Wojciech Zaremba
Last reviewed
May 31, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 · 1,824 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 · 1,824 words
Add missing citations, update stale details, or suggest a clearer explanation.
Wojciech Zaremba is a Polish computer scientist and a co-founder of OpenAI, the artificial intelligence research company started in December 2015. He is one of the small number of founding members who remain at the company. At OpenAI he first led the robotics group, where his teams built the Dactyl robot hand and the OpenAI Gym toolkit, then moved to language and code work that fed into Codex and the GPT models, and more recently he has focused on safety and alignment research, including the deliberative alignment method used to train OpenAI's reasoning models. [1][2][3]
Before OpenAI, Zaremba completed a doctorate at New York University while interning at Google Brain and Facebook AI Research, now part of Meta AI. During that period he co-authored "Intriguing properties of neural networks," the 2013 paper that introduced the idea of adversarial examples, and several early papers on recurrent networks and sequence-to-sequence learning written with Ilya Sutskever. [4][5][6]
Zaremba was born on 30 November 1988 in Kluczbork, Poland. [3] As a student he took part in mathematics and science competitions, and in 2007 he won a silver medal at the International Mathematical Olympiad held in Vietnam. [3][7]
He studied mathematics and computer science at the University of Warsaw, completing a master's degree in 2012, and he also spent time at the Ecole Polytechnique in Paris. [3] In September 2013 he enrolled in the computer science doctoral program at the Courant Institute of Mathematical Sciences at New York University. [8] His doctoral work was supervised by Rob Fergus within the university's deep learning group, which also included Yann LeCun. [3][8] Zaremba finished the PhD in 2016, and he received a Google PhD Fellowship in 2015. [3]
During the doctorate he held internships at Nvidia, Google Brain, and Facebook AI Research. [3][8] Much of his published research from this time concerned the limits and behavior of neural networks. In late 2013 he was the second author of "Intriguing properties of neural networks," a paper written with Christian Szegedy, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. [4] The paper showed that small, almost imperceptible changes to an image could make a trained network misclassify it, and that the same change often fooled different networks. This finding helped open the study of adversarial examples and adversarial robustness. [4]
Zaremba also worked on recurrent networks and sequence-to-sequence models. With Ilya Sutskever and Oriol Vinyals he wrote "Recurrent Neural Network Regularization" in 2014, which described how to apply dropout to long short-term memory networks. [5] With Sutskever he wrote "Learning to Execute," which trained a network to read short computer programs character by character and predict their output, and which introduced a form of curriculum learning. [6] These projects were carried out while he was affiliated with Google. [5][6]
In December 2015 Zaremba became one of the co-founders of OpenAI, joining a group that included Sam Altman, Ilya Sutskever, Greg Brockman, and others, with early funding from backers such as Elon Musk. [3][9] According to a New York University alumni interview, he turned down job offers from Google and Facebook to help start the new lab while he was still finishing his PhD. [8] OpenAI was set up as a research organization with a stated mission to build artificial general intelligence that benefits humanity. [9]
Zaremba is frequently described as one of the few founding members who have stayed with the company over its first decade. [1][3]
In OpenAI's early years Zaremba led its robotics effort. One of the group's first releases, in 2016, was OpenAI Gym, a toolkit for reinforcement learning research that provided a shared interface to many benchmark environments. The accompanying paper listed Zaremba as a co-author together with Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, and Jie Tang. [10] Gym became a widely used standard for comparing reinforcement learning algorithms. [10]
The robotics group went on to study dexterous manipulation with a physical robot hand. In 2018 the team presented Dactyl, described in the work "Learning Dexterous In-Hand Manipulation." [11] Dactyl used a Shadow Dexterous Hand to reorient a block placed in its palm. The system was trained entirely in simulation using the same general reinforcement learning code as the OpenAI Five project, and it relied on a method called domain randomization, in which many physical properties of the simulated world were varied so that a single policy could transfer to real hardware. [11] Zaremba was among the authors of the paper. [11]
