Large Language Models Are Prediction Engines

A large language model can write an email, explain code, translate a paragraph, and still invent a source that never existed. All of those behaviours come from the same basic job: predicting what text is likely to come next.

The results can feel like a conversation, but fluency is not the same as knowledge.

How the prediction becomes useful

During training, an LLM processes a large collection of text and adjusts many numerical parameters to model patterns in language. The text is split into smaller units called tokens. Given a sequence of tokens, the model estimates which token should follow.

Repeated one token at a time, that produces a sentence, a code block, or a long answer. The model uses the context in your prompt and conversation, but it does not look up every statement in a database of facts.

What LLMs are good at

They are useful for tasks where language patterns matter:

The surrounding application matters. A model connected to search, a calculator, or your own documents can do work that the model alone cannot do reliably.

Where they fail

An LLM can produce false information in exactly the tone it uses for a correct answer. It may also reproduce bias from training data, miss recent events, misunderstand an ambiguous instruction, or expose information placed carelessly into a third-party service.

Use it as a source of drafts and candidates, then verify anything consequential. For code, run the tests. For facts, open the source. For private data, understand where the prompt goes before pasting it.

I find LLMs most useful when the output is cheap to check. They can save time getting from nothing to a draft, but the final ten percent—judgment, evidence, and responsibility—still belongs to the person using them.