No Priors Ep. 113 | With OpenAI's Eric Mitchell and Brandon McKinzie

In this episode of No Priors, Sarah and Elad interview Brandon McKinzie and Eric Mitchell from OpenAI about their O3 reasoning model. The discussion covers the advancements in O3, including its enhanced ability to think before responding and utilize tools like web browsing and code execution to solve complex tasks. The guests explain how reinforcement learning is used to train the model to solve difficult tasks, and they explore the potential bifurcation of AI models into fast, efficient models for basic tasks and slower, more expensive models for complex tasks. They also discuss the importance of tool use in test time scaling, potential applications in research and coding, and the challenges of simulating human interaction in AI models. Finally, they touch on the future of AI development, the need for high-quality evaluation data, and the importance of exploring the distribution of responses from AI models.