How to Use This Glossary
If you have ever read an article about AI and felt lost within the first paragraph, you are not alone. The field is packed with abbreviations and jargon that can make it feel like everyone else got a handbook you never received. This glossary is that handbook.
Every term below is explained in two or three sentences of plain English — the way a patient friend would explain it over a cup of tea. Use the letter index to jump to a specific term, or simply scroll through to build your understanding one word at a time.
Tip: You do not need to learn every term. Bookmark this page and come back whenever you meet a word you do not recognise. Understanding grows one term at a time.
AGI Artificial General Intelligence
This is the idea of an AI that can do any intellectual task a human can do, across any subject, without needing to be specially trained for each one. It does not exist yet — today's AI is narrow, meaning it is good at specific things but not genuinely general-purpose. AGI is a long-term goal that researchers are working towards, and there is lively debate about whether it is years or decades away.
AI Artificial Intelligence
The broad umbrella term for any computer system that can learn, reason, or make decisions in ways that resemble human thinking. It covers everything from the recommendation system on a shopping site to the chatbot that answers your questions in full sentences. If a piece of software learns from examples rather than following a fixed set of rules, it falls under the AI umbrella.
Algorithm
A set of step-by-step instructions a computer follows to accomplish a task. In the AI world, the algorithm is the recipe that tells the system how to learn from data and make decisions. You can think of it as the method, while the data is the ingredients.
API Application Programming Interface
A way for one piece of software to talk to another without a human getting involved. In AI, an API is how a developer's app connects to a service like an LLM so it can send questions and receive answers automatically. It is the behind-the-scenes plug that lets different programs work together.
Alignment
The effort to make sure an AI system's goals and behaviour match what humans actually want. If an AI is misaligned, it might pursue an objective in a way that is technically correct but unhelpful or even harmful. Alignment research is about building guardrails so powerful AI acts in our best interests.
Annotation
The process of a human labelling data so an AI can learn from it. For example, someone might draw boxes around cars in photos and label them “car,” so the AI learns what a car looks like. Good annotation is one of the less glamorous but most important parts of building a reliable model.
Autonomous Agent
An AI system that can pursue a goal on its own over multiple steps, making decisions without a human guiding each move. Instead of just answering one question, it might plan a task, use tools, check its own work, and adjust along the way. Think of it as a digital assistant that can be given an objective and left to get on with it.
Backpropagation
The method a neural network uses to learn from its mistakes. When the system gets an answer wrong, it works backwards through its own layers, adjusting the connections that led to the error so it is less likely to repeat it. It is the core learning mechanism behind most modern deep learning.
Base Model
An AI model trained on a huge amount of general data before any specific customisation is added. It is a broad, general-purpose foundation that can then be fine-tuned for particular tasks. Most of the well-known chat AIs started life as base models before being shaped into helpful assistants.
Bias
A systematic error in an AI's output that causes it to treat some groups or situations unfairly or inaccurately. Bias usually comes from the training data — if the data reflects human prejudices or gaps, the AI learns those too. Recognising and reducing bias is one of the biggest challenges in responsible AI development.
Chatbot
A program designed to have a conversation with you in plain language, usually through text. Modern chatbots are powered by language models and can answer questions, help you write, and explain things conversationally. Older chatbots followed rigid scripts, but today's versions can genuinely understand and respond to what you say.
Context Window
The amount of text an AI can keep in mind at one time during a conversation. If a context window is small, the AI may forget what you said earlier once the conversation gets long. A larger context window means it can remember more of your chat, a longer document, or a bigger piece of code all at once.
Chain of Thought
A technique where an AI is encouraged to break a problem into smaller steps and reason through them out loud before giving a final answer. It tends to produce more accurate results on complex questions, much like a person who shows their working. You can often trigger it simply by asking the AI to “think step by step.”
Computer Vision
The branch of AI that deals with understanding images and video. It lets a computer recognise faces, read text in photos, identify objects, or describe what is happening in a scene. If a phone unlocks by looking at your face, that is computer vision at work.
Copilot
A general term for an AI assistant built into another piece of software to help you as you work. It might draft text in a word processor, suggest code in a programming tool, or summarise emails. The name comes from the idea of a helpful co-pilot sitting beside you rather than flying the plane itself.
Deep Learning
A type of machine learning that uses neural networks with many layers, allowing the system to learn very complex patterns. The “deep” refers to the number of layers, not to any kind of understanding. It is the technology behind most recent breakthroughs in image recognition, speech, and language.
