Learn how LLMs work by playing a game

There is one clock and it is always running. A right answer hands two to four seconds back, a wrong one costs five, and thinking costs whatever it costs. So the pressure comes from the topic you do not know rather than from a bar sliding past — which is a more honest way of finding the gaps than a score at the end of a quiz.

Free, and no account needed — play as a guest. Runs in the browser — nothing to install.

The fifteen topics

A run shuffles the syllabus and walks it, one question per topic, coming back round if the clock holds out:

• Generative AI — what it is and what it is not
• Large language models — what an LLM actually is
• How they work — prediction, not retrieval
• Tokens — the unit everything is counted and billed in
• Embeddings — meaning as coordinates
• Vector databases — searching those coordinates
• RAG — retrieval augmented generation
• Prompt engineering — what moves a model and what does not
• Function calling — letting a model reach outside itself
• Structured outputs — getting JSON you can rely on
• Streaming — why answers arrive a word at a time
• Context windows — the limit that explains most surprises
• Fine-tuning versus RAG — the choice people get wrong most often
• Comparing models — reading a benchmark without being fooled by it
• What it costs — how the bill is actually assembled

Why the clock is the whole design

A timed quiz usually punishes you for being slow. This one pays you for being right: time is the score and the resource, so every correct answer literally extends the run. The consequence is that a topic you know well is free and a topic you half-know is expensive — and you feel which is which within about thirty seconds, without anyone having to grade you.

Where the questions come from

A hand-written starter bank, plus questions written by one model and then judged blind by a second — the judge never sees the answer key and has to pick the right option itself. Only what survives both is banked. Every question records which road it came in by and what the judge scored it, so a bad one can be found and pulled rather than merely counted.

Questions and answers

Do I need to be technical?

No. The syllabus starts at what generative AI is. It is aimed at somebody who uses AI tools and wants to understand what is happening underneath.

Is this about one particular model?

No — the topics are the ideas common to large language models generally, not the features of any one product.

Is it free?

Yes. Every game on SULULU is free and there is nothing to buy.