Traditional education organizes knowledge around what the world is made of. In the age of AI, perhaps personal learning should also be organized around how an agent survives, understands, chooses, and brings about change within it.
Imagine waking up tomorrow morning in a country you know nothing about.
Your phone has 20% battery left. You do not understand the language. You do not know where you are, whether the area is safe, or which way to go.
Your first question probably would not be: “Does this problem belong to geography or sociology?”
You would ask:
Where am I? What does the information in front of me mean? What might happen next? What options do I have? Which is better? If I take a wrong turn, how can I find out quickly? Are the people around me friends, competitors, or simply indifferent?
Then you make a decision, walk a little way, see what happens, and revise your judgment.
Look closely, and this is what we do throughout our lives.
Finding a job, falling in love, starting a company, investing, choosing a city, assessing a news story, raising a child, looking after our bodies, navigating a conflict—reality never introduces itself by saying:
“Hello. This problem is 30% psychology, 20% economics, 15% statistics, 10% game theory, and 25% management.”
Reality simply gives you a problem.
And that exposes a revealing mismatch in education:
The way humanity produces knowledge is not the same as the way a person uses it.
1. Our map of knowledge was drawn for its producers
The familiar landscape of knowledge looks roughly like this: mathematics, physics, chemistry, biology, economics, psychology, sociology, history, philosophy, computer science, and so on.
Of course, the actual history of disciplines is far more complicated. They were not simply carved out according to their objects of study. But as an approximation, it is easy to understand the modern system this way:
Whatever exists in the world, we build a discipline around studying it.
There is matter, so we study physics. There is life, so we study biology. There are minds, so we study psychology. There are human societies, so we study sociology. There is exchange and the allocation of resources, so we develop economics. There is information and computation, so we develop computer science.
The university begins to resemble an enormous warehouse of knowledge:
World
├── Matter
│ └── Physics
├── Life
│ └── Biology
├── Mind
│ └── Psychology
├── Society
│ └── Sociology
├── Exchange
│ └── Economics
└── Information and computation
└── Computer science
There is nothing wrong with this.
In fact, this division of labor is essential if humanity is to keep producing reliable knowledge. A physicist and a historian should not abandon their very different standards of evidence merely because they both study “the world.” Mathematical proof, randomized controlled trials in medicine, and the critical evaluation of historical sources belong to different systems for establishing what we can know.
One of the great services disciplines provide is to preserve the standards by which knowledge is tested.
So I have little patience for the shallow claim that “AI makes disciplines obsolete.”
The real question is not whether we need disciplines. It is this:
Disciplines should remain part of the underlying structure of knowledge production. But should they still be a person's first point of entry into learning?
Those are two entirely different questions.
2. The world comes to us as verbs
At school, knowledge usually arrives as nouns: mathematics, history, economics, programming, psychology.
The problems of life tend to arrive as verbs.
I need to understand this. I need to judge whether it is true. I need to predict what comes next. I need to choose a course of action. I need to persuade someone. I need to design something. I need to change a system. I need to find out whether I am wrong.
This may be the first opening for redesigning the structure of personal knowledge:
Begin with what an agent needs to do in the world, rather than with what the world contains.
Suddenly, the question changes. Instead of first asking, “Which course am I missing?”, we ask:
What kind of cognitive capacity am I missing for the problem in front of me?
This goes beyond simply being interdisciplinary.
Interdisciplinary work often still means putting mathematics, computing, and biology together. What I want to explore is a more radical step:
Could we begin with action itself and redefine the top-level folders of personal knowledge?
3. First principles: what, exactly, is a human being?
Forget academic specializations for a moment. Ask a simpler question:
Why do people need knowledge?
One answer I find especially useful is:
A human being is a finite agent in a world that can never be fully known.
Three parts of that sentence matter.
First, finite. Your time, attention, memory, and computational capacity are limited. You cannot wait until you understand the entire world before you start living.
Second, never fully known. You never have access to the world's complete state. You do not know what will happen to your company, what another person really thinks, or what next year's economy will look like. Even the workings of your own body are usually something you infer indirectly from signals.
Third, agent. You are not standing outside the universe, merely watching it. Every decision you make changes the world a little. The changed world then gives you new information.
