On Explaining Things
How to Play a Board Game
There are a lot of smart people I know who are bad at explaining things. One of the settings I see this most often is when people are introducing a new board game to a friend who's never played before.
The story goes like this: the experienced player starts by going through all the rules to the game and the best strategies, often with phrases like:
- "You always/never want to..."
- "Oh, and I forgot to mention..."
- "Should I explain it one more time?"
Meanwhile, the new players have no idea what's actually going on. They ask clarifying questions, going back-and-forth with different hypothetical scenarios. Ten minutes pass. Then comes my favorite line:
- “Trust me, it'll make more sense when we play.”
This makes me want to jump off a bridge! Even though people tend to pick up on the rules as they play, I've found that most explanations focus disproportionately on the wrong things, which only adds to people's confusion. I see this everywhere. Professors who focus too much on specifics without motivating the big-picture idea. Classmates who can't explain how they solved a problem beyond just reading from their work. Even myself, when I'm halfway through writing an essay and realize my argument is wrong.
Intuitively, you'd expect that the more you know about a particular topic, the easier it would be to explain it. But this phenomenon also works in the reverse direction, where people who have been studying the topic for a while have a difficult time zooming out and remembering what it was like not to know. This is known as the curse of knowledge, and applies to every type of specialist from PhD students to Monopoly experts.
This blog post is about explaining things well. It's certainly not an exhaustive guide on communication, nor is it entirely based on empirical evidence. Many of these are simply qualities I've observed in explanations that made me go "Aha! I understand now!", as well as in the people that helped me get there.
Clear Explanations
Clarity is something I value a lot in a good explanation. But this feels somewhat ambiguous and unhelpful—what do we actually mean by the word clear?
When you're just beginning to explain something, the idea is that we should start as broadly as possible, and then narrow our scope as we progress—almost like a funnel. This is where the concepts of abstraction and distillation come into play.
- Abstraction is about intentionally obscuring the details of how something works in order to simplify things conceptually.
- Distillation is about extracting the most important and essential elements of a complex idea.
Here's an example: for a long time, I struggled to understand how large language models worked. The first time I read through the paper, I found that I was constantly getting lost in the mess of linear algebra and technical jargon—niche details that only obscured the big-picture idea. Then I came across 3Blue1Brown's series on transformers, followed by Neel Nanda's tutorial on how to build one, and the ideas clicked almost instantly. The main difference was in how the information was structured and presented; both 3Blue1Brown and Neel Nanda started with a high-level goal of what we were trying to accomplish, and then slowly narrowed our scope as we worked through the technical details.
When I'm explaining the idea behind a transformer, I don't start by saying that it's an "autoregressive, attention-based transformer model for next-token prediction." I start by saying that it takes a piece of text, and its goal is to output a prediction for the next word. Then I get into how you actually represent words using numbers ("embedding vectors"), how you determine which words are most relevant to predicting the next one ("attention"), and how the model learns by making lots of incorrect guesses, comparing its guesses to the answers, and tweaking its dials so that its guess is closer to the correct answer ("backpropagation").
There are a few important points I want to make. The first is that good explanations often involve a tradeoff between clarity (how easily something can be understood) and truthfulness (how in-depth and factual our explanation is). The explanation I just provided is technically incorrect on a few levels (models are predicting the next token, not word; attention is calculated using key, query, and value matrices; backpropagation utilizes gradient descent performed using some loss function). Yet I still think the explanation I provided is a helpful starting point, because it abstracts away the unnecessary information and distills the concept of ChatGPT into a form that could ideally be understood by a 5th grader. Of course, we should be careful we're not oversimplifying—but simplification is (I think) always necessary when it comes to learning a new topic, so long as we resolve those uncertainties at a future point.
The goal of clarity is to make an idea feel so obvious that people wonder why they didn't understand it before. It's not about bombarding people with factual information (i.e. the rules of the game), and being as clear as you can with each fact. What matters most is the actual order in which you present things, and the level of abstraction that you employ to break something down into its fundamental essence.
To be continued...