Frontier language models are still unreliable in important ways, but they have become good enough to act as unusually effective learning partners. They don’t hallucinate as frequently, and often come up with creative or original insights that allow you to use them as creative sparring partners.
Likewise, I now feel comfortable learning new material entirely through Claude Opus 4.8 or GPT-5.5 as my teacher. After spending months (or years) experimenting with these models to actively learn new subjects, I have converged on a learning method that works well for me. It allows me to absorb information more quickly and at greater depth than I have ever been able to before.
The Problem #
Most of the sequential learning methods I have used suffer from a few recurring issues:
It is difficult to contextualize information when I encounter abstract definitions before understanding how that information
fits intomy mental model. What problem is this concept solving? Why does it exist?It is difficult to learn concepts
activelyrather thanpassively. The definitions presented are often the byproduct of someone else’s cognitive process, and it becomes easy to accept them without wrestling with the ideas personally. The friction of learning shifts toward memorization and terminology rather than problem solving. Yet problem solving is the skill that carries more weight today.
Socratic Spiral Learning #
Through trial and error, I developed a variation of Jerome Bruner’s Spiral Curriculum that leverages the “instant oracle” nature of LLMs to teach almost any subject quickly and effectively.
The core idea behind Socratic Spiral Learning is to revisit the same material multiple times, descending one layer of abstraction on each pass. By first learning a high-level overview of a topic and then revisiting it with increasing depth, the learner gains a stronger understanding of context while gradually filling in gaps in understanding. Intuition is slowly reinforced by rigor and problem solving, creating a more durable foundation for the material.
| Layer | Focus / Content | Jargon & Rigor Level | Primary Use |
|---|---|---|---|
| Layer 1 | Qualitative, plain-English overview | No jargon, no formal definitions, no proofs | Building initial intuition and a mental scaffold |
| Layer 2 | Definitions, formal structure, examples | Gentle mathematical definitions, LaTeX | Learning applications, solving problems, expanding intuition |
| Layer 3 | Proofs, rigor, challenging problems | Technical vocabulary, proof steps, LaTeX | Building rigor and correcting gaps in intuition |
In practice, I generally begin by asking the LLM for a high-level overview of the material, often by pasting in a textbook chapter.
After grasping the high-level picture, I ask the LLM to generate clarifying and guiding questions that force me to think through the implementation details of the next layer myself. This directly addresses problem #2, because I am now required to actively engage my problem-solving abilities in order to construct the rigor of the next layer. This feature is the core of Socratic Spiral Learning.
None of the underlying pedagogy is new: spiral curricula, Socratic questioning, active recall, and tutoring are all old ideas. What feels new is that LLMs make this combination available on demand for almost any technical subject. Of course, like any learning method, this approach must be still be supplemented with practice problems.
Example #
I recently used this approach to learn calculus of variations for Lagrangian mechanics. The following is part of my conversation with Claude Opus 4.8.
You are my learning partner for calculus of variations. We learn this in layers, not sequentially. Read these rules and follow them for the whole session.
My background: comfortable through multivariable calculus and linear algebra; new to real analysis; not a mathematician, so guide me carefully through rigor. Default to intuition first, then rigor.
The method is a spiral. We cross the entire topic at each layer before going deeper:
Layer 1 (qualitative): plain English, minimal jargon. For every concept, tell me what it is trying to do, why it matters, and one concrete real-world example. No formal definitions or proofs yet. This is the scaffold everything else hangs on.Higher layers: definitions -> formal structure -> full rigor with LaTeX, added one layer at a time, only when I say to go deeper.
Rules for the whole session:
- No walls of text. Use headings to organize and show how concepts connect.
- Anchor everything in a concrete problem: “this concept exists to solve X.”
- Active learning, not passive. Ask me a comprehension question regularly, and wait for my answer before continuing.
- Once past Layer 1, use LaTeX freely and guide me through every step of the rigor.
- Do not flatter me or accept vague answers. If I am hand-wavy or wrong, push back and make me sharpen it.
- Never invent facts. If something is uncertain or contested, say so.
Right now do only this, then stop and wait for me:
- Build a concept map of the whole topic: every concept we will cover and how they connect. - Deliver Layer 1 for the whole map: the plain-English, example-driven overview of all of it.
Do not teach any formal definitions or rigor yet. End by asking whether the map and Layer 1 landed before we go deeper.
Here’s the whole territory laid out as a dependency map. Read it top to bottom: two motivating problems force a new idea (the functional), a trick makes it solvable, that trick produces one master equation, and everything below the equation is what it buys you.
