myynti@metsosivut.fi

"*" indicates required fields

041 498 9805
Lets start conversation ➤
ai perustermit aloittelijalle

AI Glossary for Beginners

When talking about artificial intelligence, you quickly come across terms such as LLM, prompt, token, context, RAG, agent, and MCP. Some of these terms sound technical, even though the basic ideas behind them are often fairly simple.

This AI glossary explains some of the most important terms without unnecessary technical jargon. The goal is not to cover every detail, but to give you a clear understanding of what these terms mean and how they relate to one another.

Neural Network

A neural network is a machine learning model whose structure is loosely inspired by biological neural networks.

A neural network consists of many computational units, often called neurons. They process incoming values and pass the results to other parts of the network.

A single neuron performs a fairly simple calculation. When many of them are connected and the network is trained on a broad dataset, the system can learn to recognize highly complex patterns.

Neural networks are used for tasks such as image recognition, speech processing, prediction, and text generation.

Modern large language models are also based on neural networks.

Transformer

A transformer is a neural network architecture that most modern large language models are based on.

The transformer architecture was introduced in 2017, and one of its key ideas is attention.

Attention allows the model to determine which parts of a text are relevant to one another.

For example, in the sentence:

Matt put his phone on the table because he no longer needed it.

the model needs to understand what the words “he” and “it” refer to. Attention helps the model identify these kinds of relationships between different parts of the text.

The transformer architecture made it possible to train neural networks on extensive text datasets. It is one of the key technologies behind the development of modern language models.

LLM

LLM stands for Large Language Model.

An LLM is an AI model trained to process and generate language. Services such as ChatGPT, for example, use one or more language models behind the scenes.

A language model does not process text in the same way a person does. In simple terms, it predicts what would be a suitable continuation based on the text it has received.

If you write:

The capital of Finland is…

the model can determine that a highly likely continuation is Helsinki.

The same basic principle also works with much more demanding tasks.

Large language models can write text, summarize documents, translate languages, write code, analyze information, and have conversations with users.

An LLM is not the same thing as artificial intelligence as a whole. It is one type of AI model.

Token

A language model does not usually process text directly as complete words. Instead, the text is divided into smaller pieces called tokens.

A token can be a whole word, part of a word, a number, punctuation, or another piece of text.

For example, a long compound word may be divided into several tokens, while a common short word may consist of just one.

Tokens are important because the amount of text a language model can process is often measured in tokens.

API pricing for AI services is also often based on the number of tokens processed.

The more text you send to a model and the longer the response it generates, the more tokens are processed.

Prompt

A prompt is the input or instruction given to an AI.

A simple prompt might be a question:

What is the largest lake in Finland?

A prompt can also contain much more information:

Write a 500-word blog post about search engine optimization for business owners. Use clear language and include three practical examples.

A good prompt can describe the task, goal, target audience, background information, desired style, and response format.

A prompt is not a special programming language for AI. It is simply an instruction or input given to the AI.

The prompt alone does not determine the final result. The response can also be influenced by the language model being used, additional context, available tools, and system instructions.

Context

Context refers to the information an AI model has available when generating a response.

When you have a conversation with an AI, earlier messages can form part of that context.

For example, if you first write:

My company sells tractors in Jyväskylä.

and later ask:

Write an advertisement for my company.

the model can use the earlier message and understand that the advertisement should be related to tractors and probably to Jyväskylä as well.

Context can also include documents, website content, database information, and data retrieved from external systems.

Context Window

A language model has a limited context window.

This refers to the amount of information the model can process at one time. The context window can include user messages, earlier conversation, system instructions, documents, and information returned by tools.

The size of a context window is usually measured in tokens.

Hallucination

An AI hallucination is a situation where the model produces incorrect or invented information in a convincing way.

A language model might, for example, invent a study that does not exist, give the wrong year, or provide a source that is not real.

This happens because the basic task of a language model is to generate a likely response based on the context it has been given. It does not automatically verify every claim against a reliable source.

For this reason, information generated by AI should not automatically be treated as fact.

The risk of hallucinations can be reduced by giving the model reliable sources, using search, or retrieving information from a database before generating the response.

This brings us to RAG.

RAG

RAG stands for Retrieval-Augmented Generation.

The basic idea is simple: before generating a response, the system retrieves relevant information from an external source and provides it to the language model as context.

Imagine, for example, an AI assistant that answers questions from a company’s employees.

An employee asks:

How many days are we allowed to work remotely?

Instead of trying to answer based on its training data, the system can first search the company’s employee handbook for the section about remote work.

The relevant text is then given to the language model, which generates its answer based on that information.

RAG is particularly useful when an AI needs access to company-specific, current, or frequently changing information.

Embedding

An embedding is a numerical representation of information, usually in the form of a vector.

That sounds technical, but in practice embeddings allow computers to compare the meanings of different things.

For example, the words:

car

automobile

are closer in meaning than:

car

strawberry

An embedding model attempts to place things with similar meanings close to one another in a mathematical space.

This can be used for tasks such as semantic search.

