GPT-3 shows scale alone unlocks few-shot learning
175 billion parameters let a model learn new tasks from a handful of examples in the prompt
Quick facts
- Institution
- OpenAI
- Parameters
- 175 billion
- Key capability
- Few-shot learning from prompts, no fine-tuning
What happened
Tom Brown and over thirty co-authors at OpenAI published 'Language Models are Few-Shot Learners,' introducing GPT-3, an autoregressive Transformer with 175 billion parameters, ten times larger than any earlier non-sparse language model. Rather than fine-tuning the model's weights for each new task, the paper showed that simply describing a task in the prompt and providing a small number of examples, so-called few-shot learning, let GPT-3 perform competitively on many benchmarks with no gradient updates at all. Performance on these tasks improved smoothly as the model scaled up, without any architectural changes, suggesting scale itself was driving the new capability. In September 2020, Microsoft announced an exclusive license to GPT-3's underlying model for its own products, while OpenAI continued to offer the model to others through an API.
Why it matters
GPT-3 shifted the field's dominant strategy toward scaling existing architectures with more data and parameters rather than inventing new ones, and its few-shot prompting behavior became the basis for how people would later interact with ChatGPT and other large language models through natural-language instructions alone.
How we know
The original arXiv paper documents the 175-billion-parameter scale and few-shot results directly; Microsoft's own blog post on its exclusive license independently confirms the model's scale and commercial significance at the time.
Sources
- Tom B. Brown, Benjamin Mann, Nick Ryder, et al., OpenAI. Language Models are Few-Shot Learners · Primary source (author-declared)arxiv.org · Cited as a "primary" source (no stronger domain match). · Link is live and its text matches the event's key terms (Jul 2026)
- The Official Microsoft Blog. Microsoft teams up with OpenAI to exclusively license GPT-3 language model · General sourceblogs.microsoft.com · Cited as a "reference" source (no stronger domain match). · Link is live and its text matches the event's key terms (Jul 2026)
See something wrong? . Corrections with a source get fixed fastest.
Part of a timelineArtificial Intelligence30 events · From a wartime theory of neurons to machines that write, paint, and fold proteinsView all →