Dataset
The collection of examples an AI learns from. A dataset might be thousands of labelled photos, millions of paragraphs of text, or years of weather readings. The quality and variety of the dataset largely determines how well the AI will perform.
Diffusion Model
The type of AI behind most modern image generators. It works by starting with random visual noise and gradually refining it, step by step, until a meaningful image emerges that matches your description. It is a bit like a sculptor removing material to reveal the shape hidden inside the block.
Distillation
A technique for making a large, powerful AI model smaller and faster by having it teach a more compact version of itself. The smaller model learns to mimic the bigger one's answers without needing all the original computing power. It is how many efficient models are created for use on phones and laptops.
Embedding
A way of representing words, sentences, or other content as lists of numbers so a computer can compare them. Things that are similar in meaning end up close together in this numerical space, so the AI can tell that “dog” and “puppy” are related but “dog” and “toaster” are not. Embeddings are the foundation of how AI understands meaning.
Ethics in AI
The study and practice of building AI systems that are fair, transparent, safe, and respectful of people's rights. It covers issues like bias, privacy, who is responsible when AI goes wrong, and how to avoid harm. Good AI ethics is not just philosophy — it shapes real design decisions in real products.
Explainability
The ability to understand and describe why an AI made the decision it did. Some models are “black boxes” that give an answer without any obvious reason, which is a problem in areas like medicine or law. Explainability is the effort to open that box so humans can trust and check the results.
Fine-tuning
Taking a model that has already been trained on a broad range of data and giving it extra training on a more specific set to specialise it. For example, a general language model could be fine-tuned on legal documents so it becomes a legal assistant. It is like sending a general practitioner to study cardiology — the broad knowledge stays, but a new specialism is added.
Foundational Model
A very large model trained on a vast amount of general data that can be adapted to many different tasks. Instead of building a separate AI for every job, you start from one foundation and customise it. The big language and image models you hear about are foundational models.
Few-shot Learning
When an AI learns a new task from just a handful of examples rather than thousands. You might show it three or four examples of what you want and it picks up the pattern. It is impressive because older AI needed enormous datasets to do anything new.
Generative AI
AI that creates new content — text, images, audio, video, or code — rather than just analysing or sorting existing information. It produces something that did not exist before, based on patterns it learned from its training data. ChatGPT writing an email and an image generator creating a picture from your description are both examples of generative AI.
GPU Graphics Processing Unit
A type of computer chip originally designed for drawing graphics in games, but which turned out to be excellent for AI work. GPUs can do many calculations at the same time, which is exactly what training and running large models requires. They are the engine behind most modern AI, and demand for them is enormous.
Gradient Descent
The mathematical method most AI models use to improve themselves during training. It works by repeatedly nudging the model's internal settings in the direction that reduces its errors, like a hiker walking downhill to find the lowest point in a valley. Step by step, the model gets better at its task.
Hallucination
When an AI confidently states something that is not true, often inventing facts, sources, or details that sound plausible but are completely made up. It happens because the model is generating what sounds most likely next, not checking whether something is actually correct. This is why you should always verify important information an AI gives you.
Hyperparameter
A setting chosen by the people building an AI that controls how the training process works — things like how fast it learns or how many layers the network has. Unlike the model's internal values, which it learns on its own, hyperparameters are set in advance by humans. Getting them right can make the difference between a great model and a useless one.
Inference
The moment an AI actually uses what it has learned to produce an answer. Training is the studying phase; inference is the exam. Every time you ask a chatbot a question and it replies, that is inference happening in real time.
Instruct Model
A language model that has been specifically trained to follow instructions and be a helpful assistant, rather than just continuing text. A base model might finish your sentence; an instruct model understands you are asking it to do something and tries to do it. Most chatbots you use are instruct models.
LLM Large Language Model
A type of AI that has read an enormous amount of text and learned how human language works so well that it can generate its own. It predicts what word should come next based on everything it has seen before, and by doing that repeatedly it produces whole paragraphs. ChatGPT, Claude, and Gemini are all LLMs.
Latent Space
The internal, numerical “map” an AI builds of concepts it has learned, where similar things sit near each other. It is how the model organises its understanding of the world in a form it can compute with. When an image generator blends a cat and a spaceship, it is moving through latent space between the two concepts.
LoRA Low-Rank Adaptation
A clever, efficient way to fine-tune a large model without needing to retrain the whole thing. Instead of adjusting every part of the model, LoRA trains a small add-on layer that modifies its behaviour. This makes customising big models far cheaper and faster, opening it up to smaller organisations and individuals.
