Sam McCandlish is a theoretical physicist turned AI researcher who co-founded Anthropic in January 2021 alongside Dario Amodei, Daniela Amodei, Jack Clark, Chris Olah, Jared Kaplan, Tom Brown, and Ben Mann. The founding group left OpenAI citing concerns about the pace of AI commercialization relative to safety research.

McCandlish holds a PhD in theoretical physics from Stanford University, with prior degrees in mathematics and physics from Brandeis University. He conducted postdoctoral research at Boston University before joining OpenAI in 2018.

At Anthropic, McCandlish served successively as Co-Founder and Chief Scientist, then as Chief Technology Officer, and — following an executive restructuring in October 2025 — as Chief Architect, where he concentrates on pre-training and large-scale model training. His body of research, spanning AI scaling laws, mechanistic interpretability, and AI alignment techniques, has accumulated over 100,000 citations.

Anthropic's valuation reached $380 billion in February 2026 and rose further to approximately $965 billion by May 2026. Forbes estimates each of the seven co-founders holds approximately 1.6% of the company.

Business Relationships

McCandlish works closely with Dario Amodei, CEO and co-founder of Anthropic, with whom he collaborated at OpenAI on large-batch training and scaling laws research. His research partnerships include Jared Kaplan, with whom he co-led the landmark scaling laws paper and the GPT-3 paper, as well as Tom Brown, who is also an Anthropic co-founder.

His interpretability research intersects with work by Chris Olah, another Anthropic co-founder whose mechanistic interpretability focus connects directly to McCandlish's contributions to Toy Models of Superposition. Anthropic has major investment partnerships with Amazon and Google's parent Alphabet.

Ownership

McCandlish is a co-founding shareholder in Anthropic, with Forbes estimating each of the seven co-founders holds approximately 1.6% of the company. Anthropic is structured as a Public Benefit Corporation governed by a Long-Term Benefit Trust, a legal arrangement designed to insulate board decisions from ordinary shareholder pressure.

As Chief Architect since October 2025, McCandlish concentrates on pre-training and large-scale model training, reporting to President Daniela Amodei. A new CTO, Rahul Patil (formerly of Stripe), was appointed to oversee compute infrastructure and engineering.

Family & Heirs

McCandlish keeps his personal family life out of public view.

Evidence

2018-12-01 - 😇 - McCandlish leads 'An Empirical Model of Large-Batch Training,' establishing gradient noise scale to optimize neural network training efficiency

2019-12-05 - 😇 - McCandlish co-leads 'Scaling Laws for Neural Language Models,' predicting AI capabilities based on compute, data, and model size

2020-05-01 - 😇 - McCandlish co-authors 'Language Models are Few-Shot Learners' (GPT-3), a landmark AI research paper with over 78,000 citations

2022-09-01 - 😇 - McCandlish co-authors 'Toy Models of Superposition,' advancing mechanistic interpretability research for AI safety

2022-12-01 - 😇 - McCandlish co-authors 'Constitutional AI: Harmlessness from AI Feedback,' introducing safety training methodology that reduces harmful AI outputs without human labels

2023-09-21 - 😇 - McCandlish personally designs a board-approval requirement into Anthropic's original Responsible Scaling Policy to guard against safety-test bias

2024-08-15 - 😈 - Anthropic successfully lobbies California to weaken SB 1047 AI safety enforcement, limiting attorney general ability to sue before critical harm occurs

2024-10-15 - 😇 - McCandlish served as Anthropic's Responsible Scaling Officer, personally overseeing the Responsible Scaling Policy implementation for over a year

2025-04-24 - 😇 - Anthropic launches model welfare research program under McCandlish's CTO tenure to investigate potential AI consciousness and moral consideration

2025-09-05 - 😈 - Anthropic settles for $1.5 billion after training Claude on pirated books from LibGen and Pirate Library Mirror without author consent

2025-11-29 - 😈 - Analysis documents McCandlish made misleading public statements about Anthropic's secret non-disparagement agreements in employee contracts

2026-01-27 - 😇 - McCandlish joins all seven Anthropic co-founders in pledging to donate 80 percent of personal wealth to combat AI-driven inequality and wealth concentration

2026-02-25 - 😈 - Anthropic abandons core safety promise, replacing binding Responsible Scaling Policy with non-binding framework citing competitive pressure

Analysis

McCandlish's career represents one of the most technically consequential AI researcher trajectories of his generation. The scaling laws paper he co-led at OpenAI predicted the emergence of large language model capabilities before they were built, directly informing the development of GPT-3 and subsequent models. His large-batch training work established the mathematical framework for how compute is allocated across massive AI training runs.

The tension in McCandlish's record is between genuine contributions to AI safety and institutional decisions that undermined them. He co-authored Constitutional AI and Toy Models of Superposition, advancing both alignment techniques and interpretability research.

When Anthropic first published the Responsible Scaling Policy in 2023, McCandlish personally designed its board-approval requirement for policy changes specifically to guard against the temptation to make Anthropic's own safety tests too easy. He went on to personally lead that policy for over a year as its Responsible Scaling Officer.

Yet in February 2026, Anthropic leadership, including McCandlish as co-founder and Chief Architect, abandoned the binding version of that policy, citing that "shortcomings in its two-year-old Responsible Scaling Policy could hinder its ability to compete." The USD 1.5 billion copyright settlement — stemming from Anthropic training Claude on pirated books from LibGen and Pirate Library Mirror — also occurred during the period when McCandlish led technical model development as CTO.

McCandlish's 80% wealth pledge is one of the largest philanthropic commitments made by an AI researcher. His model welfare research initiative demonstrates a willingness to engage seriously with questions about AI welfare that most in the industry have not pursued.

Questions

What specific deployment of the 80% wealth pledge is planned — which causes, timelines, and governance structures?
How does McCandlish's role as Chief Architect influence Claude model development given that external safety commitments were relaxed in February 2026?
What was McCandlish's specific role in the decision to train on pirated book datasets, given his direct oversight of model training as CTO?

Redemption Arc

McCandlish's 80% wealth pledge and the establishment of Anthropic's model welfare program represent meaningful corrective actions. Whether the pledged funds will be deployed toward addressing AI-driven inequality and harm, and whether Anthropic's safety commitments will be restored after the February 2026 rollback, remain the primary open questions.