Glossary
What is a large language model (LLM)
In short
A large language model (LLM) is a neural network trained on a huge amount of text. It writes, summarizes and analyzes text on request.
Language models · large language model
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine a friend who has read a whole library: fairy tales, textbooks, letters and recipes. He does not remember each book by heart, but he is so used to words that he easily guesses how any phrase will end. Ask him for a poem or an explanation of a problem, and he answers. Only sometimes he gets confused and says something wrong with confidence, so adults check his answers.
In plain words
It is a program that learned to write by going through billions of pages from books, websites and documents. It holds no ready answers. It looks at your question and picks, step by step, the next word that fits the meaning. This is how a coherent letter, summary or list appears. For example, you paste a long contract and ask for the payment terms: it finds them and writes them briefly. Its knowledge ends at the date when training stopped, so fresh facts must be given right in the request.
For business
An LLM handles text tasks: answers to customers, draft emails, contract review, knowledge base search and call summaries. There are two ways to pay for it. A cloud service charges for the amount of text in and out. Your own model on a server needs graphics cards and an engineer, but the data does not leave the company, which matters under 152-FZ, Russia’s personal data law. The main risks are invented facts, leaks of confidential information into someone else’s cloud, and dependence on one vendor. Before you deploy, check which tasks it gets wrong, who checks its answers and where the requests are stored. Which processes will pay back such a solution is calculated in the AI solutions map.
For a specialist
A large language model (LLM) is a neural network, usually built on the Transformer architecture, trained to predict the next token on a large corpus of text and code. The path from a raw model to an assistant includes pretraining, instruction tuning and alignment with human preferences (RLHF, DPO). Kaplan and coauthors (2020) showed that the loss falls by a power law as the number of parameters, the data size and compute grow. Brown and coauthors (2020) described GPT-3 with 175 billion parameters, which solved tasks from a few examples in the prompt (few-shot) without updating its weights. There are dense models and mixtures of experts, base and chat models, closed models and models with open weights. Quality is measured by benchmarks and, more important for deployment, by tests on your own tasks.
Do not confuse. LL.M. (Legum Magister) is a master of laws degree and has nothing to do with neural networks.
Example
A lawyer uploads twenty pages of a supply contract into the model and asks it to list the penalties and deadlines. A minute later he has a table with the clauses and page numbers, which he only needs to check.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map