Researchers from MIT and the University of Washington have developed an AI model that learns from human mistakes to improve its ability to assist us in everyday tasks. By analyzing our imperfections and previous decisions in similar situations, the AI can predict when we might make an error and offer better solutions.
The researchers trained the AI using chess games, where it planned moves and compared them to actual human moves. This allowed the AI to identify situations where humans struggled and needed to plan their next steps. In the future, the AI could intervene at these points and take over the problem-solving. The model was tested in three scenarios: chess games, inferring communicative intentions, and determining goals based on previous actions. The results were equal to or better than other projects with similar objectives. The researchers aim to apply this AI to other areas, such as reinforcement learning in robotics. By understanding human behavior and inferring goals from it, AI assistants can become more helpful collaborators, adapting to our weaknesses and offering better solutions when needed.
