When we talk about AI literacy, we often start with chatbots, large language models, or classroom tools. Yet some of the most important lessons for educators, leaders, and curious minds come from a very different corner of AI: games.
In this article, I’d like to take you on a short journey from classic reinforcement learning to DeepMind’s AlphaZero, and then back to a question that matters deeply for humans, not machines:
Why are our brains still so much better at navigating messy, real‑world futures than even the most advanced AI systems?
Understanding this story is a powerful way to build AI literacy for adults and students alike. It shows what current AI can and cannot do, and why brain‑inspired AI is still more aspiration than reality.
From Temporal Difference Learning to AlphaZero
For decades, researchers tried to apply Richard Sutton’s temporal‑difference learning—a form of model‑free reinforcement learning—to complex board games such as chess.
Model‑free methods can shine in some situations. Temporal‑difference systems performed well in certain video games and in backgammon. But they struggled badly in more complex settings that require planning and playing out possible futures. They saw only the immediate reward, not the quiet move that looks unremarkable now but wins you the game 20 turns later.
Then, in 2017, DeepMind introduced AlphaZero. This AI system reached superhuman performance not just in chess, but also in Go, an ancient Chinese board game that is vastly more complex than chess. There are trillions upon trillions more possible board positions in Go than on a chessboard.
So the natural AI literacy question is:
- How did AlphaZero succeed where earlier temporal‑difference systems failed?
- And what can that tell us about human learning and AI in education?
What Made AlphaZero Different: Model‑Based Reinforcement Learning
The key innovation is that AlphaZero learned by simulating futures.
Like its predecessors, AlphaZero was still a reinforcement learning system. Its strategies were not hand‑coded by grandmasters. They emerged from trial and error—self‑play and feedback.
However, unlike earlier efforts such as TD‑Gammon in the 1990s, AlphaZero did not rely purely on model‑free learning. It became a model‑based reinforcement learning system. It did not just guess the next move; it searched through possible future moves before deciding what to do.
The process looked roughly like this:
- After the opponent moved, AlphaZero paused.
- It picked several promising candidate moves to consider.
- For each candidate, it ran thousands of simulated games to see how things might unfold.
- It then chose the move that produced the highest fraction of wins in its imagined futures.
On paper, this is an impossible search problem. Even with Google’s supercomputers, simulating all possible futures from a Go position would take more than a million years. Yet AlphaZero runs its simulations within about half a second.
How? By prioritising. Instead of trying to simulate everything, AlphaZero simulates only about a thousand futures. It focuses its search on the branches that look most promising.
Search as a Way to Check “Hunches”
To appreciate what is going on, let’s stay with AI literacy and use human language.
In classic temporal‑difference learning, the system’s actor learns to pick a move based on a kind of hunch about the board position—no planning, just pattern recognition.
AlphaZero keeps this architecture but extends it:
- The actor still generates what it believes are the best moves.
- Instead of acting on a single “best guess,” AlphaZero takes several top candidates.
- It then uses search to verify the actor’s hunches.
It is as if AlphaZero says:
- “If you think move A is best, let’s quickly imagine how the game plays out if we choose A.”
- “Now let’s also imagine what happens if we follow your second‑best and third‑best ideas—moves B and C.”
Search is not used to check every possible future. It is used to test and refine the guesses produced by the actor‑critic system.
This is not a reinvention of Sutton’s temporal‑difference learning; it is a clever elaboration of it.
For AI literacy, this is a beautiful teaching example: modern AI often succeeds by combining fast pattern recognition with targeted planning, instead of relying on one technique alone.
Why Go Is Still Simpler Than Life
The story could end here with a triumph of machine intelligence—but that would miss an important truth about the limits of AI.
Even a complex game like Go is still vastly simpler than navigating the real world:
- Actions are discrete in Go. From any position, there are roughly 200 legal moves. In daily life, actions are continuous. You can move your body in an almost infinite number of ways; you can say different words with different tones.
- Information is complete and clean on a Go board. You see everything. In reality, information is noisy, incomplete, and sometimes misleading.
- Rewards are simple in Go: you win or lose. Real animals and humans juggle many competing goals—food, safety, belonging, curiosity—and these priorities shift over time.
So while AlphaZero is a major milestone in model‑based reinforcement learning, current AI systems remain far from being able to plan in environments with:
- Continuous action spaces
- Uncertain, partial information
- Complex, changing reward structures
This is where brain‑inspired AI gets interesting.
What Mammal Brains Do Differently
The brilliance of simulation in mammal brains is unlikely to be a single, magical search algorithm waiting to be discovered. It seems far more likely that the magic lies in flexible strategy selection:
- Sometimes we pause to simulate our options. Sometimes we don’t—and simply act on instinct. Somehow, our brains decide when each mode is appropriate.
- Sometimes we mentally explore future scenarios. At other times, we simulate the past—replaying events, wondering “what if I had chosen differently?” Our brains switch fluidly between future‑planning and past‑reframing.
- Sometimes we imagine every little detail. Sometimes we sketch only a rough outline. Our brains know when we need fine‑grained planning and when a broad sketch is enough.
How do brains do that?
This open question is not just a matter of neuroscience. It is essential to the future of AI, especially if we want AI systems that can support education, healthcare, or complex policy work in a trustworthy way.
Why This Matters for AI Literacy and Education
For AI Edutainment and our work with teachers, students, and leaders, the AlphaZero story offers several powerful teaching points:
- AI is not magic. Even superhuman game‑playing agents are built from understandable building blocks: reinforcement learning, search, heuristics, and clever prioritisation.
- Context matters. A system that dominates a well‑defined game still struggles when the world becomes continuous, noisy, and value‑laden—like a classroom or a hospital.
- Human flexibility is still unique. Our ability to switch between instinct, detailed planning, reflection on the past, and imagination of the future is something current AI cannot match.
- AI literacy should include planning. When we teach AI in schools or leadership programmes, it is not enough to talk about neural networks or prompts. We should also discuss how AI systems plan, when they cannot, and where humans must stay in the loop.