Basic AI Concepts – Terms Related to Artificial Intelligence. Part One.
You can treat the article as a map of AI concepts—a reference point worth returning to while learning from other materials on artificial intelligence.
The purpose of this article is to organize knowledge about AI in a simple and practical way. The text is intended to support learning from other sources: it presents basic concepts, how models work, and the most important areas of application.
.
We begin by explaining what artificial intelligence is and how models, neural networks, and phenomena such as hallucinations work. Next, we look at large and small language models, image and speech recognition, recommendation systems, and content generation. We then discuss how models are trained and used, what technological infrastructure they require, and where AI hubs are developing in Poland and around the world.
There are a great many terms related to artificial intelligence. In this text, I focus on the basics; individual topics will be described in separate articles and appropriately linked. Additionally, you can find a more extensive list of AI-related terms here.
Table of contents
Fundamentals of the Fundamentals
What Is Artificial Intelligence
Artificial intelligence, more broadly known as Artificial Intelligence (AI), is a collection of methods that enable machines to perform tasks previously associated with human intelligence. These include understanding language, recognizing images, and planning. Since there is no single definition of human intelligence, the description of AI is general in nature and serves mainly as context for understanding the phenomenon.
Machine Learning
A fundamental subset of AI is Machine Learning (ML). In this approach, the model is not programmed step by step; instead, it learns patterns directly from data. As a result, it can respond better to new situations.
Deep Learning
A special type of ML is Deep Learning (DL), based on multilayer neural networks. Such models scale exceptionally well to massive datasets and underpin modern solutions such as speech and image recognition systems.
The Model and How It Works
A machine learning model is a mathematical function with a set of parameters. In the case of neural networks, these are weights and biases. A weight determines the strength of the connection between neurons, while a bias introduces an offset that allows the model to fit the data better. These parameters determine how effectively the model recognizes patterns and what predictions it can generate.
Hallucinations in AI
In the context of language models, people often talk about hallucinations. These are situations in which a model generates answers that sound credible and confident but diverge from the facts. This results from the probabilistic nature of how it operates: the model predicts subsequent words based on the patterns it knows and strives to “complete” the statement. If the training data was incomplete or the instruction is ambiguous, it is easy for incorrect content to arise.
Main AI Topics
LLM – Large Language Models (large language models)
LLMs are artificial intelligence models trained on vast collections of text that can understand and generate natural language. They make it possible to answer questions, translate, summarize, and create new content. In practice, they support education, business, programming, and communication. Examples include Claude, ChatGPT, and Grok.
Small Language Models
They are simplified versions of large models that operate faster and require fewer resources. Although they have a limited ability to understand and generate text, they are highly effective for specialized tasks—for example, on mobile devices or in business applications. Importantly, they can even be run on an ordinary laptop or smartwatch. A few years ago, they represented a significant stage in development; today, their role has been partially taken over by newer, more efficient LLMs.
Computer vision (image recognition)
This is an area of AI that enables machines to analyze images and video. Systems of this kind recognize faces, objects, text, or entire scenes, as well as track movement and classify images. They have found applications in medicine (for example, analyzing X-rays), the automotive industry (autonomous cars), and everyday applications such as unlocking a phone with facial recognition. The development of this field proved simpler than creating large language models, and it paved the way for many other AI applications.
AI You Don’t See
ChatGPT was the first widely available AI system that enabled ordinary users to make practical use of the capabilities of artificial intelligence. As a result, anyone could experience something that had previously been the domain of experts. It is worth remembering, however, that alongside visible applications, there is an entire layer of “invisible” AI—operating behind search engines, recommendation systems, and business tools—that shapes our everyday experiences.
Speech Recognition (speech recognition)
This is technology that converts speech into text. It enables voice communication with assistants (Siri, Alexa), automatic subtitles in videos, and voice command functionality in devices.
Recommendation systems
These are algorithms that analyze data about users and their choices. Based on this, they suggest content tailored to their preferences: movies on Netflix, products in an online store, or music on Spotify.
Image Generation
This is a field of AI in which models create graphics based on a textual description or other data. This makes it possible to create illustrations, photographs, and graphic designs without human involvement. This technology has applications in art, marketing, and design.
AI performing specific tasks, simulations, and playing games
Artificial intelligence can be trained for highly practical purposes, such as controlling robots, but also for entertainment—for example, controlling characters in games or simulating various scenarios.
