[State of RL/Reasoning] IMO/IOI Gold, OpenAI o3/GPT-5, and Cursor Composer — Ashvin Nair, Cursor

Recorded at NeurIPS 2025, this Latent Space episode features Ashvin Nair, a researcher who contributed to OpenAI’s o1 and o3 teams before joining Cursor approximately three months prior. The conversation opens with Nair’s background—Berkeley PhD in robotics, a 2017 OpenAI internship alongside a cohort that has since spread across much of the AI industry—and explores why robotics researchers tend to adapt well to language model work, given a shared emphasis on data scrutiny and real-world grounding.

Nair offers first-hand perspective on the development of inference-time compute at OpenAI, discussing what drove the performance jumps from o1 to o3 that produced IMO and IOI gold-medal results and what the team learned about scaling test-time reasoning. He reflects candidly on whether the RL techniques that powered those advances generalize across modalities and domains, and why organizations that care deeply about one vertical—like coding—may structurally diverge from general-purpose model labs.

The discussion also covers Cursor Composer and Nair’s current work, his view that LLM agents represent a trillion-dollar market opportunity that will outpace commercial robotics by years, and his experience during the OpenAI governance crisis when Sam Altman was briefly fired—including his nuanced take on nonprofit board structures and the difficulty of getting AI governance right. A substantive inside view from someone who has worked at multiple frontier points in recent AI development.

📺 Source: Latent Space · Published December 30, 2025
🏷️ Format: Podcast