Large language model (LLM)
Quick definition
A large language model (LLM) is a neural network trained on very large amounts of text to predict the next word, which lets it understand and generate natural language. It is the reasoning engine behind modern chatbots and AI agents.
On this page
How an LLM works
An LLM reads text as tokens — pieces of words — and, for each position, computes the probability of every possible next token. Generating an answer means repeating that prediction token after token. Almost all current LLMs use the Transformer architecture introduced by Vaswani and colleagues in 2017, whose attention mechanism lets the model weigh every part of the input when producing each word.
Training happens in stages. Pre-training on a huge text corpus teaches language and general knowledge. Instruction tuning and feedback from people then teach the model to follow requests, hold a dialogue and decline some tasks. In 2020, the GPT-3 paper showed that a large enough model could perform new tasks from a few examples in the prompt, without retraining.
What an LLM does not know
An LLM's knowledge is frozen at the end of its training data, and it has never seen your business's private information: prices, stock, return policy, customer orders. Asked about them, it will often produce a plausible guess — a hallucination. It also has a limited context window, the amount of text it can take into account at once.
That is why useful business assistants rarely rely on the model alone. Retrieval-augmented generation (RAG) feeds it the relevant passages from your content, and tools let it query live systems such as a catalog or an order database.
Choosing a model
Models differ in quality of reasoning, speed, price per token, the languages they handle well, context size, and where they are hosted. For customer service, a fast, mid-sized model is often enough for routine questions, while harder cases — long policies, multi-step requests — benefit from a stronger one. Data location matters too: a European business may prefer a model hosted in the EU.
The model is only one part of the result. Instructions, grounding data, tools and evaluation on real conversations usually matter more than switching from one leading model to another.
LLMs in AI SmartTalk
AI SmartTalk assistants are built around LLMs and use EU-hosted models by default; a customer can choose to use other providers' models instead. The assistant's answers are grounded in the business's indexed content and catalog, and it uses tools to look up live data rather than asking the model to remember it.
Frequently asked questions
Is ChatGPT an LLM?
ChatGPT is an application built on top of LLMs made by OpenAI. The model generates text; the application adds a chat interface, conversation memory, tools such as web search, and safety rules.
What does “large” mean in large language model?
It refers to the number of parameters — the internal weights learned in training, from billions to hundreds of billions — and to the size of the training corpus. There is no exact threshold.
Can an LLM learn from my company's data?
Two approaches exist. Fine-tuning retrains the model on your examples, which is costly and does not keep up with changing facts. Retrieval-augmented generation leaves the model unchanged and gives it the right passages at question time, which suits prices, stock and policies better.
Are LLMs safe to use with customer data?
It depends on the provider's terms, where the model is hosted and how the application handles personal data. Check whether the provider uses your data for training, where it is processed, and whether a data processing agreement is available.