Published 2026-06-25
Tags: #education
In a previous blog post, we gave a philosophical answer to why we need education even in the age of AI – to want better. Here we give a more practical answer: Because education helps you build mental models, and readily-accessible mental models are necessary for you to (1) communicate in real time with other humans (2) steer AI more effectively (3) think more authentic thoughts.
Wikipedia’s definition is “internal representation of external reality”. I’ll expand a bit.
That is, why can’t we just rely on the mental models that exist in an AI? Here are 3 reasons – one for each kind of interaction: [with other humans, with AI, and with yourself].
Here are some practical scenarios. During a design review, a senior engineer says: “Let’s cache this. But remember to add invalidation code in the real-time write paths.” Or, your co-founder, during a brainstorming session says: “I’m not sure if the round-robin approach to fairness is enough. We also want to incorporate how many times the job has failed for each customer, and their subscription tier.”
It’s not acceptable for you to always respond with “Oh sorry give me 1 hour to learn about these topics with AI first.” Conversations that happen in real time require you to have a lot of mental models pre-built in. If you don’t have a rich library of mental models, fewer people would want to collaborate with you.
You also need mental models even for workflows where no human is waiting for your response in real time. Consider your workflow as you review an AI’s design or code output. You might argue that you don’t need to have a mental model for understanding the AI’s output beforehand, because you can simply have conversations with the AI to build the mental model on the fly. But (1) will you want to do this when you are mentally tired and you have an impending deadline? (2) do you even know about the right clarifying questions to ask, if you don’t know about the pertinent design choices for the system?
As an example, when you review code, you might quickly have spidey-senses that go like this: “Huh, there’s a for loop in this DB transaction, and … wait there are remote calls in this loop?! Whoa, this transaction is going to hold on to the locks for a long time.” As you scan through a lot of code files, you need to have this kind of reasoning happening in real time to be able to flag which sections to probe deeper with AI. Without a rich mental model of DB transactions, locks, and network calls, you would not be able to truly verify and be a responsible steward for the code generated by the AI (i.e., you produce AI slop).
In all the examples above, you need to chain thoughts together to respond to the world. Chaining thoughts requires mental models that are readily meldable to each other, as well as to your goals.
I suspect that chaining in real time, without waiting for AI, is important for generating thoughts that are closer aligned to your values. I’m not saying that we shouldn’t use AI. I’m saying that in the process of iteratively refining what you want, when it’s your turn to think and provide an input to the AI, sometimes you might want to think harder, or heck, even take a walk first, before writing the prompt. This is because this self-driven thinking is intimately tied to your personal sense of “good”. E.g. I recently went on a thinking walk to figure out what kind of app would help me keep track of my jazz repertoire. At the end of the walk, I decided that I just want a simple app that has 0 cloud components, and one that has a video-game-like interface. It’s unclear if I would have come up with a simple, personally delightful solution if I hadn’t done this thinking alone first.
Another angle to see this is to think about system inputs and outputs. If you don’t think too much before writing prompts to feed into AI, then you will probably just feed in the top K prompts that other humans have fed in, and in turn get the same small space of responses. You are unlikely to generate thoughts that are uniquely you.
Building mental models within you allows you to think for yourself, collaborate with others, and steer AI more effectively.
As for how to build them (and how AI can help or detract) … that’s for a future post!