0 × 100 = 0,
1 × 100 = 100,
and in between.

More and more often, I hear some version of the same question from students:

Why should I learn this? AI can already do it.

It is a good question. An agent can write code, TL;DR papers, draft reports, generate visualizations, and produce plausible answers to an alarming number of questions. Pretending these tools do not exist is ridiculous. Asking students to ignore them is not education. It is theatre.

But the conclusion many people draw is wrong. If AI can do more, learning does not become less valuable. It becomes more valuable.

The multiplier

A useful model is deliberately simple:

The 100 is the AI system. It is not a literal number. It stands for a tool that lets one person attempt, build, test, write, analyse, and communicate far more than before. The important variable is the first one.

Someone with no React experience can ask for an application. Someone who has never written SQL can obtain a query. Someone unfamiliar with a subject can get an explanation that sounds convincing. AI can supply implementation, syntax, and a first answer.

So You cannot mean facts memorised by heart or the number of APIs you can reproduce unaided. It means the ability to frame a problem, set priorities, test an answer, spot a bad assumption, ask a better question, connect ideas, and take responsibility for the result. In short, it means judgment.

That distinction matters because output that looks “finished” is becoming a weaker signal of competence. A student can submit elegant code, a good-looking literature review, or a convincing explanation without having the mental model needed to defend, reproduce, or correct it. The output may look like 100, but the real result is not necessarily 100.

Why learning has more leverage

The usual conclusion is that powerful AI makes studying less necessary. The logic goes the other way. If a tool multiplies your capability, every improvement in the thing being multiplied matters more.

0.2 × 100 = 20
0.3 × 100 = 30

The improvement is still 0.1. Its effect is now ten times larger. The numbers are rhetorical, but the direction is clear: a better tool raises the return on becoming better yourself.

This applies well beyond programming. A strong researcher can use AI to explore more hypotheses, compare alternatives, and spend more time on difficult questions. A strong writer can use it to improve an idea rather than just manufacture one. A good manager can test scenarios faster without mistaking plausible prose for insight. The difference is not the volume of output. It is the quality of the person directing it.

The multiplier also works in the other direction. -0.1 × 100 = -10. Weak judgment does not become harmless when paired with a powerful system. It becomes scalable. A bad architecture can be produced faster. A weak experiment can look more rigorous. A bad decision can arrive with a cleaner deck, a more confident forecast, and much better language.

AI can make you faster at being right. It can also make you spectacularly faster at being wrong. The question is not only, “Can AI do this?” It is, “What exactly is being amplified?”

Vibe coding is not ownership

Pure vibe coding deserves a harder name. If it means describing the desired result, accepting whatever appears, and turning off the part of your brain that asks what changed and why, then You is not literally zero. But it approaches zero in the variable that matters: judgment and ownership. You are not building a system. You are buying increasingly specific guesses.

That can be fine for a throwaway personal prototype. It is not a responsible default for anything that must survive users, money, private data, regulation, an attack, a production incident, or simply next month. Code you cannot explain, test, modify, or replace is a demo, not a product.

Imagine asking an agent to build a small invoicing application. In an hour, it produces a login page, a dashboard, a database, invoice downloads, and a reassuringly modern interface. Then a customer changes an identifier in a request and sees another customer’s invoice. Or an export endpoint returns data from every tenant because nobody defined the authorization model. The agent generated the happy path. It could not infer the boundary that matters: who is allowed to see what, under which conditions, and how you prove that rule still holds after the next change.

The point is not that an agent cannot write a secure endpoint. It can. The problem is that no generic prompt contains the business rule, threat model, legal obligation, data boundary, operational constraint, and failure tolerance that make an endpoint safe in its actual context. A prettier interface and ten more prompts do not solve a security model if you never understood.

Research has the same problem, but the failure is deeper than a fabricated citation. A model may produce an excellent literature review and you may still understand nothing in the papers it mention. Reading is not extracting sentences. It is knowing the claim, assumptions, method, comparison, and the reason the authors might be wrong.

Without that mental model, you do not ask the questions that matter. You do not ask whether the train-test split leaked information, whether the baseline used comparable settings, whether the metric measures the real objective, or whether a claimed state of the art came from a different protocol. The danger is not that a model always misses these questions. It is that, if you outsource the thinking, you may not know that verification is required.

As of 2026, leaving an agent alone with “build the product” and expecting it to infer the correct mental model is not autonomy. It is delegation without ownership.

Pure vibe coding is a good gift to OpenAI, Anthropic, or whichever vendor hosts the model. You provide more prompts, more context, and more inference work (more tokens accumulated both in input and output). They provide fluent output. There is nothing wrong with paying for leverage. The mistake is confusing paid inference with accumulated capability.

What gets cheaper

Some work gets cheaper. When a task mainly translates a clear specification into familiar syntax, boilerplate, screens, or API calls, it can require less time and fewer people. That is a shift in leverage, not a moral judgment.

This does not erase front-end engineering or deep API knowledge. It makes memorising the shape of a call less scarce. What remains scarce is knowing which call belongs in which architecture, what it costs, what it breaks, and how to replace it. Automating a task does not automatically erase a profession, but the pressure falls hardest on mechanical implementation after somebody else has already done the thinking. The leverage moves to people who define the real problem, build abstractions that survive change, evaluate trade-offs, and verify behaviour outside a demo.

There may be fewer people per task. The people with the most leverage are not the ones who type fastest. They are the ones who understand what should be typed, why it belongs there, and what happens when it is wrong.

Education is about changing the person

Banning AI would preserve an old ritual while ignoring the real change. These tools are becoming part of engineering, science, writing, analysis, and knowledge work. Professionals (and students as future ones) should learn to use them well.

The educational question should move away from, “Can you produce this without assistance?” It should move toward better tests of understanding: Can you explain why it works? Can you detect when it fails? Can you verify the assumptions? Can you improve the result? Can you formulate a better problem? Can you make it useful in a real context?

Some old tasks should disappear. Memorising every API, writing boilerplate by hand, and treating manual effort as a virtue are poor goals in a world with capable agents. But replacing them with nothing would be worse. The goal is not to preserve difficulty for its own sake. It is to develop mental models, abstraction, taste, evaluation, creativity, scientific judgment, and the ability to understand systems.

Every assignment produces two outputs. The first is the submitted artifact: code, report, design, or analysis. The second is the person who exists after doing the work. AI makes it easy to optimize the first while accidentally skipping the second. In education, the artifact is often disposable. The mental model remains.

The real deliverable is not the assignment. It is the updated version of the student.

Use the tool twice

AI is not only the ×100. Used well, it can also improve the number on the left side of the equation. It can be a tutor, explainer, sparring partner, code reviewer, generator of exercises, or source of immediate feedback.

The same tool supports two very different habits. In the first, AI becomes substitution: “Do this for me”. Output rises while underlying capability does not move. In the second, AI becomes acceleration: “Help me understand this. Challenge my reasoning. Show me the assumption I am missing. Give me a harder version. Do not give me the answer yet.” Output rises, and so does the person.

That choice is not limited to students. Every professional faces it now. When AI gives you a shortcut, you can consume the saved effort or reinvest it. You can let the machine replace part of your thinking, or use the time it gives back to operate at a higher level.

AI will keep getting better. The multiplier will keep getting larger. That part is probably outside your control. The number being multiplied is not.