How AI Plays Games: AlphaGo, Chess Engines, and What It Teaches Us

When Deep Blue beat Garry Kasparov in 1997, a lot of people thought AI had arrived. Chess was supposed to be the ultimate test of human intelligence — the “drosophila of AI,” as researchers called it. A computer beat the world champion. What more is there?

Quite a lot, it turned out. Deep Blue was spectacularly narrow. It could play chess at superhuman levels and absolutely nothing else. And the way it played chess — brute-force search through hundreds of millions of positions per second — wasn’t really what most people mean by “thinking.” The story of game-playing AI since then is the story of researchers realizing that winning at games is a much more interesting question than they originally thought.

Key Takeaways

• Chess engines like Deep Blue use brute-force minimax search — they don’t “think” like humans, they evaluate positions at speeds no human can match.
• AlphaGo combined Monte Carlo tree search with deep learning, requiring human game data to train — a fundamentally different approach.
• AlphaZero learned chess, Go, and shogi from scratch using only self-play, with no human data — and became world-class at all three in days.
• MuZero extended AlphaZero’s approach to games where the rules aren’t known in advance — the model learns them.
• Beating humans at narrow games tells us something about optimization and search; it does not imply general intelligence.

The Chess Era: Minimax and Brute Force

The foundation of game-playing AI was laid by Claude Shannon in 1950 — not long after the transistor was invented. Shannon described the basic approach: model the game as a tree of possible moves, evaluate each position with a scoring function, and search the tree to find the move that leads to the best position. This is the minimax algorithm.

The problem with minimax is that game trees are enormous. Chess has roughly 10^120 possible games — more than the number of atoms in the observable universe. You cannot search the entire tree. So chess engines use alpha-beta pruning, which eliminates branches of the tree that can’t possibly affect the final decision. A well-implemented alpha-beta search can evaluate positions roughly 1,000 times more efficiently than naive minimax.

Deep Blue combined alpha-beta pruning with massive parallel hardware (480 custom VLSI chips) and hand-crafted evaluation functions written by grandmasters. It could evaluate 200 million positions per second. Kasparov could evaluate maybe 3 — but each of those 3 was informed by 30 years of pattern recognition that Deep Blue couldn’t replicate. The tension between computational speed and human pattern recognition is the central story of game AI.

After Deep Blue, programs like Stockfish continued improving. Modern Stockfish evaluates positions so accurately that even at reduced depth, it plays at a level no human can approach. At this point, chess computers aren’t interesting as AI research — they’ve solved the problem well past the point of being competitive with any human.

AlphaGo: The Mountain Everyone Said Was Impossible

Go is different from chess in several critical ways. The board is 19×19 (compared to 8×8 for chess). The branching factor — the average number of legal moves available at each turn — is about 250 in Go versus about 35 in chess. This makes the game tree so vast that brute-force search can’t get you very far. And unlike chess, where material counting and position evaluation are relatively tractable, Go positions are notoriously difficult to evaluate. Computers had been trying to crack Go since the 1970s. The professional community expected it would take decades more.

DeepMind’s AlphaGo team published their result in Nature in 2016: the system beat the European Go champion Fan Hui 5–0. Four months later, it beat Lee Sedol, a 9-dan professional widely considered the world’s strongest player, 4–1 in a match watched by 200 million people.

AlphaGo used two innovations together. First, a policy network trained on 30 million moves from human expert games, which suggested promising moves (drastically reducing the effective branching factor). Second, a value network that estimated the probability of winning from any given board position — reducing the depth of search needed. Combined with Monte Carlo tree search, the system didn’t need to evaluate every position exhaustively. It focused computation on the moves most likely to matter.

The crucial limitation: AlphaGo required human game data to train. Its initial knowledge was derived from how humans play. This raised an interesting question: what would happen if you removed the human component entirely?

AlphaZero: Self-Play From Zero

The answer came in 2017. AlphaZero started with nothing but the rules of the game. No human game data. No hand-crafted evaluation functions. Just the rules, and the ability to play itself.

Within 9 hours of training on Chess, AlphaZero surpassed Stockfish — the world’s best conventional chess engine — and beat it convincingly in match play. Within 34 hours on Shogi (Japanese chess). Within 72 hours on Go, surpassing AlphaGo Master. The same algorithm, the same hyperparameters, three different games.

The style of play AlphaZero developed was qualitatively different from conventional chess engines. Human grandmasters reviewing its games noted that it played “like a human from the future” — not because it was more accurate, but because it prioritized positional factors and long-term strategic pressure in ways that felt intuitive rather than computational. David Silver and colleagues published the AlphaZero results in Science in 2018.

What this demonstrated wasn’t that AlphaZero “understood” chess the way a human does. It demonstrated that reinforcement learning from self-play, given a well-specified reward signal (win/lose) and the rules, can develop extraordinarily effective strategies without any human knowledge injection. The implications for AI research extend far beyond games.

MuZero: Learning Without Being Given the Rules

AlphaZero required knowing the rules of the game. MuZero (2019, also DeepMind) removed that constraint. It learns a model of how the environment works — what actions lead to what outcomes — purely from experience, without being told the rules in advance.

MuZero achieved superhuman performance at Go, Chess, Shogi, and 57 Atari video games simultaneously. For the Atari games, it was learning both how to play and what the rules of the game environment were at the same time. This is a significant step toward what researchers call model-based reinforcement learning, where an AI builds an internal model of its environment that it can use for planning.

Game AI vs. General Intelligence

Here’s what the history of game AI teaches us — and what it doesn’t.

Game AI is extraordinarily good at optimization within a well-specified problem. When the rules are fixed, the goal is clear (maximize a score), and the environment is fully observable, modern AI systems achieve superhuman performance. Chess, Go, Atari games, StarCraft II (Alphastar, DeepMind, 2019 Nature) — the pattern holds.

