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

What It Covers

Ashvin Nair discusses his transition from OpenAI's reasoning team to Cursor, covering the development of o1/o3 models, reinforcement learning's role in achieving IMO gold performance, and Cursor's approach to co-designing products with models.

Key Questions Answered

• •RL Generalization Limits: Reinforcement learning applied to language models excels within training distribution but generalizes poorly beyond it. The strategy requires bringing economically useful tasks into distribution rather than expecting broad generalization, fundamentally changing how products must be designed around model capabilities.
• •Reasoning Team Scale: OpenAI's o1 model started with roughly fifty to one hundred contributors, expanding to three hundred people for o3 as safety, evaluation, and product teams joined. The progression from prototype to product requires exponentially more organizational resources than initial research breakthroughs suggest.
• •Continuous Learning Advantage: Cursor ships policy updates every two hours for tab autocomplete by co-locating product and ML teams. This rapid iteration cycle proves impossible at larger organizations where product and research groups operate separately, creating competitive advantage through organizational structure rather than pure technical capability.
• •Context Over Code: Automating programming jobs requires capturing years of accumulated knowledge—hyperparameter sweep results, architectural decisions, team conversations—not just writing code. Products must ingest this full context (Slack messages, Datadog traces, design documents) to enable models to replicate senior engineer decision-making processes effectively.