and since has won at the totally human games of Poker and Diplomacy, has now been directed at a far hairier, more analog, more “human” challenge: fine art: drawing, painting, photography illustration, and sculpture.

_{NEXT :::} The Second Renaissance