Neural Network
A neural network is a computing model built from layers of simple units, each combining its inputs with learned weights and passing the result forward, so that the network as a whole maps an input to an output without anyone writing the rule. Every large language model is one.
How the weights are set
Training shows the network an input, compares what it produced with what it should have produced, and adjusts every weight a little in the direction that would have reduced the error. Repeated over an enormous number of examples, the weights settle into values that make the network good at the task it was scored on. For a language model, the task is predicting the next token of text, and the examples are text. The finished model is the set of weights. Nothing else is stored.
What a network does not contain
It holds no table of facts and no copy of any web page. Whatever it appears to know about a business is an association spread across weights, formed from however often and however consistently that business appeared in the training text, which for most small businesses is rarely. This is why a model asked a question without access to the web can answer fluently and wrongly, and why assistants add a retrieval step: search, fetch a few pages, and compose the answer from what those pages say now. The network supplies the language and the judgement; the pages supply the facts.
Deep, large, transformer
The vocabulary around the term is mostly size and shape. Deep means many layers. Parameters are the weights, and a model’s size is usually quoted as their count. A transformer is one particular arrangement of layers, built around an attention mechanism that lets every token in the input weigh every other. A large language model is a transformer trained on text at a scale where the behavior becomes general enough to hold a conversation.
A common question
If a model stores no facts, how can it answer questions about a business?
The finished model is only its weights, so whatever it appears to know about a company is an association spread across them, formed from how often and how consistently that company turned up in the training text. For most small businesses that is rarely, which is why a model cut off from the web can answer smoothly and get it wrong. Assistants add a retrieval step for that reason: they search, fetch a few pages and build the answer from what those pages say now. The weights supply the language and the judgement, the pages supply the facts.