I am in the middle of a Master’s in IT, run an ICT division, and still spend a good chunk of my weekends trying to learn things that have nothing to do with my day job. So when I tell you that asking an LLM for a bullet-point summary of a complex topic rarely sticks, I am speaking from experience. The explanation is always correct, always thorough, and always gone from my head by Tuesday.
Then I came across an approach that actually worked, and it flips the whole idea of “using AI to explain things” on its head. Instead of asking the model to explain a topic, you ask it to build a simulation of that topic — a small, interactive, RollerCoaster Tycoon-style animation that walks you through the process. You play it, you watch the pieces move, and the knowledge finally stays. It sounded almost too clever to be real, so I tested it myself.
The idea behind learning with simulations
The approach comes from Laurentiu Raducu’s essay on using LLMs to learn complex topics, which was trending on Hacker News this week. His trigger was chip manufacturing. While analyzing the bottlenecks that might slow down data center construction, he realized he did not actually understand how a chip is made — not deeply. So he asked a question most of us would never think to ask: what if there was a game that walked you through building a chip at a fab?
He tried it. It worked. The result is ChipTycoon, an isometric theme park that is secretly a chip factory. A little cart carries one silicon wafer along a road, stopping at each of twenty-two buildings that turn ordinary sand into a finished chip. The cargo on the cart changes at every stop — a heap of sand, then rough silicon, then a silver crystal, then a mirror-flat disc — so you can literally watch the product transform as it moves through the process.
The four-step flow
Here is the exact workflow Raducu uses, and it is the part you can copy for any topic you are struggling with.
Step 1: Build the foundational knowledge in plan mode
Do not ask for a summary. Open your coding agent of choice — he uses Claude Code or OpenCode — and switch it to plan mode. Ask the model to build a foundational knowledge base for your topic: the key stages, the order they happen in, what moves between each stage, and what actually changes at each step. The plan-mode constraint matters here. It forces the model to lay out the structure before it writes a single line of code, and it gives you a knowledge base you can review on its own.
This is the same discipline I wrote about in my guide on writing agent skills for AI coding agents — the quality of what an agent builds depends on how clearly you define the knowledge it works from.
Step 2: Review the accuracy of the knowledge base
Here is the step most people skip, and it is the one that protects you from the biggest weakness of AI-assisted learning. After the model builds its knowledge base, ask it to review its own work for accuracy. Then — and this is the part that matters — spot-check the result against a real source: an official document, a Wikipedia article, a textbook chapter.
I have written before about how AI assistants can hallucinate package names with complete confidence. The same failure mode applies to explanations. One commenter on the Hacker News thread put it bluntly: “How do you know if you are learning this for the first time? Very risky to learn from LLMs. Lots of ‘oh of course you’re right — what I just told you was completely wrong.'” The accuracy review is your guardrail, and it should never be skipped.
Step 3: Build the simulation with real UX requirements
Once the knowledge base checks out, ask the model to turn it into a low-poly, RollerCoaster Tycoon-style animation. But do not just say “make a simulation.” Give it concrete UX requirements, the way you would to a designer: the page needs to work on both large and small screens, it needs controls to stop and restart the flow whenever you want, and each stage needs a short readable explanation.
This is where the technique quietly teaches you something. When you force the model to visualize a process — to decide what the cart carries, what changes at each station, what a building represents — it has to make the process concrete. And you, the reader, get to watch a moving, spatial representation instead of a wall of text.
Step 4: Push it to a repo and turn on GitHub Pages
The final step is delightfully simple. Push the files to a new repository, enable GitHub Pages, and you have a shareable, interactive lesson that runs in any browser. No backend, no build step, no dependencies — ChipTycoon itself is a pure static site. Open the index.html file and it runs.
I tested it myself, so here is the honest part
I did not want to write this tutorial on trust alone, so I built a miniature version of the idea before finishing this post. I created a small six-stop simulation called Sand to Wafer Express — sand pit, furnace, crystal puller, wire saw, polisher, cleanroom gate — as a single static HTML file with a canvas and about a hundred lines of JavaScript. A cart moves along a road, the cargo label changes at each stop, and a one-line note explains what happens there.
Then I served it locally with python -m http.server, fetched the page, and confirmed the animation loads cleanly with the JavaScript syntax checked. It worked. The whole thing is a single file you could host on GitHub Pages, Netlify, or any static host in under five minutes. The point is not that my little demo is impressive — it is not. The point is that the loop is real: knowledge base, accuracy check, simulation, deploy. It takes an evening, and it works.
If you have been living on AI explanations the way I have, this is also a useful reality check on the broader habit. I wrote about the not-so-healthy side of leaning on AI for everything earlier this month, and this simulation approach feels like the healthy version of the same instinct: you are still using the model, but you are forcing yourself to actually engage with the material instead of passively reading a summary.
Making it stick even harder
Raducu suggests two upgrades that take the technique further, and both are worth trying.
Add challenges and puzzles. After the tour plays, drop in a few questions about the previous step. Answering “what happens to the quartz sand after it leaves the furnace?” forces recall, which is the part of learning that bullet points never touch. A couple of intuitive puzzles on the key transitions will do more for retention than re-reading the explanation.
Upgrade the visuals. Low-poly is charming, but it asks a lot of your imagination. If you want a more realistic representation, the author has a skill for turning pictures into 3D objects that you can map into your simulation. The sand pile becomes an actual sand pile instead of a beige blob.
What to learn first
Pick something you have been putting off. Chip manufacturing, how a transformer works, how DNS resolves a name, how a rocket engine is built — any process with clear stages and moving pieces is a perfect fit. The author has done chip fabs, LLMs as a city, rocket engines, F1 engines, and EUV lithography machines. Notice the pattern: they are all processes with a pipeline of transformations, which is exactly what a cart traveling through a park represents so well.
And if you are worried about the model getting the details wrong, you already know the fix: the accuracy review in step two, plus a real source to check against. Treat the simulation as a mnemonic, not a textbook. It is a map that makes the shape of a topic visible — the details still need verifying, the way I covered with AI-generated code and agent workflows built around plan-write-validate agents like Meta’s Muse Code.
Try it this week
You do not need a powerful machine or a fancy setup. An open-source model on your laptop can do this. Spend one evening building a knowledge base, one evening reviewing it, one evening turning it into a simulation, and one evening pushing it to Pages. Four evenings from now you will have a permanent, interactive lesson in something you previously could not wrap your head around — and you will understand it better than any summary could have taught you.
That is the part that surprised me most. The simulation is not a gimmick to make learning fun. It is a way to force the topic into a shape your brain can hold onto. Give it a try on whatever you are stuck on right now, and tell me how the cart moves.