AI Blog: Advanced Prompting Part 1

Welcome back to our weekly AI blog. Today, we start exploring Shot-Based prompting.

After learning about the basic frameworks and iteration method of prompting, today we are going to introduce [3] fundamental types of prompting.  

Let’s begin with the simplest one, Zero-Shot Prompting.

1. Zero-Shot Prompting

This technique refers to situations where you expect an AI model (e.g. ChatGPT), to perform a task without providing any examples or additional context.  The output relies solely on the AI’s pre-trained knowledge.  For example, if you have a straightforward or simple question, you can directly pose it to the AI without further explanation.

As you can see, zero-shot prompting is highly effective for simple queries due to its quick response and straightforward nature.  However, this approach has obvious drawbacks, particularly for complex or nuanced tasks, as it might generate overly general or inaccurate results.

2. One-Shot Prompting

When dealing with more complicated tasks where you expect specific and precise outcomes, one-shot prompting is a more suitable option. 

One more step than Zero-Shot prompting, this requires at least one example to guide the AI’s understanding of your expectations.  As we mentioned before, large language models (LLMs) excel at pattern recognition.  By offering the AI even a single clear example of the desired output, you can easily get the result without writing long sentences to describe your requirements, showing details or specifying the preferred output format.

This technique significantly helps the AI in understanding the nature of your task and directs it more effectively towards your intended result.  However, for highly nuanced tasks, the results may still fail to meet all your expectations…

3. Few-Shot Prompting

Few-shot prompting is a more sophisticated and thus a more reliable method, requiring multiple examples (usually two or more) without explicitly specifying the desired style. 

Usually, the more relevant examples provided, the better the AI model's comprehension of the task, leading to more accurate and consistent outputs.  One thing that needs to be noted is that providing too many examples can sometimes restrict the AI's creativity and flexibility in responses. 

Few-shot prompting offers the best results for this kind of complicated task requiring detailed accuracy and a customised plan.  We can upload the example files, paste the links and so on, to ensure clarity without overly limiting the model’s ability to generate expected responses.

Join us next time for a look at what comes with prompting like a pro!