What Is an LLM (Large Language Model)?

An LLM is a next-word predictor, trained on a lot of text to guess what comes next. It guesses one word at a time, and the paragraph you read back is that guess, repeated. Step zero happens only once per model version: the model reads a huge amount of text and learns to predict the next word of it, over and over, until the prediction is reliable. That step runs in three stages: pre-training, where it just learns to continue text and absorbs grammar, facts, and style along the way; fine-tuning, where it's shown good question-and-answer pairs so it learns the shape of a helpful assistant instead of a text completer; and reinforcement learning from human feedback, where people rank its answers and it gets nudged toward the ones they prefer.

Once a model is trained, generating the next word goes through five steps, tracked here on one running sentence: "The dog chased the cat because it stole his food." First, the sentence is sliced into tokens. A token is not the same thing as a word: "chased" splits into "chas" and "ed," and each token, including the space in front of it, is swapped for a plain number, an id. That id is the only thing that reaches the model.

Second, those ids are embedded: each one becomes a vector, a list of numbers chosen so that tokens with a similar meaning land close together. "Dog" and "wolf" cluster together, and so do "chased" and "stole," while "food" sits off on its own. Third, attention: every token looks at every earlier token and weighs how much it matters. For the word "it," the model gives "cat" a 62 percent weight, "dog" 15 percent, and "chased" barely 8 percent, so "it" moves next to "cat" on the meaning map, the one that actually stole the food.

Fourth, that attention step plus a small network runs many times in stacked layers. Early layers mostly work out grammar and word shape, middle layers add facts like who did what to whom, and late layers handle tone and intent. Fifth, prediction: the model scores every word it could say next. For this sentence, "food" scores 85 percent, "toy" 8 percent, "bone" 4 percent. A temperature dial controls the pick: low temperature sticks to the top score, high temperature is willing to gamble on a lower one.

Then the loop closes: the picked word gets appended to the sentence, and the whole five-step process runs again from step one on the slightly longer text. Predict, append, repeat. That loop is the entire machine, and it's why a model streams its answer out one piece at a time instead of handing you the whole thing at once. On its own, a plain LLM like this can't reach live information, take actions and check the results, spend extra effort on a hard problem, or answer from a document it was never trained on. Every other topic in this series is one thing bolted onto this same plain model, built to cover one of those gaps.