Learning New Concepts with ChatGPT
Pairing Artificial Intelligence with the Feynman Technique
Recent advancements in artificial intelligence have led to some pretty incredible breakthroughs, from being able to generate life-like images to being able to write human-like text from a simple prompt. ChatGPT is one of these ground-breaking innovations, a chat bot that is able to mimic human dialogue in a conversation. Within a few days of its release, the internet was abuzz with screenshots of conversations with ChatGPT. Teachers worried that students would use it to cheat. Headlines declared that it would soon replace Google. I began to wonder: how could I use ChatGPT to gain a deeper understanding of a concept I was unfamiliar with? And if so, what method would work best? Cue the Feynman Technique.
What is the Feynman Technique?
Developed by renowned physicist Richard Feynman, the “Feynman Technique” for learning involves an iterative process of studying, explaining, and simplification. Simply put, the Feynman Technique involves the following steps:
- Pick a topic. Choose a topic you want to learn more about and try to understand some component of it
- Explain the topic in simple terms. Teach this topic to another person or explain it in a way that a layperson would understand it.
- Identify the gaps in your knowledge. As you explain the topic, identify places that you struggle to understand.
- Review the concepts you didn’t understand and simplify them. Then rinse and repeat, using the newly synthesized knowledge.
The Feynman Technique addresses a common fault in most approaches to learning — it’s often easier to receive information than it is to convey information. Asking to explain a concept in simple terms reveals the underlying gaps in knowledge in ways that highlighting words from a book or writing down notes from a lecture cannot. Furthermore, it is a targeted approach to learning — you can start with a broad concept from which you have a vague understanding and gradually narrow down until you get a more fine-grained grasp of the details. I like to think of it as carving a sculpture: first you begin with the overall shape, and then begin to refine the details, until eventually you are left with a fully-rendered form.
The problems with building from the ground up
When I learned about this topic, I was incredibly excited to try it out. But I quickly realized that using the Feynman Technique on a new concept also poses its own problems:
The first problem is what I like to call the bootstrapping problem. In learning a new concept, you are, by definition, a novice. That means that you’re probably not the best judge of how good you are at explaining a topic. Additionally, you might not catch whether something in your explanation is wrong, which can prevent you from understanding later concepts that build upon simpler foundational knowledge.
The second, related to the first, is the problem of unknown unknowns. As a novice, you’re faced with the problem of not knowing what you don’t know.Even if your explanation is free from factual errors, it’s hard to know where the gaps in knowledge arise. For example, amateur guitar players might learn a few chords and think they know everything, while expert musicians know there is always more room to fine-tune one’s technique. This phenomenon can lead to early stopping and a false sense that you’ve finished your exploration, even though you’ve just scratched the surface.
The third problem is the problem of direction. Once you’ve uncovered your gaps in knowledge, you’re still left with which one to choose next. Experts in the field might know which directions are dead ends, but as a novice, you’re traveling blind. If you’re learning alone, this often means sifting through opaque language in a topic you don’t understand or needlessly going down rabbit-holes just to figure out the next step.
These three problems of bootstrapping, unknown unknowns, and direction are all easily addressed if you have someone with expertise as your counterpart in the Feynman Technique. Experts can give you feedback on your errors, help you identify your gaps in knowledge, and point you in the direction of interesting avenues of future inquiry. This version of guided learning brings out the best in the Feynman Technique and makes it a powerful tool for learning. Unfortunately, experts are often in short supply, and the Feynman Technique, for all its benefits, is a fairly one-sided exercise. The expert’s role as listener is basically to sit there and provide feedback on a topic they already know like the back of their hand. In the absence of a frequently available and extremely patient expert, however, we must turn to other solutions.
ChatGPT as the counterpart
That’s where ChatGPT comes in: the counterpart of your dreams. I’m sure I’m not the first person to think of combining the Feynman Technique and ChatGPT, but I’ve been surprised to find that there are surprisingly few resources online for how to approach this integration — and even fewer that describe why ChatGPT works as a counterpart and where it falls short. A quick search reveals a YouTube tutorial about using the technique to learn mathematics with an earlier language model (GPT3) and a scan of late-stage Twitter yielded just one post where the roles were reversed: ChatGPT as the active “explainer” and the participant as the passive listener. Before we understand how ChatGPT can be used for the Feynman Technique, we need to understand why it makes sense to use it in this context.
First, ChatGPT has an extensive knowledge and expertise of a wide range of topics, making it a good candidate for free-form exploration of topics. While it occasionally makes factual errors, I’ve found from my own experience that it performs more than well enough to be a powerful tool for guided learning. Additionally, as a conversational agent, ChatGPT is a natural candidate for the Feynman Technique, which works best when there’s some form of interaction or feedback. Finally, its widespread availability as a free online tool makes it a scalable and highly accessible technology for learning topics in-depth, beyond simple search queries.
