Mark Chen

Mark Chen is an AI researcher whose contributions at OpenAI span code generation, image generation and vision. He was a co-author of the original Codex paper, one of DALL·E 2’s primary researchers and developers, and GPT-4’s vision team co-lead and vision deployment lead. In March 2025, he stepped into an expanded role as OpenAI’s Chief Research Officer.[6][3][4][2]

Chen’s leadership responsibilities included both research and deployment within GPT-4’s vision work. OpenAI’s GPT-4 contribution record credits him as “Vision team co-lead, Deployment lead” in its Vision section.[4]

Sam Altman announced Chen’s expanded Chief Research Officer role in Leadership updates on March 24, 2025. OpenAI continued to identify him by that title at a March 4, 2026 conversation with mathematician Terence Tao at IPAM, UCLA, facilitated by James Donovan. The event concerned AI and mathematical discovery.[2][7]

Before these leadership roles, Chen contributed to research on images and code. He joined OpenAI in 2018; publications from 2020–2022 show the earlier research work that preceded his expanded appointment.[7][5][6][3][2]

Chen, Alec Radford and Ilya Sutskever authored OpenAI’s Image GPT publication, dated June 17, 2020.[5]

In July 2021, Chen was the first-listed author of Evaluating Large Language Models Trained on Code, a paper with 58 authors, including Jerry Tworek and Heewoo Jun. The team introduced Codex, a GPT language model fine-tuned on publicly available GitHub code, and studied its ability to write Python. A distinct production version powered GitHub Copilot.[6]

The paper introduced HumanEval to assess the functional correctness of programs generated from docstrings. The researchers reported that their model solved 28.8% of the problems, compared with 0% for GPT-3 and 11.4% for GPT-J. With repeated sampling—100 samples per problem—they reported solving 70.2%. These were results for the evaluated 2021 model under different sampling conditions. The team also identified difficulties with long chains of operations in docstrings and with binding operations to variables.[6]

Chen subsequently shared primary research and development credit for DALL·E 2 with Aditya Ramesh, Prafulla Dhariwal, Alex Nichol and Casey Chu. The April 2022 system card describes a model trained on image-caption pairs from public and licensed sources. It generates images through diffusion, gradually transforming a pattern of random dots into an image, and supports text-guided image generation, editing and variations of existing images.[3]