Papers · GPT-3
AI PAPER

Language Models are Few-Shot Learners

Demonstrated that a 175B parameter language model can perform tasks from just a few examples in context, establishing in-context learning as a new paradigm and proving the scaling hypothesis.

Authors
Tom Brown, Benjamin Mann, Nick Ryder +7
Institution
OpenAI
Published
May 28, 2020
Citations
40,000
Category
Scaling
Impact
foundational
PAPER EXPLAINED

What This Paper Does

GPT-3 scaled autoregressive language modeling to 175 billion parameters, trained on a filtered version of Common Crawl plus curated datasets totaling roughly 300 billion tokens. The paper’s central experiment was in-context learning: providing the model with a few examples of a task within the prompt, without any gradient updates, and measuring performance. Across dozens of NLP benchmarks, this few-shot approach often matched or exceeded fine-tuned smaller models.

Why It Matters

GPT-3 proved the scaling hypothesis. The relationship between model size and capability was not merely incremental but qualitative. Behaviors that were absent in smaller models emerged reliably at sufficient scale. Few-shot learning, arithmetic reasoning, and code generation appeared as emergent capabilities that no amount of training a smaller model could replicate.

The paper also introduced the API-as-a-product model for AI. Rather than releasing weights, OpenAI provided GPT-3 through a commercial API, establishing the business model that would eventually produce ChatGPT and generate billions in revenue. This decision shaped the industry’s economic structure.

Key Contributions

In-context learning demonstrated that large language models can adapt to new tasks without weight updates. By simply conditioning on a few examples prepended to the input, the model generalizes to new instances. This mechanism remains poorly understood theoretically but has proven extraordinarily practical.

The paper systematically documented how capabilities scale with model size across three regimes: zero-shot (no examples), one-shot (one example), and few-shot (several examples). Larger models showed disproportionate improvement in few-shot settings, suggesting that scale unlocks qualitatively different capabilities.

What Changed Because of It

GPT-3 triggered the large language model arms race. Google, Anthropic, DeepMind, Meta, and dozens of startups began training models at similar or larger scale. Global investment in AI compute infrastructure accelerated from billions to tens of billions annually. NVIDIA’s market capitalization would increase more than tenfold in the years following.

The paper also demonstrated that language models could be commercially valuable without fine-tuning. The GPT-3 API launched in June 2020 and quickly attracted thousands of developers building applications from copywriting assistants to code generators. This proved the market existed for foundation model APIs.

Connection to Current Models

GPT-3’s architecture and training approach remain the foundation of the GPT series. ChatGPT added instruction tuning and RLHF on top of GPT-3’s base capabilities. GPT-4 scaled further with multimodal inputs. The in-context learning paradigm GPT-3 established is how most people interact with AI systems today: providing context and examples within a prompt rather than training custom models.