In this podcast, Noam Brown, a research scientist at OpenAI, explores the future of Large Language Models (LLMs). He points out that while the costs of scaling pre-training are becoming increasingly high, there’s a wealth of untapped potential in enhancing test-time compute. Brown believes that focusing on this area, along with making algorithmic improvements, could help us reach Artificial General Intelligence (AGI) sooner than we think. His work on o1, a model that utilizes increased test-time compute, highlights this potential by revealing emergent reasoning abilities not seen in earlier models like GPT-4. Additionally, Brown addresses the changing role of academia in AI research and discusses the promising applications of LLMs across various fields, including social sciences and scientific research.