A loop takes shape:
Observe the world
↓
Form an understanding
↓
Predict what comes next
↓
Consider what matters
↓
Choose an action
↓
Change the world
↓
Receive feedback
↓
Update your understanding
Condensed into five words:
Observe → Model → Value → Act → Learn
This may come closer to the way an ordinary person actually uses knowledge than “mathematics—physics—chemistry—biology.”
And this perspective has a history. One of the classic abstractions in reinforcement learning is the agent–environment interface: an agent receives information about its environment, takes an action, the environment changes, and new observations or rewards follow. Partially observable Markov decision processes, or POMDPs, go further, asking how an agent can act under uncertainty when it cannot directly observe the world's true state.[3][4] In The Sciences of the Artificial, Herbert Simon emphasized the difference between describing how things are and designing how they ought to be. Intelligence concerns choice and design as well as understanding.[2]
Of course, a human being is not a reinforcement learning machine. But the abstraction gives us a remarkably useful lens.
4. A new directory begins with nine verbs—and with observation
Observe: what am I actually seeing?
Before acting, the first question is not even “What should I do?” It is:
What is happening right now?
A doctor sees symptoms, not the disease itself. An investor sees a price, not a company's “true value.” A manager sees reports, not the full state of an organization. Seeing a person's expression is not the same as reading their mind.
The first foundational capacity, then, is measuring, observing, recognizing signals, and understanding how data came into being.
It draws on statistics, measurement theory, data analysis, sensors, survey methods, interviews, information theory, and experimental design.
Notice the first shift. Statistics ceases to be merely “a course in a university mathematics department.” It becomes:
How can I infer what the world is like from limited signals when I cannot see it directly?
Suddenly, it comes alive.
5. Represent: in a world this complicated, what should I keep?
You receive an enormous amount of information every day. Most of it does not deserve a place in your representation of the current situation.
That leads to a deeper question:
In a changing world, what information counts as a “state”?
Suppose you are driving. You need to know your speed, your direction, and the positions of the cars ahead. You do not need to remember what the seventeenth tree you passed three hours ago looked like.
A good state representation does not contain everything about the world. It:
Retains the information that still matters for predicting what comes next and choosing how to act.
In an ideal Markov model, if the current state already contains the relevant information from the past needed to predict the next step, the earlier history can be folded away. In a partially observable environment, we cannot even obtain that state directly. We must maintain a “belief” about it from our observations.[3][4]
This idea reaches far beyond control theory.
A good doctor compresses symptoms. A good founder compresses user feedback. A good historian extracts structure from a mass of sources. A good mathematician searches for invariants. A good writer finds the one sentence that matters among thousands of details.
The second capacity that may become increasingly valuable is:
Representation: compressing a complex world into a structure we can think with.
This is why summarizing is so much more than turning a thousand words into a hundred. The most powerful kind of compression happens when the right representation makes a previously tangled problem suddenly tractable.
6. Model and predict: if the world keeps running, what happens next?
Once we have a state representation, we begin to ask what comes next.
This is the work of a world model. Physics builds world models. So do economics and psychology. Your rough understanding of your boss, your partner, your users, and yourself is also a collection of world models.
Often, to “understand” means:
I know not only what happened, but how things are likely to change when the conditions change.
“This product sells well” is a description.
“If the price rises by 20%, sales probably will not fall very much, because buyers are purchasing an expression of identity more than a set of functions” begins to resemble a model.
A genuinely useful model does more than explain the past. It lets you ask:
What if?
What if we change the price? Move to another city? Remove this policy? Hide the brand from customers? Do nothing at all?
Dynamical systems, probability, causal inference, economics, physics, and psychology begin to converge from very different departments on a shared question:
How will the world evolve?
7. Reason about causes: what happens if I intervene?
We need to distinguish two questions:
“What tends to happen alongside what?”
and:
“What happens if I deliberately change one of those things?”
They are not the same question.
When ice cream sales rise, drownings may also rise. It does not follow that banning ice cream will reduce drownings. A third variable may be at work: hotter weather.
Correlation can be valuable to an observer. For an agent, it is far from sufficient. Eventually, you need to ask:
What happens if I press this button?
That is why causal inference belongs near the center of a system of knowledge organized around agents.
We learn about the world to describe it, but also to discover where an intervention can actually make a difference.
8. Value: even if I know what will happen, then what?
This is the step most easily overlooked.
