Jack Clark

Who he is

Jack Clark is a co-founder of Anthropic and has held two senior roles at the company: head of policy, and later head of public benefit.^[1] In March 2026 he began leading the Anthropic Institute, a think tank launched by Anthropic to study AI.^[1] He is also a co-author of the GPT-3 paper, "Language Models are Few-Shot Learners", which was written at OpenAI.^[2]

The GPT-3 paper

The paper is usually referred to simply as the GPT-3 paper. It was first submitted to arXiv on 28 May 2020, and the latest version listed there, v4, was revised on 22 July 2020.^[2] Jack Clark is among the listed authors.^[2]

The title names the idea the work is built around: few-shot learning. In the most familiar way of training a model, you change its parameters by showing it many examples, so that it gradually gets better at a task. Few-shot learning takes a different route at use time. A handful of examples of the task — a few input-and-output pairs — are placed inside the prompt itself, and the model is asked to continue the pattern for a new input. Nothing in the model's weights changes; the examples do the work of telling it what is wanted. The approach matters because it turns a single general-purpose model into something that can be pointed at many different tasks by someone who cannot retrain it, and because it makes the model's behaviour on a task depend heavily on how the prompt is written.

Policy and public benefit

At an AI company, a policy role generally means the part of the organisation that follows and takes part in the rules, standards and public argument surrounding the technology, rather than the part that builds systems. A public-benefit role sits nearby: it is concerned with how the company's work is meant to serve people outside it, and with the commitments the company makes about that. Clark held the first of these roles at Anthropic and then the second.^[1]

Both positions are boundary roles. They require understanding what the technology actually does well enough to describe it accurately to people who do not build it, and they put someone from the research world into regular contact with the institutions — regulators, standards bodies, civil society — that shape how a technology is received.

The Anthropic Institute

A think tank is an organisation whose product is study rather than software: it examines a subject in depth and publishes research, analysis and recommendations. In March 2026 Clark began leading one such organisation, the Anthropic Institute, which Anthropic launched to study AI.^[1] The move places a co-founder of the company at the head of a body whose stated subject is the technology itself, rather than any particular product built with it.

Where he sits

Clark's record combines two kinds of work that are often kept apart. On one side is authorship of a paper that introduced a way of using large language models; on the other are the policy and public-benefit functions inside an AI company, and now the leadership of a research institute devoted to studying AI.^[1][2] Careers of this shape — from technical co-authorship to governance and institutional work — have become more common as AI systems have moved from research laboratories into general use, and they are one of the ways the field's technical and institutional conversations stay connected to each other.