As part of her insightful talk during December’s AI Meet-up, Meike Nauta provided us with some hands-on tips in steering a ‘black box’ like ChatGPT, to generate output that aligns with your objectives. As a guideline, Meike likes to use the PREPARE model (from the AI Educator):
P = Prompt. Start with what you want from the language model, for instance: “Write a summary about the history of AI in Dutch hospitals.”
R = Role. Give it a certain role, for example, if you need a medical summary, add “Act as if you are a medical expert.”
E = Explicit. Clearly explain what you are looking for in the answer, to get a better response: “Include how AI can support human doctors.”
P = Parameters. Set the right parameters and criteria: “Use an informal tone and structure it with bullet points.”
A = Ask. Ask the AI to pose clarification questions to you before proceeding, to steer the language model even further in the right direction: “First ask me some clarification questions, and then answer.”
R = Rate. Ask the AI to evaluate its output and ask for the thought process, to get a feeling of the quality of the answer: “How good do you think this output is? What could be improved?”
E = Emotions. If you add emotion to the prompt, the performance will be better, as it has probably seen these types of sentences in training data: “Please try your best, it is important to me and I greatly value your analysis.”
Using these seven points in your prompt will help you guide your model through the training data and get the answer you need.
Concluding her talk, Meike explained that she sees hybrid intelligence as the future, in which human intelligence (intuition, moral compass, common sense) and artificial intelligence (speed and quantity) are merged. Eventually, combining AI and humans will make us smarter.
Reach out to Meike or Datacation if you’d like to know more about large language models such as ChatGPT, or how to best work with it as an organization.
