What is a language model, and why doesn't it know my company?
The short answer
A language model is a program trained on a very large amount of text, learning to predict what comes next. That is how it learned to write, summarise and reason. But your contracts, prices and clients were never in that text, and its knowledge stops at a fixed date. Whatever it needs to know about your company, someone has to give it.
How it learned
A language model is the engine inside every AI assistant and every AI agent. It was built by showing it a huge collection of text: public web pages, books and other written material. During this training, the model tries again and again to predict the next piece of text, and adjusts itself each time it gets it wrong.
The scale is hard to picture. GPT-3, a model described in a well-known 2020 research paper, already had 175 billion parameters. A parameter is one of the internal numbers that the model adjusts during training. Out of this process comes the ability to write, translate, summarise and work through a problem.
What it did not learn
Your internal documents were never part of that text. Your price list, your contracts, your client history, and the exceptions your team knows by heart: the model has never seen any of them.
It also has a cutoff date. Every model stops learning at a certain date, and many model makers publish it. Anthropic, for example, lists for each of its models a “reliable knowledge cutoff”: the date up to which its knowledge can be relied on. For anything that happened after that date, and anything that was never public, the model cannot be relied on.
Why it still answers
A language model is built to produce a likely continuation of the text in front of it. Ask it about your company, and it will produce a likely-sounding answer, whether or not it knows the facts. This is where confident wrong answers come from. The question Why does my AI agent give confident wrong answers about my own business? covers this in detail.
What this means for you
Think of the model as an engine, and of the knowledge about your business as the fuel you must supply. There are two ways to supply it. You can give the model the right documents at the moment it answers, or you can train it further on your own data. The page Should I train the AI on my company’s data, or give it the data when it answers? explains the difference and which one to choose.