Everything That Works Feels Wrong

Clear teaching can create confident forgetting

Feeling that you understand something and being able to produce it when asked are not the same thing. Unfortunately, the first feeling is a poor guide to the second.

Interviews only test the second.

Good explanations feel easy to follow, and that is useful. Clear teaching helps you make sense of unfamiliar material in the first place. But ease creates a problem, because we tend to mistake fluency for learning.

If somebody explains a DCF beautifully, every step can feel obvious while it is on screen. That feeling disappears surprisingly quickly once the screen is gone and you have to reconstruct the answer yourself. This is sometimes called the fluency illusion.

The practical consequence is simple. Understanding an explanation is not evidence that you can reproduce it.

That distinction matters particularly for interview preparation. A video can help get the idea in. The interview asks whether you can get it back out, and those need different forms of practice.

Ten study techniques were tested. Two stood out.

In 2013, Dunlosky and colleagues reviewed ten widely used learning techniques across different ages, subjects and outcomes.

The rankings are uncomfortable, because the methods people naturally gravitate towards did not perform especially well.

Difficulty is sometimes the mechanism

There is a pattern in that table. The comfortable methods mostly let you recognise information. The stronger methods force you to produce it.

Robert Bjork's phrase for this is desirable difficulties. Some conditions make practice harder today and improve what remains available later.

Spacing does this. Testing yourself does this. Mixing different types of problem does this. Trying to answer before you feel completely ready does this.

The important distinction is between performance during practice and learning that survives afterwards. Those two can move in opposite directions.

A smooth session feels reassuring because everything is available to you. A retrieval session feels worse because it exposes what is not.

Exposing the gap is the useful part. If you cannot retrieve something today, you have found the thing worth working on, and rereading can hide that from you.

The finding interview candidates under-use most

Interleaving deserves more attention than it gets.

In one randomised trial across 54 classes and 787 students, Rohrer and colleagues compared two groups. One practised problems grouped by topic. The other practised the same problems, mixed together. A month later, the interleaved group scored 61% against 38%.

The reason is that mixed practice forces you to decide what kind of problem you are looking at before you can solve it. Blocked practice quietly gives you that answer. If every question on the page is about WACC, you never have to recognise a WACC problem, because the heading did it for you.

Now think about an interview. Nobody says: the next five minutes will be accretion and dilution.

You get a question. A company buys another company. Maybe the important issue is financing. Maybe it is synergies, or valuation, or accretion. Recognising the mechanism is part of the test.

Yet most interview preparation is organised by chapter. Most interviews are not.

What good learning products should optimise for

Put that evidence next to the way most study tools are designed and there is an obvious tension.

Products can easily measure minutes watched, questions completed, sessions, streaks, badges and courses finished. What you actually care about is harder to see: whether you can still produce the answer after ten days of not looking at it.

That leads to a different set of priorities.

Adapt the difficulty, not the decoration

The learning-styles idea is a useful example. People clearly have preferences. Some like diagrams, some prefer listening, some would rather read.

The unsupported leap is that teaching somebody in their preferred format reliably makes them learn more. That has not held up well when tested directly.

A more useful form of personalisation is much less glamorous. Change the difficulty.

If you keep missing terminal value, terminal value should come back. If you know the formula but cannot explain why it works, you should be asked for the explanation. If you can answer the standard version, the wording should change. If you keep getting something right, it should stop appearing every day.

That is genuine adaptation. Changing the animation is presentation. Changing the next question is learning.

What AI actually changed

There is an old result in education that has shaped decades of thinking about personalised learning. Benjamin Bloom's 1984 paper found very large gains from one-to-one tutoring combined with mastery learning, compared with conventional classroom instruction.

Later evidence has generally produced more modest effects. The underlying intuition remains powerful: individual feedback is valuable, and historically it was expensive.

AI changes that cost.

A particularly interesting example came from a 2025 randomised study in an introductory Harvard physics course, published in Scientific Reports. Students using a carefully designed AI tutor learned substantially more, in less time, than students in the comparison classroom condition.

