Wojciech Zaremba

wiki
, , , ,
1

Polish mathematician and computer scientist who co-founded OpenAI, co-discovered adversarial examples, helped build the robotic hand that solved a Rubik’s Cube one-handed, and led the teams that produced Codex, GitHub Copilot, and the RLHF infrastructure underlying ChatGPT.

Profile
Born November 30, 1988, Kluczbork, Poland
Nationality Polish
Current Institution(s) OpenAI (Co-Founder, Research Scientist)
Research Areas Deep Learning, Recurrent Neural Networks, Reinforcement Learning, Robotics, Program Synthesis, Large Language Models, RLHF
Doctoral Advisor Yann LeCun; Rob Fergus
Doctoral Thesis Learning Algorithms from Data (New York University, 2016)
Website wojzaremba.com
X / Twitter @woj_zaremba
GitHub wojzaremba
Google Scholar Wojciech Zaremba

Overview

Wojciech Zaremba is a Polish computer scientist and mathematician who co-founded OpenAI in 2015 and has remained at the organization across its entire trajectory from non-profit research lab to the developer of ChatGPT and GPT-4. A former International Mathematical Olympiad silver medalist, he brought a mathematics-first sensibility to deep learning research while interning at Google Brain and Facebook AI Research during his PhD at NYU Courant under Yann LeCun and Rob Fergus. His doctoral research on training neural networks to execute computer programs and learn algorithms was among the earliest systematic efforts to close the gap between neural network pattern-matching and symbolic computation. At OpenAI, he successively led the robotics team (whose culminating result was a robotic hand that solved a Rubik’s Cube using reinforcement learning and automatic domain randomization), the team that developed Codex and GitHub Copilot, and the human data and RLHF infrastructure that aligned ChatGPT with human preferences. He has described fostering collaboration — combining technical, research, engineering, and accessibility work — as his most important professional skill.

Early Life & Education

Zaremba was born on November 30, 1988, in Kluczbork, a small city in the Opole region of southern Poland. From an early age he showed exceptional aptitude for mathematics, chemistry, computer science, and physics simultaneously, winning local competitions across all four domains. He was a Polish Children’s Fund Scholar from 2000 to 2007, a national program identifying and supporting gifted students. In 2007, aged 18, he represented Poland at the 48th International Mathematical Olympiad held in Vietnam, winning a silver medal in one of the most competitive mathematical competitions in the world.

Zaremba studied mathematics and computer science at the University of Warsaw, and simultaneously pursued a master’s program at the École Polytechnique in Paris, graduating in 2013 with two master’s degrees in mathematics. He had also done a brief stint at NVIDIA during his bachelor years — before the deep learning era — gaining early exposure to GPU-accelerated computation.

In 2013 he began a PhD in computer science at the Courant Institute of Mathematical Sciences at New York University, working in Rob Fergus’s CILVR Lab (Computational Intelligence, Learning, Vision, and Robotics), with Yann LeCun as co-advisor. He chose NYU after concluding that, in 2012, only three universities were seriously cultivating deep learning: the University of Toronto, Université de Montréal, and NYU. He spent a year at Google Brain and another at Facebook AI Research (FAIR) as internships during his PhD. Zaremba received his PhD in 2016, having been named an OpenAI co-founder the previous year while still completing the degree. He received a Google Fellowship in 2015.

Career

NYU Courant — PhD Research (2013–2016)

Zaremba’s doctoral work, summarized in the thesis Learning Algorithms from Data, pursued a central question: can neural networks learn to emulate the algorithmic behavior of programmable computers, not just approximate statistical patterns? The work produced several influential results.

Adversarial examples (Google Brain internship, 2013). During his Google Brain internship, Zaremba co-authored “Intriguing Properties of Neural Networks” (2013, ICLR 2014) with Christian Szegedy, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. The paper discovered that imperceptible perturbations to input images could cause high-confidence misclassifications in state-of-the-art neural networks, and that these adversarial perturbations transferred across architectures — establishing the foundation of the adversarial machine learning field.

Recurrent Neural Network Regularization (FAIR internship, 2014). A paper developing principled dropout techniques for LSTMs, which became an influential practical guide for training recurrent networks at a time when LSTM training instability was a major obstacle.

Learning to Execute (2014, with Ilya Sutskever). A systematic study demonstrating that LSTMs could learn to evaluate simple computer programs from input-output examples, predicting the correct output of short programs with loops, conditionals, and variable bindings. The paper established a benchmark and methodology for studying neural network program execution that influenced the neural program synthesis literature.

OpenAI — Co-Founder and Research Scientist (2015–present)

Zaremba was announced as one of OpenAI’s eleven founding members in December 2015, turning down offers from Google and Facebook — along with, in his words, “borderline crazy” salaries — to join the non-profit. He has remained at the organization through all of its subsequent transformations.

OpenAI Gym (2016). Zaremba was among the contributors to OpenAI Gym, the open-source toolkit for reinforcement learning research that became the default benchmark environment for the RL community, standardizing comparisons across algorithms in continuous control, Atari games, and robotics tasks.

Robotics Research Manager (2015–2020). For the first five years at OpenAI, Zaremba led the robotics research program. The team’s most widely seen result was the Dactyl system (2019), which trained a five-fingered humanoid robotic hand entirely in simulation using automatic domain randomization (ADR) — a technique that automatically generates a growing distribution of randomized simulated environments — and transferred the learned policy to a physical hand that could solve a Rubik’s Cube one-handed. The Dactyl result was notable for its demonstration of sim-to-real transfer at a level of dexterity not previously achieved and for the use of LSTM-based policies rather than hand-engineered controllers. The robotics team was dissolved in 2020 amid a broader strategic pivot toward language models.