Now the plain-English tour of that map. No definitions, no proofs yet, just what each piece is for. I’ll assume you’re fluent in ordinary single- and multivariable calculus, since that’s the real prerequisite here.
The one idea behind all of it
Ordinary calculus answers: which number minimizes this function? You vary a single variable x
and look for where f(x)
bottoms out. Calculus of variations asks a bigger question: which entire function minimizes this total quantity? The unknown is no longer a point on a line. It is a whole curve, shape, or trajectory. Everything in the chapter is machinery for doing calculus when you optimize over all possible paths instead of all possible numbers.
Where the problem comes from
Two concrete problems force the idea into existence:
Shortest path between two points. You know the answer is a straight line, but you have probably neverprovenit. To do that, you must compare the lengths of all curves connecting the points and show one wins.Fermat’s principle. Light traveling through a medium takes the path of leasttime, not least distance. With a varying refractive index, that path bends in non-obvious ways. It has the same structure as the first problem, just with a different quantity to minimize.
Both are “out of all possible paths, find the special one,” and that pattern is everywhere.
The conceptual object: a functional
A functional is a machine that swallows an entire path and returns a single number: total length or total travel time. Picture feeding it different squiggly curves between two fixed points. Each curve comes back tagged with one number, and you want the curve carrying the smallest tag. Ordinary functions map number -> number; a functional maps function -> number.
The trick that makes it solvable
You cannot naively “differentiate with respect to a function.” The clever move is to dodge it. Take the curve you suspect is optimal and add a small bump to it, scaled by a knob α
. Now the functional’s output depends only on that one ordinary number. The true path sits at α = 0
, and if it is optimal, the output must be momentarily flat there. You have smuggled the problem back into ordinary calculus. That “flat at the true path” condition is called stationarity.
The payoff: one equation
When you push that condition through the algebra, the arbitrary-bump machinery disappears and leaves a differential equation that the optimal path must obey: the Euler-Lagrange equation. The recipe becomes mechanical: write your quantity as an integral, plug it into this equation, solve the resulting ODE, and out comes the optimal path.
What it buys you
Shortest path -> straight line. The method producesy = mx + b
.The brachistochrone. What shape should a frictionless slide have so a bead reaches the bottom fastest? Not a straight ramp or circular arc, but acycloid, the curve traced by a point on a rolling wheel.
The honest caveat
The Euler-Lagrange equation only guarantees the path is stationary, not that it is a minimum. It is analogous to df/dx = 0
identifying maxima, minima, and saddle points indiscriminately. Between two cities on the equator, both great-circle arcs satisfy the equation. The math finds flat spots; deciding which kind they are is a separate problem.
Scaling up, and why physicists obsess over this
Real systems need more than one coordinate. A double pendulum is described by two angles, for example. The fix is painless: one Euler-Lagrange equation per coordinate, coupled together. Replace “length” with a physical quantity called the action, demand that it be stationary, and the Euler-Lagrange equations become the equations of motion. They are equivalent to Newton’s F = ma
, but dramatically easier to use in awkward coordinates.
We’re ready to drop to the next layer on this concept. Don’t just present it. Run this loop:
- DISCOVERY FIRST. Before telling me anything new, pose 2-4 revealing questions that push me to DERIVE the next layer myself — e.g. “given the Layer 1 picture, how would you actually define / build / formalize this? What would it need to work? Where does the simple story break?” Pitch them just beyond what I currently know: hard enough to make me think, not so open-ended that I flail. Then WAIT for my attempt. Do not answer your own questions.
- REACT TO MY ATTEMPT. Tell me what I got right, where I went wrong, and WHY. Don’t just hand me the answer.
- FORMALIZE. Introduce the next layer properly: definitions, structure, and rigor, with LaTeX, built step by step.
- CORRECT THE INTUITION. Explicitly flag every place where the Layer 1 intuition was misleading or outright wrong, and explain why the formal version differs. (This matters most in areas where the naive picture breaks down — e.g. real analysis or group theory.)
- CHECK, DON’T ASSUME. End with one RETRIEVAL question that I must answer from memory, not recognition, so we can catch any illusion that I understood it simply because it read smoothly.
Good. Layer 2 is the actual derivation: turning “wiggle the path and demand flatness” into the Euler-Lagrange equation, line by line. Before I show you any of it, I want you to reconstruct it. Here are four questions, each picking up where the last leaves off. Don’t look anything up. Derive.
1. Do the differentiation. The setup gives us
and stationarity means . Carry that out: move inside the integral and apply the chain rule to . What two-term expression do you get, and what does it simplify to once you set ?
