In one of our Monday meetings, our engineer Soon demoed OpenAI’s GPT-3 text generator. He used the example of generating questions about personas using the language model. All of us present at the meeting were quite impressed with the fluency and relevance of the results.
In that same meeting, I suggested that we rotate blog writing responsibility, so that every week there’s one new blog post written by taking turns. Joke’s on me, and I got to be the second in line. Because I’m currently busy with teaching, I thought an AI model could help me.
Hence, I borrowed Soon’s idea, even stealing his prompt, to generate some questions related to personas. For the following outputs, I used OpenAI’s Playground (https://beta.openai.com/playground). As you’re about to see, the results are pretty good.
My initial prompt was “8 questions about personas”. Interestingly, GPT-3 auto-completed this first to “8 questions about personas that you should consider when creating a persona.” before starting to list those eight questions. I generated the questions iteratively five times using this expanded prompt. You can see the results below.
FIRST RUN:
8 questions about personas that you should consider when creating a persona.
- What needs does your persona have?
- What are your persona’s goals?
- What is your persona’s role in their organization?
- What is your persona’s primary motivation?
- How does your persona make decisions?
- What is your persona’s attitude towards risks?
- What is your persona’s level of technical expertise?
- What is your persona’s preferred method of communication?
SECOND RUN:
8 questions about personas that you should consider when creating a persona.
- What are the demographics of your persona?
- What is your persona’s occupation?
- What is your persona’s education level?
- What is your persona’s income level?
- What are your persona’s interests?
- What are your persona’s hobbies?
- What is your persona’s favorite type of entertainment?
- What is your persona’s favorite type of food?
THIRD RUN:
8 questions about personas that you should consider when creating a persona.
- What are the demographics of your persona?
- What is your persona’s occupation?
- What is your persona’s level of education?
- What is your persona’s income level?
- What are your persona’s hobbies and interests?
- What is your persona’s relationship status?
- What is your persona’s religious affiliation?
- What is your persona’s political affiliation?
FOURTH RUN:
8 questions about personas that you should consider when creating a persona.
- What are the persona’s goals and objectives?
- What motivates the persona?
- What does the persona need or want?
- How does the persona behave?
- What does the persona think?
- What does the persona feel?
- What does the persona value?
- What does the persona fear?
FIFTH RUN:
8 questions about personas that you should consider when creating a persona.
- What demographics does your persona represent?
- What is your persona’s age?
- What is your persona’s gender?
- What is your persona’s income level?
- What is your persona’s education level?
- What is your persona’s occupation?
- What is your persona’s interests?
- What is your persona’s lifestyle?
…now, what do we learn from this? A couple of things:
- First, there is a lot of repetition between the runs, e.g., concerning demographics. This is not a bad thing: in contrast, it shows stability in that the model has clearly learned some stable characteristics of questions that most commonly belong to the persona creation process. In fact, one could argue that the reappearing questions are the ones one should focus on, as these keep popping up in the model’s thinking (which is based on the writings of the world).
- Second, new information keeps emerging as well, which suggests that one should rely on multiple runs when interrogating a language model, in order to reach saturation of what can be learned from the model. Below, I’m bolding the new information aspects appearing in each iteration:
FIRST RUN:
8 questions about personas that you should consider when creating a persona.
- What needs does your persona have?
- What are your persona’s goals?
- What is your persona’s role in their organization?
- What is your persona’s primary motivation?
- How does your persona make decisions?
- What is your persona’s attitude towards risks?
- What is your persona’s level of technical expertise?
- What is your persona’s preferred method of communication?
SECOND RUN:
8 questions about personas that you should consider when creating a persona.
- What are the demographics of your persona?
- What is your persona’s occupation?
- What is your persona’s education level?
- What is your persona’s income level?
- What are your persona’s interests?
- What are your persona’s hobbies?
- What is your persona’s favorite type of entertainment?
- What is your persona’s favorite type of food?
THIRD RUN:
8 questions about personas that you should consider when creating a persona.
- What are the demographics of your persona?
- What is your persona’s occupation?
- What is your persona’s level of education?
- What is your persona’s income level?
- What are your persona’s hobbies and interests?
- What is your persona’s relationship status?
- What is your persona’s religious affiliation?
- What is your persona’s political affiliation?
FOURTH RUN:
8 questions about personas that you should consider when creating a persona.
- What are the persona’s goals and objectives?
- What motivates the persona?
- What does the persona need or want?
- How does the persona behave?
- What does the persona think?
- What does the persona feel?
- What does the persona value?
- What does the persona fear?
FIFTH RUN:
8 questions about personas that you should consider when creating a persona.
- What demographics does your persona represent?
- What is your persona’s age?
- What is your persona’s gender?
- What is your persona’s income level?
- What is your persona’s education level?
- What is your persona’s occupation?
- What is your persona’s interests?
- What is your persona’s lifestyle?
As can be seen, in all of the reiterations, new aspects keep emerging, although I’d argue there’s a clear effect of decreasing marginal returns (i.e., the same things keep popping up, with slightly different phrasing). The point is that using the language models several times, rather than just once, we can “sample” the “real” population, in this case what has been written about the topic. This can be quite a powerful technique for multiple use cases, of which there will probably be many in the future.