Cyn’s AI Glossary
Common AI Terms Explained (Foundations-First)

Understanding AI gets much easier once the key terms make sense. This page is a foundations-first glossary used across AI with Cyn — written in plain English, grounded in real-world AI use, and focused on clarity over hype.

Use this reference alongside our FAQs, Local AI in Practice pages, and AI Regulation explainers whenever you see a term you’re not sure about. Where a topic is more advanced, we label it clearly and link to deeper explanations.

Beginner tip: If a term is marked (Advanced), you can skip it for now and come back later.

AI - Artificial Intelligence


The simulation of human intelligence processes by machines, enabling them to perform tasks such as learning, reasoning, and problem-solving.  Discover the 5 different levels of intelligence of AI by CLICKING HERE and the different Framework of intelligence of AI by CLICKING HERE

AI Agent

An AI system that can take actions or complete multi-step tasks using tools or external data, rather than only generating text responses.

AI Ethics

Principles guiding the responsible development and application of AI technologies to prevent harm and ensure fairness, accountability, and transparency in AI systems.

Algorithm

A set of rules or instructions given to an AI system to help it perform specific tasks or solve problems.

Automatic Speech Recognition (ASR)

Technology that converts spoken language into written text for transcription and voice-based applications.

Bias

The tendency of AI systems to produce unfair or prejudiced outcomes due to training data that may reflect societal biases.

Chatbot

An AI program that simulates human conversation through text or voice interactions, often used in customer service and information retrieval.

Citations (Showing Sources) or Transparency (Explaining AI Outputs)

RAG improves user trust by providing citations and references for the information generated, allowing users to verify claims easily. This transparency is vital in applications where accuracy is critical, such as healthcare or finance.

Closed-Source Model

An AI model that runs on a company’s servers and does not provide public access to its internal weights or training details.

Context Window (How Much the AI Can Remember)

The context window is the amount of text an AI model can consider at one time, including previous messages and documents.

Once the limit is reached, older information may be forgotten or ignored.

CPU vs GPU (Where the AI Runs)

AI models can run on either a CPU or a GPU.

CPUs work everywhere but are slower, while GPUs process many operations in parallel and are much faster for AI workloads.

Deep Learning

A specialized form of machine learning that uses neural networks with multiple layers to analyse various factors of data.

It's particularly effective in image and speech recognition.

Embeddings (Turning Meaning Into Numbers)

Embeddings are numerical representations of text that capture meaning.

They are used in semantic search, RAG, and vector databases.

Diffusion Model

A type of AI model commonly used for image generation, which creates images by gradually refining noise into structured visuals based on prompts.

Fine-Tuning (Adjusting a Pre-Trained Model)

Fine-tuning is the process of taking an existing AI model and making small adjustments so it performs better at a specific task.

Instead of training a model from scratch, fine-tuning uses a smaller, task-specific dataset to refine its behaviour.

This is more advanced than basic AI use and is often optional when working with local AI systems.

GANs & VAEs (Advanced Generative Models)

GANs and VAEs are advanced types of AI models designed to generate new data, such as images or sounds.

They are mainly used in research and specialised creative applications and are not required for most everyday local AI use.

Beginner note: You can safely skip this term until later.

Generative AI

A type of AI focused on generating new content, such as text, images, or music, often using models like GANs (Generative Adversarial Networks) or VAEs (Variational Autoencoders).

Grounded Answers (AI Using Real Sources)

Grounded answers refer to AI responses that are based on real, external information rather than guesses.

This usually means the AI looks up facts from documents, databases, or websites before generating an answer, making results more reliable and trustworthy.

Hallucination

A phenomenon where AI models generate plausible-sounding but incorrect or nonsensical information. RAG aims to reduce hallucinations by grounding outputs in verified external data sources.

Incremental Updates (Keeping AI Knowledge Fresh)

Incremental updates allow an AI system to receive new information without retraining the entire model.

This makes AI systems easier to maintain and update as information changes over time.

Inference (When the AI Is Running)

Inference is the process of using a trained AI model to generate outputs, such as answers or images.

This is different from training — inference is what happens every time you ask the AI a question.

