Probabilistic Nature of Language Models

Most people first meet a language model through a chat window. You type a question, and text comes back that looks like an answer. It is natural to assume the model is listening to you the way a person would, understanding your request and replying to it. That assumption shapes how people use these systems, and it is also the source of many surprising results. This article starts from that assumption, tests it, and then builds a more accurate picture of what the model is doing when it writes text.

The central idea is that a language model can be treated as a completion machine. It receives a piece of text, called the prompt, and produces the text that most plausibly follows it. The model does not look at your prompt and decide to answer it. It asks, in effect, what text usually comes after text like this, given everything in the prompt.

The misconception to correct is this: a model that produces fluent replies must be engaging with your request as a conversation partner. Completion is a different activity. The same mechanism can produce a helpful answer, a pointless restatement, or a second question, depending on what the prompt looks like and what usually follows text like it.

Conversational behavior is a separate layer. The source notes that shaping a model to respond appropriately to a user's request is handled in a later stage called post-training. The basic completion behavior comes first, and conversational behavior has to be taught on top of it.

A conversation requires the model to recognize that you asked something and to respond to that request. A completion does not require that at all. A raw completion model simply extends the text it has been given. This is why a question can come back as another question: the text that usually follows a question in similar writing is not always an answer.

Consider a prompt that is not a famous quotation: 'The lighthouse keeper climbed the stairs and'. The model's continuation depends on which words most commonly follow sentences like that. The model does not need a story about lighthouses to extend the sentence. It needs only to predict the next text that fits. The diagram below traces the process as a loop: the model predicts a next piece, that piece joins the text, and the model predicts again from the longer text.

Because the model estimates which text is likely to come next, it does not produce one fixed answer by rule. Its estimates are probabilities. Many different continuations can be plausible for the same prompt, and the one that appears in the output is a continuation that is likely given the input. The same mechanism that makes these systems flexible is what makes their output uncertain.

A completion is a prediction, and predictions can be wrong. A continuation can read smoothly and still be factually incorrect, because fluency and correctness are separate properties. The model is optimizing for text that fits what usually follows, and that is a different target from text that is true.

The source describes this probabilistic character as what makes language models both exciting and frustrating to work with. The limitation is not a bug in one particular model. It follows from the framing itself: a system whose job is to continue text will sometimes continue in ways that serve you poorly. Knowing this is the first step toward writing prompts that steer it.

Once you accept that a model completes text, many jobs can be restated as completion problems. The source highlights translation, summarization, coding, and solving math problems, and also spam classification. The method is the same in each case: write the prompt so that the most likely continuation is the output you want.

Translation works this way. The prompt 'How are you in French is ...' leaves a gap that a plausible continuation can fill with 'Comment ça va'. The model is not given a translation rule to execute. It is continuing a pattern in which a phrase is followed by its French rendering, so the translation appears as the completion.

Spam detection follows the same pattern. Consider a prompt that asks 'Is this email likely spam?' and then includes the email text. A plausible completion is 'Likely spam'. Seen this way, the language model becomes a spam classifier, because the label it writes is simply the continuation that fits the question.

The source names summarization, coding, and math as further examples of the same reframing. The principle does not change: the prompt should make the desired output the text that most plausibly comes next. The output is still a prediction, so it is still subject to the reliability limits described above.