The important part is not simply that the tutor used AI. Look at how it worked. The system had the correct material available. It controlled how much information appeared at once. It asked questions rather than handing over answers. It checked understanding continuously.

It was designed as a tutor, which is very different from putting a general-purpose chat box beside a textbook.

A chat window is not automatically a tutor

Open-ended assistants have a natural tendency to help. That sounds like an advantage, and during retrieval practice it can become a problem.

You give a half-complete answer and the assistant helpfully completes it. You miss a point and it says you were essentially correct. You ask whether you are ready and it gives you encouragement.

All of those interactions feel good. None of them tells you accurately what you can produce alone.

For finance interviews there is a second problem: a model without a fixed marking scheme can quietly invent its own standard and then grade you against it.

So constrain it. Give it the rubric. Tell it not to fill in missing points. Make it distinguish what you said from what a correct answer would contain. Ask it to quote your own words as evidence for its judgement. When you are wrong, make it leave the gap visible before it teaches you anything.

The interesting use of AI in learning is not that it can explain almost anything. It is that it can listen to almost anything you produce and compare it against a standard. That used to require a person.

Two things that used to be expensive

The first is marking what somebody actually says.

Traditional self-study leans heavily on multiple choice, because multiple choice is easy to grade. An interview does not give you four buttons. It asks you to walk somebody through a DCF, and the relevant question is whether you can construct the answer yourself. Free-response practice is now far easier to grade at scale.

The second is remembering what you forgot.

A system can track dozens of concepts individually. It can know that you missed merger accounting three times and have WACC cold, bring one back tomorrow, and leave the other alone for three weeks. None of that is especially difficult technology any more.

What remains difficult

Deciding what somebody should learn in the first place.

What are the eighty concepts that actually matter? Which ones depend on which others? Which questions genuinely appear in interviews? What separates a technically correct answer from a strong one?

That is curriculum design, and better models do not automatically produce better curricula.

The next generation of useful learning products may therefore be differentiated less by which model they use, and more by whether somebody did the unglamorous work of deciding what is worth learning.

The approach we are trying

That thinking is what sits behind L3 Primer.

A topic is broken into the handful of concepts it actually contains. You learn each one briefly, then answer before you feel completely ready, then explain the full idea from memory. Your answer is checked against the points a strong answer should contain, and the things you miss come back later.

Once you know several topics, the questions begin to mix together, so you no longer know what is being tested until you recognise it yourself.

The short video is the easy part. The loop afterwards is where the learning happens.

Six prompts to make your own revision harder

You do not need L3VLUP to use any of this. Most of the useful changes are constraints on how you study and how you use an assistant.

The common principle is the same in all six: make the assistant reveal the gap before it helps you close it.

They are built for six different moments.

• When something feels easier than it probably is
• When your revision plan is too comfortable
• When you want an assistant to behave like a tutor
• When you keep making the same mistake
• When you think a topic is finished
• When somebody is selling you a new learning method

I have just finished studying the topic below and I feel like I understand it. Test whether I actually do.

TOPIC: [e.g. how a DCF gets from cash flow to share price]
WHAT I THINK I KNOW: [three or four sentences, written from memory with nothing open]

Do this in order:
1. List, one line each, the specific claims my summary actually makes.
2. List what a complete explanation contains that my summary does not. Be specific about what is missing, not about how well I wrote.
3. Ask me one open question about the largest gap. Then stop and wait.

Rules:
- Do not praise the summary and do not tell me it is a good start.
- Do not fill in a gap because I obviously meant it. If I did not write it, it is missing.
- Do not teach the topic at any point. I will go back to the source for that.

Here is how I am revising. Rewrite it so it is harder in the ways that improve retention, and easier nowhere.

