# Zero-Shot / Few-Shot **Zero-shot:** The model performs a task it has never been explicitly trained or given examples for — you simply describe what you want, and the model does it. "Translate this sentence to Japanese" with no translation examples provided. The model's pre-training on internet text gives it enough general capability to handle novel tasks from description alone. **Few-shot:** The model is given a small number of examples (typically 1–5) of the desired input-output pattern, then asked to continue the pattern with a new input. "Q: What is 2+2? A: 4. Q: What is 3+5? A: 8. Q: What is 7×3? A:" The model infers the pattern from examples and extrapolates. ## Why the terminology matters Zero-shot and few-shot capability is what distinguishes [[Foundation Model]] from all previous AI systems. Traditional machine learning required extensive task-specific training: thousands of labeled examples to teach a model to classify emails, translate languages, or identify objects. Foundation models can perform these tasks zero-shot — from a natural-language description alone. This shift is what made foundation models commercially viable. A single model that can handle any text task zero-shot replaces dozens of task-specific models, each requiring its own training pipeline, labeled data, and maintenance. The vocabulary of "zero-shot" and "few-shot" captures this capability transition. GPT-3's few-shot capability (demonstrated in its 2020 paper) was the result that convinced the AI research community that scaling was the path forward — a large enough model could learn to perform novel tasks from just a few examples, without any fine-tuning. This was the empirical foundation for the [[Scaling Laws]] investment thesis that has driven hundreds of billions of dollars in AI development. ## See Also - [[Foundation Model]] — the paradigm zero-shot capability defines - [[Prompt Engineering]] — how few-shot examples are used in practice - [[Fine-Tuning]] — the alternative when zero/few-shot isn't sufficient