Artificial Intelligence Glossary

Lesson 1 – Intro to AI: General terms

Artificial Intelligence (AI)

A field in computer science that deals with creating systems that perform tasks that usually require human intelligence: learning, understanding, problem-solving, and decision-making.

Algorithm

A defined sequence of steps for solving a problem or performing a task. In AI: the rules and computations the model uses.

Model

A mathematical/computational representation created during training and used for predictions or inference on new data.

Training

The process in which an algorithm learns from data: it adjusts parameters so that the model improves performance on a given task.

Inference

Using a trained model to give a prediction or answer for new input, without updating the parameters.

Data

The information the system learns from or processes – text, images, numbers, user input, etc.

Parameter

A value the model learns during training (e.g. weights in a neural network). Large models contain billions of parameters.

Narrow AI

AI systems that excel at one task or a limited domain (face recognition, translation, games) – as opposed to AGI.

Artificial General Intelligence (AGI)

Human-level intelligence – the ability to understand, learn, and perform across a wide range of tasks and domains like a human.

Artificial Superintelligence (ASI)

Theoretical: a system smarter than humanity in every relevant domain. Does not exist today.

Automation

Performing tasks without human intervention. AI is used to automate repetitive processes and create content.

Lesson 2 – History of AI

The timeline, dates, and historical concepts (Turing test, Dartmouth conference, Turing machine, etc.) appear on the History of Artificial Intelligence page.

Lesson 3 – Machine Learning

Machine Learning (ML)

A branch of AI in which computers learn from examples and data and improve performance without explicit programming for each case.

Supervised Learning

Training on examples labeled with a tag: the model learns to associate input with desired output.

Unsupervised Learning

Training on unlabeled data – the goal is to find patterns, clusters, or structure in the data.

Reinforcement Learning (RL)

Learning through interaction with an environment: the agent receives rewards and improves its policy.

Label

The correct answer or classification attached to an example in supervised training (e.g. "cat" for an image).

Feature

A characteristic or measure extracted from the data and used as input to the model (color, size, word frequency, etc.).

Regression

The task of predicting a continuous numerical value (price, temperature, time) based on input features.

Classification

The task of assigning input to one of a fixed set of categories (spam/not spam, type of animal, etc.).

Clustering

Grouping items into similar groups without predefined labels – a typical method in unsupervised learning.

Overfitting

When the model "memorizes" the training data and generalizes less well to new data.

Underfitting

When the model is too simple and does not capture the patterns in the data well.

Cross-Validation

An evaluation method: split the data into parts, train on some and test on the rest, and average the results.

Training Data

The examples used to update the model’s parameters during training.

Test Data

Data held aside and not used in training, used to evaluate model performance.

Learning Rate

The step size when updating parameters in optimization. Too large – instability; too small – slow training.

Batch

A subset of examples processed together in one parameter update during training.

Epoch

One pass over all training data during training. Training usually includes several epochs.

Loss Function

A measure of how far the prediction is from the correct value. Optimization tries to minimize the loss.

Backpropagation

An algorithm for computing the gradient of the loss with respect to parameters, so they can be updated by gradient descent.

Gradient

The direction of greatest change of a function. In training – the direction in which parameters are updated to reduce loss.

Gradient Descent

An optimization method: update parameters in the direction opposite to the gradient to reduce the loss.

Lesson 3 – Neural networks and deep learning

Neural Network

A structure of nodes (neurons) and layers that compute complex functions. The basis for most Deep Learning models.

Deep Learning

Machine learning using neural networks with multiple layers; enables complex representations and advanced tasks.

Convolutional Neural Network (CNN)

A network especially suited to images: uses convolutional layers to detect local patterns.

Recurrent Neural Network (RNN)

A network that handles sequences (text, time) using internal memory; LSTM and GRU are common variants.

Transformer

An architecture based on the attention mechanism, without recurrence. The basis for GPT, BERT, and most large language models.

Attention Mechanism

Allows the model to "look" at relevant parts of the input (e.g. previous words) when producing output.

