How I use LLMs to learn complex topics
TL;DR
A developer learns unfamiliar subjects like chip fabrication in three steps. A coding agent writes a knowledge base and reviews it for accuracy. Then it turns the knowledge base into a low-poly simulation in the style of Rollercoaster Tycoon. He claims the result is "100% accurate and free of hallucinations".
The article
- The author finds LLM explanations hard to follow, too simple and full of emojis. He wanted a game format instead. In a game, each concept maps onto an object on screen and sticks better.
- He works in an agent's plan mode, in Claude Code or OpenCode. The agent first writes a knowledge base on the topic, then reviews it for accuracy. Next it builds a low-poly, Rollercoaster Tycoon-like simulation with controls to pause the flow. The simulation fits small and large screens. He publishes each one on GitHub Pages.
- His main example, ChipTycoon, follows a cart from a pile of sand to a finished chip delivered to a data center. The cart changes shape at each manufacturing step. He calls the result "100% accurate and free of hallucinations". He says it works better for him than search results or a bulleted LLM answer.
- For more realistic visuals he suggests turning photos into 3D objects with a skill he wrote. For retention he suggests quizzes and puzzles about earlier steps.
- He links similar simulations of rocket engines, LLMs, F1 engines and EUV machines. After the HN thread he released a skill file for the method. He also released a course product called Learnscape.
Why it matters
The post takes three minutes to read. Click through to ChipTycoon and the other simulations. Decide for yourself how much a low-poly simulation teaches, since that is the thread's main argument.
The discussion
A big, mostly skeptical thread. It soon dropped the simulations and argued about whether LLMs help people learn at all. Nearly everyone attacked the "100% accurate" claim. Plenty of commenters shared workflows that do work for them. Almost all of those keep a book, a problem set or the learner's own effort in charge.
Use the LLM next to a real source, not as the source
The most-endorsed pattern was to read a book yourself and ask the LLM about the passages you get stuck on. Commenters used LLMs to find books, fed PDFs into projects, worked through Baby Rudin line by line, or treated the LLM like office hours. Several who had built elaborate LLM learning tools returned to "just read the damn book and go to the AI when you've got questions." One critic said that softening Rudin with an LLM means you never learn to absorb hard material.
It's much better to feed the book to the LLM and ask questions as you read along, instead of asking the LLM to basically write a custom book for you from scratch.
In both cases, the LLM isn't providing the curriculum or guiding what I learn. The textbooks, papers and coding tasks are. But now I can pick these things up much more efficiently.
LLM prose is exhausting to learn from
The top thread complained that LLM output is long, monotonous and padded, so the useful part drowns, and it named Opus 5 among the worst. Suggested fixes were "be concise", style transfer from sample text, Simplified Technical English, caveman-style skills, and a system prompt that limits answers to four concepts. Others said these barely help or make the answers worse. Some blamed the verbosity on LLMs not knowing what the listener already knows, others on the LLM hedging against its own hallucinations.
It's just long. It just doesn't shut up. It's overly verbose. And you can't tell it to be concise or you degrade its quality.
It's dense along the wrong axis and verbose Along the wrong axis.
Self-review doesn't make it "100% accurate"
Dozens of comments quoted the accuracy claim, calling self-review "turtles all the way down" and noting that a novice can't spot the errors left. One camp held that a second run makes it unlikely the same hallucination survives, but critics replied that the errors aren't statistically independent. Commenters also found real mistakes, like a conveyor belt running backwards in the rocket engine simulation. The author granted that LLMs may have little training data on secretive chip manufacturing, but held that high-level introductions are fairly safe.
It’s just a wild lack of insight: “I’m using the LLM to teach me something I don’t know about, I definitely have the knowledge base to spot any errors that might remain!”
That rests on a false-assumption that the errors are statistically independent events, and have nothing to do with the shared nature of the judges.
Feeling like you learned vs. learning
The second-largest thread asked for proof, meaning problems people can now solve that they couldn't before, and compared LLM learning to edutainment on YouTube. People answered with Rust, Terraform, CUDA and hacking cheap routers. Skeptics replied that code tells you plainly when you're wrong, while in biology or quantum mechanics you can't check a single source against anything. As for the simulations, many called them a shallow recipe or a bulleted list in costume that hides how complex chip fabrication really is.
In my experience, LLMs are really good for taking up your time and making you feel like you're learning, in the same way that many of the popular educational videos on YouTube are fun to watch and don't really teach you anything.
You've learned something when you have a mental model that makes correct predictions. Until then you've memorized it at best…
Make the learner do the work
Many found LLMs most useful with the roles reversed, the LLM quizzing the learner or asking one-sentence Socratic questions during a walk. Others had it write exercises without solutions, or point students to the right section of the source like a lab partner. One commenter called Socratic skills naive next to learning-science research, which another identified as Graesser's AutoTutor and its pump, hint, prompt and assertion pattern. Many held that struggling is how you know you're learning, so the LLM should clear roadblocks and not hand over answers.
TLDR: Actual human tutoring sessions were recorded and analyzed and Socratic questioning was barely used at all.
If the follow up is just "continue", "go on", "next" or a non-sequitur then it is smell of a stall. If it is challenging or filling a gap in the answer then it is progress.
Consensus
An LLM can speed up learning as a patient tutor next to a trusted source and real practice. It can't be the only source. No self-review step guarantees its output is right, least of all for someone new to the topic.
Dissent
A sizable minority reported gains they could test, like picking up new languages, CUDA or homelab hardware. They said current LLMs rarely hallucinate on undergraduate material. They called the skepticism outdated or a refusal to use a better tool.