Will computers ever achieve the holy grail of artificial general intelligence (AGI)—an intelligence that matches or surpasses human abilities across virtually all cognitive tasks? Experts disagree not only on the feasibility but also on the desirability of such an outcome. Optimists envision an era of abundance. Pessimists fear an existential threat.
One case study suggests AGI may be closer than widely believed. In 2017, Google DeepMind’s AlphaZero taught itself more about chess in four hours than humans had managed to uncover in 1,500 years. That’s remarkable in itself, but the truly amazing part is that AlphaZero accomplished this with a level of style and creativity that even the best human players can’t understand, much less emulate.
AlphaZero’s success raises a provocative question: if a computer can teach itself a complex domain like chess in hours, what does that imply about how close we might be to machines that can teach themselves anything?
The implications of AlphaZero’s methods go far beyond board games. For example, in 2024, chemists using a similar self-learning approach won the Nobel Prize for Chemistry. The National Institutes of Health report that their findings promote the development of new vaccines, enhance disease prevention, support personalized medicine, and generally deepen our knowledge of how life works at a molecular level.
From Chaturanga to AlphaZero
Humans struggled to solve the mysteries of chess for 1500 years. Chaturanga (Sanskrit for “four divisions of the military”) originated in India around the 6th Century, then spread to Persia, then to the Islamic world, and finally into Europe by around the 9th century. Creative players including Ruy López (16th century), Wilhelm Steinitz (19th), Bobby Fischer (20th) made contributions to understanding the game.
At its highest levels, chess is an extraordinarily difficult game. The number of possible chess positions is astronomically large due to the combination of 32 pieces, 64 squares, and the complexity of the rules. The number of possible positions is estimated to be larger than the number of atoms in the universe.
Analysts have written thousands of books about the game. The John G. White Chess and Checkers Collection at Cleveland Public Library alone has over 32,000 chess books and over 6,000 bound volumes of chess periodicals. No doubt, many books and periodicals have eluded their grasp.
Despite the complexity of the game, humans were able to develop a high level of mastery. Chess-playing computer programs started out slowly, playing like total novices, but after decades of development, they improved dramatically.
Why AI Developers Experiment with Chess
From Alan Turing to Claude Shannon and John von Neumann, early computer scientists gravitated toward chess. Turing even designed a chess-playing algorithm in 1948, before digital computers existed. Computing each move by hand took even a mathematical genius like Turing half an hour to compute.
Chess served the same role for computer scientists that fruit flies (Drosophila melanogaster) did for geneticists: a simplified system with enough complexity to reveal general principles. Just as breakthroughs in fruit fly research translated into broader biology, insights from computer chess informed AI progress more generally.
The earliest digital chess-playing computers were pathetically weak. They got better—much better. The 1997 victory of IBM’s Deep Blue over World Champion Garry Kasparov was a milestone, but that system depended on brute-force calculation—searching 200 million positions per second—plus a database of human-coded strategies.
Just a few years after Kasparov’s defeat, the Deep Blue successor Stockfish computer would consistently win 100 out of 100 games against the very best human opponents. Stockfish dominated humans, but it could not compete with AlphaZero in their 100-game 2017 match. AlphaZero did not lose a single game.
More striking was its style—sacrifices and strategies that seemed alien. Grandmaster Peter Heine Nielsen, coach of two world chess champions, told the BBC: “I always wondered how it would be if a superior species landed on Earth and showed us how they play chess. I feel now I know.”
How AlphaZero Surpassed Human Comprehension
AlphaZero was a product of London-based DeepMind (now a Google subsidiary). Founder Demis Hassabis—a former chess prodigy—aims big. His company’s mission is nothing less than to “solve intelligence” and then use it “to solve everything.” Hassabis’s team began by letting AI systems master 1970s arcade games like Pong, Space Invaders, and Breakout without being told the rules. The same learning framework then tackled more complex domains.
The AlphaGo project astonished the world by defeating top human champions at Go, a game long thought beyond the reach of machines. Its successor, AlphaZero, applied similar self-learning methods to chess—with extraordinary results.