Imagine a superintelligent AI that can predict the consequences of every choice in your life with extraordinary accuracy. It tells you that job A pays the most, job B is the most secure, job C offers the most status, and job D gives you the most freedom.
Excellent. Now what?
It still has not told you:
Which one you should choose.
“What will happen?” is a descriptive question. “What is worth bringing about?” is a normative one.
A person is more than a prediction machine. A person also needs some structure of values.
Does money matter? Freedom? Love? Family? Creation? Status? Curiosity? Safety? How much of the present is it reasonable to sacrifice for the future?
This is why I do not think philosophy, the humanities, literature, and history automatically become less useful as AI becomes more capable. They may become more important.
Machines may get better and better at answering, “How can this goal be achieved?” That forces us back to the older question:
Why is this goal worth achieving?
Many questions about AI alignment eventually encounter a similar boundary. An optimization system must optimize something; the difficult question is often what we actually want that something to be. Stuart Russell describes the classical model of AI in terms of systems choosing actions, on the basis of their observations, to achieve objectives. The next question is where those objectives come from.[9]
The same holds for a person's life. AI can help you optimize it. Before you begin, you would do well to know:
What must not be optimized away.
9. Decide and act: given this world, what should I do?
Traditional education is very good at asking what the world is like. We train much less systematically for another question:
Given that the world is like this, what do I do now?
Knowing that climate change exists is different from designing energy policy. Knowing that users like a certain product is different from building it. Knowing that a team has communication problems is different from changing its working practices.
Here we encounter optimization, control theory, operations research, game theory, engineering, product design, programming, and project management.
They share an orientation:
Finding ways to act under constraints, as well as explaining the world.
Simon distinguished the natural sciences from the “sciences of the artificial.” The former seek to understand things as they exist; the latter necessarily involve design, purpose, and the transformation of a present situation into a desired one.[2]
A learner in the age of AI cannot stay forever at “I understand.” A second question must become a habit:
What follows from that? Which variable can I change now?
10. Coordinate: there are other people in this world
If you were the only agent in the universe, things would be much simpler.
The complication is that the world is full of other agents. They too observe, maintain models, pursue goals, and act.
Cooperation, competition, negotiation, trust, incentives, power, institutions, culture, and language enter the picture at once.
You may think a proposal is “obviously best,” while someone else disagrees completely. That need not mean they are foolish. It may mean:
You are optimizing for different things.
The business owner cares about cash flow. Employees care about workload. Customers care about prices. Governments care about employment. Investors care about growth. Society cares about externalities.
A real problem is rarely finished when someone finds a mathematically optimal solution. The question is also:
How can agents with different world models, interests, and values create a system that works together?
Psychology, sociology, political science, history, game theory, and communication flow back into the same river.
11. Learn: after acting, how do I discover that I was wrong?
This is the final link, and the most important.
The biggest difference between reality and an exam is that:
Reality does not include an answer key on the last page.
You make a business decision and sales rise. Did the strategy work? Perhaps. Or perhaps a holiday happened to arrive at the right time.
You begin exercising every day and your mood improves. Was exercise responsible? Perhaps your stress happened to fall at the same time.
The real world always mixes together noise, delay, chance, and hidden variables. An agent has to keep running this loop:
Predict
↓
Act
↓
Observe the discrepancy
↓
Update the model
↓
Act again
That is learning.
It is also the scientific method, Bayesian updating, feedback in control systems, testing in engineering, iteration in a startup, and what we ordinarily call reviewing our experience.
From this perspective, making mistakes is not the most frightening possibility. This is:
A system keeps making mistakes, with no reliable feedback channel to tell it that it is wrong.
12. A new map of personal knowledge
Putting these ideas together gives us something like this:
World
↓
① Observe / Measurement
↓
② Represent / Representation
↓
③ Model / Prediction
↓
④ Reason about causes / Intervention
↓
⑤ Value / Evaluation
↓
⑥ Decide / Decision
↓
⑦ Act / Control
↓
⑧ Coordinate / Coordination
↓
World
↓
⑨ Learn / Feedback
└──→ Update everything above
Notice what has happened. The old disciplines have not disappeared. They have been mounted in new places.