How (language models) work
Even a superficial understanding of how language models work makes it easier to use them. Building a model from scratch is very difficult (more difficult than image recognition), but adapting one through fine-tuning, RAG, or other techniques is already within the reach of many people.
Image generation is a separate field of AI in which models create graphics based on a textual description or other data. This makes it possible to obtain illustrations, photographs, or graphic designs without the involvement of a human artist. This technology has found applications in art, marketing, and design.
Transformer
A Transformer is a general, abstract term for a method of processing data based on layers, the Attention mechanism, and Feed-Forward networks. It is not a physical element like a transistor or logic gate. It can be understood as a language model architecture rather than a single component. It was this concept that led to the creation of systems such as ChatGPT.
Layer
A layer is a computational block that transforms an input tensor into an output tensor. Lower layers capture relationships between words, while higher layers capture relationships between sentences or paragraphs. In Transformer models, we encounter various layers, for example those responsible for the attention mechanism, vector transformations, or value normalization. Each operates like a mathematical function: it takes data as input and returns a result.
Token (Token)
Tokens are the smallest units of text processed by the model. They can be individual characters, parts of words (subwords, syllables), and sometimes entire words.
Embedding Matrix
Language models do not operate on ordinary numbers, but on multidimensional vectors. For example, the number assigned to the word “cat” (e.g., 5432) is converted into a vector [0.12, -0.87, 0.33, …]. This mapping is performed by an embedding matrix, which assigns an appropriate vector to each token.
Parameters
Parameters are numbers that control a model’s behavior. In neural networks, these are weights and biases that determine how the model learns to recognize patterns and which signals it considers important.
Tensor
A tensor in artificial intelligence is a multidimensional array of numbers. It generalizes mathematical objects such as matrices and vectors and serves as the fundamental way of storing data in models.
Neural Network
A neural network is a system modeled on how the brain works. It consists of layers of interconnected “neurons” that process data step by step. This enables the network to recognize patterns and perform complex tasks, such as understanding language, analyzing images, or making decisions.
Neuron
A neuron in artificial intelligence is a basic computational unit. It receives input data, multiplies it by assigned weights, adds a bias, and then passes the result through an activation function. In practice, this means a simple mathematical equation that determines whether the signal is passed on.
.
Temperature
Temperature is a parameter in language models that controls the randomness of responses. A higher temperature increases the diversity and creativity of generated content, while a lower one makes responses more predictable and conservative. At the end of the model’s operation, numerical values (logits) are transformed into a probability distribution using the softmax function, which amplifies the differences between the results.
.
AI Concepts Related to Image Recognition)
Unlike language models, most people use image recognition functionality entirely passively, treating the model as a black box on an “either it works or it doesn’t” basis.
Image recognition is simpler than text generation, and anyone with a little time can understand how it works. However, it requires basic programming skills.
Understanding how image recognition works is useful not only as an exercise that helps you better understand artificial intelligence. In many situations, AI can also be used for quality control.
Convolution (convolution)
A fundamental operation in convolutional networks. It involves sliding a filter (a small matrix) over an image and calculating values. This makes it possible to detect edges, shapes, and other visual features.
Filter / kernel
A matrix of numbers used in convolution. It determines which image features are enhanced and which are suppressed (for example, horizontal or vertical edges or textures).
Pooling
An operation that reduces the resolution of an image in a network. “Max pooling” is most commonly used; it selects the largest value from a small section of the image, preserving key information and reducing the amount of data.
Feature Map
The result of applying filters to the image. Each feature map shows which parts of the image were activated by a given filter—for example, areas with edges or a repeating pattern.
AI Concepts Related to Training Artificial Intelligence.
Training artificial intelligence consumes the most energy and computing power. Once trained, the model itself no longer requires such extensive resources, but creating connection matrices between words is a costly and time-consuming process. Knowledge of how to collect and clean data so that it can be useful for training models is highly practical and often more important than knowledge of training techniques themselves.
Fine-tuning
Further training of the model on a smaller, specialized dataset to improve its performance in a specific application, such as medicine or law.
Training Corpus
A vast collection of texts from which the model learns. These may include books, articles, websites, or technical documents. The quality and diversity of the corpus directly translate into the model’s capabilities.
Pre-training (pretraining)
The first phase of training a model on very large and general datasets. It enables the model to learn the basic rules of language before being adapted to specific tasks.
Instruction Tuning
The process of fine-tuning a model using sets of instructions and question-and-answer examples so that it responds more in line with users’ expectations.