2. Clean up the awkward term. Your answer to (1) will contain multiplying . That’s a problem: is an arbitrary bump, and having both and its derivative loose in the integral blocks you from extracting a clean condition on . What standard tool converts the into an ? When you apply it, a boundary term pops out — write it down, and tell me precisely what must hold at and for it to vanish. Then the real question: why are we entitled to assume that endpoint condition in the first place?
3. Make the final leap honest. Suppose you’ve now got everything in the form
holding for every admissible bump . You want to conclude on the whole interval. But notice: for a single fixed , that integral being zero tells you essentially nothing about . So what is the phrase “for every ” actually buying you? Argue it by contradiction — if an adversary insists at some point, what specific would you construct to break their claim?
4. (Stretch) Pin down what “stationary” really certifies. In Layer 1 I called the optimal path “flat, like a minimum.” But only tests the variation along one direction at a time — we then demand it for all of them. Compared to an honest minimum in ordinary multivariable calculus, is this condition equivalent, weaker, or stronger? Concretely: what does it fail to rule out, and how does that connect to the globe / two-great-circles caveat from Layer 1?
Take a real crack at all four — rough is fine, I want to see your actual reasoning, including where you get stuck. I won’t answer them for you; once you’ve attempted, I’ll tell you exactly where you’re right, where you’re off, and why, and then we formalize.
Failure Modes #
The biggest failure mode of Socratic Spiral Learning is the illusion of learning. Reading LLM-generated explanations passively can create a strong feeling of understanding without genuine mastery. In addition, shallow first passes can introduce misconceptions that later passes must actively correct.
Some concepts even have intuitive explanations that are partially or entirely wrong. For that reason, it is important to prompt the model to explicitly identify where an intuition breaks down and where the formal treatment differs from the simplified story.
Useful Prompts #
These are the primary prompts I use in my own learning. Feel free to adapt them to your own needs.
Layer 1
You are my learning partner for [TOPIC / CHAPTER]. We learn this in LAYERS, not sequentially. Read these rules and follow them for the whole session.
My background: [BACKGROUND — e.g. comfortable through multivariable calculus and linear algebra; new to real analysis and group theory; not a mathematician, so hold my hand through rigor]. Default to intuition first, then rigor.
The method is a spiral. We cross the ENTIRE topic at each layer before going deeper:
-
Layer 1 (qualitative): plain English, minimal jargon. For every concept, tell me what it’s trying to DO, why it matters, and one concrete real-world example. No formal definitions, no proofs yet. This is the scaffold everything else hangs on.
-
Higher layers: definitions -> formal structure -> full rigor with LaTeX, added one layer at a time, only when I say to go deeper. Rules for the whole session:
-
No walls of text. Use headings to organize and show how concepts connect.
-
Anchor everything in a concrete problem: “this concept exists to solve X.”
-
Active learning, not passive. Ask me a comprehension question (conceptual or a small problem) regularly, and WAIT for my answer before continuing.
-
Once past Layer 1, use LaTeX freely and guide me through every step of the rigor.
-
Don’t flatter me or accept vague answers. If I’m hand-wavy or wrong, push back and make me sharpen it. I want honest correction, not validation.
-
Never invent facts. If something is uncertain or contested, say so. Right now do ONLY this, then STOP and wait for me: (a) Build a CONCEPT MAP of the whole topic — every concept we’ll cover and how they connect (a dependency sketch is ideal). (b) Deliver LAYER 1 for the whole map: the plain-English, example-driven overview of all of it.
Do NOT teach any formal definitions or rigor yet. End by asking whether the map and Layer 1 landed before we go deeper.
Next Layer
We’re ready to drop to the next layer on [CONCEPT, or “the whole map”]. Don’t just present it. Run this loop:
DISCOVERY FIRST. Before telling me anything new, pose 2-4 revealing questions that push me to DERIVE the next layer myself — e.g. “given the Layer 1 picture, how would you actually define / build / formalize this? What would it need to work? Where does the simple story break?” Pitch them just beyond what I currently know: hard enough to make me think, not so open-ended that I flail. Then WAIT for my attempt. Do not answer your own questions.
REACT TO MY ATTEMPT. Tell me what I got right, where I went wrong, and WHY. Don’t just hand me the answer.
FORMALIZE. Introduce the next layer properly: definitions, structure, and rigor, with LaTeX, built step by step.
CORRECT THE INTUITION. Explicitly flag every place where the Layer 1 intuition was misleading or outright wrong, and explain why the formal version differs. (This matters most in areas where the naive picture breaks down — e.g. real analysis or group theory.)
CHECK, DON’T ASSUME. End with one RETRIEVAL question that I must answer from memory, not recognition, so we can catch any illusion that I understood it simply because it read smoothly.