An AI-generated 50s comic-book-style robot dressed like a magician holds a wand.

It’s not magic, it’s math

We will start our exploration of generative AI at the beginning…that is, the beginning of an interaction with a generative AI tool. The prompt.

Prompt engineering–defined as “the process of structuring text that can be interpreted and understood by a generative AI model”–rose to prominence in 2023 as humans grappled with how to most effectively generate outputs from generative AI. Being able to effectively prompt and interact with generative AI is increasingly seen as a marketable skill and is even emerging as a field of study and work.

Why does a generative AI prompt matter? Part of the explanation for why prompting matters connects to the title of this activity, “it’s not magic, it’s math.” It feels like magic when a generative AI tool responds to your question with a helpful and human-sounding response, but really it’s just fancy math.

Generative AI creates human-sounding language by analyzing massive amounts of text with complex formulas and making predictions about what word should come next. Your cell phone’s auto-complete feature is a very simple version of this. That software guesses what comes next based on a few simple factors (your text history, the previous word, a few letters). LLMs use hundreds of factors to decide what word comes next. Prompting generative AI well helps improve context and scope for the mathematical inferences that generative AI is making to create sentences.

What is involved in good prompting? Being specific in your prompt and providing examples of good words or phrases or ideas can reduce the number of assumptions or jumps in logic that the tool has to deal with; this will, in turn, help the tool to use its computational resources more effectively. Some folks recommend using the RTRI approach to prompting:

Role – Tell the tool what or who to act as
Task – Give a summary or description of what the AI needs to do
Requirements – Add requirements that specify what the output should include, contain, be, etc.
Instructions – Include what the AI should do to act on or complete the prompt

Some experts recommend using chain-of-thought prompting which guides the generative AI tool through what you want in a step-by-step way. Some experts argue that adding more emotional language and entreaties will significantly improve a generative AI tool output, saying that “the emotional words you are using will trigger correspondence to patterns associated with how humans use words. In turn, the generative AI will leverage those patterns and respond accordingly.” There is no magical approach to prompting, so try a few approaches and see if you can identify what works best for you and what you need.

How much does your prompt matter? Generative AI models have an emergent ability for in-context learning, meaning that they “learn” from information provided in prompts and subsequent interactions within a conversation. As you’ll see in this week’s activity, a good first prompt is helpful but interaction and feedback is even more important. As you interact with a generative AI tool, you can guide how and what the tool learns from you and help it to improve its outputs.

Let’s get to this week’s activity.

Activity: Write a poem with AI

Step 1: Pick the generative AI tool you want to use.

Some suggestions: ChatGPT 3.5, Bard, Claude 2.

If you are not interested in using a generative AI tool, or if you have concerns about your data, skip down to the Nope! options below.

Step 2: Construct your prompt

Write a prompt asking the tool to write a poem about Middlebury (or your home school, or favorite city).

Need some help? Here’s an example: “Write a poem for me about Middlebury College that captures its unique geographic location and its historic buildings.” You could, of course, choose to focus your prompt on different characteristics or on another part of the institution, for example: “You are a Middlebury Institute alum. Write a poem about the Middlebury Institute that emphasizes the impact of its alums and expert faculty on the world. The poem should use at least two languages and should reference environmental impact.”

Step 3: Improve your prompt

Given what we said earlier about how prompts shape how the generative AI tool responds, try different ways to improve your prompt.

For example:

“That’s a good start, but your poem is missing several key components of the Middlebury (or your chosen school) experience, such as ____”
“Middlebury (or your chosen school) is a place that means a lot to me. Can you make this poem more ____ and include ____” (insert your own ideas in those blanks)?

Step 4: Ask the tool to improve its own work

Remember that these tools can and will improve their outputs through interaction with you. Ask the generative AI tool you’re using to improve the poem you’ve co-created.

You could ask it to:

“Review this poem and make edits that would improve the emotional appeal of the poem”
“You are an expert in epic poetry. Rewrite this poem to make it more like an epic poem.”
Imagine you are Pablo Neruda. How would you rewrite the poem?

Step 5: Reflect on the poem

How does this poem look to you? What’s good? What’s missing? What changes could you make to the prompt/interaction to improve? What would make this output better? How connected is this output to your human experience of Middlebury or the school of your choice? Keep prompting and interacting until you’ve created something that you’re ready to share with others.

Excellent job! You’ve completed the first activity in our Digital Detox. Ready to take this to the next level?

Step 6: Share your output

This Detox will encourage you to share your activity results with others who are participating in the Detox. We hope this will provide participants with opportunities to learn with each other. You can do that below.

Note: Posts will be moderated, so you should see your post appear within 24 hours of submission.