RLHF (Reinforcement Learning from Human Feedback)
A method in which people evaluate the model’s responses, and their evaluations are used for further training. This enables the model to respond better in line with users’ preferences.
Overfitting (overtraining)
A situation in which a model memorizes training data instead of learning to generalize from it. This results in very good performance on known examples but poor performance on new data.
Batch and epoch
A batch is a portion of data on which the model learns in a single step. An epoch means one pass through the entire training set. The number of epochs and the batch size have a significant impact on the quality of training.
Learning rate
A parameter determining how large a step the model takes when updating its weights during training. Too high causes chaos, while too low significantly slows down learning.
Regularisation
A set of techniques (e.g., dropout, weight constraints) that prevent overfitting and improve the model’s ability to generalize.
Training Tokenizer
An algorithm that divides text into the smallest units (tokens) used by the model. When training, it must be designed carefully and adapted to the language and data the model will work with.
Using AI
Concepts related to the use, testing, and evaluation/assessment of artificial intelligence.
Command (Prompt)
A prompt is an instruction or question that you give to a model. The response generated by artificial intelligence depends on its content and precision. Current language models are good enough that general skills for communicating with AI are usually more useful.
AI API
An interface that enables the integration of artificial intelligence models with applications and systems. This allows AI to operate as part of a larger solution.
Prompt Engineering
It is the ability to formulate prompts so that the model provides more accurate, more useful answers. In practice, it involves testing different variants and choosing the best one.
Prompt engineering is particularly important in programs that do not have a highly developed language module, such as advanced image generators.
Jailbreaking
Bypassing safeguards and censorship in language models.
Zero-shot learning
A situation in which a model solves a task it has not encountered before, based solely on the description in the prompt.
Few-shot learning
Here, the user provides several examples of a solution in the prompt, and the model learns the pattern and responds better to subsequent questions.
Chain-of-thought
This is a method in which the model first lays out its reasoning step by step and only then provides the final answer. This makes it more accurate when dealing with complex problems.
Context window
This is the amount of text (measured in tokens) that the model takes into account in a given interaction. If you exceed the limit, the model “forgets” the earlier context.
Hallucination
This is what the situation is called when a model invents facts and presents them as if they were true, even though they do not exist in the data.
RAG (Retrieval-Augmented Generation)
A technique in which a language model combines text generation with searching a knowledge base or documents. This makes the responses more up-to-date and reliable.
Latency
This is the time that elapses from sending a query to receiving a response from the model. Low latency means fast operation.
Data drift
Data drift is a phenomenon in which the data fed into a model over time differs from the data on which the model was trained. This causes a decline in prediction quality because the model no longer accurately reflects reality.
AI Ecosystem
Large AI factories have energy demands comparable to those of major cities. Talking about artificial intelligence is not only about software, but also about the energy and scientific infrastructure behind it, etc.
.
GPU Servers
Specialized machines equipped with graphics cards that significantly accelerate the computations required to train and run artificial intelligence models.
Cloud Computing
Services such as AWS, Azure, and Google Cloud enable the use of computing power and storage without the need to invest in costly on-premises infrastructure.
Edge computing
Processing data closer to where it is generated, for example on a smartphone, camera, or other endpoint device. This makes it possible to reduce latency and the costs associated with transmitting data to the cloud.
Data center
Data centers housing servers and the infrastructure supporting the storage and processing of the vast datasets essential for the operation of AI systems.
AI Frameworks
Programming libraries such as TensorFlow and PyTorch enable the structured and repeatable creation, training, and optimization of artificial intelligence models.
Databases
Systems designed for storing and quickly providing access to data used both for training and serving AI models. Examples include PostgreSQL, MongoDB, and specialized vector databases such as Pinecone.
MLOps
A collection of practices and tools supporting the full lifecycle of artificial intelligence models: from training, through deployment, to monitoring and updates in a production environment.
Containerization
Technologies such as Docker and Kubernetes make it possible to run AI models consistently, regardless of the hardware or operating system environment.
SDK Software Development Kit
An SDK (Software Development Kit) is a set of tools for developers: libraries, documentation, examples, and tools needed to create applications for a given platform.
Examples: Android SDK for writing mobile applications, Java SDK (JDK) for Java programs, OpenAI SDK for using AI models in applications.
Monitoring and Observability
Systems used to monitor model performance, including accuracy, errors, and costs. They also support the detection of data drift and potential misuse.
