O3 and the Next Leap in Reasoning with OpenAI's Eric Mitchell and Brandon McKinzie
Eric Mitchell and Brandon McKinzie, two researchers behind OpenAI's O3 reasoning model, join No Priors to discuss what makes O3 distinct from prior O-series releases and from GPT-style foundation models. They frame O3 as a model trained to "think carefully before responding" via reinforcement learning with a focused objective on solving hard tasks, rather than next-token prediction over a broad pre-training corpus. The big differentiator, they argue, is the integration of tool use with test-time scaling: browsing, code execution, and image manipulation dramatically extend productive chain-of-thought, whereas pure reasoning without tools tends to plateau or degrade into unproductive internal monologue. The conversation explores product implications, including Deep Research (browsing, synthesizing, and charting on the user's behalf), steerability between fast and slow modes, and the desire for models that accurately calibrate their own uncertainty and effort. They highlight coding and accelerating AI research itself as the most exciting near-term applications, while noting that real-world embodied tasks (e.g., robotics) introduce hard constraints around latency and environment uncertainty that purely disembodied reasoning avoids. On the research frontier, they discuss multi-agent RL training, interactive training with human-in-the-loop interventions, and the looming convergence of robotics foundation models with general reasoning models. They emphasize that high-quality uncontaminated evals are underappreciated and increasingly scarce as models saturate them, and advise users to send the same prompt multiple times to understand the distribution of model behavior rather than judging from a single sample.
Original sources
Podcast No Priors
Episode:O3 and the Next Leap in Reasoning with OpenAI’s Eric Mitchell and Brandon McKinzie ↗
YouTube No Priors
Video:No Priors Ep. 113 | With OpenAI's Eric Mitchell and Brandon McKinzie ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.