Wojciech Zaremba is an OpenAI co-founder and Head of AI Resilience at the OpenAI Foundation. His career spans neural-network research, robotics and language-model teams, and work on the institutions and safeguards needed as AI becomes more capable.[1][3][2]
Start with Resilience in the age of AI →
X 
GitHub Background
Zaremba studied mathematics and computer science at the University of Warsaw and pursued master’s studies at École Polytechnique before joining NYU’s computer science PhD program in September 2013. His dissertation, Learning Algorithms from Data, is dated May 2016. NYU’s interview describes research internships at Google and Facebook before he joined OpenAI. This background connects abstract questions about learning with the engineering needed to make neural networks perform useful tasks.[4][5]
Career
At OpenAI, Zaremba says he led robotics, CodeGen, and language teams, and established the human-data team that helped shape ChatGPT’s behavior. The robotics program’s Dactyl system learned object manipulation in simulation by practicing under varied physical conditions, then transferred that behavior to a real hand. He also says he initiated OpenAI’s democratic-input project. Its 2023 call funded experiments in public deliberation about AI behavior; the announcement explicitly described those early experiments as nonbinding.[3][8][9]
Contributions
Zaremba co-authored Intriguing Properties of Neural Networks, which demonstrated that small, carefully chosen changes to an image could make a classifier give the wrong answer even when the change was imperceptible to people. That exposed a reliability problem that ordinary test accuracy could miss. With Ilya Sutskever, he also wrote Learning to Execute. The researchers trained recurrent networks to read short programs and predict their outputs. A training curriculum substantially improved results.[6][7]
Ideas
The OpenAI Foundation announced Zaremba as Head of AI Resilience in March 2026. In his June 2026 essay, he argues that safe adoption requires an ecosystem beyond model developers. His examples include defenses against biological misuse, cybersecurity for institutions with fewer resources, independent model evaluation, and research on effects on young people. The central idea is that stronger models and stronger social defenses must develop together. The essay lays out priorities and intended investments; it should be read as a program agenda, not proof that the risks have been resolved.[1][2]
Tools, writing & talks
Writing, research & talks
Start here
Resilience in the age of AIopenaifoundation.orgAn accessible explanation of Zaremba’s current work. It connects AI risks to concrete institutions, including cybersecurity support, independent testing, and public-health defenses, rather than leaving safety as an abstract property of a model.
The clearest nontechnical entry to his current role, with specific explanations of safety work beyond model training.
Author[2]A research paper showing that apparently strong image classifiers can be fooled by carefully constructed changes invisible to people. It gives technical readers an early, concrete account of adversarial examples and why average accuracy is insufficient.
Co-author[6]- Learning to Executearxiv.org
An experiment in teaching recurrent networks to predict the outputs of short programs. It helps explain the connection between program structure, memory, and training curricula; basic neural-network knowledge is useful.
Co-author with Ilya Sutskever[7] - Learning dexterityopenai.com
OpenAI’s illustrated explanation of Dactyl, which transferred simulated practice to physical object manipulation. Read it for the mechanics of varying simulated conditions and the limits of transfer to a real robot.