An engineer published his method for learning complex topics with LLMs. Instead of asking a model to explain something to him, he has it build an interactive visual simulation. Hacker News voted it 560 points.

Laurentiu Raducu analyzes data center bottlenecks and wanted to understand chip manufacturing. He opened Claude Code in plan mode and gave it a three-step flow: build a knowledge base on the topic, review it for accuracy, then transform it into a low-poly animation in Rollercoaster Tycoon style. The result is ChipTycoon, a playable web page where you follow a cart from sand to finished chip.

Why this matters to you. Most people use LLMs for learning like they’re a talking book: ask, read, follow-up question. Raducu’s method shifts perspective. Have AI build a concrete representation of what you’re studying, and the knowledge sticks because you saw it in motion, not just read it.

What matters beyond the simulation: deciding what to have AI represent forces you to break the topic into concrete steps. The quality of what you get depends on how much you already know what you want to see. This is the same principle we discussed on August 4th: expertise matters more than prompting tricks.

If you want to try it: Raducu’s blog post describes the flow in detail and links the other simulations he built (rocket engines, EUV machines, F1 engines).

In detail

Raducu starts from a recognizable problem: the way LLMs explain things is hard to follow. Too simplified, too terse, sometimes loaded with emoji. For complex technical topics, standard explanations don’t stick.

His method uses a coding tool (Claude Code or OpenCode) in plan mode, a function where the agent plans before executing. The flow has three phases. First, the agent builds a knowledge base on the topic: gathers foundational concepts and organizes them. Second, the agent reviews its own base to verify accuracy. Third, the agent transforms that base into a low-poly simulation, a web page with simple graphics in Rollercoaster Tycoon style, with controls to pause and adapt to small and large screens. Everything ends up in a GitHub repository with Pages enabled, so accessible from a browser.

The concrete use case is ChipTycoon. You follow a visual cart that starts from collecting quartz sand, passes through the furnace, and arrives at the finished chip delivered to a data center. The low-poly graphics lose detail, but the conceptual flow stays visible. Raducu has built other simulations with the same method: rocket engines, EUV machines, F1 engines, and how LLMs themselves work.

The limits. The accuracy review step is the weak point. The agent verifies itself, and LLMs can be too confident in wrong answers. If the knowledge base contains an error, the simulation reproduces it faithfully. The low-poly design requires imagination to understand what happens to materials in intermediate steps, as Raducu himself admits. And the method requires a coding agent with plan mode, not a simple chat.

To improve memorization, Raducu suggests adding challenges and quizzes to the simulation. Answering questions about a previous step in the process helps retain knowledge. He also suggests transforming images into 3D objects and mapping them in the simulation to get a more realistic representation.

The signal for readers here: the value isn’t in the specific tool, it’s in the method. Breaking down a topic, asking AI to represent it in concrete and interactive form, then using it as a visual reference is a pattern that works even without reaching 3D simulations. Even a simple HTML page with an animated diagram generated by an agent serves the same purpose. The playbook learning a new tool in an afternoon follows the same logic: use AI as an active tutor, not a passive encyclopedia.