How AI Learns: Our Hive-mind AI Experiment

Cebiso Mhlanga
Jun 10
2 min read

Updated: Jun 12

We hand crafted an interactive experiment that allows users from all over the world to collectively train AI to play Tic Tac Toe.

Ever wondered how AI actually learns? Here's a simple exercise we built so you can experience that process first-hand, by playing Tic Tac Toe against an AI that initially knows nothing, not even the rules. Easy win, right? Your opponent is an AI that learns entirely through reinforcement learning, improving after every match played by human opponents worldwide. Our studio embarked on this intriguing experiment to show just how quickly AI, starting with no knowledge at all, can learn through repetition.

(Note: Scroll past the interactive experience to read the rest of the article)

With each match, the AI analyses its moves, assessing the outcomes to shape future decisions. It gradually recognises winning patterns, avoids mistakes, and steadily evolves its strategy. But just how many games does it take for an AI to master Tic Tac Toe? Interestingly, perfection in Tic Tac Toe can be reached by AI after thousands of games. Yet the fascinating part is that the AI doesn't need to be perfect to outperform the average human. Often, it achieves that in significantly fewer games.

This principle extends beyond board games. It's one reason why studies suggest self-driving cars are likely to become significantly safer than human drivers. If your car has lane assist, you've already seen a glimpse of that future in action.

"It's one reason why studies suggest self-driving cars are likely to become significantly safer than human drivers"

Of course, giving this AI all the rules of Tic Tac Toe in one go would let it master the game in mere fractions of a second. But our experiment thrives on discovery, curiosity, and authentic learning. By deliberately withholding instructions, we've turned a simple childhood pastime into an exploration of how artificial intelligence adapts when learning from scratch.

To deepen player engagement, we introduced several captivating features: a timer creating windows of opportunity for mistakes by both the AI and human players, calming sound design to enhance the experience, and seven visually engaging board design themes. Additionally, players can see a live counter displaying the number of training sessions the AI has completed, as well as a real-time view counter showing how many people around the world are actively training the AI at that very moment. This immersive approach demonstrates the power of reinforcement learning and encourages players to reflect on their own learning journeys.

Ultimately, this experiment underscores a profound insight: AI is us. All our mistakes, and all our wins.

-Ceb

Cebiso M Studio is a Milton Keynes based production house that specialises in interactive experiences, animation & advanced visualisations.

www.cebisom.com