What an LLM actually is — explained with a matatu stage

Draft outline. This article hasn't been written yet — what follows is the outline Isaac will write from. [ISAAC TO PROVIDE: final text]

The analogy

A matatu stage: dozens of routes, conductors shouting destinations, a tout who's memorized which vehicle to point you to based on the first word you say. Nobody at the stage "understands" your final destination in a deep way — but the system routes you correctly almost every time because it has seen millions of similar requests before.

Mapping the analogy to an LLM

• The stage = the model. It doesn't "know" facts the way a person does — it's a very well-trained routing system.
• Millions of past trips = training data. The tout is good at this because they've seen the pattern thousands of times, not because they reasoned it out from scratch.
• "Which matatu goes to Rongai?" = a prompt. Change the wording slightly and you still get routed correctly — the system generalizes.
• Getting on the wrong matatu sometimes = hallucination. Confident, fast, occasionally wrong — because it's pattern-matching, not verifying.

Why this matters practically

Once you see an LLM as a very fast, very well-practiced router instead of a thinking machine, you know exactly when to trust it (well-worn routes) and when to double check (unusual destinations, exact facts, numbers).

Close

A plain-language definition to leave the reader with, plus a pointer to the next "Learn AI" piece on prompting.