07 Prompt engineering

A good prompting of the AI should be as detailed as possible, and provide the necessary context. Many users at UiO have already had a year of experience with prompting the UiO Chat GPT. This is a more advanced AI that can remember the conversation. In this workshop, you have practiced deciding the role of the AI. An economist or a pirate. In your future interactions with various AIs, review outputs critically and refine your prompts iteratively. Use the model’s strengths to handle complex tasks by breaking them down into simpler questions, like for instance «what is the best cure of Ibux and Paracetamol, if you have headache?». Use follow up questions. Ask the model to elaborate on something you want to know more about. Context and information is important in order to make use of the AI.

Conversation and the Limits of Computational Architecture

When conversations with AI extend too long, the model might begin to seem confused or produce unusual language. Smaller models with fewer parameters may struggle more to maintain coherence, leading to grammatical deterioration in extended dialogues. As interactions with the AI continue, especially without precise prompting, a model might gradually shift its understanding of entities or topics, resulting in a drift in focus or coherence. This issue is partly due to the inherent limitations of the model’s context window, which dictates how many tokens the AI can process at once. GPT-4.0o, has a current context window of 128 000 tokens for text. When this limit is exceeded, older parts of the conversation may be truncated to accommodate new inputs, potentially causing the AI to lose track of earlier components in the conversation. To mitigate these effects and improve coherence, maintain precision in prompting and summarize longer discussions. A practical solution is to ask the model to sum up the conversation, use this as a starting point for a new dialogue, and then delete the old dialogue, alternatively store it locally for later reference.