Microsoft is imparting autonomous Android usage capability to GPT-4.

Scientists from Microsoft Research and Peking University may have discovered a breakthrough in getting ChatGPT, an AI language model, to operate autonomously within an operating system. While ChatGPT excels at tasks like drafting emails or writing poems, it struggles when it comes to manipulating an operating system. In order to understand why this is the case, the team developed a training environment called AndroidArena that mimics the Android OS. They found that the key challenges were understanding, reasoning, exploration, and reflection. During their research, they also stumbled upon a method to increase the model’s accuracy by 27% by providing it with automated information about its previous attempts. This discovery could have significant implications for improving AI assistants.