Simplified Transformer Architecture

The diagram and steps below describe a simplified version of the process of how transformers are used when training LLMs to understand language.

1. Tokenise: the input text is tokenised
2. Embed Meanings: the tokens are turned into lists of numbers that encode their possible meanings
3. Find word connections: the transformer finds connections between the words represented as numbers using lots of thinking steps, stacked on top of each other all working at the same time.

Using those connections, the transformer then guesses which word is likely to come next.