Activation Function

A non-linear function after a neuron layer (ReLU, Sigmoid, Softmax) that allows the network to learn complex patterns.

Dropout

A regularization technique: randomly "turn off" some neurons during training to reduce overfitting.

Normalization

Standardization of intermediate layer values (e.g. Layer Norm) for training stability and faster convergence.

Diffusion Model

A model for generating images (and video/audio): learns to remove noise gradually and build an image from scratch.

Generative Adversarial Network (GAN)

Two networks: a generator creates examples, and a discriminator tries to tell real from fake. They compete until quality improves.

Variational Autoencoder (VAE)

A generative model that takes input, compresses it to a "latent space," then reconstructs; good for both generation and compression.

Lessons 4–5 – Language, text, and language models

Natural Language Processing (NLP)

A branch of AI that deals with understanding, analyzing, and generating text and human language.

Natural Language Understanding (NLU)

A subfield of NLP: interpreting intent, sentiment, entities, and structure in written or spoken language.

Natural Language Generation (NLG)

Generating new text by a system – answers, summaries, translations, marketing, and more.

Large Language Model (LLM)

A very large text model (e.g. GPT-5, Claude, Gemini) trained on huge amounts of text, capable of completion, Q&A, and generation.

GPT – Generative Pre-trained Transformer

The name of OpenAI’s model series (GPT-3, GPT-4, GPT-5, GPT-5.2) – large transformer-based language models. As of 2026 the flagship is GPT-5.2 with built-in reasoning.

Token

The basic unit of text the model works with – a word, subword, or character. Tokenization splits text into tokens.

Tokenization

The process of splitting text into units (tokens) that the model receives as input.

Context Window

The maximum number of tokens the model can attend to – longer text is truncated or summarized.

Embedding

A vector representation (array of numbers) of a word, sentence, or item – so that similar meaning is close in space.

Vector

An ordered list of numbers. In AI used to represent embeddings, features, and internal state of models.

Cosine Similarity

A measure of similarity between two vectors by the angle between them; common for comparing embeddings and semantic search.

Few-Shot Learning

Showing a few examples in the prompt so the model adopts a style or format without further training.

Zero-Shot

Performing a task with no examples in the prompt – only an explanation or instruction. Large language models can do this.

Transfer Learning

Using a model trained on one task as a starting point for training or adapting to another task, usually with less data.

Fine-tuning

Further training on an existing model with specific data, usually with a low learning rate, to adapt it to a task or domain.

Base Model

A general-purpose model trained on broad data before fine-tuning or direct use via prompts.

Sentiment Analysis

Identifying the tone or emotion in text – positive, negative, neutral – often as classification or scoring.

Lesson 4 – AI in daily life: Images, video, and vision

Computer Vision

A field in AI that deals with understanding images and video: object recognition, segmentation, description, OCR, and more.

Image Classification

Assigning an image to one category from a given set (e.g. "dog," "car").

Object Detection

Locating objects in an image and returning position (bounding boxes) and class for each object.

Segmentation

Dividing the image into regions – each pixel or region is assigned a class (person, background, vehicle, etc.).

Text-to-Image

Generating an image from a text description; examples: DALL-E, Midjourney, Stable Diffusion.

Video Generation

Creating videos by AI from text, images, or both – e.g. Sora, Runway.

Speech-to-Text (STT)

Converting speech audio to written text.

Text-to-Speech (TTS)

Converting text to human-like voice – narration, voice assistants, accessibility.

Multimodal

A model or system that handles multiple types of input/output: text, images, audio, video.

Modality

The type of information the system processes – text, image, sound, video – each is a different modality.

Lessons 5–6 – Prompts and tools

Prompt Engineering

The art and skill of formulating instructions (prompts) for AI to get accurate and relevant answers.

Prompt

The text the user supplies to the model – a question, instruction, or context – on which the model bases its response.

System Prompt

An internal instruction that sets the model’s "role" or style (e.g. "You are a professional assistant"). Usually not shown to the user.

