The One-Sentence Definition, Unpacked

Lesson 1 of 4 in What Is a Large Language Model?.

Here is the whole field in one sentence:

A large language model is a neural network trained on enormous amounts of text to do one thing — predict the next Token — and everything else it appears to do emerges from that.

That sentence sounds too small for the technology it describes. It is not. It is the honest, complete definition, and this lesson takes it apart clause by clause. Once each clause is solid, most of the mystery around LLMs turns into engineering — which is exactly the state of mind the rest of this academy assumes.

“A neural network…”

Not a program full of if-then rules, and not a stored collection of facts. A neural network is one enormous mathematical function with millions to billions of adjustable numbers inside it — the parameters, also called Weights. Feed numbers in, numbers come out, and the parameters shape the mapping. Change the parameters and you change what the function does. In modern LLMs, that function is almost always a Transformer — an architecture the whole second domain of this academy opens up piece by piece.

“…trained on enormous amounts of text…”

Nobody types knowledge into an LLM. Training shows the network vast amounts of real text — web pages, books, code — one stretch at a time, asks it to predict what comes next, and nudges the parameters slightly whenever the prediction is poor. Repeat this astronomically many times and the parameters end up encoding the patterns of the training text. The word large in the name mostly points at two dials: the parameter count and the amount of text seen during training.

“…to predict the next token”

A Token is a small piece of text — a common word, or a fragment of a rarer one — drawn from a fixed list called the Vocabulary. Given everything so far, the model produces a score for every token in that list: a full probability distribution over what could come next. That is Next-token prediction, the single trained objective. There is no second, secret objective — no “answer the question” module, no “be truthful” module. Those behaviors, where they exist, ride on top of prediction.

“…everything else emerges from that”

Nobody programmed translation, coding, or summarization into the model. Predicting the next token extremely well across the breadth of human text turns out to require absorbing grammar, facts, idioms, code conventions, and styles of argument — so those capabilities show up as side effects of the one objective. This Emergence is the strangest and most consequential fact about LLMs, and the Capabilities and Hard Limits module examines it — including where it runs out.

One prediction gives you one token — so how does a model write a paragraph? By running the loop below over and over. The model predicts a token, the token is appended to the text, and the grown text goes straight back in as input for the next prediction. Every answer you have ever seen an LLM produce was built this way: one token at a time, each chosen in light of all the ones before it.

The generation loop — every LLM product runs this

  1. Your text

    The prompt: everything the model will condition its prediction on.

  2. Tokenizer splits it into tokens

    A Tokenizer chops the text into vocabulary pieces and hands the model their IDs. Try it yourself in the Tokenizer Playground.

  3. The model reads all tokens so far

    One pass through the neural network, conditioning on the entire sequence to date.

  4. A probability for every vocabulary token

    The output is a full distribution — tens of thousands of candidate next tokens, each with a probability.

  5. Sampling picks one token

    A Sampling rule chooses a single token from the distribution — sometimes the most probable one, sometimes a plausible alternative.

  6. Token appended to the text

    The chosen token joins the sequence and becomes part of the input for the next round.

  7. Stop token or length limit?

    Generation ends when the model emits a special stop token or hits a configured maximum.

  8. The full response

    What looks like one answer is the trace of many loop iterations.

From probabilities to a token

The model’s raw outputs are not probabilities but unnormalized scores called Logits — one real number per vocabulary entry. The Softmax function turns that score vector into a proper probability distribution: exponentiate every logit, then divide each by the sum, so higher scores become larger shares of a total that sums to one.

Then something has to choose. Always taking the single most probable token (greedy decoding) is one option, but production systems usually sample from the distribution — often after reshaping it with parameters like temperature, which flattens or sharpens the probabilities before the draw. That is why the same prompt can produce different answers on different runs: the model’s distribution was identical; the dice landed differently.

Two things worth internalizing now, long before the details arrive in the Inference & Serving domain. First, the model commits to one token at a time — there is no plan for the sentence written down anywhere, only a distribution for the very next piece. Second, the probabilities are the model’s entire output; everything a product shows you downstream is built from a sequence of draws from these distributions. The full journey of one token through the network — and what happens to the logits at the end of it — is traced step by step in One Token’s Journey.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.