Sam McCandlish

Overview

Sam McCandlish is a co-founder of Anthropic, the AI company, and has served as its chief technology officer ^[1][5]. He was also Anthropic's first Responsible Scaling Officer, the person responsible for overseeing the company's Responsible Scaling Policy ^[1].

His public record is mostly a record of papers and roles rather than of interviews or speeches: he appears as an author on several of the technical reports that shaped how large language models were built and governed in the early 2020s.

Scaling laws

In January 2020 McCandlish was second author, with Jared Kaplan, of "Scaling Laws for Neural Language Models" ^[2]. A scaling law is a description of how some measurable quantity — most often a model's loss, the error it makes on the task it is trained for — changes as the size of the model, the amount of training data and the compute spent on training grow. The practical appeal is predictive: if the relationship between scale and loss is smooth, then results measured at one scale can be used to reason about what a much larger model would do, before that model is built.

That framing turned model development partly into a forecasting problem, and it is one reason later systems were designed at sizes that would once have seemed impractical.

GPT-3 and Constitutional AI

McCandlish was a co-author of the GPT-3 paper, "Language Models are Few-Shot Learners", published in 2020 ^[3]. The paper's contribution statement credits him and Jared Kaplan with predicting that a giant language model should show continued gains — that is, with the bet that simply making the model much larger would keep paying off rather than flattening out ^[3].

In December 2022 he was a co-author of "Constitutional AI: Harmlessness from AI Feedback" ^[4]. The approach trains a model toward harmlessness using feedback generated by AI against a written set of principles, rather than relying only on human labels for every judgement. It became one of the methods associated with aligning models to stated rules.

The Responsible Scaling Policy

Anthropic's Responsible Scaling Policy is organised around AI Safety Level Standards: graduated safety and security measures that become more stringent as model capabilities increase, starting at ASL-1 for models with very basic capabilities and progressing through ASL-2 and ASL-3 ^[1]. The idea is that the obligations a developer accepts should scale with what a model can do, so that stronger systems carry heavier requirements.

As Anthropic's first Responsible Scaling Officer, McCandlish oversaw that policy ^[1]. On 15 October 2024 Jared Kaplan succeeded him in the role, while McCandlish remained chief technology officer ^[1].

Why it matters

McCandlish's record sits at the junction of two things that are usually discussed separately: how to make models bigger and better, and how to decide when a model is capable enough to require stricter safeguards. His scaling-law and GPT-3 work belongs to the first; his authorship of the Constitutional AI paper and his time as Responsible Scaling Officer belong to the second ^[2][3][4][1].