Video narration
When an LLM answers you, does it write the whole response at once—or find it somewhere online? Think about phone autocomplete: it uses text already typed to predict what comes next. An LLM does this on a much larger scale. During training, it sees many examples and adjusts huge numbers of internal settings, learning patterns in how language fits together. Internally, text becomes numbers. Your question is broken into tokens—pieces that can be whole words, parts of words, or punctuation. From the context, the LLM calculates probabilities for next tokens, chooses one, adds it back, and predicts again. Repeating that loop builds the answer piece by piece. But fluent prediction is not certainty: it can sound right while being wrong, because the model generates a likely sequence rather than checking truth. Neither: an LLM builds responses by repeated next-token prediction from context.