OpenAI's Mark Chen & Jakub Pachocki on GPT-5, RL, and the March Toward an Automated Researcher
OpenAI's Chief Research Officer Mark Chen and Chief Scientist Jakub Pachocki joined a16z's Anjney Midha and Sarah Wang to discuss the GPT-5 launch, which they framed as the company's effort to bring reasoning into the mainstream by merging its instant-response GPT series with its longer-thinking o-series. They acknowledged that many traditional evals are now saturated and described a shift toward benchmarks that measure actual discovery and economically meaningful tasks, with math, programming competitions, and hard-science problem solving as the current frontier. The central research target they articulated is building an "automated researcher" capable of autonomously discovering new ideas over long time horizons, with current models operating reliably at roughly one-to-five-hour reasoning windows. They credited reinforcement learning as the most versatile and productive lever in their stack, said RL on natural language is finally delivering on long-standing ambitions, and argued that compute—not data—remains the binding constraint on progress. They also discussed the just-released GPT-5 Codex, which rebalances compute spend between easy and hard coding tasks and is tuned for messy real-world software environments. On culture, both emphasized protecting fundamental research from product pressures, maintaining conviction-driven long-term roadmaps rather than reacting to competitor launches, and portfolio-managing compute across bets. They described the shift from "vibe coding" to "vibe researching" as inevitable, and reflected personally on their own competitive-programming roots and what it feels like to watch models surpass their abilities.
Original sources
Podcast The a16z Show
Episode:From Vibe Coding to Vibe Researching: OpenAI’s Mark Chen and Jakub Pachocki ↗
YouTube The a16z Show
Video:From Vibe Coding to Vibe Researching: OpenAI’s Mark Chen and Jakub Pachocki ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.