A plain summary, so you can get the gist here without leaving.
In 2020, OpenAI showed that a very large language model with 175 billion parameters could perform new tasks just from a few examples written into the prompt, without any retraining.
What it is
GPT-3 is a large language model trained on a vast amount of text to predict the next word. What made it notable was not only its size but a new way of using it. Instead of fine-tuning it for each task, you simply describe the task and show a few examples in the prompt, and the model follows along.
This is called few-shot learning. The model learns what you want from the handful of examples sitting right there in the input, then applies the pattern to your real question.
The core idea
A model trained broadly enough can pick up the shape of a new task on the fly. Show it a couple of translations or a couple of question-and-answer pairs, and it infers the rule and continues it, all without changing its internal weights.
This flips the older workflow. Before, adapting a model meant gathering labeled data and retraining. With few-shot prompting, you adapt the model by writing good instructions and examples, which is far faster and more flexible.
Why it matters
GPT-3 made prompting the main way people interact with large models. The idea that you can steer a single general model with words alone is the basis of how most people use AI today.
For builders, this is the moment the prompt became the interface. Knowing how to give clear instructions and good examples is now a core skill, and this paper is where that approach was shown to work at scale.
- Published in 2020 by OpenAI.
- A 175-billion-parameter model trained to predict the next word.
- Performs new tasks from a few examples placed in the prompt.
- No retraining needed, the model adapts from the prompt itself.
- Established prompting as the main way to use large models.
OpenAI
New to this? Come build with us.
Reading is good. Building with people is better. Our drop-ins are free and open to total beginners.