AI Models Play Pac-Man to Prove They're Smarter Than You

Imagine AI models playing Pac-Man against each other. That's exactly what a developer did to compare six popular decision-making AIs: jev 1.13, Kev 4B, Clef, Clef Flash, GPT-6 Luna, and Laya. They all respond in milliseconds, so they can play the game in real time. The results? jev 1.13 came out on top with an average score of 2,750, followed by GPT-6 Luna at 2,568 and Clef Flash at 2,538. The worst was Laya, scoring only 639. The leaderboard is public, and the code is open-source, so you can run your own AI and see how it does.

Why should you care? These AI models are designed to make quick decisions, just like in Pac-Man. That skill is crucial for real-world applications: self-driving cars deciding when to brake, chatbots handling customer complaints, or even robots navigating warehouses. The faster and smarter the AI, the better these services work. Plus, since the test is open, anyone can experiment and improve AI, which could lead to better tools for everyone.

The catch? Playing Pac-Man is a simplified game. Real life is messier, with unpredictable situations and ethical dilemmas. An AI that excels at Pac-Man might not handle a tricky driving scenario or a nuanced conversation. Also, the test only ran each model 100 times, so there's a margin of error. Some models might tie in performance. But it's a fun way to see how AI is progressing.

You can even play Pac-Man yourself against these AIs as ghosts. The project is open-source, so if you're curious, you can try it out. It's a peek into how AI is learning to make decisions faster and better, which will eventually affect many aspects of our lives.