What is an LLM? The language model behind ChatGPT, Claude and Gemini

When you write to ChatGPT and it replies, there is nobody on the other side thinking. There is an LLM guessing, very well, which word comes next.

LLM stands for Large Language Model. In one sentence: it is a program that has read an enormous amount of text and has learned to predict the most likely continuation of whatever you write to it.

How it works, without the maths

Think of the suggestion on your phone keyboard, the one that offers you the next word. An LLM does the same thing, but on a gigantic scale and taking the whole conversation into account, not just the last two words.

During training, the model reads books, websites, manuals and code. It does not memorise the pages: it learns how words fit together. From there come abilities that look like reasoning, such as summarising a contract or writing a payment reminder email.

The word "large" matters. It is the biggest models, trained on more text, that can hold long conversations and follow complicated instructions.

An example that concerns you

You ask an LLM: "Write a polite reply to a customer who is complaining about a late delivery". It knows nothing about your customer or your warehouse. But it knows what thousands of apology emails look like, and it composes a plausible one.

This is the weak point. Plausible does not mean true. If you ask it for the delivery date and you have not given it one, it may invent one with the same confidence with which it writes the rest. People in the field call this a hallucination.

Why it really matters to you: the same engine under different bodywork

The leading models are built by a few companies, such as OpenAI, Anthropic and Google. Training one costs a huge amount, so almost nobody starts from scratch.

Many of the AI tools offered to you, from the email assistant to the chatbot for your website, are an interface built on top of one of these models. The supplier uses it on a pay-as-you-go basis and resells you the result under its own brand.

There is nothing wrong with that: a good product adds integrations, your own data, support. But you need to know it, because sometimes you pay a subscription to get, more or less, what you would get from ChatGPT or Claude directly.

What to do now

The next time a supplier presents you with a "proprietary AI", ask a simple question: which LLM is underneath? A clear answer is a good sign. A vague answer is not.

Then ask yourself what they add compared with the bare model. If the answer is "a nicer interface", try the original tool for a few weeks first.

And do not trust an LLM with facts you have not supplied yourself: prices, deadlines, customer data. Use it to write and to tidy up, and check the content yourself.

If you want to understand which model sits behind the tools you are evaluating, at AIR we often do exactly this comparison with business owners. But you can ask the right question on your own, starting tomorrow.