GlossaryKind of term: Manufacturing termLarge language model (LLM)
What is a large language model?
A large language model (LLM) is an AI model trained on vast amounts of text that generates text by repeatedly predicting a likely next word fragment.
Technically an LLM is a large neural network. The word fragment it predicts is called a token, and each prediction builds on what is already there. That lets it summarise an email, pull details such as quantities and delivery dates out of a request, draft a text or answer a question in plain language. Many recent models handle images as well as text.
What does a large language model depend on?
A language model does not calculate the way an estimate does: it generates an answer that fits the input statistically. A number that looks plausible can therefore be wrong without the text showing it; in the field this is called hallucination. How reliable an output is depends on the task, the input, the instructions and the system around the model. When that system has the model call a calculation or consult a database or document, the number comes from that source, and the source can be shown with it.
How does a large language model work in practice?
In an estimate, what matters is where a number comes from. A language model can name a cycle time or a price per kilo that looks right with no machine, material, setup or source behind it. It is useful for the work around the price: sorting a request, finding quantities and delivery dates, drafting a covering note. Time, rate and quantity come from a calculation you can check. Whatever a model drafts goes to a customer only once someone has read and approved it.
What is a large language model often confused with?
Costing software calculates with times, rates and quantities by fixed rules; a language model generates text. A search engine works differently: a search engine points to existing sources, a language model generates an answer from the model itself, unless the system around it retrieves sources and passes them in. And an agent builds on a language model: it uses the model to carry out steps in other software.
Frequently asked questions
Can a large language model calculate a cost?
Not reliably on its own. It can name a figure that looks plausible with no calculation behind it. If the system around it calls a calculation step, the figure comes from that calculation; then check the inputs and the source.
Why does a wrong answer from a language model sound so convincing?
Because the model generates text that fits the input rather than looking up a fact. A wrong number or a non-existent standard can read as convincingly as a correct one. A source or a calculation step alongside makes checking possible.