What is Large Language Model?
Also called LLM.
A large language model is a statistical model trained on very large amounts of text to predict the next piece of text in a sequence. That single objective, applied at scale, produces systems that can answer questions, write and summarize documents, translate, and generate code. The model stores what it learned in billions of numeric weights rather than in a searchable database of documents.
Training works by hiding the continuation of a text passage and asking the model to guess it, then nudging its internal weights whenever the guess is wrong. Repeating this across trillions of words forces the model to encode grammar, facts, styles, and reasoning patterns, because all of them help with prediction. Nothing in the process stores sentences verbatim, which is why a model can produce fluent text about a topic and still be wrong about it.
Scale matters in three directions at once: the amount of training data, the number of parameters, and the compute spent on the training run. Increasing all three together has reliably improved capability, which is the observation behind the so-called scaling laws. Gains are smooth in aggregate but lumpy in practice, so a larger model can be dramatically better at one task and roughly equal on another.
The most common misunderstanding is treating a language model as a search engine. It has no index to consult at answer time and no built-in sense of whether a claim is true, so it can state incorrect details in confident, well-formed prose. It also has no memory between separate conversations unless something outside the model stores that history and supplies it again with the next request.
Language models are one family within the broader class of foundation models, and nearly all current ones are built on the transformer architecture. Everything a model reads or writes is first cut into tokens, and the amount it can consider at once is bounded by its context window. Behavior after pre-training is shaped by instruction tuning and by feedback-based methods rather than by the original prediction objective.
Key points
- Trained to predict the next token, not to look up verified facts.
- Knowledge sits in numeric weights, not in a searchable document index.
- Capability rises with more data, more parameters, and more compute.
- Fluent output is not evidence of accuracy.
- Stateless by default: each request carries whatever history it needs.
In practice
Imagine typing 'The invoice is attached and payment is due within' into a model. It looks at every preceding word, ranks all possible continuations, and finds that 'thirty' scores far higher than 'purple'. It emits that piece, adds it to the sequence, and repeats the whole calculation. The same loop, run over a longer prompt, is what produces a full paragraph, a summary, or a block of working code.