A large language model (LLM) is an AI model that has learned to predict the next piece of text in a sequence. Everything else follows from that single ability: summarising, translating, classifying, programming and answering questions can all be phrased as “continue this text sensibly”. It learns from very large amounts of text, without anyone teaching it grammar, rules or facts one by one.
The term is younger than the technology behind it. Language models have existed for decades. They became large and practically useful only with the transformer architecture and with training runs that need serious compute clusters.
How an LLM works
Text becomes numbers
A model never sees letters. It sees tokens, fragments of words that have been translated into numbers beforehand. Each token gets a position in a high-dimensional space where similarity of meaning shows up as spatial closeness. After that the model only ever does arithmetic.
The architecture
The technical core is the transformer, presented in a Google paper in 2017. Its trick is called attention: while processing one token, the model may look at every other token in the text and weigh which ones matter right now. That is why an LLM can hold a reference across several paragraphs, where older approaches lost it after a couple of sentences. Underneath sits ordinary deep learning, a very deep neural network trained through machine learning.
Two training stages
Pre-training runs a single task: guess the next token, billions of times over. After that the model knows language but does not yet do what it is told. The second stage turns it into an assistant. It learns from examples of instruction and good answer, and people rate answers so the model reinforces the preferred patterns. That second stage explains why two technically similar models can differ noticeably in tone and in how well they follow instructions.
The numbers on the data sheet
| Figure | What it describes | Why it matters |
|---|---|---|
| Parameters | Number of trained weights in the network | Rough hint at capacity and at the hardware needed to run it |
| Context window | How many tokens fit into one call, prompt and answer together | Decides how much document, history or code the model can see at once |
| Knowledge cut-off | How recent the training data is | Anything later is known only if you supply it |
| Modalities | Whether images, audio or video are handled alongside text | Determines whether a scan or a screenshot can be used directly |
Bigger is not automatically better. Smaller models answer faster, cost less and run on modest hardware. For classifying, extracting and rewriting they are often enough.
What it is good for
- Reshaping text: shortening, rewriting, changing register, forcing output into a fixed format.
- Extracting and sorting: pulling invoice data out of PDF attachments, sorting support requests by topic, turning free text into structured fields.
- Search across your own material: the model is handed the relevant passages from your documents and answers from them. The method is called retrieval augmented generation.
- Development support: code suggestions, test cases, migration scripts. The extreme version of this is vibe coding.
- Conversational interfaces: an AI chatbot for recurring questions, sensible wherever the answer comes from maintained sources.
One pattern runs through all of these. Language models are strong where a task processes language, has many valid variants, and where a mistake shows up before it gets expensive.
Where it runs out
Invented detail. A model produces plausible continuations, not verified statements. Wrong answers therefore sound exactly as convincing as right ones, see AI hallucination.
No knowledge of your organisation. Internal processes, prices, contracts and stock levels are not in the training data. If you need them, you have to supply them.
Answers that vary. The same question can be answered two different ways. Processes with a fixed expectation need checking steps around the model.
Data protection. Whatever goes to an external model leaves the building. With personal data that question belongs at the start of a project, not at handover.
Attackable through its input. Instructions can be hidden inside supplied material, see prompt injection. A model does not reliably distinguish your instruction from a sentence in an uploaded document.
Someone else’s model or your own
| Criterion | Model behind an external API | Self-hosted open model |
|---|---|---|
| Getting started | An API key, first answer in minutes | Hardware, operations, updates |
| Data flow | Content leaves your environment | Stays entirely in house |
| Cost | Billed per token processed | Fixed infrastructure, independent of volume |
| Control | The provider can change or retire models | The version stays frozen until you move it |
For the second route, look at open-source LLMs. Legally both land in the same place: the EU AI Act asks what the system is used for, not how it is hosted.
Frequently asked questions about large language models
Does an LLM understand what it writes? It has no world model in the human sense and no intent. What it has is a very dense statistical map of how language works over subject matter. For many tasks that is indistinguishable from understanding. For others the difference surfaces the moment the question falls outside what was learned.
How do I teach a model about my company? In most cases not through training at all, but by supplying context. The relevant passages are placed alongside the question and the model answers from them. That is cheaper, faster to update and easier to audit than a training run of your own.
What is the difference between an LLM and generative AI? Generative AI is the umbrella term for models that produce new content, images and music included. A language model is the variety that works with text.
How much work is a production deployment? Calling the model is the smallest part. The work sits in the data, in reviewing the output, in permissions and logging, and in deciding what happens when an answer is wrong.
Conclusion
A large language model is a tool with a very broad edge and a very sharp one. Broad, because a surprising number of tasks can be phrased as a text task. Sharp, because the model never announces that it is guessing. Projects rarely fail on the model itself and almost always on what gets built around it: clean data, clear ownership of the review step, a defined way of handling errors. How we build software, and where AI belongs in it, is described on our page about custom software development. The route from a model to an application that survives daily operation is described on our page about AI integration into existing software. What holds up in practice and what does not is written up in the article AI in software development. For an assessment of your own case there is a free consultation.