A simple AI-controlled robot completely mastered a marble maze puzzle in a matter of hours. It was breathtaking to watch the CyberRunner robot grow from a shaky novice to a smooth and sophisticated expert - and like humans, it even found ways to cheat. This is compelling proof of AI’s ability to master complex real-world skills.

Just months after we saw an AI-piloted drone absolutely overwhelm a human world champion in a circuit race, researchers at ETH Zurich have revealed a real-world physical demonstration of AI's amazing learning abilities.

The goal of this study: the rolling marble maze game. You've definitely played this game; as you roll marbles through a wooden maze, a pair of knobs allow you to tilt the board, trying to avoid falling into a hole.

The CyberRunner AI robot is equipped with two electric motors to drive the knobs, a camera to view the board, and an AI brain running a "deep reinforcement learning system" that can send up to 55 control instructions per second. It was informed of the objectives of the exercise and conducted the experiment on its own.

Sure, early attempts were jerky and wobbly, but after 6.06 hours of training, the AI ​​smashed the fastest record of an "extremely skilled human player" by 6% better than the previous record, displaying a fluid, practiced confidence that's dizzying:

Remarkably, the AI ​​quickly discovered shortcuts it could use to cut off entire sections of the maze and proceed at a much faster pace—the researchers had to enter the maze and specifically tell it not to cheat.

The CyberRunner team says they will soon make the entire project open source, allowing anyone to build and train it at home.

"We believe this is an ideal testbed for real-world machine learning and artificial intelligence research," Professor Rafael-André said in a press release. "Before CyberRunner, only institutions with huge budgets and custom experimental infrastructure could conduct research in this field. Now, for less than $200, anyone can engage in cutting-edge AI research with applications in the physical world. Furthermore, once thousands of CyberRunners appear in the real world, it will be possible to conduct large-scale experiments and conduct parallel learning on a global scale. The ultimate goal of citizen science!"

It's just a cute little game, but it shows something much bigger. 2023 is definitely a watershed and tipping point for artificial intelligence systems, and their integration with the robotic world is driving equally astounding advances in the embodied world. Watching these machines learn in real time, whether starting from scratch or by observing others, becoming increasingly capable and mastering new skills faster and faster, feels like a terrifying privilege.

2023 is about to pass, and a far-reaching technological revolution seems to be a foregone conclusion. The seeds of human beings' elimination have been sown. Now it's time to see how quickly the value of human wisdom and human labor will approach zero. I don’t think we are still 20 years away. Even 10 years at this crazy accelerated rate seems too long.

What an unusual time this is.