# Tom Brown

> ~1990– · AI Researcher
>
> **Recorded contribution:** GPT-3 lead author — "Language Models are Few-Shot Learners"

## How to use this dossier

Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory.

## 1. Historical orientation

AI researcher Tom B. Brown was lead author of the 2020 GPT-3 paper, Language Models are Few-Shot Learners. The paper showed that a 175-billion-parameter autoregressive Transformer could perform many tasks from instructions or demonstrations in its prompt without gradient updates. Public biographical data is limited and name collisions are severe, so this dossier does not repeat the registry's approximate birth year or a mismatched Wikipedia page. GPT-3's contribution was an empirical demonstration that task adaptation could occur inside the context window. The trained parameters remained fixed, while instructions and examples changed the probability distribution over continuations; increasing scale made this behavior useful on a wider set of evaluated tasks.

## 2. The problem inherited

NLP systems usually required task-specific labeled data and fine-tuning, limiting one model's ability to adapt immediately to new text tasks.

## 3. The central contribution

Brown led the GPT-3 paper and its large-scale evaluation of in-context zero-, one-, and few-shot behavior.

## 4. Reconstruct the mechanism

1. Train a decoder-only Transformer to predict the next token across a very large text corpus.
2. At inference, place task instructions or demonstrations in the context window.
3. Condition subsequent token probabilities on that prompt without changing model weights.
4. Evaluate across tasks while comparing scale, prompting conditions, calibration, and failure.

## 5. What changed downstream

- GPT-3 made in-context learning and prompt design major research and product paradigms.
- Its scale accelerated investment in foundation models, APIs, and analysis of emergent capabilities and risks.
- The result shifted interface design toward prompts and demonstrations, encouraged general-purpose model APIs, and motivated research on in-context learning, instruction following, and foundation-model evaluation.

## 6. Attribution, limits, and uncertainty

- The paper has 31 authors plus infrastructure and data contributors; lead authorship is not sole creation of GPT-3.
- Personal identity details beyond paper authorship remain unresolved here. Few-shot benchmark performance does not guarantee truth, reasoning, safety, or contamination-free evaluation.
- Prompt sensitivity, data contamination, memorization, and benchmark selection complicate claims of learning; a fluent completion can imitate a demonstrated format without acquiring the underlying rule.

## 7. Reconstruction lab

Evaluate a small language model on the same classification task with zero, one, and four demonstrations. Shuffle labels and order to test whether apparent learning is robust. Create adversarial demonstrations with superficial correlations and compare rule transfer to format imitation, reporting variance across prompt wording, ordering, and random seeds.

## 8. Evidence trail

- [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165) — NeurIPS / arXiv
- [GPT-3 research release](https://openai.com/index/language-models-are-few-shot-learners/) — OpenAI

---

*Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*