What is a large language model?

A large language model is a transformer network with billions of parameters trained to predict the next token across an enormous body of text. The word large is doing real work: the same architecture at a hundred times the scale acquired capabilities that were absent at smaller sizes, including following instructions and performing tasks it was never explicitly trained on.

How is an LLM trained?

In three stages. Pretraining predicts the next token across trillions of tokens of text, which is where nearly all the knowledge and capability comes from. Instruction tuning then trains on examples of requests and good responses, turning a text continuer into something that answers questions. Preference alignment finally uses human comparisons between candidate responses to shape tone, helpfulness and refusal behaviour.

What is RAG and when do I need it?

Retrieval-augmented generation searches your own documents for passages relevant to a question and places them in the model's context before it answers. You need it whenever answers must reflect information the model was not trained on: internal documentation, current data, or anything that changes. It reduces fabrication and makes answers citable, and it cannot fix a question your documents do not answer.

Should I fine-tune or use RAG?

Fine-tuning teaches form: a house style, an output format, a domain vocabulary, a consistent tone. Retrieval supplies facts. Teams routinely reach for fine-tuning when they need facts and are disappointed, because fine-tuning on documents does not reliably make a model able to quote them. Start with prompting, add retrieval, and fine-tune only when the remaining gap is about style or format.

An agent is a language model given tools it can call — search, code execution, a database, an API — and a loop in which it decides what to call, sees the result, and decides again. This lets it perform multi-step work rather than only answering, and reliability drops with the number of steps because errors compound, which is why narrow agents with few tools work far better than open-ended ones.

What is prompt injection?

A model cannot distinguish instructions you wrote from instructions embedded in the content it reads. Text in a web page, a document or an email can therefore redirect its behaviour, and if the model has tools that send email or query a database, that becomes a security issue rather than a curiosity. There is no complete fix, so systems must be designed so a compromised model cannot do serious damage.

Large Language Models: what they are and how to use them properly

An LLM is a next-token predictor, as lesson 3 of the generative AI course explained. That sentence is accurate and explains almost nothing about why one can pass a bar exam and still fail to count letters in a word.

What this course sets out to do: give you an accurate mental model of what these systems are, what the three training stages did to them, and what the surrounding architecture — retrieval, tools, agents — is actually for.

What this course does not do: teach you to build production LLM systems. The premium catalogue covers RAG, fine-tuning, LangChain and agents hands-on.

What you are about to discover

Course contents

#	Lesson	Main goal	Time
1	What makes a model large	Scale, parameters, and what emerged from it	9 min
2	The three training stages	Pretraining, instruction tuning, preference alignment	9 min
3	Retrieval: giving it your knowledge	What RAG fixes, how it works, where it fails	9 min
4	Tools and agents	Function calling, loops, and why reliability decays	9 min
5	Choosing, costing and securing	Model selection, token economics, evaluation, prompt injection	9 min
6	Recap and FAQ	Synthesis, a decision guide, and 12 common questions	6 min
7	Quiz and attestation	Validate what you learned with 5 corrected questions	3 min

Is this course for you?

You are building something with an LLM and want to know which architecture the problem needs.
You are evaluating vendors and want to distinguish substance from packaging.
You keep hearing "RAG", "agent" and "fine-tune" used loosely and want them precise.
You need to explain to a security or legal colleague what the actual risks are.

Recommended first: Generative AI and NLP.

What you will be able to do at the end

Explain what scale changed and what emergence does and does not mean.
Describe the three training stages and which one is responsible for which behaviour.
Decide correctly between prompting, retrieval and fine-tuning for a given problem.
Explain how tool use works and why long agent chains degrade.
Estimate token cost and pick the smallest model that passes your tests.
Design around prompt injection rather than assuming it can be filtered out.

Estimated time

Around 50 to 60 minutes of reading.

Prerequisites and next steps

Prerequisites: Generative AI. Deep Learning helps for lesson 1.

Natural continuation:

MLOps — running and monitoring this in production
Cloud AI — where these models are hosted and what it costs
Ethics of AI — accountability when the output affects someone

Frequent questions, answered in one line

Does an LLM understand what it says?

It has learned statistical structure deep enough to translate, summarise, write working code and follow multi-step instructions, and it holds no beliefs, intentions or model of the world. The argument about whether that constitutes understanding is largely about the word; the operationally important consequence is that it cannot notice when its own answer is impossible.

Are open models good enough to use instead of hosted ones?

For a great many tasks, yes, and the gap has narrowed considerably. Hosted frontier models remain ahead on the hardest reasoning tasks. Open models win on data control, cost at volume and freedom from a vendor's roadmap, which for regulated environments is often decisive.

Will a bigger model fix my problem?

If the problem is difficult reasoning, sometimes. If it is missing information, no — supply the information. If it is fabrication, no. Reaching for a bigger model is the most common way to spend money without fixing anything.

Can I run one locally?

Yes. Quantised models of several billion parameters run acceptably on a modern laptop, and larger ones on a consumer graphics card. Quality is below the frontier and often sufficient, and nothing leaves your machine.

Want to build LLM systems rather than read about them?

The premium catalogue covers RAG, fine-tuning, LangChain, agents and local deployment, with a verifiable certificate after a 40-question examination. Included in every paid plan.

Ready? Start with lesson 1 →

What you are about to discover​

Course contents​

Is this course for you?​

What you will be able to do at the end​

Estimated time​

Prerequisites and next steps​

Frequent questions, answered in one line​