Big systems.
Small steps to understanding.
An interactive guide that connects a single chat message to the GPUs, networks, and facilities needed to serve millions of people.
The pieces are hard to connect.
Learning about LLM systems often means jumping between model architecture, GPU specifications, networking terms, and serving software. A beginner needs a connected explanation of why each component exists.
Teach it like the alphabet.
Each lesson introduces one idea with three illustrated steps, a plain-language explanation, the real technical term, and a short question. The A–Z journey moves from a request to a whole service, then explores prompting, retrieval, fine-tuning, evaluation, safety, tools, multimodal inputs, and efficiency.
Simple on the surface. Explicit underneath.
- Progressive explanation. Familiar pictures come before technical vocabulary. Learners can revisit any step and get immediate quiz feedback.
- Inspectable assumptions. The capacity calculator shows how traffic becomes replica counts, server cabinets, and power. Benchmark throughput is explicitly hypothetical until replaced with a measurement.
- Two different networks. Separate lessons distinguish rack-scale NVLink connections from cluster leaf–spine networking. A topology calculator makes port-count assumptions visible.
- A small technical footprint. Plain HTML, CSS, and JavaScript run without a backend or build step. Arithmetic lives in a separate module with six automated tests.
- Accessible interaction. Semantic controls, keyboard focus states, reduced-motion support, pause controls, and responsive layouts support different ways of learning.
Try one meaningful change.
Open lesson O, “Build for 100 million.” Keep the default demand and change hypothetical replica throughput from 2,000 to 500 output tokens per second. Required GPUs rise from 99,208 to 396,832.
The lesson: a user count alone cannot size a fleet. The model, workload, latency target, and measured serving capacity determine the answer.
Open the lessons →What has been checked
Six automated tests cover demand arithmetic, whole-replica rounding, the slower-throughput scenario, input validation, and network pod sizing. Browser checks covered lesson navigation, quiz feedback, the live capacity scenario, and a phone-sized network diagram.
What this project demonstrates
This is an educational interface and capacity-modeling tool. “100 million users” is a scenario, not a claim of deployed traffic. It does not train or host an LLM, reproduce OpenAI’s private infrastructure, or provide a deployment-ready equipment list. Hardware references and calculation boundaries are available within the learning experience.
Start with one little question →