| The agent's question | Traditional knowledge to draw on |
|---|---|
| What am I seeing? | Statistics, measurement, data analysis, experiments |
| What information is worth keeping? | Mathematics, representation learning, information theory, abstraction |
| What happens next? | Physics, dynamical systems, economics, probability |
| What happens if I intervene? | Causal inference, experimental design, analysis of mechanisms |
| What is worth pursuing? | Philosophy, ethics, literature, history, aesthetics |
| What should I choose? | Decision theory, optimization, operations research |
| How can I actually change it? | Engineering, control, programming, design |
| How can we do it together? | Psychology, sociology, politics, game theory |
| How can I tell that I am wrong? | Scientific method, statistics, feedback, reflection |
The point is not to invent nine new disciplines. It is to give existing knowledge a second index.
The disciplinary index asks, “Where does this piece of knowledge belong?” The agent's index asks:
When will I need it?
The National Academies' review of learning science emphasizes that experts differ from novices in more than the number of facts they possess. Experts organize knowledge more effectively, allowing them to identify relevant information and connect new information to existing frameworks. The review also discusses how cognition can be distributed across environments, tools, and other people. Knowing where to find a resource is itself part of competence.[1]
I find myself increasingly drawn to this sentence:
Disciplines are the folders of human knowledge. They should not have to be the navigation bar of a human mind.
13. Why does AI make this urgent?
Because the economics of knowledge are changing.
In the past, to draw on a piece of knowledge, you first had to learn it, remember it, or spend a long time finding someone who understood it. Education naturally emphasized moving large amounts of knowledge into people's heads in advance.
That made sense. Retrieval was expensive. So were explanation, translation, programming, and searching the literature. Consulting an expert cost even more.
Large language models are rapidly lowering many of those costs. At any moment, you can ask:
“Explain adverse selection.”
“I cannot follow this proof. Give me a geometric intuition instead.”
“Write some Python to check this hypothesis.”
“Give me five competing explanations.”
“Take the opposing side and attack my proposal.”
The cost of calling on knowledge is falling.
An important question in learning is shifting from “How much can I remember?” toward:
Do I know what to call on now?
Then it moves further. Can I judge whether the answer is reliable? Have I defined the problem correctly? Which variable have I missed? Why am I optimizing this objective? What must be tested in the world itself?
This does not make foundational knowledge unimportant. Quite the opposite. Without a basic grasp of a field's structure, you may not even notice where AI has gone wrong.
The National Academies' review likewise emphasizes that prior knowledge helps people recognize relevant information, tackle more complex problems, and accelerate subsequent learning.[1]
What is happening is not “internal knowledge → zero.” It is closer to:
A person's internal models + external knowledge systems + knowledge drawn on dynamically through AI
together forming a larger cognitive system.
14. The trap: using AI well is not the same as having learned
A randomized mathematics experiment involving nearly a thousand high school students, published in 2025, offers a striking warning.
Students given ordinary GPT-4-style assistance performed substantially better during practice. But when AI access was removed, this group scored about 17% lower on subsequent unaided exams than a control group that had never received access. When the AI was instead designed to support learning through teacher-informed hints and safeguards against simply giving answers, the negative effect was largely mitigated.[6]
Another randomized study, published in Scientific Reports in 2025, found that a carefully designed AI tutor helped university physics students achieve larger immediate learning gains on the material tested, in less time, than an in-class active learning lesson.[7]
The two findings are interesting together. They do not establish that “AI makes people less intelligent.” Nor do they establish that “AI tutors always beat teachers.” They suggest something more useful:
Whether AI helps learning depends not only on the capability of the AI, but on how cognitive work is divided between the person and the system.
If AI always defines the problem, chooses the variables, forms the explanation, designs the solution, and judges the result for you, an absurd situation becomes possible:
Your work becomes more impressive while your own abilities fail to keep pace.
15. The crucial capacity may be taking charge of the whole loop
A 2025 study of 319 knowledge workers collected 936 examples of generative AI use. In these self-reports, greater confidence in AI was associated with less effort devoted to critical thinking. The researchers also found that critical thinking was shifting toward checking information, integrating responses, and overseeing tasks.[8]
This was a survey. It does not demonstrate that AI causes a decline in thinking ability. But it points to an important shift:
Human cognitive work may be moving from personally producing every step to managing the entire process of thought.
Increasingly, I think “AI-native learning” should mean more than using ChatGPT to learn calculus. The deeper change is:
Redesigning which parts of the loop a person must take responsibility for.
