GPT-5 Launch Day: Inside OpenAI's Post-Training, Agents, and Coding Breakthroughs
Recorded on the day of GPT-5's launch, this episode features OpenAI researchers Christina Kim (Core Models post-training lead) and Isa Fulford (Deep Research and ChatGPT Agent post-training lead) in conversation with a16z's Sarah Wang. They discuss what makes GPT-5 feel like a step change—particularly in coding (with Michael Truell's endorsement calling it the best coding model available), creative writing, and front-end web development, where the team invested heavily in data curation and reward modeling rather than architectural changes. A major thread is how agent products like Deep Research and ChatGPT Agent informed GPT-5's development. Fulford explains the reinforcement learning insight: training data sets built for frontier agent models get fed back into flagship reasoning models, creating a self-reinforcing loop. The researchers also describe how post-training is treated as an art—balancing competing rewards (helpfulness vs. sycophancy) and intentionally resetting model behavior after the GPT-4o sycophancy issues. Hallucination reductions are attributed to chain-of-thought reasoning allowing models to pause before responding. On agents, Fulford defines them as systems that do useful work asynchronously on a user's behalf, with the long-term vision of a chief-of-staff-like assistant. Both researchers identify high-quality, realistic RL environments and task data as the key bottlenecks—areas they believe are ripe for startup collaboration. They introduce mid-training as a way to update knowledge cutoffs and extend model intelligence without a full pre-training run. Closing reflections touch on OpenAI's integrated research/product culture, the importance of 'taste,' and why GPT-5's biggest significance may be making their smartest model freely accessible.
Original sources
Podcast The a16z Show
Episode:GPT-5 and Agents Breakdown – w/ OpenAI Researchers Isa Fulford & Christina Kim ↗
YouTube The a16z Show
Video:GPT-5 and Agents Breakdown – w/ OpenAI Researchers Isa Fulford & Christina Kim ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.