What Counts as Knowledge?

I recently met two young people, both fresh out of college.

One had a master's in computer science from a prestigious university, excellent grades, and more practice with algorithm problems than most people I know. The other had an undergraduate degree from an ordinary school, in a field that wasn't even a close match for the job. They interviewed for the same role. The company gave each a small assignment: take a real business need and, in two days, deliver something that actually worked.

The second candidate got the offer.

I've thought about that for a while. A traditional professor might say the first candidate plainly had more knowledge, in a more systematic and rigorous form. That may be true. The company was also right to choose the other person. So what do we mean when we talk about "knowledge"?

My answer is this: real knowledge is the ability to get something done under real constraints and judge whether the result is any good. AI has pushed the price of the part that can be recited or tested on an exam close to zero.

1. Knowing the answer is worth less than it used to be

Start with a simple point that's easy to overlook: knowledge is a kind of product, and its price depends on how scarce it is.

For a long time, knowing something was valuable in itself. You knew how to derive a formula, interpret a regulation, or use a programming language when other people didn't. That knowledge gave you an edge. Much of the university system, from courses and exams to degrees, grew around that scarcity: put knowledge into students' heads, then use exams to check that it stuck.

The cost of getting information has changed. Anything that can be written down, searched, or repeated can now be produced by an AI in seconds, often with a reasonably accurate answer. It doesn't get tired or forget, and the answer may cost nothing. Something that took you four years to memorize can be retrieved by someone else in four seconds.

The thing losing value isn't knowledge itself. It's the kind of knowledge we test by asking people to produce an answer from memory.

The reason is straightforward. An exam typically asks for a standard answer, recalled without outside help. That happens to be a task AI handles well. If a model can ace an exam, the exam is testing something a machine can now supply cheaply. The test hasn't become easier; what it measures has become less scarce.

You can see this in courses that haven't changed in years. Their content may be correct: how to use office software, work with a particular database, or follow a textbook built around one medium. But students finish and find that a chat window can supply much of what they learned. The material hasn't become false. Its value has changed.

That shift affects individuals even more directly than it affects courses. If your main advantage is "I know this," you're in a market where that advantage is getting cheaper. Faster than most people want to admit.

2. Knowledge has to survive delivery

So what hasn't lost value?

Back to the two candidates. What did the second one do? As I understand it, he spent half a day clarifying three points with the business team. He turned a vague request into clear requirements. Then he broke the assignment into smaller tasks and used AI on each one. He ran every piece himself, fixed what failed, and delivered something usable after two days. It wasn't polished, but he knew what was fragile and what still needed work.

The process came down to four abilities: explain the problem, break it down, make something, and test it.

To explain the problem is to turn a vague idea into instructions another person or a machine can act on. It sounds easy. Often it's the hardest part. People struggle to do the work because they haven't yet decided what they want.

To break it down is to divide a large problem into smaller ones, then understand their order and dependencies. That takes a structural understanding of the problem itself.

To make something is to turn an idea into a real object or result. AI is taking on much of the routine work involved.

To test it is to decide whether the result is correct, useful, and good enough to use.

AI is lowering the cost of the first three. The cost of the last one is rising. Once generation becomes cheap, the amount of output grows sharply; someone still has to decide whether each piece can be trusted. If a tool can give you a hundred answers a second, the person who can tell which answer is sound becomes the bottleneck.

I think of this as a scarcity of acceptance: companies increasingly want to know not just "Can you do the work?" but "Can I trust what you deliver enough to use it?" AI may help with the first. The second still depends on your judgment.

That helps explain the gap between someone who can pass a test and someone who can deliver. In the first case, knowledge sits in a person's head; others can't see or verify it. In the second, it appears in something concrete that can run, be checked, and be used. That makes it possible to test and price.

3. The stronger AI gets, the more the fundamentals matter

You might reach this conclusion: if AI can do everything, there's no need to study the basics. Just learn how to direct the tool.

That conclusion is wrong, and risky.

Take a company's financial statements. Feed them to an AI, and within minutes you get a polished write-up. Some of its judgments will be sound; others may only sound plausible. It might mistake a one-time gain for an operating improvement, read an accounting change as real business growth, or describe an industry's competitive structure fluently while missing the constraint that matters.

Someone who understands accounting and the industry may spot the problem at once. Someone who doesn't may simply admire the write-up.

Both people used the same tool and got the same output. The difference was whether they had the grounding to judge it.

Code works the same way. AI can write hundreds of lines in a minute. Whether that code belongs in production depends on whether the person reviewing it understands how the system works, where its boundaries lie, and what could make it unsafe. If you don't, you're handing a client something you can't assess.