MY CURRENT PLAN: [what you actually do, honestly]
TIME I HAVE: [hours per week]
DEADLINE: [date]

Make these four changes and name where you applied each:
1. Replace re-reading, re-watching and highlighting with retrieval. Same material, but I produce it instead of reviewing it.
2. Split any long single-topic session across separate days. Keep the total time identical.
3. Mix topics inside a session once I have more than two, and never put two of the same topic in a row.
4. Move every self-test earlier than is comfortable. Before I feel ready, not after.

Then tell me in one sentence which part of the new plan will feel worst, and why that is the part doing the work.

Do not add hours. If something has to go, say what goes and why.

For the rest of this conversation you are my tutor on one topic, under fixed rules. Confirm you have them, then begin.

TOPIC: [e.g. LBO mechanics]
WHAT I ALREADY KNOW: [one or two lines]
WHERE I NEED TO GET TO: [e.g. answering this out loud, unprompted, in an interview]

RULES:
1. One idea per message. Never more than one new thing, never more than about 120 words.
2. Never hand me an answer I could produce myself. Ask first. If I am stuck after two attempts, give me the next step only, never the conclusion.
3. End every message with a question I have to answer.
4. When I am wrong, say so and say exactly what is wrong. Do not soften it, and do not quietly rewrite my wrong answer into a right one.
5. When I am right, move on. No praise paragraphs.
6. If I ask you to just tell me, refuse once and ask again. If I ask twice, tell me, then have me say it back in my own words.
7. Every fifth exchange, ask me to summarise everything so far from memory.

Start by asking what I already believe about this topic. Explain nothing yet.

The same thing keeps going wrong and repeating it is not working. Find what is actually missing.

WHAT I KEEP GETTING WRONG: [the concept or question]
WHAT I SAID THE LAST TIME I GOT IT WRONG: [as close to verbatim as you can]

Do this:
1. List every concept this one sits on top of. The things I would need to already understand for it to make sense.
2. Write one short open question per prerequisite, answerable in a sentence, that would show whether I have it.
3. Ask them one at a time, most upstream first. Wait for each answer.
4. Stop at the first one I cannot answer cleanly. That is where I go back to, not the thing I thought was the problem.

Do not explain any of the prerequisites. Finding where the floor gives way is the whole job.

Ask me this topic in a form I have not seen it in.

TOPIC: [e.g. accretion and dilution]
HOW I HAVE PRACTISED IT: [e.g. always as "walk me through it", always in the same order]

Write three questions that need the same understanding but do not look like the practised version:
- One that gives me the answer and asks me to work backwards to the assumption that produced it.
- One set in a context I have not studied, where the same mechanism applies.
- One where the obvious method is the wrong one, and the skill is noticing that before starting.

Give me the three questions only. Hold everything else until I have attempted all three.

Then mark each on two things separately: did I identify the right approach, and did I execute it. Those are different failures and I need to know which one I made.

I am about to change how I study based on the claim below. Tell me what is actually known.

THE CLAIM: [e.g. "match the material to your learning style", "highlighting helps", an app's pitch]

Answer in this order:
1. State the claim precisely enough that it could be wrong. Vague versions are unfalsifiable, so give me the version a study could fail.
2. The strongest evidence for it: study design, roughly how many people, and how long after training the outcome was measured.
3. The strongest evidence against it, on the same three terms.
4. Whether the effect, if it is real, showed up on immediate performance or on delayed retention. Say which, because most study advice looks good on the first and vanishes on the second.
5. Your verdict in one line: well supported, mixed, or widely believed and unsupported.

If nobody has tested it properly, say that instead of reaching for adjacent research that sounds close.

What none of this fixes

There is one important limit. Everything above is about content: knowing it, remembering it, producing it without help. That is necessary for an interview and it is not sufficient.

A marking rubric can tell you that your DCF answer missed terminal value. It cannot easily tell you that the answer was painful to follow. It cannot hear that your conclusion disappeared underneath six caveats, or see that you became defensive after one challenge, or notice that you knew the answer and sounded as though you did not.

Those are different problems. Spacing and retrieval get you to the point where knowledge is no longer the bottleneck. After that, how the answer lands starts to matter.