AI can retrieve information quickly, offer several explanations, generate examples, simulate different perspectives, write code, organize material, help construct an initial model, search for counterexamples, and create exercises.
But you must gradually learn to judge:
Which questions are worth asking.
Which information deserves belief.
Which aspects of the situation deserve to be tracked.
Which assumptions are quietly determining your conclusions.
Which outcomes are worth pursuing.
Which things can only be learned by going out into the world.
That is where the division of labor between people and AI becomes truly interesting.
16. How would I redesign my own learning?
I would stop maintaining a list called “The 100 Courses Everyone Must Take.” Instead, I would build three layers of knowledge.
Layer one: foundations that must live in your own mind
These include language and expression, basic mathematics and a sense of quantity, probability and uncertainty, logic and argument, basic computational skills, scientific method, a sense of history, and elementary frameworks for understanding people.
They are not there to help you compete with AI. They form:
The coordinate system with which you examine the world.
Without that coordinate system, you cannot even locate an error.
Layer two: the nine capacities of the agent's loop
Every so often, ask yourself:
Can I observe? Can I identify the state that matters? Can I build a model? Can I distinguish correlation from causation? Do I know what I am optimizing? Can I make a decision? Can I turn a plan into something real? Can I coordinate with other people? Can I revise my model in response to feedback?
Those nine questions come closer to measuring real growth than “How many courses did I finish this year?”
Layer three: at least one deep line of expertise
Here I want to resist a seductive illusion of the AI era:
“Since I can look anything up, knowing a little about everything is enough.”
It is not.
Real interdisciplinary ability needs a deep anchor. Only by going far enough into one field do you encounter what rigorous evidence looks like, what a genuinely difficult problem is, what an elegant but false explanation feels like, and what sounds reasonable to an outsider but immediately troubles an expert.
The knowledge structure of the future should not be “3% of every field.” It should look more like this:
Wide connections
────────────────────────────────────────
│
│
│
│
│ Deep expertise
│
│
Perhaps it is not even the conventional T shape. Perhaps it is a deep well, continually extending sensors into the surrounding world.
17. A more concrete example: should I leave my job?
Suppose someone begins seriously asking, “Should I quit?”
Traditional categories can make this feel bewildering. Economics brings opportunity cost; finance, cash reserves; psychology, risk preferences; statistics, the probability of succeeding in a new job; sociology, identity and social networks; philosophy, the question of a life worth living; career development, transferable skills.
Follow the course catalog, and you might spend six months reading before you begin.
Use the agent's framework, and you can begin now.
Step one: observe
What exactly makes me want to leave? Is it the work itself, or one recent project? Income? My manager? Stalled growth? Exhaustion?
Hold off on explaining. Observe first.
Step two: represent
Which variables actually matter: income, growth, time, stress, location, identity, family responsibilities?
Do not track a hundred variables. Identify the state that matters.
Step three: predict
If I stay, what might life look like in a year? What if I change jobs? What if I take three months off?
Step four: investigate causes
Is the company really causing my unhappiness? Or is it a combination of sleep, relationships, and prolonged fatigue? Would changing companies solve the problem?
Step five: clarify values
What do I actually want: more money, more time, status, freedom, learning, stability?
Step six: decide
Given those goals and constraints, what options are feasible?
Step seven: act
Must I resign with nothing lined up? Could I first speak with three people, send five applications, try a side project, take two weeks off, or run a small experiment?
Step eight: coordinate
What does my partner think? What risk would my family bear? Could my manager adjust the role?
Step nine: learn from feedback
After these small experiments, which of my original judgments no longer hold?
This is learning.
It does not require finishing your education before starting your life. Knowledge unfolds as life calls for it.
The best role for AI here is more than announcing, “Based on your circumstances, I recommend resigning.” It can help identify missing variables, generate competing hypotheses, construct a simple decision model, challenge your reasons, organize interviews, and explore scenarios.
Then it should send you back into the world to obtain information that is real.
18. The person most at risk may know AI very well
I worry about a different kind of person.
They are excellent at using AI. They can generate anything: a report in ten minutes, code in half an hour, a business plan overnight, apparent fluency in an unfamiliar field within a day.
They seem to possess unlimited ability.