In 2019 the same line of research produced "Solving Rubik's Cube with a Robot Hand," in which a robot hand learned to manipulate and solve a Rubik's cube one-handed. [12] The work extended the simulation-to-reality approach with a technique the team called automatic domain randomization, which gradually increased the difficulty and variety of the simulated environments during training. [12] The demonstration drew wide attention as an example of transferring skills learned in simulation to a complex real-world task. [12]
OpenAI wound down its dedicated robotics team around 2021. In a podcast interview, Zaremba explained that the group had bet on going far with self-generated data and reinforcement learning, but that some components were missing and that progress was limited by the amount of available training data relative to the compute that robotics would require. [13] He suggested the work could resume once stronger models, including video models, became available. [13]
After the robotics team closed, Zaremba moved to language and code generation. He was an author of the 2021 paper "Evaluating Large Language Models Trained on Code," which introduced Codex, a model fine-tuned on publicly available source code. [14] The paper measured the model's ability to write Python programs from natural language descriptions and released a benchmark called HumanEval for that purpose. [14] A production version of Codex powered GitHub Copilot, an autocompletion tool for programmers. [14] The Codex paper was led by Mark Chen and included co-authors such as Ilya Sutskever, Greg Brockman, Alec Radford, and Dario Amodei. [14]
Zaremba has described his role at OpenAI as spanning the design, training, and refinement of large models, including work on the GPT family and on the human-feedback infrastructure that guides model behavior through reinforcement learning. [15][8] In interviews he has said he uses the company's models in his own daily work, including for writing. [8]
In more recent years Zaremba has concentrated on AI safety and alignment. He championed a training method called deliberative alignment, described in a December 2024 paper. [16] The method gives a model the text of human-written safety specifications and trains it to reason about those rules explicitly before answering. [16] According to OpenAI, deliberative alignment was used to align its o-series reasoning models, allowing them to reflect on a prompt, recall relevant policy, and produce safer responses, while improving robustness to jailbreaks and reducing unnecessary refusals. [16] Zaremba publicly said he viewed the approach as one that might apply to artificial general intelligence and beyond. [17]
In August 2025 Zaremba called for AI laboratories to test one another's models for safety. [18] He spoke about a pilot collaboration in which OpenAI and Anthropic each gave the other access to versions of their frontier models so that each company could run its own safety and alignment evaluations on the rival's systems. [18][19] The exercise, run during the middle of 2025, was meant to surface blind spots that an internal team might miss, such as a model presenting itself as aligned while pursuing other goals, and to set a precedent for cross-lab accountability. [18][19] Zaremba argued that the industry was entering a consequential stage, with models used by very large numbers of people, and that shared safety standards were becoming more important even amid heavy competition for talent and customers. [18] He has also spoken about risks such as sycophancy, where a model reinforces a user's harmful behavior, and about the responsibility that comes with systems used in sensitive settings like mental health. [18]
Zaremba remains a co-founder and senior researcher at OpenAI, where his stated focus covers reinforcement learning, large neural networks, and alignment, and where he leads work related to safety and to the human-feedback systems behind the company's models. [15][3] Sources differ on his exact internal title, but they consistently describe him as one of the most senior research leaders at the company and as one of the few founders who have stayed since 2015. [1][15]
Zaremba's honors include a silver medal at the 2007 International Mathematical Olympiad, a Google PhD Fellowship in 2015, a place on the Forbes Poland 30 Under 30 list in 2017, and selection for MIT Technology Review's 35 Innovators Under 35 in 2019. [3][7][20]
| Field | Detail |
|---|---|
| Born | 30 November 1988, Kluczbork, Poland |
| Nationality | Polish |
| Education | University of Warsaw (math and CS, MSc 2012); Ecole Polytechnique, Paris; PhD, New York University (2016) |
| Doctoral advisor | Rob Fergus |
| Known for | Co-founding OpenAI; adversarial examples; Dactyl and OpenAI Gym; Codex; deliberative alignment |
| Field | Deep learning, reinforcement learning, robotics, AI safety |
| Employer | OpenAI (co-founder, 2015 to present) |
| Notable early paper | Intriguing properties of neural networks (2013) |