3.1. Introduction
We interact with generative artificial intelligence by giving a language model (ChatGPT, GEMINI, CLAUDE, DALL-E, etc.) a short instruction or description to guide the responses it generates. This instruction or description is called a prompt.
By providing a specific prompt, we can influence the type of responses generated by the model. The choice of prompt is decisive, as it directly affects the nature and quality of the response generated. Experimenting with different prompts can help you obtain more accurate and useful results for the task you want to carry out.
Example of the procedure used by a GAI model to respond to a text prompt
Once the GAI model has been trained, generating a text response to a written prompt involves the following steps:
- The prompt is divided into smaller units (called tokens).
- Statistical patterns are used to predict likely words or phrases that could form a coherent answer to the question.
- Patterns of words and phrases found in the big data model (stored in neural networks) are identified.
- GAI uses these patterns to calculate the probability that specific words or phrases will appear in a given context and generate predictions.
- Starting with a random prediction, GAI uses these probabilities to predict the next likely word or phrase in its response.
- Predicted words or phrases become readable text.
- The readable text is filtered to remove any offensive content.
- These final steps are repeated until the response is complete.
- The answer is reprocessed to improve it (format, punctuation, etc.).