Artificial intelligence researchers have been refining large language models that exhibit remarkable capabilities, challenging our understanding of learning and cognition. The latest model, GPT-4, can solve novel and difficult tasks spanning mathematics, coding, vision, medicine, law, and psychology without any special prompting, and often surpasses human-level performance. This paper reports on the investigation of an early version of GPT-4, which we argue is part of a new cohort of language models that exhibit more general intelligence than previous AI models. We explore its limitations and discuss the challenges ahead for advancing towards deeper and more comprehensive versions of artificial general intelligence, including the possible need for pursuing a new paradigm moving beyond next-word prediction. We conclude with reflections on societal influences of this technological leap and future research directions.
