ChatGPT and other big AI chatbots aren’t ones for holding anything back. If you ask them a simple question, you’ll frequently get multi-part answers, complete with bullets and emojis.
Those giant AI-generated answers can be overwhelming. If your intent is to actually learn something rather than being on the receiving end of a term paper, ChatGPT’s lengthy replies might be little more than a waste of tokens.
For example, when I asked ChatGPT (the “instant” version) to teach me something about neural networks, a key AI concept that I’m still wrapping my head around, it spat out an 800-word answer in 15 parts–no kidding. I didn’t get far before my eyes started to glaze over.
Now, I’ve already written about an AI prompt that helps you learn 80 percent of a concept by distilling the top 20 percent of what you need to know about it (the “80/20” prompt). This time, I have a prompt that takes a different approach to learning: making an AI teach you about a topic by asking you questions about it.
Here’s the prompt:
I want to learn [Concept]. Do not explain it to me yet. Instead, ask me a series of questions—one at a time—to help me construct the understanding myself. Only provide the full explanation once I have correctly identified the core principles through my own reasoning.
Yes, we’re talking about the Socratic method, a teaching method that involves asking a student a series of probing questions rather than spoon-feeding the answers. The Socratic method prompt is a popular way to turn a lecture-prone AI into a classically-trained guide.
When I appended the Socratic method prompt to my original “teach me about neural networks,” I got this response:
Perfect. I won’t explain neural networks yet. We’ll build the idea together, one question at a time.
Question 1:
Imagine you’re trying to build a computer program that can recognize whether a photo contains a cat.
What kinds of things do you think the program could look for in the image to help decide “cat” versus “not cat”?
In its answer, ChatGPT is kicking off a classic lesson about how neural network-powered AI systems can be trained to recognize images of cats. (I’m still struggling through the relevant chapter in Melanie Mitchell’s Artificial Intelligence: A Guide for Thinking Humans, which is required reading for those interested in the history of AI.)
After a few more questions and answers, we went from discussing clues that would distinguish cat photos from other images (“whiskers”) to the concept of weights in AI model training.
And you know what? I actually learned something.