If a user searches for:

car maintenance in winter

the system may also find a document discussing “winter vehicle maintenance,” even if it does not contain exactly the same words.

Embeddings are commonly used in RAG systems to find documents and passages that are relevant to a user’s question.

AI Agent

An AI agent is a system that does more than simply answer a single question. It can carry out a task through multiple steps.

For example, you might ask a standard language model:

What will the weather be like in Helsinki tomorrow?

Without access to current weather information, the language model may not be able to answer reliably.

An agent, however, can use a weather service, retrieve the forecast, and generate an answer based on the information it receives.

A more advanced agent can, for example:

The “brain” of an agent may be an LLM, but the agent as a whole usually contains other components as well.

This is why an LLM and an agent are not the same thing.

MCP

MCP stands for Model Context Protocol.

It is an open standard that allows AI applications to connect to external tools, data sources, and services.

An AI agent might, for example, need access to:

Without a shared standard, every AI application and external service would need to be connected separately using its own integration.

MCP aims to provide a common way for AI applications to discover and use the tools and data sources made available to them.

You can think of it somewhat like the USB standard. USB is not a computer or a device connected to one. It is a common way for different devices to communicate with each other.

Similarly, MCP is not a language model or an AI agent. It is a protocol that allows different systems to connect with one another.

MCP also does not mean that an AI automatically gets access to everything. Permissions and security still need to be configured separately in the application.

Fine-Tuning

Fine-tuning means taking an already trained AI model and training it further for a particular purpose.

Training a large language model from scratch is usually not practical. It requires extensive data and substantial computing power.

Instead, an existing model can be trained further using a smaller dataset.

Fine-tuning can be used to teach a model particular response styles, tasks, or patterns of behavior.

Fine-tuning and RAG solve somewhat different problems.

If the goal is to give a model access to company-specific information that changes regularly, RAG is often a natural solution.

If the goal is to change how the model behaves or performs a particular task, fine-tuning may be more suitable.

Inference

The life cycle of an AI model can be simplified into two stages: training and use.

When a model learns from data, this is called training.

When a trained model receives new input and generates a response or prediction based on it, this is called inference.

So when you send a prompt to an AI service and receive a response, inference is taking place.

The term is commonly used in technical descriptions of AI services and when discussing costs. Training a model can be very expensive, but continuously running a model also requires computing power.

Multimodal AI

Multimodal AI can process more than one type of information, or modality.

A text-only model processes text. A multimodal model, on the other hand, may be able to process:

For example, you can give an AI a picture of a broken device and ask what it can see. You can also provide a document containing both text and charts.

Multimodality makes AI useful for a much wider range of tasks than text processing alone.

Temperature

Temperature is a setting available in some language models and APIs that can influence the randomness and variation of generated responses.

In simple terms, a lower temperature usually leads to more predictable responses.

A higher value can introduce more variation and produce more creative or unexpected responses.

For example, if you ask the model ten times:

Come up with a name for a café.

the answers may vary more when a higher temperature is used.

The exact behavior depends on the model and service being used, and temperature is not adjustable by the user in every modern AI service.

Q-Learning

Q-learning is a machine learning method that belongs to reinforcement learning.

The method was introduced in Chris Watkins’ 1989 doctoral thesis, making it considerably older than today’s large language models.

In Q-learning, an agent learns by trying different actions in its environment and observing the consequences.

A simple example could be a character moving around in a computer game.

The agent might be able to choose between actions such as:

These actions lead to different situations and rewards. A good outcome may give the agent a positive reward, while a bad outcome may result in a negative one.

Over time, the agent learns which actions are useful in different situations.

This is where the method gets its name. Q-learning learns so-called Q-values, which describe the expected value of taking a particular action in a particular situation.

Q-learning is a good example of reinforcement learning.

However, it is important to distinguish it from modern LLM-based agents. An AI agent does not automatically use Q-learning, and ordinary LLM agents generally do not learn from every task they perform through Q-learning.

How Do AI Terms Relate to Each Other?

The relationship between these terms is easier to understand through a simple example.

When you type a question into an AI service, the text you provide is the prompt.

The text is converted into tokens, which the model processes.

The prompt, earlier conversation, and other available information form the model’s context.

The response may be generated by an LLM, which is based on a neural network and usually uses a transformer architecture.

When the trained model generates a response, inference is taking place.

If the system needs to retrieve relevant documents before answering, it can use RAG. Embeddings can help it find the most relevant documents and passages.

If the system needs to do more than answer a question and must also perform tasks or use external tools, it can be described as an AI agent.

The connection between an agent and external tools can be implemented using technologies such as MCP.

If the behavior of a model needs to be adjusted through additional training, fine-tuning can be used.

If a model can process images and audio in addition to text, it is a multimodal AI model.

And although modern language models may feel like a new technology, many important ideas in artificial intelligence are much older. Q-learning, for example, was developed in the late 1980s.

AI terminology changes quickly, and you do not need to understand every technical detail. Once the key concepts are familiar, however, it becomes much easier to understand how new AI services work, what they can do, and where their limitations lie.