As AI becomes an inescapable component of education, instructors will need to answer the question: What should students keep inside their heads and what can live externally in the machine?

For the answers, I have been looking to computer science caching theory, which deals with how systems determine what to store locally versus externally. Much has been written about how the computer is a bad analog for the brain, though much of this debate may stem from a simple semantic disagreement. Regardless, it is useful to think of AI as functioning as an external brain with both memory and cognitive abilities. In transhumanism, this is called the “exocortex,” and it is not even remotely science fiction given how much of our minds and abilities we store in our smartphones.

I’ve spent the last decade determining what information is included or excluded in accredited and nonaccredited classes, as both the co-founder of MasterClass and later as the founder of Outlier.org. There is frequently more a student should learn in a class than there are teaching hours to convey it. This is where computer science will help us not with our computers, but with our minds.

Caching Theory

In computer science, caching theory studies what information to store in faster, smaller local memory (the cache), versus what to store in slower, larger, external memory. The right choices can mean the difference between a sluggish system and high performance.

Our brain is like the local cache, and the external storage could be everything from the contents of a book to a Google search to a language translation smartphone app.

Relying on external tools for skills and memory is a hot debate, and has been for millennia. Even Plato had Socrates criticize writing in Phaedrus, “For this invention will produce forgetfulness in the minds of those who learn to use it, because they will not practice their memory.”

Books are an excellent example of an external storage device, and we can use caching theory to determine what we should memorize from a book versus what we can look up later. For the purposes of this framework, books and AI are functionally interchangeable.

In caching, the primary determination of a good caching strategy is how many “cache hits” you have versus “cache misses.” A cache hit occurs when you need to access some information and it is in the local cache. A miss is when it is not in the local cache and you need to access external storage to retrieve or recreate it. To optimize for cache hits, you look at the frequency and urgency of accessing the information.

Frequency

How frequently does a student need to access a piece of information? A frequently accessed piece of knowledge that should be in our local cache (our brain) is our multiplication tables. Many times a day, we might need to know how many items at $2.99 we can buy for $20, or even how many bottles of wine you should get for a party if 20 people show up. If we had to look up our multiplication tables in a book or use a calculator every time we encountered this problem, it would be so tedious and time-intensive, we’d resort to guessing, or avoid situations where we needed these answers.

Urgency

Urgency is exactly what it sounds like: How quickly do you need a piece of information? This is especially important when it is necessary to continue a task, or if the task itself is time-bounded. If you trip walking down the stairs, it is extremely important to have access to the classic, but embarrassing, grab-on-to-the-handrail-and-land-on-your-posterior protocol. Imagine if you tripped, then had a cache miss on this protocol and had to look up what to do in a book, via ChatGPT or on Google long after you were in a crumpled heap at the bottom of the stairs. Many social skills fall into this category, as does any information you’re expecting to use in conversation, whether that’s a student being cold-called in a class, or in collaboration with peers.

These two categories give us the beginning of a framework for testing information to see if it should be held in cache versus in the exocortex. Take any learning objective in your course, and test it with frequency and urgency. Does it make sense for your student to store it in their brain? Long ago, we learned how to do imperfect square roots by hand. This has gone the way of the dodo, and now we assume that if you are doing square roots, you are probably someone who has access to a calculator.

The Romans and Plato’s fellow Greeks used the “Method of Loci” or memory palace style of memorization to hold huge amounts of information in their brains. This, too, has gone the way of square roots by hand. Technology improves. We adapt. We change. And we do our best to prepare the next generation for the future.

Existence

Education needs two more categories beyond frequency and urgency. The first is existence: Should you know this information exists in the external storage? When learning computer science, we call these “trailheads.” You don’t need to know Dijkstra’s shortest path algorithm, but you do need to know that it exists in case you need to implement it. We should know a recipe for tiramisu exists, but it’s not necessary to know it by rote.

World Models

The final category is “world models,” the internal simulations of how the world works from physics to economics, politics to social protocol, and computer systems to animal behavior. The more a student can predict the world, the better.

A New York Uber driver once told me that as a 13-year-old, he learned about inertia the hard way when he hung off the side of a subway car on a dare and found out that a last-second jump didn’t save him from a bad injury when the train entered a tunnel. World models are important, especially in physics. He needed that particular knowledge urgently, but hopefully not frequently.

This gives us four easy questions to ask about each learning objective in our courses:

  1. Does your student need this information frequently?
  2. Will they need it urgently?
  3. Do they need to know it exists?
  4. Is it a world model they should internalize?

Now, when we bring the capabilities of AI into this framework, these questions still apply, but the answers shift. Do I internally need to know how to program an app? For some applications, I can “know it” externally. Well over a year ago, I released a simple app that was entirely created by ChatGPT.

How about your students? Do they need to know another human language? Language can have the same frequency, urgency, existence and world model categories applied to it. AI translation is excellent for asynchronous communication, but what about synchronous communication?

If you’ve tried real-time audio translation, you will have experienced the frustration of waiting for AI audio translation to kick in while you try to remember the facial expressions of your conversation partner from several seconds ago. Urgency dictates local caching of the language. Meanwhile, the percentage of the language that is cached (learned) can be dictated by frequency. Many language learning programs (without academic literature support) will advertise that knowing (caching) only 5,000 words will get you through most conversational settings. Lower-frequency words can be stored in a dictionary.

Application

How does one apply these categories to a syllabus or list of learning objectives to figure out what they should teach in a course? There are four options for each learning objective.

  1. Leave it as is.
  2. Make it augmented.
  3. Consider it an optional curiosity.
  4. Make students aware it exists.

The first is the easiest. For information that is needed frequently, urgently or is a world model, use the first option: Leave it as is.

The second option is to teach the objective with the assumption it will be an augmented piece of knowledge. This is how square roots are done today. We assume students will use a cognitive augmentation, like a calculator, to perform the task. Whether relying on a book or an AI, the student will not be expected to perform the task unassisted.

The third result is that this learning objective is an optional curiosity. Navigating by using paper maps is now in this category for most students. Outside of specialized situations, the vast majority of people will be using GPS as their augmentation rather than maps. It is a fun extra credit assignment to learn orienteering with a map and a compass.

The final option is to simply make your students aware of the knowledge—in other words, that it exists. Almost everything that falls under optional curiosity should get a mention as extant.

Step through each learning objective, categorize it, then figure out if it should remain as it is. With those four categories for knowledge, and four actions an instructor can take as a result, the existing content in our courses will shrink. That will create space to teach more new things that must be learned.

There is no shortage of knowledge students need, but in the age of AI we need to prioritize what they should hold in their brain, and what should stay external, whether it’s declarative or procedural, accessible through books or an advanced LLM.

Plato was right to have Socrates be concerned about the technology of writing in Phaedrus. It is undeniable that the proliferation of the written word and the skill of reading has changed the way we think and know as a species. Artificial intelligence will cause similar changes. Not all will be positive. But if we are diligent about how we prioritize knowledge and learning for the next generation, and in our own minds, it will prove an extraordinary tool that will grant unprecedented cognitive leverage to us all.

Aaron Rasmussen is the co-founder of MasterClass, the founder of Outlier.org and serves on the Carnegie Foundation for the Advancement of Teaching’s Board of Trustees. All views are his own.

Next Story

More from Teaching