---
title: "LLM: definition and example – Carl-Uwe Würfel"
description: "LLM (large language model) explained: how large language models produce answers, what they are good for in business, where they go wrong, and how to choose one."
url: https://carl-wuerfel.de/en/glossary/llm
author: Carl-Uwe Würfel
---

1. [Home](https://carl-wuerfel.de/en)
  2. [Knowledge](https://carl-wuerfel.de/en/knowledge)
  3. [Glossary](https://carl-wuerfel.de/en/glossary)
  4. LLM

# LLM

An LLM (large language model) is an AI model that processes and generates text based on learned language patterns, for example to summarize, draft or sort information.

## On this page

  1. How a language model is made
  2. How an answer comes together
  3. What language models are good for
  4. What choosing a model comes down to
  5. Typical failure modes
  6. What running one in a business involves
  7. Where the term stops working
  8. Frequently asked questions
  9. Sources

 _Figure: Model size and product name alone say little about whether a language model does a specific task well; only testing it on the use case in question shows that. A fluent answer can be factually wrong._

**Example**

A model suggests which department an incoming inquiry belongs to, such as a billing question going to accounting, and explains the suggestion based on the text. A person checks the routing before the inquiry moves on. Before rollout, several models were compared on a sample of several hundred real inquiries that had already been routed; the team chose the cheapest model that reached the accuracy target set in advance.

An LLM calculates probabilities for which piece of text comes next and builds its answer one piece at a time. It splits input into [tokens](https://carl-wuerfel.de/en/glossary/token), small units of words or word fragments, and takes into account everything that fits into its [context window](https://carl-wuerfel.de/en/glossary/context-window). “Large” refers to two things: the huge amounts of text the model learns from, and the number of its parameters, the weights that are set during training.

Technically, today's widely used language models are built on the transformer architecture, which researchers at Google introduced in 2017 in the paper “Attention Is All You Need.” A [chatbot](https://carl-wuerfel.de/en/glossary/chatbot) is the application that uses an LLM. The model itself has no interface, no memory beyond the current conversation and no access to company data unless the application supplies it.

For executives, fit matters more than the technology. Model size and product name alone say little about whether a language model does a specific task well. Only testing it on the [use case](https://carl-wuerfel.de/en/glossary/use-case) in question shows that. The choice of model therefore follows the task, not the brand name.

## How a language model is made.

An LLM is built in several training phases. In the first, pretraining, the model learns from very large amounts of text which word is likely to come next in which context. In doing so, it picks up statistical patterns of language, facts and lines of reasoning without anyone giving it rules.

A model fresh out of pretraining continues text, but it does not follow instructions yet. Further phases train it to complete tasks, answer helpfully and decline certain requests. A well-known method is training with human feedback, which OpenAI described in 2022 in its InstructGPT paper: people rate answers, and the model is adjusted toward the better-rated ones.

Two properties that matter in daily use follow from this. First, every model has a knowledge cutoff where its training data ends. It knows nothing about what happened after that. Second, a model does not look anything up. It answers from what it learned and from what is in its input. Current or internal knowledge only gets in through the application, for example through a search ([RAG](https://carl-wuerfel.de/en/glossary/rag)) or connected tools.

## How an answer comes together.

With every request, the model receives an input that can have several parts: a fixed framing instruction from the application (the [system prompt](https://carl-wuerfel.de/en/glossary/system-prompt)), the conversation so far, any attached documents and the actual question. All of it has to fit into the context window, whose size varies by model.

The model generates the answer token by token. At each step it calculates how likely each continuation is and picks one. Because an element of chance is involved, the same question can produce differently worded answers. Settings such as temperature narrow that range, but in practice it rarely disappears entirely.

Newer models can generate intermediate steps before the actual answer in order to break a task down. Providers call this reasoning or extended thinking, for example. It often improves results on multi-step tasks, but it costs extra tokens and time.

What matters for operations: providers typically bill model usage by the token, with separate rates for input and output. Long inputs, long answers and many intermediate steps make a call more expensive and slower.

## What language models are good for.

Language models are strong at tasks where text has to be understood, reshaped or produced. That includes summarizing long reports, minutes or call notes. It includes drafting text that a person then reviews and approves, such as replies to customer inquiries, product descriptions or internal announcements. And it includes sorting: routing inquiries to the right department, classifying documents by type, picking up the tone of feedback.

Another field is pulling details out of unstructured text, such as invoice data from emails or deadlines from contracts, and moving them into fixed fields. Translation, along with writing and explaining code, is also a common use.

These tasks share one thing: the result can be checked before it goes out. Less suitable are tasks where a mistake slips through unnoticed or has serious consequences, such as exact calculations without a connected calculation tool, statements about facts that are not in the input, or decisions with legal effect made without human review.

## What choosing a model comes down to.

Model size and product name alone say little about fitness for a specific task. A smaller, cheaper model can handle a simple routing task just as well as a provider's largest one. Selection therefore follows testing on the use case in question: the same real cases, several models, results compared.