Knowledge Base

A structured repository of information that RAG can utilize to fetch relevant data during the generation process. This can include databases, documents, and other resources that provide contextually appropriate information for generating responses.

Large Language Models (LLMs)

AI models trained on vast amounts of text data that can generate human-like text based on input prompts. Examples include GPT-3 and BERT.

Latency (How Fast the AI Responds)

Latency refers to the delay between sending a request and receiving a response from the AI.

Lower latency means faster, more responsive interactions.

Model Checkpoint (A Saved Version of a Model)

A checkpoint is a saved state of an AI model.

Different checkpoints of the same model may behave differently.

Model Loader (How AI Software Loads Models)

A model loader is the part of an AI application that loads and runs an AI model.

You usually don’t interact with it directly — it works behind the scenes to make sure the model runs correctly.

Multimodal Model

AI that can understand and process more than one type of input — such as text, images, or audio — within the same system.

Natural Language Processing (NLP)

A field of AI focused on enabling computers to understand and process human language, facilitating tasks like translation, sentiment analysis, and text generation.

Open-Weight Model

An AI model whose trained parameters (weights) are publicly available to download and run locally, subject to its license terms.

Parameters (What Makes a Model Large or Small)

Parameters are the internal values an AI model learns during training.

Models with more parameters are usually more capable but require more memory and processing power.

Prompt (What You Ask the AI)

A prompt is the input you give an AI model, usually as text. It tells the model what task to perform or how to respond.

Small changes in wording can significantly affect the output, which is why prompting is an important skill when working with AI.

Quantization (Making Models Smaller/Faster)

Quantization is a technique used to make AI models smaller and faster by reducing numerical precision.

In practice, this helps large models run on consumer hardware with less memory and power.

Reinforcement Learning

A type of machine learning where an agent learns to make decisions by taking actions in an environment to maximize cumulative rewards.

Retrieval-Augmented Generation (RAG)

A technique that combines the strengths of generative AI with external information retrieval systems to improve the accuracy and relevance of AI-generated responses. It allows LLMs to access up-to-date information from various sources, enhancing their output quality.

Sampling (How the AI Chooses Words)

Sampling refers to the method an AI model uses to select the next word or token.

Different sampling strategies affect accuracy, creativity, and consistency.

Semantic Search (Meaning-Based Search)

Semantic search focuses on understanding the meaning of a question rather than matching exact keywords.

This allows AI systems to return more relevant results, even if the wording doesn’t match exactly.

Temperature (How Creative the AI Is)

Temperature controls how predictable or creative AI responses are.

Lower values produce more factual, consistent answers, while higher values increase variation and creativity.

Text-to-Speech (TTS)

Technology that converts written text into spoken audio using AI-generated voices.

Tokens (How AI Splits Text)

Tokens are the small pieces of text that AI models use to process language.

Words, parts of words, or symbols can all be tokens. AI reads and generates text token by token.

Throughput (How Much AI Can Produce)

Throughput describes how much output an AI system can generate over time.

Higher throughput allows faster processing of longer documents or multiple requests.

Transfer Learning (Reusing a Trained Model)

A method where a pre-trained model is adapted for a new but related task, which saves time and resources compared to training a model from scratch.

Variational Autoencoders (VAEs)

They are a class of generative models in machine learning that utilize deep learning techniques to generate new data samples and perform various tasks such as anomaly detection and noise removal.

Vector Database (How AI Finds Similar Information)

A vector database stores information in a way that allows AI to find similar content quickly.

It is commonly used in systems like RAG to retrieve relevant documents based on meaning, not exact wording.

VRAM (GPU Memory for AI Models)

VRAM is the memory available on a graphics card.

AI models loaded on a GPU must fit into VRAM, which is why GPU memory often limits which models can run locally.

Web Loaders (Extracting Website Text for AI)

Web loaders are tools that collect text from websites and prepare it for use by AI systems.

They are commonly used when building chatbots or search tools that need to reference online content.

Web User Interface (Web UI)

The visual and interactive component of a website or web application that allows users to engage with the content and functionalities provided. It serves as the connection point between users and the digital system, facilitating navigation, information access, and interaction.