That's why I say acceptance is becoming scarce, and why it depends on solid knowledge of the field. The two ideas are sides of one change: knowledge that can be copied cheaply is losing value; knowledge that lets you judge the copy is gaining it.

The difference is between knowing an answer and understanding a problem. Someone else can supply an answer. Understanding can't be handed over so easily. It means knowing why the problem is hard, where mistakes tend to occur, and how an answer that looks right might fail. You get that through sustained, systematic study, not a few rounds in a chat window.

AI amplifies the judgment you already have. With sound judgment, it can extend your reach. Without it, you may make mistakes faster, and make them more convincingly.

That's why I find the phrase "mastering AI" a little misleading. What you master is a field or a skill. AI is a lever you use to do that work. A longer lever is useless without a fulcrum.

4. Knowledge has a half-life

If foundational knowledge matters, should we try to learn everything? Of course not. Time is limited, so we need to decide what deserves it.

I find it useful to think of knowledge in three layers, each with a different half-life.

The outer layer is tools: how to prompt a particular model, call a framework's interface, or combine features in a product. It changes quickly. A best practice from six months ago may already be obsolete. Learn what you need when you need it; don't build your whole education around it.

The middle layer is method: how to define a vague problem, break a large one into parts, design a test, and get closer to a goal through iteration. This lasts much longer. A strong engineer from twenty years ago and a strong engineer today still share these habits.

The inner layer is the field itself: basic economics, how financial assets are priced, the mechanics and systems thinking behind engineering, and statistical inference. These foundations rarely expire. They're also what make sound evaluation possible.

Once you see the layers, the trade-off becomes clearer. Spend most of your time on tools and you're chasing the fastest-moving target. By the time you catch it, it may have changed. Learn tools without methods or a field, and each technology shift can wash away another part of your advantage.

For educators, that means we can't simply chase the newest tools. A framework taught today may be old technology in three years. Methods and fundamentals deserve more of the time because students can use them for life and adapt faster when the tools change.

That doesn't mean ignoring tools. Courses need a way to stay current, so students encounter living practice rather than a decade-old textbook. The materials should change; the underlying structure should hold.

The same question applies to your own learning: how much time do you spend on tools, and how much on methods and the field you work in? If nearly all of it goes to the first layer, the return on your effort may be low.

5. What you test matters more than what you cover

Here's one more point, and the one I think is easiest to miss: what someone learns depends more on how their learning is tested than on the content alone.

People put effort where it will be measured. If the test is an exam, they memorize, practice questions, and work out what the examiner likes. If the test is a real deliverable, they have to ask what problem it should solve, how to solve it, and what it takes to make it work.

The same material, tested in different ways, can produce very different people.

That's why I'm skeptical of reforms that simply add AI material to a course. If students hear a few lectures on AI but still take a conventional final exam, they'll learn the exam material: the names of model architectures, perhaps a few prompting tricks. They may still have no idea how to use the technology to get something done at work.

Change the content but keep the test, and the result is unlikely to change.

Assessment matters for another reason: it determines how quickly people get feedback. An exam gives a score, often long after the work, without showing exactly where the student went wrong. A deliverable gives immediate feedback. Does it run? Does the user find it useful? You can try it and see. That density of feedback is a major source of learning.

There's a qualification. Education can't be judged entirely by whether a company wants the result. Some things students learn at university help them understand the world, form judgments, and become thoughtful people. Those things may not turn into a deliverable right away. Reducing education to job training would be a loss.

That doesn't contradict my point. If education is meant to cultivate understanding and judgment, it should be less dependent on testing what a machine can reproduce from memory. Real understanding means being able to apply an idea in a new setting, explain it, make a judgment, and produce something. A real deliverable may be a better test of that than recall alone.

Back to the two candidates

I want to return to the two young people at the beginning.

I don't think the first candidate was unqualified. Quite the opposite: he had a strong foundation. If he changed how he tested his own learning, he could catch up quickly. The second candidate had gaps too. Without filling them, he'd soon hit a ceiling.

For young people, my advice is concrete: stop measuring yourself by how much you've studied. Pick a small, real goal. Make something, hand it over, and let someone use it. You'll discover which parts you understand and which only feel familiar. That discovery can teach you more than a hundred exam papers.

If you follow investing and AI, take an inventory of what you know. Which parts are "exam knowledge," and which can survive delivery? Knowing what a price-to-earnings ratio means is exam knowledge. Taking a particular company and explaining why its earnings are, or aren't, durable, then being willing to put real money behind that judgment, is closer to delivery.

AI can help with the first. The second has to be yours. This week, choose one thing you've wanted to understand and turn it into something you can show someone else: an analysis, a small tool, or an essay someone could challenge.