The simple explanation
An LLM — a large language model — is a prediction machine trained on vast amounts of text. Training adjusts billions of internal numbers so the model gets good at one job: given some text, guess what comes next. Ask it a question and the same machinery runs — your prompt plus everything written so far becomes the context, the model scores possible next tokens, picks one, and repeats. The answer emerges from thousands of tiny predictions chained together.
Two things follow from that. First, the model doesn’t “look up” facts the way a database does; it produces the most likely continuation of what you wrote. Second, training and answering are separate: training changes the model’s internal numbers using many examples, while answering just uses those learned numbers — it does not retrain the model or learn from your conversation.
See this idea move.
The How an LLM Works experiment walks through this concept step by step — press run and watch it happen. Everything is simulated in your browser; no real AI runs.
A concrete example
Type “The cat sat on the” and pause. Several endings fit: mat, rug, floor, windowsill. The model picks among plausible continuations, and picking mat says nothing about a real cat — it’s a bet about which ending best matches the patterns the model has seen.
Now swap in a question: “What is the capital of France?” The model still isn’t looking anything up; it’s producing the most likely continuation, which — for a fact repeated millions of times in its training — is usually “Paris.” Same mechanism, different luck. Famous facts come out right because they were everywhere; rare facts are where the luck runs out.
Where people get misled
The first trap: fluency reads as confidence, and confidence reads as truth. A model can write a smooth, detailed, completely wrong answer with perfect grammar. The second trap is treating it like a mind — it has no intentions, no beliefs, no memory of “deciding.” It’s pattern continuation shaped by your prompt. The third: assuming that because it answered well once, it understands. It predicts, and prediction is not understanding.
The honest limits
It has a knowledge cutoff and no live lookup of its own. It can be confidently wrong about anything rare, new, or precise — dates, names, numbers. It struggles with exact counting and multi-step arithmetic. And fluent text is never proof of anything. Use it to explain, summarize, and draft — then check the claims that matter.