Large language model
A large language model, or LLM, is an AI system trained on vast amounts of text that writes and answers questions by predicting likely words.
- LLM
- Language model
What it means
A large language model learns patterns from a very large amount of text. Given some words, it predicts what is likely to come next, a piece at a time. That is how it writes sentences, answers questions and summarises. Google’s glossary gives Gemini and GPT as examples.1
The key word is likely. A model writes what usually follows, not what it has checked. So it is good at fluent copy, and liable to state a plausible detail that is not true.
With product data, a model is safest rewriting facts it is given, not supplying new ones. A title built from a row’s own values can be checked against the row; a capacity the model made up cannot.
Think of it like this
It is like the predictive text on a phone, scaled up enormously. It is very good at guessing what usually comes next, which is not the same as knowing what is true.
For example
Given “Type: freestanding, Material: acrylic, Length: 1,700 mm, Colour: white”, an LLM can write “A freestanding acrylic bath, 1,700 mm long, in white.” Asked for more, it may add “luxurious” or a guarantee the row never mentions. That is why every claim needs checking against the data.
Why it matters
Google counts copy a language model wrote as generative AI: such titles belong in structured_title and descriptions in structured_description, labelled trained_algorithmic_media.23 Record which copy a model wrote, so it can be checked and sent correctly.
Not to be confused with
A large language model is one kind of generative AI, the kind that works with text. Image generators are generative AI too, but they are not language models.1
Sources
- Google for Developers. Machine Learning Glossary.
- Google Merchant Center Help. Title [title] and structured title [structured_title].
- Google Merchant Center Help. Description [description] and structured description [structured_description].