Tom Brown

Overview

Tom Brown is an AI researcher whose name appears on several widely read papers in large language models and learning from human feedback. Wikipedia lists him among the co-founders of Anthropic ^[5]. He is best known as the first author of "Language Models are Few-Shot Learners", the May 2020 paper that introduced GPT-3 ^[1].

Learning from human preferences

One of his earlier co-authored papers is "Deep reinforcement learning from human preferences" from 2017 ^[3]. The problem that line of work addresses is practical: for many tasks it is easier for a person to say which of two behaviours is better than to write down a numeric reward function. The approach is to collect human comparisons between pairs of trajectories, fit a model of the human preference, and then train a reinforcement learning agent against that learned preference model instead of a hand-specified reward. This idea of learning a reward from human judgements later became a standard ingredient in aligning language models with what people want.

Scaling laws and GPT-3

In January 2020 he was a co-author of "Scaling Laws for Neural Language Models", submitted to arXiv on 23 January 2020 ^[2]. That paper studied how the performance of a language model changes as model size, dataset size and the amount of compute used for training grow, and described regular patterns in that relationship. Such regularities are useful because they let researchers estimate, before spending the compute, roughly what a larger model will achieve.

He then led the author list of "Language Models are Few-Shot Learners", submitted to arXiv on 28 May 2020, where the arXiv record lists him first among 31 authors ^[1]. The paper introduced GPT-3 and argued that a sufficiently large language model can perform a new task from only a handful of examples placed in its prompt, without any change to its weights — the setting known as few-shot learning. Instead of training a separate model for each task, the same model is given a short description and a few demonstrations and then asked to continue. The paper reported experiments across many tasks using this in-context approach ^[1].

Anthropic and Constitutional AI

He is also a co-author of "Constitutional AI: Harmlessness from AI Feedback", submitted to arXiv on 15 December 2022, where the record lists him among 51 authors ^[4]. The method replaces some of the human labelling in preference-based training with feedback from the model itself, guided by a written set of principles — the "constitution" — so that a model can be trained to decline harmful requests with far less human annotation of harmful content. Separately, Wikipedia lists him among the people who co-founded Anthropic ^[5], the organization with which that work is associated.

Why it matters

The thread running through these papers is a shift in how language models are built and steered: from training a model for one task, to prompting one large model with examples; from hand-written rewards, to rewards learned from human or AI judgements. Brown's contributions sit at each of these steps — learning from preferences, describing how capability grows with scale, demonstrating few-shot behaviour at scale, and reducing the human cost of making a model safer ^[1][2][3][4].