HOW LLMs WORK FOR THE NON-SPECIALIST

WHAT A LARGE LANGUAGE MODEL IS, WITHOUT MATHEMATICS An LLM is a computer program that has learned the patterns of written language by studying an enormous quantity of text, and that can read some text and then write more text that fits. Let us unpack the three words in the name. Large means two things. First, the model was trained on a very large amount of text, such as books, articles, websites, and computer code. Second, the model itself is large, containing billions or even trillions of adjustable numbers called parameters. A parameter is one of the adjustable dials inside the model that were tuned during training. Language means that the model works with human languages, and, in many cases, with programming languages too. Model means that it is a stand-in, a compact numerical description of the patterns found in the training text. A helpful everyday comparison is autocomplete on a phone. When you type "see you", your phone may suggest "soon" or "tomorrow." It does this because it has seen many messages and knows which words often follow. An LLM is like autocomplete that has read a vast library, and that has learned not just which word follows which, but patterns of grammar, facts, styles, reasoning steps, and the structure of arguments and programs. So when you ask an LLM a question, you are, in effect, giving it a beginning and asking it to continue in the way a helpful answer would continue. The process by which the model learns is called training. During training, the model reads huge numbers of passages with the next piece hidden, tries to guess it, and is nudged to guess better. After a long and costly training process, the dials settle into values that make the model good at continuing text. Many modern models then go through additional training, often called fine-tuning or alignment, in which people show the model examples of helpful, honest, and safe answers so that it behaves like a useful assistant rather than a raw text continuer. After training is complete, the dials are frozen. When you use the model, it does not learn from your conversation unless the builders retrain it. This single fact explains many behaviors that surprise beginners, and we will return to it when we discuss statelessness.