Codex and GitHub Copilot (2020–2021). Following the robotics team’s closure, Zaremba led the development of Codex, a GPT model fine-tuned on publicly available code from GitHub. The resulting model, described in “Evaluating Large Language Models Trained on Code” (2021), solved 28.8% of problems on the HumanEval benchmark — which measured functional correctness for synthesizing programs from docstrings — compared to 0% for GPT-3. Zaremba publicly announced the Codex launch in August 2021. A production version of Codex became the engine powering GitHub Copilot, the AI code completion tool that became one of the most widely used AI-assisted developer tools globally, with tens of millions of users.

GPT Models and Human Data / RLHF (2021–present). Zaremba subsequently led teams responsible for the human feedback infrastructure — the data collection, labeling operations, and reinforcement learning from human feedback (RLHF) pipeline — that shaped the behavior and safety of the GPT model series underlying ChatGPT. This included managing the Human Data area that produced the preference data used to train reward models and fine-tune GPT-4 via RLHF, a process fundamental to the conversational quality and instruction-following behavior of ChatGPT. At the time of his NYU Courant interview in 2023, he described leading this team and using ChatGPT daily for his own writing and structuring.

Key Contributions

  • Adversarial Examples (ICLR 2014) — “Intriguing Properties of Neural Networks,” with Christian Szegedy, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. First systematic discovery and characterization of adversarial perturbations that cause confident misclassification in neural networks, and their transferability across architectures. Founded the adversarial machine learning subfield.

  • LSTM Dropout / Recurrent Neural Network Regularization (2014) — Developed practical dropout methods for LSTMs that reduced overfitting and stabilized training, becoming standard practice for recurrent network training and one of the most applied technical results from his doctoral period.

  • Learning to Execute (arXiv 2014, with Ilya Sutskever) — Demonstrated that LSTMs could be trained to evaluate short computer programs, establishing a benchmark methodology for the neural program synthesis field and directly informing Zaremba’s later work on Codex.

  • OpenAI Gym (2016) — Contributing author to the open-source reinforcement learning benchmark environment that became the standard evaluation framework for the RL research community.

  • Solving Rubik’s Cube with a Robot Hand / Dactyl (2019) — Co-led the OpenAI Robotics team effort that trained a physical five-fingered robotic hand to solve a Rubik’s Cube using policies learned entirely in simulation via automatic domain randomization — a demonstration of sim-to-real transfer at unprecedented dexterity.

  • Codex / GitHub Copilot (2021) — Led development of the Codex language model, fine-tuned on GitHub code, which powered GitHub Copilot and introduced AI-assisted code completion to tens of millions of software developers. One of the most widely deployed practical applications of GPT-class models.

  • RLHF Infrastructure for ChatGPT — Led the human data and reinforcement learning from human feedback pipeline at OpenAI that aligned GPT-4 and ChatGPT with human preferences, managing the data labeling and reward model training operations foundational to ChatGPT’s conversational quality.

Awards & Recognition

  • Silver Medal, International Mathematical Olympiad (2007) — Representing Poland at the 48th IMO in Hanoi, Vietnam; one of the highest competitive mathematics achievements available to a high school student.
  • Polish Children’s Fund Scholar (2000–2007) — Polish national scholarship for gifted students across multiple STEM disciplines.
  • Google Fellowship (2015) — North American Google PhD Fellowship, awarded during his NYU doctoral studies.
  • Forbes Poland 30 Under 30 (2017) — Listed among the most influential Polish people under 30 by the Polish edition of Forbes magazine.
  • MIT Technology Review Innovators Under 35 — Named among influential innovators for contributions to AI.

Key Relationships

  • Ilya Sutskever — The closest research collaborator of Zaremba’s doctoral years; co-authored “Learning to Execute” and several related program synthesis papers; OpenAI co-founder and former chief scientist. Their shared interest in neural computation and algorithmic learning defined the intellectual core of Zaremba’s PhD work.
  • Rob Fergus — Primary PhD advisor at NYU Courant’s CILVR Lab; described Zaremba as “dazzling” during his doctoral years and as clearly destined for a “big role in the future of AI.”
  • Yann LeCun — PhD co-advisor at NYU; the broader intellectual framework of LeCun’s group — deep learning for perception and representation — shaped the environment in which Zaremba developed.
  • Christian Szegedy — Co-author on the adversarial examples paper during the Google Brain internship; the Szegedy-Zaremba collaboration produced one of the most consequential safety-relevant results in deep learning history.
  • Sam Altman — OpenAI CEO and co-founder; the founding partnership through which Zaremba has operated throughout his professional career.
  • Greg Brockman — OpenAI co-founder and former president; close colleague across the organization’s technical infrastructure and research programs.

Personal Style

Zaremba is unusual among AI researchers in combining deep mathematical foundations — Olympiad-level problem solving, dual master’s degrees in mathematics — with an explicit commitment to the organizational and human dimensions of research. He has identified “fostering collaboration” as his core professional skill, and has spoken publicly about the misconception that AI progress results from single clever ideas rather than the combination of engineering, research, product, and accessibility work. His stated research vision is expansive and optimistic: he describes AI as comparable to electricity in its civilizational impact, and has discussed specific application domains — AI-assisted therapy, personalized medicine, interdisciplinary scientific synthesis — with more concrete enthusiasm than is typical for a researcher at his technical level. He is multilingual (Polish, English, French), and has connected language learning to his broader belief that non-typical backgrounds produce adaptability. He remains on the advisory boards of the Qualia Research Institute and Growbots, and has been open about personal interests ranging from sleep research to AI-mediated conflict resolution in personal relationships.

References