What is a large language model?
The kind of AI behind chat assistants, explained without the jargon: what it is, how it produces text, and where it falls short.
The short version
- → An LLM is a program that predicts the next chunk of text, one step at a time.
- → It learned those predictions by reading a huge amount of writing, not from a database of facts.
- → It is fluent by design, which is also why it can be confidently wrong.
LLM stands for large language model. It is the technology behind chat assistants like ChatGPT, Claude, and the models Berges AI runs. At its core it does one deceptively simple thing: given some text, it predicts what text should come next.
That is the whole trick. Everything a chat assistant appears to do, answering questions, writing an email, explaining code, comes out of predicting the next piece of text again and again until the response is finished.
It predicts the next token
A model does not read or write whole words. It works in tokens, which are word fragments. It looks at everything so far and assigns a probability to every possible next token, picks one, adds it, and repeats. Fluent paragraphs are the result of doing this thousands of times very fast.
The "large" is about scale
These models are large in two ways: the amount of text they trained on, often much of the public internet, and the number of internal values, called parameters, that store what they learned. Modern models have tens or hundreds of billions of parameters, and the largest now pass a trillion. Scale is what turned next-word prediction from a party trick into something genuinely useful.
It has no live memory or database
A base model does not look anything up. Its "knowledge" is patterns baked into its parameters during training, frozen at the point training stopped. That is why it has a knowledge cutoff and why, without extra tooling, it cannot tell you today's news. Techniques like retrieval and tool use are how you give it fresh or private information.
Why it sounds sure even when it is wrong
The model optimizes for plausible-sounding text, not for truth. When it does not know something, it does not go quiet, it generates the most likely-looking answer anyway. That failure mode is called hallucination, and it is the single most important thing to keep in mind when you use one.
An analogy
Think of an extremely well-read autocomplete. Phone autocomplete guesses your next word from a little context. An LLM guesses the next fragment from everything written so far, having absorbed patterns from billions of pages. Same idea, vastly more context and training.
Go deeper
Are LLMs stochastic parrots?
Where the phrase came from, what it gets right about next-token prediction, what researchers have since found inside the models, and what to do with the answer. 15 min read →
Where Berges AI fits
Berges AI runs several of these models and adds a thin layer on top, pipelines and house rules, that shapes how they respond: the rules ask for no filler and for saying so when unsure, and the pipelines set up the kind of task. The model does the predicting; the layer keeps it useful.
Related concepts
How transformers work
The neural network design that powers almost every modern language model, and the one idea, attention, that made it work. Read →
What are tokens?
Models do not see words. They see tokens, the fragments that pricing, context limits, and speed are all measured in. Read →
How AI models are trained
The process that turns a random network into something useful: read a lot, adjust, repeat, then teach it to behave. Read →
Things people ask
Is an LLM the same as AI?
No. AI is the broad field. An LLM is one kind of AI, focused on language. Image generators, recommendation systems, and self-driving perception are all AI too, but they are not LLMs.
Does an LLM understand what it says?
It has no beliefs or awareness. It models statistical patterns in language extremely well, which can look like understanding, but there is no inner comprehension behind it. Useful, but not a mind.
Why does it sometimes make things up?
Because it is built to produce likely text, not verified text. When it lacks a real answer it still generates a plausible one. Always check anything factual or high-stakes.