Yet gradually they lose their sense of when to doubt, when a problem has not been properly defined, when text cannot settle a question, when they should speak to a real person, when the data themselves cannot be trusted, and when the objective they are optimizing is absurd.
This may be a new form of ignorance in the age of AI:
Having answers, but losing control of the cognitive loop that gives them meaning.
19. What would genuinely AI-native learning mean?
If I had to offer a definition today, it would be this:
AI-native learning redesigns a person's cognitive system on the assumption that external intelligence is readily available. It asks which knowledge must be internalized, which can be called on externally, which judgments must remain our responsibility, and which insights can only be acquired through real action.
That is different from using AI to assist an existing educational process.
An internet-native business was never merely a paper directory moved onto a webpage. Likewise, AI-native learning cannot simply be a textbook placed inside ChatGPT.
It should begin again with a question:
If an external intelligence can already explain, retrieve, program, simulate, and generate whenever we need it, would we design a person's education the way we do now?
I suspect we would not.
20. Disciplines will remain. Their place may change.
I am not announcing that mathematics, history, and biology departments should be dismantled. I think we need real experts more than ever.
AI will produce more and more material that looks like knowledge. Deep disciplinary traditions preserve the distinctions we need: what counts as proof, what counts as evidence, what can be replicated, what makes an explanation good, and which mistakes people already made a century ago.
Disciplines should remain. But a person may need another interface.
The university's underlying structure can still be:
Mathematics / Physics / History / Biology / Philosophy / Computing …
A person's navigation can become:
Observe
Represent
Model
Reason about causes
Value
Decide
Act
Coordinate
Learn
Organize the underlying system by the sources of knowledge.
Organize the interface by the needs of action.
The two maps are not enemies. A strong learner can move between them.
21. Perhaps education should cultivate a world model and the capacity to act
We used to say that education was about acquiring knowledge. Later, we said it was about developing abilities. But “abilities” is still vague.
If I were to define education today, I might say:
Education gradually builds a sufficiently good world model within a person, while developing their capacity to choose actions, test results, and keep updating themselves even though that model will always be incomplete.
This includes knowledge and judgment, mathematics and values, theory and action, oneself and other people.
AI has not made learning meaningless. Quite the opposite. As answers become easier to obtain, we can see with unusual clarity that:
Knowing the answer is only a small part of intelligence.
There is also knowing where to look. Knowing what matters. Knowing why. Knowing what to do. Knowing why you want to do it. And knowing, once the results arrive, when to admit:
My model of the world was wrong.
A final thought: do not build your mind into a library
When knowledge is scarce, the most valuable person may be the one with the most books. When retrieval is scarce, it may be the person who remembers the most.
But when an external intelligence can retrieve, explain, translate, calculate, and generate for us whenever we ask, a person's value may no longer depend primarily on how many files are stored in their head.
What we need to build may be a new operating system.
One that knows when to observe, when to model, when to doubt, when to calculate, when to ask another person, when to call on AI, when to turn it off, when to act, and when to stop and ask again:
Why am I trying to get there in the first place?
If I had to compress this entire essay into one sentence, it would still be:
Disciplines are the folders of human knowledge. They should not have to be the navigation bar of a human mind.
Traditional education has largely organized knowledge around “What is the world made of?” Personal learning in the age of AI may need an additional map:
How can a finite person see, understand, choose, and act in a changing, incompletely knowable world full of other agents—and continue learning after getting things wrong?
If we really begin to redesign learning around that question, we may only just be approaching what an AI-native education could mean.
Appendix A · A simple mathematical model of the agent's knowledge structure
The discussion above can be expressed through a deliberately rough but suggestive mathematical framework.
Suppose the world has a hidden state at time :
The agent usually cannot observe directly. Instead, it receives an observation:
It must construct an internal state or belief from the information available so far:
The agent also has a world model:
which asks: if the world is roughly like this, and I take action , what might happen next?
Prediction alone is insufficient. The agent needs some account of value:
or, more generally, an evaluation of a future trajectory:
Decision-making can then be written as:
where is a policy.
