Wojciech Zaremba

Research

Zaremba's early papers, several with Ilya Sutskever, appeared in 2013 and 2014. "Intriguing properties of neural networks" (2013), with Christian Szegedy and others, showed that imperceptible perturbations can cause a network to misclassify an image; "Recurrent Neural Network Regularization" (2014), with Sutskever and Oriol Vinyals, showed how to apply dropout to LSTMs; and "Learning to Execute" (2014) trained recurrent networks to evaluate short computer programs.⁵⁶⁷

Zaremba co-authored the OpenAI Gym toolkit for reinforcement learning research (2016) and robotics work including domain randomization for transferring networks trained in simulation to the real world (2017) and "Solving Rubik's Cube with a Robot Hand" (2019). Zaremba was the last-listed author of the 2021 paper introducing Codex and the HumanEval benchmark for code generation.³⁸⁹⁴

OpenAI Foundation

In March 2026, the OpenAI Foundation announced that Zaremba, a co-founder of OpenAI, was joining as Head of AI Resilience. The program's initial areas are AI's effects on young people, biosecurity and AI model safety, including support for independent testing and evaluation and for foundational safety research. Zaremba set out the program's approach in the June 2026 essay "Resilience in the age of AI".¹²

Work and contributions

• HumanEval Open data

  Last-listed author of "Evaluating Large Language Models Trained on Code" (2021), the paper that introduced Codex and released the HumanEval evaluation set for measuring the functional correctness of programs synthesized from docstrings.⁴

• OpenAI Gym Published research

  Co-author of the 2016 OpenAI Gym paper, which describes a toolkit for reinforcement learning research with a collection of benchmark problems that share a common interface.³

• Resilience in the age of AI

  Author of the OpenAI Foundation's June 2026 essay, which covers bio-resilience, cyber-resilience, AI model safety and AI's impact on young people.²

Timeline

1. Dec 2013Co-authored "Intriguing properties of neural networks", on imperceptible perturbations that cause misclassification.⁵
2. Sep 2014Posted "Recurrent Neural Network Regularization" as first author with Ilya Sutskever and Oriol Vinyals.⁶
3. Oct 2014Posted "Learning to Execute" with Ilya Sutskever as first author.⁷
4. Jun 2016Co-authored the OpenAI Gym paper introducing a toolkit for reinforcement learning research.³
5. Mar 2017Co-authored the domain randomization paper on transferring networks from simulation to the real world.⁸
6. Oct 2019Co-authored "Solving Rubik's Cube with a Robot Hand".⁹
7. Jul 2021Co-authored the Codex paper that introduced the HumanEval evaluation set.⁴
8. Mar 2026Joined the OpenAI Foundation as Head of AI Resilience.¹
9. Jun 2026Published the OpenAI Foundation essay "Resilience in the age of AI".²

Other documented roles

Roles the sources document that are not confirmed as current.

• Co-founder; co-author of OpenAI's 2016 Gym paper and 2021 Codex paper, OpenAI¹³⁴

Where they work

• OpenAI Foundation 1 person in the wiki

Related people

• Shengjia ZhaoLanguage models · Reinforcement learning
• Koray KavukcuogluLanguage models · Reinforcement learning
• Nathan LambertLanguage models · Reinforcement learning
• Paul ChristianoSafety and alignment · Reinforcement learning
• Sami JaghouarReinforcement learning · Language models
• Leandro von WerraLanguage models · Reinforcement learning
• Yulia TsvetkovLanguage models · Safety and alignment
• Evan HubingerSafety and alignment · Language models
• Samuel R. BowmanSafety and alignment · Language models
• Yejin ChoiLanguage models · Safety and alignment