Sam McCandlish co-founded Anthropic and is its Chief Architect; McCandlish focuses on large-scale model training and leads pretraining, research productivity and RL infrastructure. According to Anthropic's leadership page, McCandlish previously built and led a team at OpenAI focused on the science and scaling of machine learning. McCandlish is first author of "An Empirical Model of Large-Batch Training" (2018), written with an OpenAI affiliation, and an equal-contribution author of "Scaling Laws for Neural Language Models" (2020), whose contribution statement says Jared Kaplan and McCandlish led the research.123
At a glance
- Current role
- Co-Founder and Chief Architect
- Company
- Anthropic
- Sector
- Company
- Fields
- Training systems, Language models
- Education
- Stanford
- Work listed
- 3
- Sources cited
- 11
- Profile reviewed
Career
At OpenAI, McCandlish built and led a team studying the science and scaling of machine learning, whose work on scaling laws Anthropic credits with helping lead to large models such as GPT-3. McCandlish then co-founded Anthropic and served as its CTO until October 2025, when Rahul Patil became CTO and McCandlish became Chief Architect.14
As Chief Architect, McCandlish focuses on large-scale model training, continuing to lead pretraining and overseeing research productivity and RL infrastructure. As Responsible Scaling Officer, McCandlish also helps shape and implement Anthropic's Responsible Scaling Policy.14
Research
McCandlish's OpenAI papers include "An Empirical Model of Large-Batch Training" (2018), on the largest useful batch size, and "Scaling Laws for Neural Language Models" (2020), which McCandlish co-led with Jared Kaplan. McCandlish also co-authored the GPT-3 paper "Language Models are Few-Shot Learners" (2020), "Scaling Laws for Autoregressive Generative Modeling" (2020) and the 2021 Codex paper that introduced HumanEval.23567
At Anthropic, McCandlish co-authored alignment research such as "A General Language Assistant as a Laboratory for Alignment" (2021), "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback" (2022) and "Constitutional AI - Harmlessness from AI Feedback" (2022).8910
Work and contributions
Scaling Laws for Neural Language Models Published research
Equal-contribution author of the 2020 paper on power-law scaling of language-model loss; its contribution statement says Jared Kaplan and McCandlish led the research.3
An Empirical Model of Large-Batch Training Published research
First author of the 2018 paper, written with an OpenAI affiliation, on predicting the largest useful batch size for training across domains.2
HumanEval Open data
Co-author of the 2021 Codex paper that introduced HumanEval; the paper lists McCandlish with Anthropic AI and notes the work was performed while at OpenAI.711
Timeline
- Released "An Empirical Model of Large-Batch Training" as first author, with an OpenAI affiliation.2
- Co-led "Scaling Laws for Neural Language Models" with Jared Kaplan.3
- Co-authored "Language Models are Few-Shot Learners", the GPT-3 paper.5
- Co-authored the Codex paper that introduced the HumanEval benchmark.7
- Co-authored "A General Language Assistant as a Laboratory for Alignment" at Anthropic.8
- Co-authored Anthropic's paper on training a helpful and harmless assistant with RLHF.9
- Co-authored "Constitutional AI - Harmlessness from AI Feedback".10
- Moved from CTO to Chief Architect at Anthropic as Rahul Patil became CTO.4
Other documented roles
Roles the sources document that are not confirmed as current.
Where they work
- Anthropic 34 people in the wiki
Related people
- Tom BrownAlso at Anthropic
- Rohan TaoriAlso at Anthropic
- Samuel R. BowmanAlso at Anthropic
- Nicholas CarliniAlso at Anthropic
- Azalia MirhoseiniAlso at Anthropic
- Christopher RéTraining systems · Language models
- Sami JaghouarTraining systems · Language models
- Arthur MenschLanguage models · Training systems
- Stephen RollerLanguage models · Training systems
- Yonghui WuLanguage models · Training systems
Suggest a correction or removal
Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.
References
11 sources · profile reviewed Oct 21, 2026 · data updated Oct 21, 2026
- Leadership at Anthropic (Co-Founders, full bios)
- An Empirical Model of Large-Batch Training (arXiv 1812.06162)
- Scaling Laws for Neural Language Models (arXiv 2001.08361)
- Rahul Patil joins Anthropic as Chief Technology Officer
- Language Models are Few-Shot Learners (arXiv 2005.14165)
- Scaling Laws for Autoregressive Generative Modeling (arXiv 2010.14701)
- Evaluating Large Language Models Trained on Code (arXiv 2107.03374)
- A General Language Assistant as a Laboratory for Alignment (arXiv 2112.00861)
- Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback (arXiv 2204.05862)
- Constitutional AI: Harmlessness from AI Feedback (arXiv 2212.08073)
- openai/human-eval repository (README citation; MIT License)