# Sanmi Koyejo

Co-founder, Virtue AI. TIME100 AI 2026, Thinkers. He co-wrote the award-winning paper arguing that AI's sudden 'emergent abilities' can be a measurement illusion. ([neurips.cc](https://neurips.cc/virtual/2023/poster/72117))

- Education: MS, PhD, Electrical engineering, University of Texas at Austin, 2013; BS, Electrical engineering, New Jersey Institute of Technology, 2005 ([linkedin.com](https://www.linkedin.com/in/sanmi-koyejo-984754))
- Career: Meta (Research Scientist, 2026); Virtue AI (Co-founder and Head of AI, 2024–2026); Stanford University (Associate Professor of Computer Science, from 2022); DeepMind (Research Scientist, 2018–2024); University of Illinois Urbana-Champaign (Assistant Professor, 2016–2021); Stanford University (Research Associate, 2014–2016) ([linkedin.com](https://www.linkedin.com/in/sanmi-koyejo-984754), [startupfortune.com](https://startupfortune.com/meta-cuts-the-ai-safety-team-it-acqui-hired-from-virtue-ai-four-months-ago/), [axios.com](https://www.axios.com/2025/04/15/virtue-ai-lightspeed-walden-catalyst-funding), [cas.stanford.edu](https://cas.stanford.edu/people/sanmi-koyejo), [csl.illinois.edu](https://csl.illinois.edu/news-and-media/32288))
- TIME profile: https://time.com/collection/time100-ai/2026/sanmi-koyejo/
- Profile: https://www.autumn.ai/resources/most-influential-people-in-ai/sanmi-koyejo

- Co-wrote a NeurIPS 2023 award-winning paper arguing that sudden emergent abilities in large language models can be an illusion created by how they are measured. ([neurips.cc](https://neurips.cc/virtual/2023/poster/72117))
- Is a founding member of Black in AI and has served as its board president since 2021. ([csl.illinois.edu](https://csl.illinois.edu/news-and-media/32288))
- Earlier work mapped how brain regions interact; his brain encoding and decoding maps were built into the public Neurosynth toolbox. ([csl.illinois.edu](https://csl.illinois.edu/news-and-media/32288))
- Meta hired him and his Virtue AI co-founders in June 2026 for AI security, then parted ways with the team four months later. ([startupfortune.com](https://startupfortune.com/meta-cuts-the-ai-safety-team-it-acqui-hired-from-virtue-ai-four-months-ago/))
- His startup's red-teaming platform tested AI models against more than 600 attack vectors and 1,000 risk categories. ([startupfortune.com](https://startupfortune.com/meta-cuts-the-ai-safety-team-it-acqui-hired-from-virtue-ai-four-months-ago/))
- Argues AI models should meet the same performance standards as clinical-trial biomarkers and audits. ([time.com](https://time.com/collection/time100-ai/2026/sanmi-koyejo/))