Learning with It

For the first time in history, anyone can have a private tutor: available at any hour, patient beyond any human limit, able to explain the same thing eleven different ways without a trace of irritation.

And the overwhelmingly common use of it is: ask a question, read the answer, close the tab. That's a reference book. A reference book is useful, but nobody ever got a personal tutor and used them as a dictionary.

Why tutoring is the right frame

In 1984 Benjamin Bloom published the result that has haunted education ever since: students taught one-to-one with mastery learning performed about two standard deviations better than students in a conventional classroom — the difference between an average student and the top few percent. The 2-sigma problem was the gap between knowing this and being unable to afford a tutor for everyone.

Two honest caveats. The exact 2-sigma figure has never been reliably reproduced, and later meta-analyses put the tutoring advantage lower — real and large, but not miraculous. And what produced the effect wasn't the tutor's knowledge. It was the loop: constant checking, immediate correction, difficulty matched to the student, and nothing moving forward until the current piece is solid.

That distinction decides everything about how to use AI here. Knowledge alone gives you a textbook that talks. The loop is what teaches — and the loop is exactly what people don't ask for.

A bad tutor explains. A good one asks.

The early evidence is encouraging on this point. In a 2024 study at Harvard, Gregory Kestin and colleagues had physics students learn a topic either in a well-run active-learning class or with an AI tutor built to work this way — asking, checking, pacing. The AI-tutored students learned more, in less time, and reported higher engagement. The design of the tutor was the thing being tested, not the model.

Six modes worth having

Default behaviour is "lecturer". You have to ask for anything else.

Examiner. The most valuable and least used. "Ask me questions on this topic one at a time. Don't tell me the answer until I've tried. After each answer, tell me what's missing, then adjust the difficulty." This is retrieval practice — the strongest learning operation there is — and the model is unusually well suited to running it forever without getting bored.

Feynman partner. "I'm going to explain this to you. Interrupt wherever I'm vague, and ask the question that would expose whether I actually understand." You generate; it marks. Explaining is how you find out what you only think you know, and this makes it available without a study partner.

Explainer, on your level. Its real advantage over a textbook: "explain this three ways — as an analogy, formally, and through a worked example", then "now assume I understood none of that and try a different angle." Just remember that a clear explanation produces the feeling of understanding either way — always close it and reproduce it before believing yourself.

Error analyst. For anything practical, this is where the value concentrates. Not "what's the right answer" but: "here's my solution and my reasoning. Where exactly did the reasoning go wrong, and what wrong belief would produce this specific mistake?" The misconception behind an error is worth more than the correction.

Problem setter. "Give me five problems that would reveal whether I actually understand this, ordered by difficulty, and don't show solutions." Then, straight from the chunking article: ask for a worked example first, then a partially worked one where you fill in steps, then a bare problem. Scaffolding removed on purpose is how the skill assembles.

Opponent. "Argue against my conclusion as strongly as you can." The counter-case habit from article 3, aimed at your thinking.

One more with an immediate payoff: card generator. Have it turn what you just learned into question-answer cards for spaced repetition, using the rules from that article — one idea per card, questions that demand retrieval rather than recognition, nothing you don't already understand. Ten minutes of work turns into a review schedule that lasts months.

What breaks it

Asking before attempting. The rule from the previous article, and it's the difference between a tutor and a crutch.

Reading the summary instead of the source. It saves the time and removes the material — you get the shape of the knowledge with none of the detail that makes it usable, and, worse, the confidence that you've covered it.

Trusting it in a field where you can't evaluate it. This is the sharp one, and it's the exact situation of every beginner: thin terrain plus a confident tone plus a student with no way to detect the difference. It's the dangerous quadrant from article 9, pointed straight at the person least equipped for it.

The fix is structural, not vigilance: take the skeleton from a source that can be checked — a real course, a textbook, a curriculum someone stands behind — and use the model as the tutor working on top of it. Then the model's job is what it's genuinely best at: explaining, questioning, drilling, catching your errors on material whose correctness isn't its responsibility. And when it contradicts your source, that's not a tie to resolve by preference; the source wins until you've verified otherwise.

Studying inside a chat. The conversation isn't a record of progress — it's a desk that gets wiped. What you learned has to end up somewhere you'll see again: notes, cards, your own written summary. Otherwise you'll have an excellent afternoon and nothing in a month.

In practice

Ask for a mode, not an answer. "Quiz me", "mark my explanation", "find my error", "give me problems".

Always attempt first, then compare.

Never accept an explanation you haven't reproduced. Close it, say it in your own words, check the gaps.

Take the structure from a source you can trust, and use the model on top of it.

Convert every session into cards or notes. The chat remembers nothing, and neither will you.

When it disagrees with your textbook, the textbook wins until you can verify.

Check yourself

Close the article and answer in your own words:

1. What was Bloom's 2-sigma result, what's the honest caveat, and what actually produced the effect?
2. Why is "a bad tutor explains, a good one asks" the operative principle here?
3. Name the six modes and what each is for.
4. Why is the error analyst mode more valuable than simply being told the right answer?
5. Why is a beginner in an unfamiliar field the person most at risk, and what's the structural fix?
6. Why isn't a chat a record of what you learned?

In short

• Tutoring works because of the loop — constant checking, immediate correction, difficulty matched to you — not because the tutor knows things. Ask for the loop.
• Bloom (1984) put one-to-one tutoring around two standard deviations above classroom teaching; the figure hasn't been reliably replicated, but the effect is real and large.
• A 2024 Harvard study (Kestin et al.) found students learned more in less time with a properly designed AI tutor than in a well-run active-learning class.
• Six modes: examiner, Feynman partner, explainer on your level, error analyst, problem setter, opponent. Default is lecturer, so you have to ask.
• Turn every session into spaced-repetition cards — one idea each, questions that demand retrieval.
• Failure modes: asking before attempting, summaries instead of sources, trusting it where you can't evaluate it, and studying inside a chat that forgets.
• Beginners are the most exposed: take the skeleton from a checkable source and use the model as the tutor on top.
• A bad tutor explains. A good one asks.