John Schulman

Career

John Schulman completed a PhD in computer science at UC Berkeley advised by Pieter Abbeel, working on robotics and reinforcement learning; the Abbeel group lists the December 2016 thesis "Optimizing Expectations: From Deep Reinforcement Learning to Stochastic Computation Graphs". Schulman then co-founded OpenAI, where Schulman led the creation of ChatGPT and from 2022 to 2024 co-led the post-training team that developed models for ChatGPT and the OpenAI API.¹²

After OpenAI, Schulman did research on the Alignment Science team at Anthropic. Schulman is now a co-founder and the chief scientist of Thinking Machines Lab, and is a credited author of the lab's September 2025 Connectionism post "LoRA Without Regret".¹⁵

Research

Schulman's early deep reinforcement learning work includes Trust Region Policy Optimization (2015) and Generalized Advantage Estimation, presented at ICLR 2016, followed by Proximal Policy Optimization (2017). Later co-authored papers listed on Schulman's publications page include WebGPT (2021), the grade school math verifiers paper (2021), studies of scaling laws for reward model overoptimization (2022), and "Let's verify step by step" (2023).³⁴⁶

Open-source work

Schulman's code page lists projects built with collaborators: TrajOpt (2012–2013), a C++ framework with a Python API for robot trajectory optimization; the Computation Graph Toolkit (2015), an automatic differentiation library; and, at OpenAI, OpenAI Gym and OpenAI Baselines (2016), Gym Retro (2018) and the Procgen Benchmark (2019).⁷

Work and contributions

• Proximal Policy Optimization (PPO) Published research

  First author of the 2017 paper proposing a family of policy gradient methods that optimize a surrogate objective with multiple epochs of minibatch updates.⁴

• Trust Region Policy Optimization (TRPO) Published research

  First author, with Sergey Levine, Philipp Moritz, Michael I. Jordan and Pieter Abbeel, of the 2015 paper describing the TRPO algorithm for optimizing large nonlinear policies.³

• LoRA Without Regret Published research

  Credited author, "in collaboration with others at Thinking Machines", of the September 2025 Connectionism post studying whether and when LoRA can match full fine-tuning.⁵

Timeline

1. Feb 19, 2015Posted Trust Region Policy Optimization with Sergey Levine, Philipp Moritz, Michael I. Jordan and Pieter Abbeel.³
2. 2016Co-authored the OpenAI Gym paper and released OpenAI Baselines with collaborators.⁶⁷
3. Dec 2016Completed a UC Berkeley PhD thesis on deep reinforcement learning and stochastic computation graphs.²
4. Jul 20, 2017Posted Proximal Policy Optimization Algorithms as first author.⁴
5. 2019Released the Procgen Benchmark for generalization in reinforcement learning with collaborators.⁷
6. 2022Began co-leading the OpenAI post-training team, which developed models for ChatGPT and the OpenAI API.¹
7. Sep 29, 2025Published "LoRA Without Regret" with colleagues at Thinking Machines Lab.⁵

Other documented roles

Roles the sources document that are not confirmed as current.

• Research on the Alignment Science team, Anthropic¹
• Cofounder; co-led the post-training team (2022–2024), OpenAI¹²

Honors and recognition

• Best Vision Paper, ICRA 2013 · IEEE International Conference on Robotics and Automation, 2013⁸
• C.V. Ramamoorthy Distinguished Research Award · UC Berkeley EECS, 2016⁸
• 35 Innovators Under 35 · MIT Technology Review, 2018⁸

Where they work

• Thinking Machines Lab 13 people in the wiki

Related people

• Danqi ChenAlso at Thinking Machines Lab
• Stephen RollerAlso at Thinking Machines Lab
• Mira MuratiAlso at Thinking Machines Lab
• Kenneth LiAlso at Thinking Machines Lab
• Liliang RenAlso at Thinking Machines Lab
• Weizhu ChenLanguage models · Reinforcement learning
• Younes BelkadaLanguage models · Reinforcement learning
• Oriol VinyalsLanguage models · Reinforcement learning
• Leandro von WerraLanguage models · Reinforcement learning
• Albert JiangLanguage models · Reinforcement learning