What is a transformer model?

What it is

First, the text is split into small pieces called tokens, which are often whole words or parts of words. The central step is called attention. For each token, the model weighs how strongly it relates to the other tokens in the text, which is how it works out what a word like "it" refers to. Finally, it scores every possible next token, picks one of the likely ones, adds it, and repeats.

Before it can do any of this, the model is trained. It is shown very large amounts of text and adjusts itself, many times over, until its predictions improve. That is why it writes smooth, natural prose, and it is also the source of its main weakness. The model produces plausible text and does not look up verified facts, so it can state something false in the same calm voice as something true. Tools such as Claude and ChatGPT from OpenAI are built this way.

Why it matters to a business

• Fluency is not accuracy: well-written output can still be wrong, so facts, figures and quotes need checking.
• Your data: with a hosted model, your text is sent to the provider to be processed, so decide what may be shared.
• Cost: usage is usually measured in tokens, so longer inputs and outputs cost more.
• Human review: a knowledgeable person should read anything that reaches a customer or a record.

How this shows up in our work

In our AI and automation work we choose the model for each task, keep a person in the loop for output that matters, and help clients decide what data may be sent. For a longer view of tools that act on their own, read about agentic AI.

Learn more

• Transformer on WikipediaA technical overview of the design, useful for readers who want more depth.
• Large language model on WikipediaA neutral overview of how these models are built and used, useful as a starting point.