What game AI is not is general. AlphaZero cannot look at a crossword puzzle. It cannot understand a news article about chess. It cannot apply its chess knowledge to a new problem that shares structural similarities. This is why researchers distinguish carefully between narrow AI (systems that perform a specific task) and general AI (systems that can transfer knowledge flexibly across domains).

System	Approach	Required Human Data	Transferable?	Year
Deep Blue	Alpha-beta + hand-crafted eval	Expert knowledge in code	No	1997
AlphaGo	Policy + value network + MCTS	30M human games	No	2016
AlphaZero	Self-play reinforcement learning	None (rules only)	No	2017
MuZero	Model-based RL, learns rules	None	No	2019
GPT-4 / LLMs	Transformer, language modeling	Internet-scale text	Partially	2023+

The last row is interesting. Large language models like GPT-4 do show something closer to transfer — they can apply knowledge about chess to answering a question about strategic thinking in business, for example. Whether this constitutes genuine generalization or sophisticated pattern matching across a massive training distribution is genuinely debated in the research community (see the ongoing discussion about how large language models work).

How to Teach Your Kid About Game AI

Ages 5–8: Play against the computer

Chess.com’s “Computer” feature lets you play against an AI at any skill level. Set it to the lowest level and beat it. Then ask: “Is this computer trying to trick me? Does it have feelings when it loses?” The point isn’t the chess — it’s building the intuition that AI behavior comes from computation, not intention.

Ages 9–12: Minimax on paper

Tic-tac-toe has a small enough game tree to draw by hand. Have your child draw out all possible moves from a given position, label wins/losses/draws, and work backward to find the optimal move. This is minimax without a computer. When they’ve done it, explain: “Deep Blue was doing this, but for chess, and 200 million times per second.” The algorithm becomes less magical when you’ve done it yourself.

Ages 13+: Run a reinforcement learning toy

The OpenAI Gym (now Gymnasium, gymnasium.farama.org) provides simple game environments for reinforcement learning experiments in Python. A teenager who has done some coding for kids can run a simple Q-learning agent on CartPole — a classic RL benchmark — in an afternoon. Seeing the agent improve through self-play makes the AlphaZero concept concrete.

The question to ask: “If AlphaZero can teach itself to be the world’s best chess player in 9 hours with only self-play, why can’t we use the same approach to solve homework problems?”

What to Watch For Over the Next 3 Months

Month 1: When your child hears “AI is smarter than humans” in any context, practice a three-question response: What game or task specifically? What are the rules of that task? Is the AI able to transfer that skill to anything else? This framework cuts through almost every AI-versus-humans claim.

Month 2: Watch one of the AlphaGo documentary clips on YouTube — there’s a full documentary available. Watch Sedol’s reaction to Move 37 in Game 2, when AlphaGo played a move that no human would have considered. Then discuss: Was that move creative? Or was it optimization? Is there a difference?

Month 3: Find a game your child is interested in that has a free AI opponent — chess, Go (many apps), even Tetris or Connect Four. Play against it at a level that challenges them. Then research together: how does the AI for this specific game work? Is it rule-based? Machine learning? Some combination?

Frequently Asked Questions

Did Deep Blue actually “think” when it beat Kasparov?

Not in any meaningful cognitive sense. Deep Blue evaluated positions by running a minimax search through a game tree at ~200 million positions per second. It didn’t understand chess, experience strategy, or feel pressure. The fact that it beat the world champion revealed the power of exhaustive computation in a constrained problem space — not machine cognition.

Why is Go harder for AI than chess?

The branching factor is about 7 times higher (250 vs. 35 average legal moves per turn), making brute-force search computationally infeasible much earlier in the game. Go board positions are also much harder to evaluate with simple heuristics — the value of a position often depends on subtle whole-board considerations that don’t reduce to countable material or simple patterns.

Can my kid get better at chess by playing against AI?

Yes, with caveats. Playing against a strong AI teaches pattern recognition but can also instill a style of play that’s hard to use against humans — humans play differently than optimizers. A better approach is using AI analysis after games to understand why specific moves were stronger or weaker, then practicing those positions against human opponents.

Is AlphaZero’s self-play the same as how kids learn?

Not really. AlphaZero plays millions of games against itself and updates based on a pure win/lose signal. Children learn through much richer feedback — language from parents, emotional stakes, physical consequences, social context, curiosity. AlphaZero is also learning within a perfectly specified, fully observable environment. Real learning almost never happens in conditions that clean.

---

About the author Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Silver, D., et al. (2016). “Mastering the Game of Go with Deep Neural Networks and Tree Search.” Nature, 529, 484–489. https://doi.org/10.1038/nature16961
2. Silver, D., et al. (2018). “A General Reinforcement Learning Algorithm that Masters Chess, Shogi, and Go Through Self-Play.” Science, 362(6419), 1140–1144. https://doi.org/10.1126/science.aar6404
3. Schrittwieser, J., et al. (2020). “Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model.” Nature, 588, 604–609. https://doi.org/10.1038/s41586-020-03051-4
4. Vinyals, O., et al. (2019). “Grandmaster Level in StarCraft II Using Multi-Agent Reinforcement Learning.” Nature, 575, 350–354. https://doi.org/10.1038/s41586-019-1724-z
5. Campbell, M., Hoane, A. J., & Hsu, F. H. (2002). “Deep Blue.” Artificial Intelligence, 134(1–2), 57–83. https://doi.org/10.1016/S0004-3702(01)00129-1
6. Shannon, C. E. (1950). “Programming a Computer for Playing Chess.” Philosophical Magazine, 41(314), 256–275. https://doi.org/10.1080/14786445008521796