Explain it at three levels

Difficulty is a dial you can set

One of the new things these tools offer is control over the level of an explanation. A textbook is written at one level for everybody. A model will write the same content at any level you name, as many times as you like, and switching between levels is where a lot of the understanding actually happens.

The technique: ask for the same idea three times, deliberately.

1. At the level below yours. Simple language, no jargon, an everyday analogy. This is where you find out whether you can hold the shape of the idea at all.
2. At your level. Correct terminology, the standard treatment, the notation your course uses.
3. At the level above. What the simplification hid, where the standard treatment breaks, what a specialist would object to in the version at level 2.

The third one is the valuable one and almost nobody asks for it. It tells you which parts of what you learned are true and which are teaching approximations, and that distinction is frequently exactly what separates a good answer from an excellent one in an exam.

An example

For why does ice float:

• Level below: water molecules line up in a lattice with gaps when they freeze, so the same amount of water takes more space.
• Your level: hydrogen bonding forces a tetrahedral arrangement in ice Ih, lowering density relative to liquid water at 0 degrees; density maximum at about 4 degrees.
• Level above: this is specific to ice Ih at ordinary pressure; there are more than a dozen other ice phases, several denser than liquid water, and the density anomaly is a consequence of the directional bonding, not a general property of solids.

The third level is where you learn that solids are denser than liquids, except water is itself a simplification.

Analogies, and their expiry date

Ask for an analogy and you will get one. They are useful and they all break. The important habit is to ask where.

Give me an analogy for how a neural network learns. Then tell me exactly where the analogy stops being true and what someone would get wrong if they took it too far.

Every analogy in science teaching has a failure point that a marker will probe. Electrons orbiting like planets: fails at quantisation and at what an orbital actually is. The cell as a factory: fails on scale, on the absence of a manager, and on the fact that everything is bumping around at random. Getting the failure point in the same breath as the analogy stops it becoming a misconception you then have to unlearn.

The Feynman move, and its honest version

The well-known technique is: explain the topic in plain language as if to somebody who does not know it, notice where you stumble, go back to the source for those bits, repeat.

The AI version people usually describe is ask the model to explain it simply. That is the wrong half. The valuable half is you doing the explaining. The correct use is:

I am going to explain X to you as if you are a bright 15-year-old who knows no biology. Do not help me. Afterwards, tell me: which parts were vague, which were actually wrong, and which I skipped over because I do not understand them.

Now the model is a listener and a critic, and you have done the production. This is one of the best single uses of a language model in study, and it inverts what most students do with it.

Levels for reading, too

The same dial works on material you are struggling with. Paste the difficult paragraph and ask for it at the level below, then read the original again. The point is not to replace the original: it is to get enough of a foothold that the original becomes readable. Then read the original.

This matters for dense primary material: a paper’s methods section, a legal judgment, a philosophical text. The simplified version is a ladder, not a destination, and the thing you will be examined on is the original.

The limit

Level control is only as good as the model’s grasp of the subject, and at the level above it is generating what a specialist objection sounds like. On mainstream material that is reliable enough to be useful. On the frontier of a narrow field it is not, and a research supervisor will spot the difference immediately. Use level three to know what questions to ask, then ask a human.