After an action, the environment changes:
and the agent receives another observation and updates its belief:
The loop is:
This resembles the basic idea of a POMDP: an agent that cannot directly observe the full state of the world maintains a belief over possible states and chooses actions accordingly.[4]
We can map different kinds of knowledge onto this structure:
| Mathematical object | Meaning for personal learning |
|---|---|
| What I can observe | |
| How reliable my measurements are | |
| What I currently believe about the world | |
| My world model | |
| Causal model | What happens when I intervene |
| What I consider worthwhile | |
| My decision rule | |
| Action | What I actually do |
| Prediction error | Where the world tells me I was wrong |
| Update | Learning |
The interesting shift is this:
Knowledge becomes more than a collection of facts about the world. It becomes the set of models an agent needs to complete the loop.
Of course, human beings are much more complicated than a POMDP. Our goals change. Our values are not a fixed reward function. We invent new actions, redefine the state space, reconsider what we care about, and face other agents with goals of their own.
The claim is not that a human being is a POMDP agent. The framework is conceptual scaffolding. It helps us ask, more precisely, what function each kind of knowledge serves in action.
Appendix B · An AI-native learning routine you can begin today
When you encounter a question you truly care about, set aside “Which book should I read first?” for a moment. Begin with these ten steps.
Step 1 · Write down the question
Before asking AI, write it yourself: what, exactly, do I want to know?
Step 2 · Describe your observations
Which things are facts? Which are feelings? Which information came from someone else?
Step 3 · Define the state
If I could track only five variables, which would I keep?
Step 4 · List three competing explanations
Do not look only for material that confirms your view. Ask AI to help generate explanations that challenge it.
Step 5 · Build a world model
If explanation A is right, what should I observe? What if explanation B is right?
Step 6 · Find evidence that distinguishes the models
Do not gather information indefinitely. Ask: which observation would best distinguish these hypotheses?
Step 7 · Clarify what you value
What am I trying to optimize: speed, money, freedom, accuracy, experience, long-term growth?
Step 8 · Take the smallest useful action
Design a low-cost experiment you can reverse.
Step 9 · Get feedback from reality
Speak to people. Build a product. Run code. Observe data. Carry out an action.
Do not let AI simulate reality and then mistake its simulation for the world itself.
Step 10 · Update your model
Finish by writing three sentences:
I used to think …
Reality showed me …
So now I think …
Over time, those three sentences may bring you closer to learning than a collection of ten thousand saved articles.
References
[1] National Academies of Sciences, Engineering, and Medicine. How People Learn II: Learners, Contexts, and Cultures. Washington, DC: The National Academies Press, 2018. DOI: 10.17226/24783. See especially Chapter 5, “Knowledge and Reasoning.”
[2] Simon, Herbert A. The Sciences of the Artificial. 3rd ed. MIT Press, 1996; reissued 2019. DOI: 10.7551/mitpress/12107.001.0001.
[3] Sutton, Richard S., and Andrew G. Barto. Reinforcement Learning: An Introduction. 2nd ed. MIT Press, 2018. See especially the agent–environment interface and Markov decision processes.
[4] Kaelbling, Leslie P., Michael L. Littman, and Anthony R. Cassandra. “Planning and Acting in Partially Observable Stochastic Domains.” Artificial Intelligence 101(1–2): 99–134, 1998. DOI: 10.1016/S0004-3702(98)00023-X.
[5] Roediger, Henry L., and Jeffrey D. Karpicke. “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention.” Psychological Science 17(3): 249–255, 2006. DOI: 10.1111/j.1467-9280.2006.01693.x.
[6] Bastani, Hamsa, et al. “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences 122(26), e2422633122, 2025. DOI: 10.1073/pnas.2422633122.
[7] Kestin, Greg, et al. “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports 15, 17458, 2025. DOI: 10.1038/s41598-025-97652-6.
[8] Lee, Hao-Ping, et al. “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” CHI 2025. DOI: 10.1145/3706598.3713778.
[9] Russell, Stuart. “Human-Compatible Artificial Intelligence.” 2019. Related ideas are also developed in Human Compatible: Artificial Intelligence and the Problem of Control.
[10] Dewey, John. Democracy and Education. 1916.
Author's note. The claim that traditional education organizes knowledge around what the world is made of is a conceptual simplification used to examine the organization of modern disciplines, not a rigorous historical claim about how they developed. “Observe–Model–Value–Act–Learn” and the nine capacities discussed here are a synthetic framework proposed in this essay for thinking about personal learning with AI. They are not a standardized model accepted across educational science.