The key insight was removing humans from the learning loop. As Ethan Mollick explains in his Substack One Useful Thing, traditional engines encoded centuries of chess heuristics: “control the center,” “protect the king,” “passed pawns are valuable.” AlphaZero ignored all of the human wisdom accumulated over centuries. Starting only with the rules, it trained entirely through self-play until it discovered principles on its own. The world’s best human players marvel at its ability to sacrifice material for dynamic advantages, conduct long-term attacks, and create new approaches beyond human understanding.
Deep Blue embodied brute force. AlphaZero embodied efficiency: using deep neural networks, it evaluated far fewer positions but understood them better. Once trained, it uses energy more efficiently, a significant advantage at a time when less efficient AI apps are straining global power grids.
As Garry Kasparov put it, “AlphaZero programs itself, so its style reflects the truth. It’s the embodiment of the cliché, ‘work smarter, not harder.”
The crucial point is this: AlphaZero was not “programmed to play chess well.” It was programmed to learn. Its intelligence emerged not from human experts telling it how to play, but from its own iterative experience.
AlphaZero Beyond the Chessboard
Could AlphaZero’s methods propel progress in AGI? Some evidence suggests yes. Hassabis and collaborators used AlphaZero-style techniques to create AlphaFold2, which solved the 50-year-old challenge of predicting protein structures from amino acid sequences. In 2024, Hassabis shared the Nobel Prize in Chemistry for this breakthrough—hailed by the committee as “the greatest benefit to humankind.”
Despite this stunning example of real-world benefits from the AlphaZero approach, skeptics warn against overgeneralizing. Leading legal tech analyst Jordan Furlong argues: “The false equivalence between closed systems like chess and open systems like organizations is deeply problematic. It assumes because agents trained on outcomes work in chess-like contexts, they’ll scale across complex enterprises. That leap ignores context, intent, and implicit norms.”
Chess differs from real life in one crucial respect: it is a perfect information game. Every piece, every move is visible. Real-world decision-making, from poker to real estate markets, involves hidden information, bluffing, and unpredictable variables. While computers can play poker well enough to often beat good human players, they have not mastered the game in the same way as they have solved the game of chess. Real life is messy.
Relationship Between AlphaZero and Large Language Models
The development of large language models like ChatGPT is another little-understood factor. AlphaZero used “machine learning.” Large language models (LLMs) work in a completely different way, studying vast amounts of data from the Internet and manipulating it.
Each approach has strengths and weaknesses. AlphaZero lacks the ability to interpret natural language or navigate human cultural nuance. It cannot generalize outside the well-defined rules of a game.
LLMs lack the ability to reliably generate new, testable knowledge outside their training data. They often imitate patterns of reasoning rather than discover original strategies.
Is it possible that hybrid models combining self-teaching and large language models could surpass either? Researchers are already exploring this approach.
This possibility amplifies another concern: alignment risks. If machines can rapidly teach themselves, ensuring their goals remain aligned with human values becomes more urgent. AlphaZero was safe because its only goal was winning chess games. A self-learning AGI pursuing misaligned goals could be far less benign.
Navigating An Uncharted Future
The AlphaZero project stands as a watershed in AI history—demonstrating how self-learning systems can surpass centuries of accumulated human knowledge in mere hours. The success of AlphaFold2 proves the approach can extend well beyond games to produce breakthroughs with profound benefits.
The road to AGI — and whether its net effects will be positive or negative — remains uncertain. Whether these technologies usher in utopia or dystopia depends less on algorithms than on how humanity governs them. As Hassabis himself observed: “Powerful technologies, and AI is no different, are neutral in themselves. How we decide to share the gains is also going to determine whether it is good or bad for the world.”
Whether AGI becomes our servant or our overlord depends not just on engineers, but on the laws we write and the values we protect. The next 20 years—not the next 1,500—may determine the fate of intelligence itself.
About the Author: Jerry Lawson’s background spans both IT and chess. He founded New Strategies in Legal Tech LLC and regularly contributes articles on IT topics to LLRX.com. He holds a Class A rating (1800+) from the United States Chess Federation and competed regularly in chess tournaments until 1995, when he realized the truth of Raymond Chandler’s observation that “Chess is as elaborate a waste of human intelligence outside an advertising agency.” He continues to follow chess news, though not with the same level of diligence that he follows the Washington Nationals.