What is the difference between NLP and an LLM?

NLP is the field, and an LLM is one kind of tool within it. Natural language processing covers every technique for getting computers to work with language, from spam filters to translation. A large language model is a general-purpose kind of model that now handles most of those tasks from a plain-language instruction.

Do we need to train our own language model?

Almost never. Training a large model from scratch costs far more than most companies would ever recover. The usual route is to use an existing model and give it your own material through retrieval, and to fine-tune only for a narrow, high-volume task with the data to support it.

Why does a language model make things up?

Because it is trained to produce likely text, not true text. When the answer is not in what it learned or what it was given, it still produces something fluent. Giving it the source material and asking it to answer from that reduces the problem a great deal; checking what matters removes the rest.

Do language models work in Norwegian?

Yes. The large models write and understand Norwegian well, though less reliably than English for specialist vocabulary and dialect. For tasks that depend on Norwegian specifically, such as transcribing speech, models trained on Norwegian data, like the National Library's NB-Whisper, are worth testing alongside the general ones.

Machine Learning

A Beginner's Guide to NLP and LLMs

Natural language processing is the field that lets software work with human language, and a large language model is its most capable tool so far. What both are, how an LLM works, where it fails and how a business should start.

By Kjell Rune Monsø6 min readAugust 28, 2025

A person taking notes in a notebook at a desk, with a laptop turned slightly away.

Natural language processing (NLP) is the branch of computer science that lets software read, interpret and produce human language. A large language model (LLM) is the most capable NLP tool built so far: a neural network trained on very large amounts of text to predict what comes next, which turns out to be enough to summarise, translate, answer questions and write code. Every chatbot, AI agent and "ask the document" feature of the last few years is an LLM underneath.

This guide explains both in plain terms, where they fail, and how a business should start.

What is natural language processing?

Natural language processing is the work of turning language into something a computer can act on, and back again. For decades that meant separate, narrow systems, each built for one task:

Classification: is this email spam, is this ticket about billing or delivery.
Named entity recognition: finding the people, companies, places and dates in a text.
Sentiment analysis: whether a review is positive, negative or mixed.
Machine translation: from one language to another.
Summarisation: a long text made short.

Each of those used to need its own model, its own labelled training data and its own team. The shift of the last decade is that one large model can now do all of them, and many more, from an instruction written in plain language.

What is a large language model?

A large language model is a neural network trained on a very large body of text to predict the next piece of text, called a token, given everything before it. A token is a word or a fragment of one. A model reads and writes in tokens, which is why prices and limits are quoted in them.

Predicting the next token sounds modest. At the scale of billions of parameters and trillions of tokens of training text, it forces the model to learn grammar, facts, the structure of an argument and the conventions of code, because all of them make the next token easier to guess. The best-known model families today include GPT from OpenAI, Claude from Anthropic and Gemini from Google, alongside openly available models such as Meta's Llama and Mistral's.

How does an LLM work?

An LLM is built and used in three steps, and each one explains something about how it behaves.

Pretraining
The model learns language by predicting text. It reads a large corpus and adjusts its parameters every time it guesses the next token wrong. Almost everything it "knows" comes from here, which is why it knows nothing after its training cut-off unless it is given newer material.
Instruction tuning
The model learns to follow instructions. It is trained further on examples of requests and good answers, and on human judgements of which of two answers is better: the reinforcement learning from human feedback that Ouyang et al. described in 2022. This is the step that turns a text predictor into an assistant.
Use
The model answers from its context. At run time it sees your instruction plus whatever you put in front of it: a document, a customer record, the results of a search. What it is given at this step matters more than which model you picked.

The architecture under all of it is the transformer, introduced by Vaswani et al. in 2017. Its central idea, attention, lets the model weigh every word in a passage against every other when working out what one of them means, which is what lets it keep track of a long document rather than only the last few words. BERT, from Google in 2018, showed how well a transformer could understand text. The generative models that followed showed how well it could produce it.

Where do LLMs go wrong?

LLMs go wrong in predictable ways, and each has a known mitigation.

They make things up. A model produces fluent text whether or not it is true, a failure the research literature calls hallucination. The fix is to give the model the source material, ask it to answer from that, and check anything that matters.
They do not know recent or private facts. Pretraining stops at a date, and it never included your company's documents. The standard answer is retrieval-augmented generation, or RAG, described by Lewis et al. in 2020: fetch the relevant passages from your own data first, then have the model answer from them.
They sound certain when they should not. An LLM rarely says "I don't know" unprompted. Design the task so that it can, and route uncertain cases to a person.
They are weaker in smaller languages. Most training text is English. Norwegian works well in the large models, specialist vocabulary and dialect less so, which is why the National Library of Norway's AI-lab publishes Norwegian models such as NB-BERT and the speech model NB-Whisper.

How should a business start with LLMs?

Start with one bounded task where the model's output can be checked, not with a platform.

Pick a task with volume and a verifiable answer. Sorting incoming requests, drafting replies from a knowledge base, extracting fields from documents.
Give the model your own material. Most business value comes from retrieval over your own documents, not from the model's general knowledge.
Measure against today. Time per case, error rate, and how often a person has to correct the output.
Keep a person on the consequential step. Let the model draft, and let a human approve anything that is sent, paid or deleted.

When the model does more than answer, when it calls tools and completes the task itself, it is an agent. What that changes, and where it pays, is covered in How will agentic AI improve my business? and Where AI agents actually earn their keep.

Sources

Vaswani et al., Attention Is All You Need, 2017. arxiv.org
Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, 2018. arxiv.org
Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, 2020. arxiv.org
Ouyang et al., Training language models to follow instructions with human feedback, 2022. arxiv.org
Ji et al., Survey of Hallucination in Natural Language Generation, ACM Computing Surveys, 2023. arxiv.org
National Library of Norway, AI-lab, Models, read 23 September 2026. ai.nb.no

About this article. Written by Kjell Rune Monsø, co-founder and partner at Apps. First published in 2025 and rewritten in September 2026: the original named GPT-3 and LaMDA as current models, which they no longer are. The explanations now rest on the papers that introduced each technique rather than on product announcements, so they should age more slowly than the model names.

FAQ

What is the difference between NLP and an LLM?: NLP is the field, and an LLM is one kind of tool within it. Natural language processing covers every technique for getting computers to work with language, from spam filters to translation. A large language model is a general-purpose kind of model that now handles most of those tasks from a plain-language instruction.
Do we need to train our own language model?: Almost never. Training a large model from scratch costs far more than most companies would ever recover. The usual route is to use an existing model and give it your own material through retrieval, and to fine-tune only for a narrow, high-volume task with the data to support it.
Why does a language model make things up?: Because it is trained to produce likely text, not true text. When the answer is not in what it learned or what it was given, it still produces something fluent. Giving it the source material and asking it to answer from that reduces the problem a great deal; checking what matters removes the rest.
Do language models work in Norwegian?: Yes. The large models write and understand Norwegian well, though less reliably than English for specialist vocabulary and dialect. For tasks that depend on Norwegian specifically, such as transcribing speech, models trained on Norwegian data, like the National Library's NB-Whisper, are worth testing alongside the general ones.