how-toreal-worldai-trends

LLMs as learning tools. A developer shares his technique for mastering difficult subjects.

A practical guide on using large language models to break down and understand complex topics, with strategies for deeper learning beyond surface-level answers.

August 10, 2026 · Updated November 8, 2026

LLMs as learning tools. A developer shares his technique for mastering difficult subjects.

The prompt looked reasonable on paper: "Explain how transformers work." The model returned 800 words of technically accurate prose that, somehow, left the reader knowing almost nothing they did not already know. The vocabulary was correct. The concepts were present. But nothing had actually been explained, because the question had been too broad, the context too thin, and the follow-up loop was never started. That is the way many people use LLMs to learn, and it is also why so many walk away from the experience thinking the tools are better for writing emails than for acquiring knowledge.

Why the "explain it to me" prompt consistently fails

Think of an LLM like a very well-read tutor who has read every textbook but has no idea what you already know. If you walk up to that tutor and say "explain distributed systems," they will start from what feels like the middle, because they have no map of your existing knowledge. What they need first is a constraint: where do you start, what do you already understand, and what specific part is causing friction right now.

The method described in Laurentiu Raducu's post on how he uses LLMs to learn sidesteps the problem by never asking for a plain explanation. He has a coding agent in plan mode build a foundational knowledge base on the topic, asks the model to review that knowledge base for accuracy, and then has it build an interactive, low-poly simulation of the topic in the style of RollerCoaster Tycoon, published on GitHub Pages. His examples include how chips are made and how LLMs work. A lighter version of the same idea works in any chat window: do not start with "explain this," start with "interrogate me first." Before any explanation starts, the model asks what you already know, finds the gap, and then targets that gap specifically. This flips the direction of information flow at the start of the session.

The analogy that clarifies this: imagine a GPS that, before giving directions, confirms your starting location. An LLM that starts explaining without knowing your baseline is giving directions from a location it assumed, not the one you are at. The explanation might be perfect for someone else's starting point and useless for yours.

A second structural problem is the single-session assumption. Complex topics have layers, and one conversation thread is usually only one layer deep. A better use pattern treats each session as a single concept unit, not a complete subject coverage attempt. You learn what a transformer's attention mechanism does in one session. You do not try to go from zero to deploying a fine-tuned model in one afternoon.

The actual costs of using LLMs for learning

There is a version of this conversation that treats LLM-assisted learning as essentially free. That is only true in the narrowest sense. The token costs are low. The time costs are not.

The time cost comes from session design. A well-structured learning session with an LLM takes preparation: you need to know what you are trying to learn this session, what you already know, and what format of explanation works best for you. For a working professional with 45 minutes of focused time, that setup cost is not trivial. If you skip it, you get the "explain transformers" problem described above.

Cost type High-friction approach Structured approach
Token spend High (long vague threads) Lower (targeted sessions)
Time per concept Longer, with little retention Shorter, with active recall
Follow-up friction High (lost context across sessions) Low (session-scoped questions)
Model required Any Any capable of multi-turn dialogue

On model choice: the approach works across most current models. Claude and ChatGPT both handle the multi-turn diagnostic conversation well. If you use the API, Claude Haiku 5.5 at $0.10/$0.50 per million tokens costs very little for sustained learning sessions. Claude Sonnet 5.5 at $2.00/$10.00 per million tokens handles longer conceptual threads more reliably when the topic has real depth. The model is not the variable that matters most here. The session design is.

There is also a buy-in cost that rarely gets mentioned: you have to actually do something with what you learned. An LLM can generate a project idea, a worked example, a quiz, a set of review questions. But if your workflow ends at "I read the explanation and it made sense," retention is low. The tools exist to support active recall. Building them into the session is a choice, not a default.

What the Hacker News discussion revealed about where this goes wrong

The post drew more than 500 comments on Hacker News, and the skeptical ones are the most useful. One commenter argued that overviews like these are closer to pop science than real learning and proposed a teach-back test: if you cannot explain the idea yourself afterwards, you did not learn it. Another warned that models have no real concept of truth, so an explanation can sound right while the reasoning behind it is wrong. Both point at the same gap: a session can feel productive without leaving anything durable behind. The missing component is usually a retention mechanism and a continuity structure between sessions.

