The AI world has its own language. And it is a language that evolves fast. Over the past two years alone, terms like “RAG”, “multimodal” and “agentic” have gone from being niche terminology reserved for machine learning researchers to appearing in board meetings, tender documents and onboarding materials.
The problem is that there is no single good place to look them up in a professional context. Existing glossaries tend to be either too technical or too shallow to actually be useful when you need them most.
This list is here to help. Here are 50 of the most commonly used words, abbreviations and concepts in the AI world today, explained in plain language without unnecessary jargon.
Core concepts
Artificial Intelligence (AI)
An umbrella term for computer systems that perform tasks normally associated with human intelligence, such as understanding language, recognising images or making decisions.
Machine Learning (ML)
A subset of AI where a system learns from data without being explicitly programmed rule by rule. Instead, it discovers patterns on its own.
Deep Learning
A type of machine learning that uses artificial neural networks with many layers. Particularly effective for images, audio and text.
Neural Networks
Computational structures loosely inspired by the human brain. They consist of nodes (neurons) connected in layers that process and pass on information.
Algorithm
A precise, step-by-step description of how to solve a problem. In AI: the set of rules governing how a model learns and makes decisions.
Dataset
The collection of data used to train, validate or test an AI model. The quality of the dataset is critical to the quality of the model.
Training
The process by which a model adjusts its internal parameters by processing large amounts of data, with the goal of improving at the task at hand.
Inference
When a trained model is used to produce answers or predictions on new inputs. This is what happens every time you ask ChatGPT a question.
Language models and text AI
Large Language Model (LLM)
A model trained on enormous amounts of text to understand and generate human language. GPT-4, Claude and Gemini are all LLMs.
Generative AI
AI systems that produce new content, such as text, images, code or audio, rather than simply classifying or analysing existing content.
Prompt
The instruction or question you send to an AI model. Writing good prompts, known as prompt engineering, is a skill in itself.
Prompt Engineering
The practice of formulating instructions to an AI model in a way that produces precise, useful and relevant responses.
Context Window
The amount of text a model can hold in memory and take into account within a single conversation. A larger context window means the model can work with longer documents.
Token
The smallest unit text is broken into for a model to process. Roughly equivalent to three to four characters. Pricing for AI APIs is typically based on token count.
Temperature
A setting that controls how creative or predictable a model is. Low temperature gives precise, consistent responses. High temperature gives more variety and more risk of errors.
Hallucination
When an AI model produces information that sounds plausible but is incorrect or fabricated. One of the most important challenges with current language models.
Embeddings
A mathematical representation of text or other data as numbers. Used to calculate semantic similarity, for example whether two sentences mean the same thing.
RAG (Retrieval-Augmented Generation)
A technique where the model retrieves relevant information from an external knowledge base before generating a response. Reduces hallucinations and keeps answers more up to date.
Fine-tuning
Continuing to train an existing model on a specific dataset to make it better at a particular task or tone, for example adapting a model to a specific industry.
Models and architecture
Transformer
The dominant architecture behind modern LLMs. Uses a mechanism called attention to understand relationships between words regardless of their distance in the text.
Multimodal
A model that can process multiple types of input, for example text and images combined. GPT-4o and Gemini Ultra are examples of multimodal models.
Parameters
The internal values in a model that are adjusted during training. “70 billion parameters” is a measure of size and capacity.
Foundation Model
A large, general-purpose model trained on broad and varied data that can then be adapted to many different tasks. GPT, Claude and Llama are foundation models.
Open-source Model
A model whose weights are publicly available. Llama from Meta and Mistral are examples. Offers more control but requires more technical expertise.
Weights
The numerical values stored in a trained model. These are what constitute the model’s knowledge. Publishing the weights is what makes a model open.
Latency
The time from a request being made to the first response arriving. Important in real-time applications. Lower latency means faster responses.
Agents and automation
AI Agent
An AI system that can plan and carry out multi-step tasks autonomously, not just answer a single question but take a series of actions to reach a goal.
Agentic AI
A collective term for AI systems that act with a degree of autonomy, planning, choosing tools and executing tasks without asking for confirmation at every step.
Tool Use / Function Calling
The ability for a model to call external functions or APIs, such as searching the web, running code or reading a file, as part of completing a task.
Orchestration
Coordinating multiple AI agents or models so they work together towards a shared goal. Essential in complex AI workflows.
MCP (Model Context Protocol)
An open standard from Anthropic that allows AI models to connect to external data sources and tools in a structured way. Think of it as USB-C for AI integrations.
Workflow
A sequence of automated steps where AI is involved as one or more components. For example: fetch data, analyse, generate report, send email.
Training and optimisation
RLHF (Reinforcement Learning from Human Feedback)
A training method where human evaluations of model responses are used to adjust and improve it. Central to the development of ChatGPT and Claude.
Supervised Learning
Machine learning where the model is trained on pairs of input and expected output. Humans have labelled the data in advance.
Unsupervised Learning
Learning without labelled data. The model finds structures and patterns on its own. Used for tasks such as clustering data or detecting anomalies.
Overfitting
When a model learns the training data too well and performs poorly on new, unseen data. Like memorising the answer sheet instead of understanding the subject.
Benchmark
A standardised test for measuring and comparing model capabilities. MMLU and HumanEval are examples. Useful, but not always representative of real-world performance.
Compute
The processing power, CPU, GPU and TPU, required to train and run AI models. One of the most important limiting factors in AI development today.
GPU (Graphics Processing Unit)
Graphics chips that turned out to be exceptionally well suited for parallel matrix multiplication, the core of neural network computation. Nvidia dominates this market.
Safety and ethics
AI Safety
The field concerned with ensuring that AI systems behave as intended, do not cause harm and are aligned with human values, especially as systems become more capable.
Alignment
Ensuring that an AI model acts in accordance with human intentions and values. One of the major unsolved problems in AI research.
Bias
Skewed tendencies in a model that reflect skews in the training data. Can lead to systematically unfair treatment of certain groups.
Prompt Injection
An attack where malicious text in input attempts to manipulate an AI model into ignoring its instructions and doing something unintended.
Guardrails
Constraints and safety measures built into an AI model to prevent unwanted behaviour, such as producing harmful content or revealing sensitive information.
EU AI Act
The EU’s regulatory framework for artificial intelligence, adopted in 2024. Classifies AI systems by risk level and sets requirements for transparency, testing and human oversight.
Use and implementation
API (Application Programming Interface)
An interface that allows programs to communicate with each other. Using Claude or GPT via API means integrating the model directly into your own application.
Copilot
An AI assistant that works alongside a user rather than replacing them. Microsoft Copilot in Word and GitHub Copilot in code are typical examples.
Vector Database
A database purpose-built for storing and searching embeddings. A key component in RAG systems. Examples include Pinecone, Weaviate and pgvector.
System Prompt
A hidden instruction that defines how an AI model should behave in a given application, covering tone, role and constraints. The user typically does not see it.
Chain-of-Thought (CoT)
A prompting technique where the model is asked to think step by step rather than jumping straight to an answer. Improves accuracy on reasoning tasks.
This is not a static list. The AI field moves fast, and new terms emerge every month. We will update this glossary regularly as the landscape continues to evolve.
Want to learn more about how Frontkom can help your organisation adopt AI in a structured and responsible way? Feel free to get in touch.
