Large language model

Most LLMs use the transformer architecture introduced in 2017. During pre-training, the model reads a very large collection of text and adjusts its parameters to get better at predicting what comes next. It is then usually tuned with instruction data and human feedback so that it follows requests, declines harmful ones and answers in a helpful format. When you send a prompt, the model generates its reply one token at a time.

For buyers, the LLM inside a product drives much of its quality, speed and cost. Two writing tools built on the same model may perform similarly, while switching to a cheaper model can cut costs but lose nuance. Many product pages name the model they use; checking that, along with the context window and data-use terms, tells you more than marketing copy does.

Example: when you ask an assistant to turn meeting notes into a summary, the LLM reads the notes as tokens within its context window and predicts a summary that fits your instructions.

Limits: unless it is connected to search or documents, an LLM does not look facts up, so it can state false information confidently (hallucination). Its knowledge stops at a training cutoff, and its output can vary between runs.