Generative artificial intelligence, particularly large language models (LLMs) like ChatGPT, can be a powerful tool when wielded appropriately, especially given the breadth of its potential application to various jobs, tasks, and contexts. Prompting these models effectively and efficiently is a skill that can greatly increase productivity and marketability.
In his recent white paper, Professor Kennedy shares his approach of using personas to create group interactions and dynamics to deliver practical advice akin to insights and recommendations that an advisory board may provide in the business world.
What Generative AI is being used for in:
Generative AI is revolutionizing diverse sectors. The following are some examples:
- The classroom: topic-generation, drafting, and initial research for writing assignments along with exposing students to LLMs to gain experience and supplement the research process with AI.
- Professional settings: breaking barriers to entry for jobs where good writing is essential. Generative AI helps organize thoughts and level the playing field for non-native English speakers.
- Government: summarize Congressional hearings. Boosts efficiency of legislative office aides and interns by distilling transcripts of lengthy and numerous meetings into digestible summaries. (See this example from Marci Harris)
- Legal practice: generate potential arguments from opposing counsel and initial evaluations of cases and theories.
What is prompting:
Prompting is the language of interaction with AI models:
- How users interact with the models, which can take the form of questions, instructions, or creating examples for the AI to follow, to name just a few.
- The models respond to text input with a continuation of what is most likely to come next, based on the data they were trained on.
- Prompting is a skill rather than a science. Crafting specific, descriptive prompts often leads to more useful results, but can take some trial and error.
- Ultimately, prompting is a process. It may take several entries to hone the prompt into a desired output. In fact, the models will advise how best to prompt them if asked.
Using AI as your advisory board:
- Persona-Context-Request-Output (PCRO): a basic approach for higher-quality inquiries
- How it works: Give the AI a persona, or perspective, from which to frame the response. Next, provide context that further narrows how the inquiry should be filtered by the AI. Then, make a request followed by a specific output that details how the result should be structured.
- The prompt: “Assume that you are an experienced and expert personal trainer with specialized training in low-impact exercises and exercise as part of a weight loss program (Persona). Also assume that I am a reasonably fit older adult male who exercises regularly (Context). What would be a great low-impact workout regimen to help me lose weight? (Request) Create a recommended workout plan for the next thirty days that includes reasonable objectives (Output).”
- Group Advisory Layer (G-A-L) Method: this modified PCRO method is useful for solving the “blank page” problem by generating ideas to get a project started.
- How it works: Create, or have AI generate, between 5-8 distinct and diverse personas that will be used to frame how the AI will respond to prompts. Then, provide context and specify a type of group interaction: round-robin discussion, debate, or devil’s advocate, for example. Depending on the interaction type, ask the AI to generate multiple rounds of responses; typically, about 6 rounds produces more substantive ideas to work with. However, over-generating can lead the AI to irrelevant or strange topics.
Keep in Mind
- Conducting independent research and fact-checking is important because the current LLMs can generate inaccurate information, are limited by the information they were trained on, and don’t fact-check themselves.
- For instance, ChatGPT 3.5 (the free version) was trained on information up to 2021, so prompting it using ideas, research, or information after 2021 may lead to slightly less accurate results. On the other hand, a paid subscription to GPT Plus comes with plugins that extend the search to real-time information on the web.
- The data sent to these services is typically aggregated and de-identified before it is used to further train the models. As a general rule, personal and confidential information should not be entered. Familiarizing yourself with the privacy policy of whatever platform you use is always a good idea.
Source: You can learn more about Professor Dennis Kennedy’s approach in this article.