OpenAI's GPT-4 Surpasses Human Performance in Theory of Mind, Identifies 16 Million Features

Jun 6, 2024, 05:30 PM

OpenAI's GPT-4 has achieved a significant milestone by exceeding human performance on higher-order Theory of Mind (ToM) tasks, according to a recent paper. The study highlights that GPT-4 and Flan-PaLM reach adult or near-adult level performance on ToM tasks up to the 6th order. Additionally, OpenAI has developed methods to find understandable patterns in GPT-4's brain using sparse autoencoders, which aim to make AI models more trustworthy and controllable. These methods have allowed researchers to identify 16 million features in GPT-4, scaling better than previous work.