A simple fix: use the model to generate a summary at the end of each session that is scoped to "what did I just learn, in my own terms, with the gaps I still have." That summary becomes the opening context for the next session. Claude and ChatGPT both have memory features now, but they decide what to keep; a summary you save yourself is a record of your progress that you control. Raducu's own suggestion for retention points the same way: add questions about earlier steps to the simulation.

Gemini Notebook (formerly NotebookLM) handles part of this differently, by letting you ground a conversation in documents you upload. For learning from a specific textbook, paper, or specification, that grounding changes what the model can actually help with. For open-ended conceptual learning, it is less relevant. The right tool depends on whether you are learning from a fixed corpus or building understanding from scratch.

The discussion also split along a familiar line: some commenters use the model as an interactive tutor, asking follow-up questions until the idea clicks, while others describe it as a faster search engine for niche material. Both are valid uses, but they need different prompts. The tutor pattern is the one that builds understanding, and it is the one the rest of this guide is about.

Three approaches to LLM-assisted learning compared

Approach Session structure Retention mechanism Suitable for Main failure mode
Passive reading (ask, read, close) Single turn None built in Quick lookups No retention between sessions
Socratic dialogue (model interrogates first) Multi-turn, scoped to one concept End-of-session summary prompt Complex conceptual topics Requires discipline to scope sessions
Document-grounded Q&A Multi-turn, anchored to uploaded source Built into notebook structure Learning from specific texts Breaks down for topics with no good source material

One-sentence verdicts: if you are learning a concept from scratch with no source material, use the Socratic dialogue approach. If you are working through a specific paper, textbook, or spec, document-grounded Q&A is faster. If you are doing a quick lookup and do not need to retain it, passive reading is fine and everything else is overhead.

For a longer comparison of tools suited to research and learning tasks, see Claude vs. Perplexity and Gemini Notebook vs. Perplexity.

A decision framework for structuring your next learning session

Before you open a chat window, answer three questions. What is the single concept you want to understand by the end of this session? What do you already know that connects to it? What would a correct explanation allow you to do that you cannot do right now?

If you can answer all three, start the session with: "I want to understand [concept]. Here is what I already know: [your baseline]. Before you explain anything, ask me questions to find where my understanding actually breaks down." This gives the model what it needs to be useful.

If you can answer the first question but not the other two, spend five minutes writing out what you do know before starting. The act of writing it surfaces the actual gaps faster than any prompt will.

If you cannot answer the first question at all, you do not have a learning goal yet. You have a subject area. Break it down: what is one specific thing within that subject that you would be able to explain to someone else after this session? That is your scope.

At the end of each session, run one more prompt: "Summarize what I just learned in terms I could use to teach it, and list the two or three things I should look at next." Save that output somewhere. Use it to open the next session. This is the continuity mechanism many learners skip, and it is the single change most likely to close the gap between sessions that feel productive and knowledge that actually compounds.

You can apply this across any model with decent multi-turn capability. Perplexity works well for learning sessions where you want source citations alongside explanations. Claude handles nuanced conceptual scaffolding well. The model matters less than the structure. See also our post on separating LLM hype from reality for context on what these tools can and cannot do reliably.

Back to the transformer explanation that taught nothing

The 800-word response that left the reader knowing almost nothing was not the model's fault. It was a structural problem. The prompt gave no baseline, asked for no diagnostic step, and set no scope. What came back was the model's best guess at an average explanation for an average reader who does not exist.

Run the same topic through the framework above: "I want to understand how attention works in a transformer. I understand that words in a sentence have relationships to each other, and I understand basic matrix multiplication. Before explaining anything, ask me questions to find where my understanding breaks down." That prompt does not guarantee a perfect session. But it gives the model enough to actually help, and it forces the reader to know what they are walking in with. The difference between a learning session that compounds and one that evaporates is almost always that specificity, applied before the first explanation ever arrives.

Some links in this article are affiliate links. Learn more.