ANI (Artificial Narrow Intelligence)
This is known as narrow artificial intelligence. It operates within a defined scope and performs specific tasks, such as image recognition, text translation, or e-commerce recommendations. It is effective but cannot go beyond its designated field. Additionally, ANI can be divided into, for example, schematic and complex ANI. This distinction is not widely accepted, but it is very practical. Schematic ANI is, for example, a face-swapping filter, while complex ANI is an image generator in which we can create different images.
AGI (Artificial General Intelligence)
This is artificial general intelligence. In theory, it matches humans in the ability to learn, understand, and solve problems across many different fields. For now, it remains a concept, and research into it is still at an early stage.
ASI (Artificial Superintelligence)
This is a hypothetical level of artificial intelligence surpassing human capabilities in all fields. ASI would not only perform tasks faster and more accurately, but would also create new ideas, technologies, and strategies on its own. For now, it is solely a concept considered in theories and forecasts concerning the future of AI.
Singularity (technological singularity)
This term describes a hypothetical moment when artificial intelligence reaches a level surpassing the human mind and begins to improve itself autonomously. According to this vision, technological development will then become exponential, while social, economic, and cultural changes will be difficult to predict. Singularity is often associated with the concept of ASI and inspires both hopes and concerns.
AI Hubs
AI Hubs are cities and regions where companies, universities, and institutions jointly advance artificial intelligence through talent, investment, and infrastructure.
In Poland, they include: Warsaw, Kraków, Wrocław, Poznań, the Tricity, and Silesia.
In Europe, the key cities are: London, Berlin, Paris, Amsterdam, Zurich, and Helsinki – with London standing out as a global AI hub bringing together science, business, and finance.
The Impact of AI on Society
Work Automation
AI is replacing some tasks performed by people, which increases efficiency but raises concerns about job losses.
Digital inequalities
Access to AI is not equal—larger companies and developed countries are gaining an advantage over smaller and poorer ones more quickly.
Disinformation
Generative models can create false content (deepfakes, fake news), influencing public opinion and politics.
Privacy
AI analyzes vast amounts of personal data, which entails the risk of abuse and privacy breaches. A certain degree of caution should be exercised when using AI models that are not hosted on your own hardware.
Algorithm Ethics
Decisions made by AI may be biased if the model was trained on incomplete or prejudiced data.
Impact on Education
AI is changing the way we learn and work with knowledge—it makes information more accessible, but it can also weaken critical thinking.
New Work Models
The development of AI is contributing to the emergence of new professions and specializations (e.g., prompt engineer, AI auditor).
Public trust
The way AI is implemented affects whether people trust the technologies and institutions that use them.
Practical Things and Summary
This article is intended for people who already have a general understanding of how artificial intelligence works. If you are encountering these concepts for the first time, the texts below may be useful to you.
Useful AI Programs
Alongside popular AI tools such as ChatGPT and Midjourney, there are many lesser-known but highly practical solutions.
An example of a lesser-known yet highly practical tool is Notebook LM by Google. It is a kind of intelligent research and note-taking assistant that works based on your own documents. You can upload files to it (such as PDFs or Google Docs) and then ask questions about their content, receive summaries, and even generate audio overviews. What’s more, Notebook LM can create quizzes, flashcards, and interactive educational materials, making it an excellent tool both for academic work and for organizing knowledge in everyday life.
Another example is Gamma App, a presentation-generation tool that combines artificial intelligence with a simple interface. Just enter a topic, and the application creates a coherent slide structure while ensuring a clear layout and aesthetic appeal.
ElevenReader from Eleven Labs is a brilliant tool that lets you listen to documents (and the audio quality is much better than in ChatGPT or Gemini)
Stages of Implementing AI in a Company
Understanding the stages of implementing AI, makes it easier to develop an action plan. Of course, this may look different in each organization, but the following stages often occur.
- Security fundamentals – establishing rules about what must not be done (e.g. not sharing confidential data with public models, verifying AI-generated content, labeling materials).
- Learning and experimentation – testing tools, learning about prompting, sharing experiences, and making initial practical attempts.
- Tool procurement – selecting appropriate licenses and team tools, and introducing basic security procedures.
- Simple internal solutions – automated note-taking, searching for procedures, implementation support, and local models for working with sensitive data.
- Advanced implementations – creating proprietary systems, customer solutions, and large AI-based projects once the team has gained sufficient experience.