Effective conversations with generative AI models (II)

February 19, 2024

In the previous post we learned how generative AI models, such as GPT, Copilot or Bard, represent a significant advance in natural language processing and understanding.

As we saw, getting accurate and relevant answers from these models depends largely on the way we communicate with them and the prompts we use to explain our needs and guide their response, allowing them to be creative in order to generate quality content.

Having notions about their fundamentals helps to simplify the formulation of questions and allows us to explore and learn about the capabilities of these advanced models. Also to better understand how they work, what they can do, what they cannot do, what is better not to tell them.

Productive conversations with Copilot and ChatGPT

Microsoft Copilot (Bing Chat)

Asking the right questions influences the quality of the response. Microsoft proposes seven recommendations for building effective instructions when conversing with Copilot, including:

Explain the desired outcome and the type and format of the expected response, including examples.
Provide the model with context and relevant details and explain the goal and purpose of the request.
Indicate a tone for the text and the intended audience of the response and request further information or clarification of your response whenever necessary.
Maintaining a polite and curious interaction to encourage a dynamic and interactive conversation with the model, providing constructive feedback and acknowledging helpful responses.

Providing context is key to getting accurate responses.

OpenAI's ChatGPT

OpenAI proposes six tactics for building better prompts for ChatGPT, including:

It recommends placing instructions at the beginning of the prompt and using markers such as ### or """ to distinguish instructions and context. This makes it easier for the model to accurately understand the requested task.

>> ###Summarize the following text in key points: "Artificial intelligence is transforming…"|

It agrees in emphasizing the need to be specific, descriptive, and detailed in the instruction, both in context, expected result, length, format and style. This helps the model to generate responses that are more aligned with the user's needs.

>> Write a 500-word article on…|

Provide specific examples of the desired output format.

>> Write a poem following this format: Title: The Star and the Moon Stanzas: 3 Verses per stanza: 4 Rhyme: ABAB Theme: The nocturnal nature|

It is recommended to start with the Zero-Shot approach, which does not use previous examples,

>> Generate a list of tips for improving personal productivity|

And if necessary move towards the Few-Shot approach, with examples to guide the model

>> Rank the feeling of these reviews: ###Example 1: Review: "I loved this restaurant. The food was delicious and the service exceptional." Feeling: positive ###Example 2: Review: "Disappointing experience. I expected more variety on the menu and better taste." Feeling: Negative Review to rate: "Prices are a bit high, but the quality of the food justifies it. I would go back again." Feeling: [Your answer] |

It advises avoiding ambiguous descriptions and providing clear, concise instructions to increase the accuracy of responses.

>> Write an executive summary for a market report. The summary should be exactly 150 words, focusing on the report's main findings and recommendations|

It agrees that it is more effective to provide positive and specific instructions on what is expected than to indicate what to avoid.

Less effective ❌:

>> When writing the article for the blog, don't use technical lingo or complicated terms. Don't make the article too long or include opinions |

More effective ✅:

>> Please write an 800-word article for our blog about current trends in technology. Use clear and accessible language for a general audience |

Conclusion

All recommendations underline the importance of being clear and concise with instructions. It is also important to structure prompts appropriately to improve interaction with generative AI models.

Properly constructing instructions and prompts effectively improves the quality and accuracy of responses.

Applying these rules and recommendations helps to simplify question formulation, examine, and learn about the capabilities of these advanced models, and better understand both what they can and cannot do.

Practicing and mastering prompting not only enriches the user-IA interaction, but also expands its usefulness to the user, the possibilities and more creative and efficient use of generative AI models.

◾ CONTINUING THIS SERIES