23Aug02 | How to talk to ChatGPT

Principles

Principle I: Write clear and specific instructions
Principle II: Give the model time to “think”

‣

Principle I:

‣

Principle II:

How to reduce model hallucinations?

Find relevant info/quotes from the text
Ask it to use those quotes to answer questions

Iterative Prompt Development

Iteratively analyze and refine your prompts:

Try something
Analyze where the result doesn’t give what you want
Clarify instructions, give it more time to think
Refine prompts with a batch of examples

Example:

‣

1/ Generate a marketing product description

‣

2/ Limit the number of words/sentences/characters

‣

3/ Ask it to focus on aspects relevant to intended audience

‣

4/ Ask it to extract information and organize it in a table

In practice, you’d end up with a prompt like this only after multiple iterations. No way around it.

‣

Rendered HTML output for the last prompt:

Summarizing

Several techniques to get the most out of GPT’s summarizing capabilities:

‣

1/ Summarize with a word/sentence/character limit

‣

2/ Summarize with a focus on shipping and delivery

‣

3/ Summarize with a focus on price and value

‣

4/ Try "extract" instead of "summarize"

‣

5/ Batch

Inference

In just a few minutes you can make multiple systems for making inferences from text that would’ve taken many days or weeks for a skilled ML developer!

‣

1/ Infer Sentiment

‣

2/ Infer Emotions

‣

3/ Infer Information

‣

4/ Infer Topics

Note: Output of LLMs is pretty inconsistent, so to be able to use its output in a prod env, we’d need to force a JSON output, and then parse it.

Transforming

‣

1/ Tone Transformation

‣

2/ Format Conversion

‣

3/ Proofread & Correct

Also there’s translation!

Expanding

‣

Generate a longer piece of text from a small one

Temperature

‣

Represents the degree of exploration / randomness of the model.

23Aug02 | How to talk to ChatGPT | For Devs