learn me a gpt
Pages in Beginners

19 Nov 2026

Likelihood and truth

The training target is the next token in the text, so a fluent continuation is not a check that the continuation matches the world.

A GPT is trained to predict the next token of text, so it learns which continuations are likely, not which are true. Here likelihood means how probable the model rates a continuation, given the text before it. A fluent continuation, one that reads smoothly, is not a check that it matches the world.

The target is the next token

During training, the model reads real text and practises guessing what comes next. The only answer it is ever marked against is the next token that actually followed in that text.

Here is one training step, using the sentence “The cat sat on the mat.”

  1. The model is shown the start: “The cat sat on the”.
  2. It produces a distribution: a share of probability for every token in its vocabulary, with all the shares adding up to 1.
  3. The share it gave to mat, the token that really came next, is checked.
  4. The model’s numbers are nudged a little so that mat gets a larger share in that spot next time.

This repeats across a very large amount of text. At no step is a sentence labelled true or false. If the training text contains a mistake, the target is still the token that followed the mistake.

So the base objective is next-token prediction on text. That objective does not compare a sentence with any outside list of facts.

Fluency is a property of the continuation

Fluent text is grammatical, reads naturally, and fits the style of the words before it. Likely tokens tend to produce fluent text, because the training text was mostly written that way.

Take the prompt “The capital of Australia is”. Two possible continuations:

  1. “Canberra.” This is true.
  2. “Sydney.” This is false. Sydney is the largest city, but not the capital.

Both continuations are grammatical. Both sound like a sentence a person might write. Sydney appears near the word Australia in a great deal of text, so it can receive a real share of probability in this spot. The model’s number measures how well a continuation fits patterns in text. Whether the claim is true is a question about the world, and the world is not part of the calculation.

This is why a fluent answer can still be false. Confident wording, a tidy structure, and correct grammar are all properties of the text. None of them is evidence that the content was checked.

When a GPT does give a true answer, it is because the true continuation was likely. That often happens when the fact appears many times, consistently, in the training text. Likely and true often agree. They are still two different things.

A comparison: someone who has read a thousand travel brochures can write a convincing one about a town they have never visited. The comparison stops at self-awareness. That writer knows they are guessing, and the model computes no separate signal that marks a guess.

A database is a separate system

A database is an organised store of facts that a program can look up, such as a table of countries and their capitals. Looking something up returns the stored entry, or nothing.

A base model, the model as it comes out of next-token training, does not consult a fact list. While it computes, its only input is the tokens in the context window. Its only output is a distribution for the next token. There is no lookup step in between.

promptlikely continuationless likely continuationfact listnot connected to the model
Likelihood is measured against text, not against a fact list