# Sam McCandlish

> Sam McCandlish is an American theoretical physicist turned AI researcher, Anthropic co-founder, and Chief Architect. His work at OpenAI helped establish empirical laws for predicting how neural networks improve with scale; at Anthropic, his remit now covers pretraining, research productivity, and reinforcement-learning infrastructure, connecting the mathematics of frontier models to the practical discipline of building them responsibly.

- **Role:** Co-Founder and Chief Architect at Anthropic
- **Organizations:** Anthropic, OpenAI, Stanford University, Boston University, Brandeis University
- **Nationality:** American
- **Education:** B.S. and M.S. in mathematics and physics, Brandeis University, Ph.D. in theoretical physics; dissertation titled 'Depth Perception in Holography', Stanford University
- **Known for:** Led the 2018 gradient-noise-scale research that offered a practical predictor of the largest useful batch size across several machine-learning domains., Co-authored the 2020 neural language-model scaling-laws paper, which documented power-law relationships across more than seven orders of magnitude., Co-authored the GPT-3 paper, 'Language Models are Few-Shot Learners.'

## Career timeline

- **2017** — Completed his Stanford Ph.D. in theoretical physics under Eva Silverstein, defending 'Depth Perception in Holography.'
- **2017-2018** — Conducted postdoctoral research at Boston University.
- **2018** — Joined OpenAI and led research showing that the gradient noise scale predicts the largest useful training batch size.
- **2020** — Co-authored 'Scaling Laws for Neural Language Models' and 'Language Models are Few-Shot Learners,' the GPT-3 paper.
- **2021** — Co-founded Anthropic with former OpenAI colleagues.
- **2022** — Contributed to Anthropic research including Constitutional AI and work on whether language models know what they know.
- **2023-2024** — As Anthropic CTO, helped develop and initially implement the company's Responsible Scaling Policy; he later handed the Responsible Scaling Officer role to Jared Kaplan.
- **2025** — Moved from CTO to Chief Architect, retaining leadership of pretraining and adding research productivity and reinforcement-learning infrastructure to his remit.

## Achievements

- Led the 2018 gradient-noise-scale research that offered a practical predictor of the largest useful batch size across several machine-learning domains.
- Co-authored the 2020 neural language-model scaling-laws paper, which documented power-law relationships across more than seven orders of magnitude.
- Co-authored the GPT-3 paper, 'Language Models are Few-Shot Learners.'
- Co-founded Anthropic in 2021 and served as its Chief Technology Officer before becoming Chief Architect.
- Helped develop and oversee the initial implementation of Anthropic's Responsible Scaling Policy.
- Contributed to research on Constitutional AI, calibration, scalable oversight, red teaming, sycophancy, and interpretability.

## Latest updates

- **2026-09** — Anthropic's leadership page continues to list McCandlish as Co-Founder and Chief Architect, focused on pretraining, research productivity, and reinforcement-learning infrastructure.
- **2026-03** — Forbes added McCandlish to its billionaire profile database, estimating his wealth from his Anthropic stake; private-company estimates can change substantially.
- **2025-10** — Anthropic announced that Rahul Patil would become CTO and McCandlish would move to Chief Architect, deepening his focus on large-scale model training.
- **2024-12** — Appeared with Anthropic's co-founders in the company's recorded conversation about scaling, safety, the founding story, and the future of AI.

## Links

- Website: https://www.anthropic.com/company/leadership
- LinkedIn: https://www.linkedin.com/in/sam-mccandlish
- Twitter/X: https://x.com/samsamoa
- GitHub: https://github.com/samsamoa
- YouTube: https://www.youtube.com/watch?v=om2lIWXLLN4

---

Profile page: https://yespress.io/sam-mccandlish
Published by YesPress — https://yespress.io
Last updated: 2026-09-26