Cost per call and speed decide the matter for tasks that run often. The size of the context window determines how much material a model can process at once. Data protection and contract terms govern where data is processed and whether the provider may use inputs for its own training. And there is a choice between models available only through a provider's API and open-weight models that can also run on your own servers.

Providers release new models regularly and retire older ones. Anyone who hard-wires applications to one model has to rebuild them with every change. A separate layer between applications and models makes switching easier and lets each task go to the model that does it best and at the lowest cost.

## Typical failure modes.

The best-known failure mode is [hallucination](https://carl-wuerfel.de/en/glossary/hallucination): a fluent answer can be factually wrong. The model then makes up facts, figures, quotes or citations that sound convincing but are not true. A language model produces likely continuations; it does not check for truth. Where it lacks knowledge, it fills the gap with something plausible.

A second failure mode is outdated knowledge. Questions about prices, legal status or product versions get answered from the state of the training data unless the application supplies current sources.

Third, results vary. The same task can come out differently on two runs, and a small change in the instruction can change the result a lot. Fourth, models pick up biases from their training data, such as one-sided assumptions about professions or groups of people.

Fifth, language models are vulnerable to [prompt injection](https://carl-wuerfel.de/en/glossary/prompt-injection): text from outside sources, such as an email or a web page, can contain instructions the model follows. This matters most when a model can also act through connected tools. All five failure modes can be reduced, but not eliminated.

## What running one in a business involves.

Using an LLM in a company takes more than buying access. First, there needs to be a rule for review: no AI result moves on unchecked.

Then come rights and access. It has to be clear who may use which model with which data, and confidential data only goes to services whose contract terms allow it. It also makes sense to log every call with its cost.

On the legal side, the EU [AI Act](https://carl-wuerfel.de/en/ai-act) distinguishes between providers and deployers. Obligations for providers of general-purpose AI models have applied since August 2, 2025, and usually fall on the model makers. Companies that use a language model are generally deployers. For them, the main points are measures for staff AI literacy under Article 4 and the transparency obligations under Article 50, for example for published content. The GDPR continues to apply alongside it.

Finally, training belongs in the picture, ideally with real tasks from each team's own work. Only people who have seen the failure modes in their own workflow review results effectively.

## Where the term stops working

A fluent answer can be factually wrong; this kind of error is called a hallucination. Model size and product name alone say little about whether a model fits a specific task. An LLM is distinct from a chatbot, which uses it as an application, from RAG, which supplies it with knowledge from documents, and from an AI agent, which connects a model to tools so it can carry out steps on its own.

## Frequently asked questions

What is the difference between an LLM and AI?

AI is the umbrella term, and an LLM is one specific kind of AI. [Artificial intelligence](https://carl-wuerfel.de/en/glossary/kuenstliche-intelligenz) covers many methods, such as image recognition, forecasting models or recommendation systems. An LLM is a large language model specialized in text. Every LLM is AI, but not all AI is an LLM.

What is the difference between an LLM and ChatGPT?

ChatGPT is an application, and the LLM is the model behind it. ChatGPT is OpenAI's chat service, which uses the company's language models and adds an interface, conversation history and tools.

What does “large” mean in large language model?

“Large” refers to the size of the model and its training data. An LLM has a very large number of parameters, the weights set during training, and learns from very large amounts of text.

Why does an LLM give different answers to the same question?

Because at each step the model picks one of several likely continuations, and an element of chance is involved. Settings such as temperature narrow that range. For tasks that need consistent results, clear instructions, fixed output formats and testing across many runs help.

Does an LLM learn from what employees type in?

A single request does not change the model. What it takes into account during a conversation sits in its context window and is gone afterward unless the application stores it. Whether a provider may use inputs to train future models is governed by its contract and privacy terms, which should be reviewed before rollout.

What is an open-weight model?

An open-weight model is a language model whose trained weights are publicly available for download. It can run on your own servers, so data does not have to leave the company. In return, the operator takes on hardware, maintenance and security, and license terms vary from model to model.

Does the EU AI Act apply to using an LLM?

Yes, though the obligations depend on the role and the specific application. Obligations for providers of general-purpose AI models usually fall on the model makers. Companies that use a model mainly have to ensure AI literacy among their staff and observe transparency obligations; if a model is used in a high-risk area such as hiring, further obligations apply.

## Further reading

  * [**AI glossary**](https://carl-wuerfel.de/en/glossary)
  * [**AI automation**](https://carl-wuerfel.de/en/ai-automation)
  * [**Token**](https://carl-wuerfel.de/en/glossary/token)
  * [**Context Window**](https://carl-wuerfel.de/en/glossary/context-window)
  * [**Hallucination**](https://carl-wuerfel.de/en/glossary/hallucination)

## Sources

  1. [Vaswani et al.: Attention Is All You Need, 2017](https://arxiv.org/abs/1706.03762)
  2. [Ouyang et al.: Training language models to follow instructions with human feedback, 2022](https://arxiv.org/abs/2203.02155)
  3. [European Union: Regulation (EU) 2024/1689 (AI Act), 2024](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)