John Schulman

John Schulman studied physics at Caltech and did a PhD in robotics and reinforcement learning at Berkeley under Pieter Abbeel. In 2015 Schulman published Trust Region Policy Optimization, a way to take large but safe steps when training a policy, and in December of that year, before finishing the doctorate, became one of the eleven founding members of OpenAI.

The simpler successor, Proximal Policy Optimization, appeared in July 2017, and OpenAI said at the time that PPO had "become the default reinforcement learning algorithm at OpenAI." It later became the standard optimizer in reinforcement learning from human feedback. Schulman led the reinforcement learning team that fine-tuned GPT-3 into InstructGPT and then ChatGPT, released on 30 November 2022. After Jan Leike resigned in May 2024, Schulman took over OpenAI's alignment science work as well.

Three months later, on 5 August 2024, Schulman left for Anthropic, writing of a wish "to deepen my focus on AI alignment" and return to hands-on research, and adding, "I'm not leaving due to lack of support for alignment research at OpenAI." The stay lasted about six months. In February 2025 Anthropic confirmed the departure, and Schulman became chief scientist of Mira Murati's new Thinking Machines Lab. There Schulman's work has centered on fine-tuning: the lab's Tinker service, the September 2025 study "LoRA Without Regret," and the open-weight Inkling models released in July 2026. On a September 2026 panel on recursive self-improvement, Schulman argued that bottlenecks in judgment and research would keep slowing any takeoff, and that "the last job for humans" would be "defining the objective and deciding what we actually want."