Sam McCandlish is the co-founder of Anthropic, an enterprise at the forefront of developing safe and advanced artificial intelligence systems. Before Anthropic, McCandlish worked with OpenAI, contributing to its research on neural network performance and safety mechanisms. ^{[1] [6]}

Early Life & Education

Sam McCandlish pursued his academic interests initially through a B.S. and M.S. in Mathematics and Physics at Brandeis University, followed by a Ph.D. in Theoretical Physics from Stanford University. His doctoral research delved into the complexities of quantum gravity and tensor networks, equipping him with a foundation in systematic modeling that he later applied to machine learning and AI. ^[1]

Career

OpenAI

Sam McCandlish was a Postdoctoral Fellow at Boston University from 2017 to 2018, where he was affiliated with the Simons Bootstrap Collaboration.

McCandlish joined OpenAI in 2017 and served as a Member of Technical Staff from May 2018 to December 2020. During this period, he participated in the AI Safety Fellowship, established the Science of AI team, and led the early development of Codex. His research examined the performance and scaling behavior of neural networks, with a focus on language models and large-scale training methods.

His research included studies on scaling laws for neural language models and large-batch training. He co-authored work on the gradient noise scale framework, which has been applied in machine learning research, including image classification and game environments. Publications from this period include Language Models Are Few-Shot Learners, Scaling Laws for Language Models, Scaling Laws for Neural Language Models, Scaling Laws for Transformers, An Empirical Model of Large-Batch Training, as well as OpenAI's Science of AI and AI and Efficiency projects.

Anthropic

In January 2021, McCandlish co-founded Anthropic with former OpenAI colleagues, including Dario and Daniela Amodei. The company conducts research on artificial intelligence with an emphasis on AI safety and interpretability.

At Anthropic, McCandlish has served as Chief Scientist, Chief Technology Officer, and, since 2025, Chief Architect. His work has included pre-training and the management of large-scale model training. He has also been involved in the development of the Claude family of language models and Anthropic's Constitutional AI methodology, which uses AI-generated feedback as part of the model alignment process rather than relying solely on human preference data.

His research at Anthropic has focused on scaling laws, AI safety, and mechanistic interpretability. It has also included work related to Constitutional AI and model alignment methods.

McCandlish and other Anthropic co-founders pledged to donate 80% of their personal wealth to initiatives related to AI risks and inequality.

Anthropic's research and safety practices have been the subject of public discussion. Some commentators have questioned whether the company's safety framework adequately addresses the societal implications of increasingly capable AI systems, while others have discussed the transparency and effectiveness of its internal governance and safety processes. ^{[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11]}