Wojciech Zaremba

Overview

Wojciech Zaremba is a computer scientist from Poland who co-founded OpenAI in 2015. After 2020 he led teams working on GPT models, GitHub Copilot and Codex. He was born on 30 November 1988 in Kluczbork, Poland.^[1]

Education and early research

Zaremba was already a strong competitor in mathematics as a student. In 2007 he won a silver medal at the International Mathematical Olympiad, representing Poland.^[1] He then studied mathematics and computer science at the University of Warsaw. He went on to New York University, where he earned a PhD in 2016. His doctoral supervisors were Yann LeCun and Rob Fergus. During his doctoral studies he spent time at Google Brain and at Facebook AI Research, two industrial research groups working on machine learning.^[1]

In 2013 he co-authored the paper “Intriguing properties of neural networks,” which became foundational work on adversarial examples.^[1] An adversarial example is an input, such as an image, that has been changed slightly in a way a person would hardly notice, yet causes a neural network to give a wrong answer. Work on this phenomenon raised questions about how reliable neural networks are, and it became one of the active areas of machine learning research.

Main contributions

Zaremba's work covers three areas that are easy to separate in a list but closely linked in practice.

The first is adversarial examples, described above. The second is robotics. At OpenAI he led the robotics research group until 2020. One of its results was a robotic arm that could solve a Rubik's Cube.^[1] The task combines perception, precise manipulation and planning, which makes it a demanding test for a learning system acting in the physical world.

The third is code. In 2021 he co-authored the paper “Evaluating Large Language Models Trained on Code,” the research paper behind Codex.^[1][2] A code model is a language model trained on programming code. It learns the patterns of how programs are written and can help produce or explain code.

Career

Zaremba was one of the co-founders of OpenAI, which was founded in 2015.^[1] He led its robotics research until 2020.^[1]

After 2020 he led teams working on GPT models, GitHub Copilot and Codex.^[1][2] GPT stands for Generative Pre-trained Transformer, a family of large language models that learn to predict text and can be adapted to many tasks. GitHub Copilot is a tool that suggests code to programmers inside their editors.

Influence

Zaremba's career connects three threads of modern AI: the reliability of neural networks, learning in physical robots, and language models for code. Each thread has grown into a large body of research. His papers and projects are part of the record of how those early ideas developed, and the Codex paper documents the research behind the Codex model family.^[2]