Chain of Thought (CoT)

Asking the model to explain its reasoning steps before the final answer – improves accuracy on complex tasks.

Temperature

A parameter in language models: high – more varied/creative answers; low – more stable and "safe."

Top-p / Nucleus Sampling

A sampling method: choose from words whose cumulative probability is up to p. Affects diversity of responses.

Chatbot

An automated conversation system with users – usually based on a language model, e.g. ChatGPT.

AI Agent

A system that receives a goal, plans steps, uses tools (search, code, calculator), and carries out complex tasks.

Prompt Injection

Exploiting a malicious prompt that tries to make the model ignore instructions or perform unwanted actions.

Retrieval-Augmented Generation (RAG)

Combining search over a repository (documents, database) with a language model – retrieve relevant passages and add to the prompt to improve accuracy and freshness.

Knowledge Base

A structured repository of information (documents, Q&A, facts) that AI systems can search or rely on.

Knowledge Cutoff

The date until which the model’s training or knowledge was updated – after that the model may not "know" new events.

Hallucination

When the model produces information that sounds plausible but is incorrect or non-existent – a key issue when using language models.

Lesson 7 – Ethics, safety, and impact

Bias

Systematic tendency in a model or in data that leads to unfair differences in treatment (gender, race, age, etc.).

Transparency

The degree to which we know how the system works, what data it uses, and how decisions are made.

Explainability

The ability to explain why the model reached a particular conclusion or decision.

Accountability

Defining who is responsible when an AI system errs or causes harm – developers, operators, regulators.

AI Safety

Research and practice aimed at ensuring AI behaves as intended and reduces unintended risks and harm.

Alignment

Aligning the system’s goals with human values and desires – preventing dangerous or unwanted behavior.

Deepfake

Content (video/voice/image) created or altered by AI so that it appears authentic – carries risks of fraud and deception.

AI-Generated Content (AIGC)

Any content – text, image, video, voice – created with AI tools.

AI Detection

Attempts to detect whether text or content was generated by AI – tools and ideas exist but are not always reliable.

Watermarking

Embedding a hidden or visible marker that flags or identifies AI-generated content.

Technological Singularity

A hypothetical scenario in which AI develops so rapidly that its implications for humanity become unpredictable.

Autonomy

The degree to which a system acts independently – decisions and actions without human intervention at every step.

Lesson 8 – Future and more techniques

Prediction

An estimate or forecast the model produces for new input – a value, class, text sequence, etc.

Data Leakage

When information from test data or from the future "leaks" into the training phase and gives an unrealistic estimate of performance.

Validation Data

Data held aside during training for choosing models, parameters, or stopping – not used directly for the final parameter update.

Optimization

The process of searching for parameters or decisions that achieve an optimal value (e.g. minimum loss) – in training usually gradient-based.

Vector Space

A space in which each item is represented as a vector; used for comparisons, search, and embeddings.

Role Playing

In prompts: instructing the model to act as a certain "character" or expert (doctor, lawyer, teacher).

Constraints

In instructions: limits on the answer – length, language, style, what must not be mentioned, etc.

Token Cost

The charge or usage units based on the number of tokens sent and/or generated in a model API.

Stop Sequence

A string that, when the model generates it, stops generation – e.g. end of paragraph or a specific format.

Content Filtering

Detecting and blocking or modifying harmful, illegal, or inappropriate content before showing it to the user.

API – Application Programming Interface

An interface that allows other programs to send requests to the model (e.g. chat, completion) and receive responses.

Open-Source Model

A model whose code and/or weights are available to the public for use, research, and adaptation.

Proprietary Model

A model owned by a company and accessible only through services and APIs the company provides.

Compute

Computing resources (processing units, memory, time) required to train or run models.

Intent

In dialogue systems: the intention or goal the user is trying to achieve (order, fix, learn, etc.).

Real-Time

Processing that responds very shortly after data is received – important in chat, speech recognition, autonomous vehicles.