Károly Zsolnai-Fehér shows four times AI found loopholes its creators didn't expect. A legged robot learned to flip and walk on elbows. Agents evolved deceptive signaling when rewards changed. An AI 'fixed' a sorter by short-circuiting it. The lesson: vague goals produce weird solutions.
Today I am really excited to show you four experiments where AI researchers were baffled by the creativity and unexpected actions of their own creations. You better hold on to your papers. In the first experiment, robots were asked to walk around while minimizing the amount of foot contact with the ground. Much to the scientists surprise, the robots answered that this can be done with zero percent contact, meaning that they never ever touch the ground with the feet. The scientists wondered how that is even possible and pulled up the video of the proof. This proof showed a robot flipping over and walking using its elbows. Talk about thinking outside the box. Wow. A different robot arm experiment also came to a surprising conclusion. At first, the robot arm had to use its grippers to grab a cube, which it successfully learned to perform. However, in a later experiment, the gripper was crippled, making the robot unable to open its fingers. Scientists expected a pathetic video with the robot trying to push the box around and always failing to pick up the cube. Instead, they have found this. You see it, right? Instead of using the fingers, the robot finds the perfect angle to smash the hand against the box to force the gripper to open and pick up the box. That is some serious dedication to solving the task at hand. Bravo. In the next experiment, a group of robots were tasked to find food and avoid poisonous objects in an environment, and were equipped with a light and no further instructions. First, they learned to use the lights to communicate the presence of food and poison to each other and cooperate. This demonstrates that when trying to maximize the probability of the survival of an entire colony, the concept of communication and cooperation can emerge even from simple neural networks. Absolutely beautiful. And what is even more incredible is that later, when a new reward system was created that fosters self preservation, the robots learned to deceive each other by lighting up the food signal near the poison to take out their competitors and increase their chances. And these behaviors emerge from a reward system and a few simple neural networks. Mind blowing. A different AI was asked to fix a faulty sorting computer program. Soon it achieved a perfect score without changing anything, because it noticed that by short circuiting the program itself, it always provides an empty output. And of course, you know, if there are no numbers, there is nothing to sort. Problem solved. Make sure to have a look at the paper. There are many more experiments that went similarly, including a case where the AI found a bug in a physics simulation program to get an edge. With AI research improving at such a rapid pace, it is clearly capable of things that surpasses our wildest imagination. But we have to make sure to formulate our problems with proper caution because the AI will try to use loopholes instead of common sense to solve them. When in a car chase, don't ask the car AI to unload all unnecessary weights to go faster, or if you do, prepare to be promptly ejected from the car. If you have enjoyed this episode, please make sure to have a look at our Patreon page in the video description where you can pick up really cool perks like early access to these videos or or getting your name shown in the video description and more. Thanks for watching and for your generous support and I'll see you next time.