How to use ChatGPT for the Feynman Technique
That brings me to the actual implementation — how do you actually use ChatGPT to learn with the Feynman Technique? The following section describes how I personally implemented the technique; it’s possible that there are many other ways to do this effectively, but this is what worked for me.
It’s important to note that it’s actually a bit awkward to begin using the Feynman Technique with ChatGPT. If you try to ask it to “try out the Feynman Technique on [topic]” you will typically end up being on the passive (listening) end of the dialogue as ChatGPT goes through a somewhat verbose description of what the Feynman Technique is and how it could be applied to your topic. Asking for a step-by-step walkthrough also results in the same failure mode.
Instead, I’ve found it helpful to start out with a simple question: “Can we hold a conversation where I try out the Feynman Technique and you listen and give critiques”?
Below are excerpts of my conversation with ChatGPT, where I try to learn more about the topic of snowflakes. I’ve edited a few sections for brevity and clarity.
Me: Can we hold a conversation where I try out the Feynman Technique and you listen and give critiques
ChatGPT: Sure! I’d be happy to listen to you practice the Feynman Technique and provide critiques… [proceeds to describe the Feynman Technique] … Would you like to choose a topic to practice the Feynman Technique on now?
Me: Yes please. I have chosen the topic of Snowflakes and I will try to summarize in my own words. Are you ready?
ChatGPT: Yes, I’m ready! Go ahead and summarize the concept of snowflakes in your own words.
Me: Snowflakes are crystalline structures formed when the temperature outside is below freezing. Snowflakes often take on symmetrical patterns, often in the form of six-fold symmetry. How did I do? Could you ask me some questions to help me understand this further?
ChatGPT: Great! You have provided a good summary of the basic concept of snowflakes. Here are a few questions that might help you further understand and explain the concept: How do snowflakes form? What factors influence the shape and symmetry of snowflakes? How do the physical properties of snowflakes, such as their size, shape, and structure, affect their behavior? How do snowflakes contribute to the formation of snow and ice? Can all snowflakes be seen with the naked eye, or are some too small to be visible?
[provides brief context on how to use these questions in the Feynman Technique]
Me: [begins answering the questions above]. The size, shape, structure, and behavior of snowflakes vary — you know, to be honest, I’m not really sure what affects these things. Could you verify what I said is true, and help me out where I’m struggling?
ChatGPT: Yes, what you have explained is generally correct. Here are a few additional points that might help you further understand the concept of snowflakes…
ChatGPT goes onto give more depth to each of the questions, and I ask for directed guidance on more questions that will deepen my knowledge of the topic. For example, later in the conversation I learn how the structure of snowflakes influences their ability to absorb and reflect light (those with a more complex / irregular structure scatter light in more directions, making them appear brighter). I won’t belabor the point. As the example above demonstrates, ChatGPT can be a powerful tool to learn new concepts in a directed way.
Of course, the method isn’t perfect. As mentioned before, integrating this technique with ChatGPT has common failure modes — setting the right framework for practicing the technique early on in the conversation seems to be important. I’ve also noticed that asking for increasingly specific feedback often results in somewhat repetitive responses (such as “give more specific examples in your explanation” even after examples are added). Importantly, when using ChatGPT — as in any language model — it’s necessary to consider its broader implications on the spread of fake news, perpetuation of existing cultural biases, and the ownership of intellectual property. It’s hard for anyone, let alone a model of human conversation, to strike the proper balance of nuance when describing complex topics. While ChatGPT is a good place to begin your journey to mastery of a topic, it’s certainly not the end. Ultimately, the goal is to reach a level of well-earned confidence in a topic to be able to consult actual experts and peer-reviewed sources.
Despite its drawbacks, I’ve found that using ChatGPT to learn new concepts has been an eye-opening experience. It has taken a self-learning exercise that used to be awkward, due to the bootstrapping, unknown unknowns, and direction problems, and turned it into a fairly fun conversation. More importantly, I found the experience to be much more collaborative and active than my earlier interactions with ChatGPT, where I typically asked a few random questions before eventually getting bored. Here, I was finding myself more and more excited about learning, even for a concept as seemingly simple as “what is a snowflake”.
In conclusion, using ChatGPT to learn new concepts has been a highly effective and enjoyable way for me to expand my knowledge and understanding. The interactive and collaborative nature of the conversation helps to keep me engaged and motivated, and the ability to ask questions and get immediate responses has been particularly useful in clarifying any misunderstandings and deepening my understanding of the material. Overall, I would highly recommend using ChatGPT as a tool for learning new concepts to anyone looking to improve their knowledge in a fun and interactive way.
That last paragraph, by the way? Not my own — but you can probably guess the author. Happy 2023 everyone, here’s to a year full of learning.