The 2020 GPT-3 paper is the source of every few-shot tip

Start here

Almost every prompting tip that says 'show it a couple of examples first' traces back to one documented finding. On 28 May 2020, OpenAI posted Language Models are Few-Shot Learners, describing GPT-3, a 175-billion-parameter model, and reporting that 'scaling up language models greatly improves task-agnostic, few-shot performance.' In plain terms: put a few worked examples of a task in the prompt itself, with no retraining, and the model gets noticeably better at that task. That mechanism, examples in the prompt rather than examples used to retrain the system, is what every 'give it a template' prompting tip is exploiting.

What the documents say

The paper reports GPT-3 at '10x more than any previous non-sparse language model,' and states it can generate news articles humans have difficulty distinguishing from ones written by people -- a capability the authors flag as a concern, not only a triumph. It also says plainly that 'GPT-3's few-shot learning still struggles' on some datasets, and notes methodological issues from training on large, uncurated web text. OpenAI's own OpenAI API announcement of 11 June 2020 turned the research into a product: a general-purpose 'text in, text out' interface a developer could 'program' by 'showing it just a few examples of what you'd like it to do.' That post says the API launched in 'private beta rather than general availability' and that OpenAI would 'terminate API access for obviously harmful use-cases.'

Check this

You can check the few-shot mechanism on almost any current chat tool: ask for something in an unusual format with no example, note the result, then repeat the request after pasting in two or three examples of the output shape you want. The shift between those replies is the paper's finding at work -- pattern-matching to examples already in view, not the model learning something new about the world. If the content is still wrong after the format improves, that is the paper's own distinction showing itself: formatting help and factual reliability are separate things.

What holds and what fails

This is an editorial extension of what the sources report: few-shot examples reliably steer tone, structure, and format, since that is a pattern the model can match within a single prompt, and this holds across most everyday writing tasks. It fails as a fix for missing facts, since the paper itself reports uneven performance across tasks even at scale, and no number of formatting examples supplies information the model was never trained on. The boundary is simple: use examples to control how an answer looks, and a separate check, a source or a second opinion, to control whether it is true.

• Try a request with zero examples, then with two or three, and compare only the format, not your trust in the content.
• Read the paper's own limitations before assuming scale alone fixes every kind of task.
• Keep formatting examples and factual verification as two separate steps.

The API announcement's own phrase, an interface you 'program' with examples, is a reminder that a prompt is closer to a template than a conversation with a colleague who already knows what you mean.