Machine Learning
The branch of AI where computers learn from data rather than being explicitly programmed for every situation. Instead of writing rules by hand, you show the system lots of examples and it figures out the rules itself. It is the technique behind almost everything we call AI today.
Model
The actual “brain” an AI uses after it has finished training — the learned patterns stored in a form that can be used to make predictions or generate content. When people talk about “GPT-4” or “Claude,” they are referring to models. You can think of a model as the finished product of the training process.
MoE Mixture of Experts
A design where instead of one giant model answering every question, several smaller specialist models work together and only the most relevant ones activate for each request. This makes large models much more efficient because not every part needs to run every time. Many of the biggest modern models use this approach.
Multimodal
An AI system that can work with more than one kind of input or output — for example, text, images, and audio all at once. A multimodal model could look at a photo and describe it in words, or listen to speech and write a summary. It is a step towards AI that interacts with the world more the way humans do.
Neural Network
A type of AI model loosely inspired by the way neurons in a brain connect to each other. It is made up of layers of simple processing units that pass information along, and it learns by adjusting the strength of the connections between them. Despite the brain analogy, it is really just a very effective maths system — it does not think the way you do.
NLP Natural Language Processing
The field of AI focused on understanding and producing human language. It covers everything from translating text to figuring out whether a review is positive or negative to powering a chatbot's conversation. NLP is the reason computers can finally work with words in a genuinely useful way.
Narrow AI
AI that is good at one specific task but cannot transfer its skill to unrelated areas. A chess-playing AI is brilliant at chess but cannot write a poem or drive a car. Every AI that exists today is narrow AI, despite how impressive some of it has become.
No-code AI
Tools that let people use or build AI-powered features without writing any programming code. You might build a chatbot or train a custom model by dragging boxes and filling in forms. It is part of a wider trend to make AI accessible to non-technical people and small businesses.
Open Source Model
An AI model whose underlying code and weights are made freely available for anyone to download, study, and modify. This contrasts with closed models where only the company that built them can use or change them. Open source has accelerated AI research enormously, though the very largest models are often still kept closed.
Overfitting
When a model learns its training data so precisely that it memorises the specifics rather than understanding the general pattern. It performs brilliantly on data it has seen before but poorly on anything new — like a student who has memorised past exam answers but cannot handle a fresh question. Avoiding overfitting is a central concern in training good models.
Parameters
The internal values a model learns during training that determine how it behaves — the “knowledge” stored inside it. When you hear a model described as having “70 billion parameters,” that is a measure of how much it can potentially learn and represent. More parameters generally mean a more capable model, but also a more expensive one to run.
Prompt
The instruction or question you type to tell an AI what you want. A good prompt is clear, specific, and gives enough context for the AI to understand your intent. Writing better prompts is one of the most useful practical skills in the AI era, and it is something anyone can learn.
Prompt Engineering
The practice of crafting prompts carefully to get the best possible results from an AI. Small changes in wording, structure, or added context can dramatically change the quality of the output. It is less about tricking the AI and more about communicating clearly what you actually want.
Pre-training
The first, big stage of training a model on a vast amount of general data before any specific task is added. It is where the model picks up its broad understanding of language, images, or whatever it works with. Fine-tuning comes afterwards, building on the foundation pre-training created.
Quantization
A technique that reduces the precision of the numbers inside a model so it takes up less memory and runs faster, with only a small loss in quality. It is how large models are squeezed onto laptops and phones instead of requiring massive data centres. Think of it like compressing a high-resolution photo — it looks almost the same but the file is much smaller.
RAG Retrieval-Augmented Generation
A technique where an AI looks up relevant information from a separate source — like your company documents or a knowledge base — and uses it to answer your question, rather than relying only on what it learned during training. This makes answers more accurate and up to date, and reduces hallucination. It is how many business AI assistants work.
Reasoning Model
A model specifically designed to work through complex, multi-step problems rather than just answering quickly. It spends more time “thinking” before responding, which improves results on maths, logic, and planning tasks. The trade-off is that it takes longer and costs more to run.
Reinforcement Learning
A type of machine learning where an AI learns by trial and error, receiving rewards for doing well and penalties for doing poorly. It is how systems learn to play games like chess and Go at a superhuman level. The AI explores, discovers what works, and gradually gets better through feedback.
RLHF Reinforcement Learning from Human Feedback
A training method where humans rate different AI responses so the model learns which kinds of answers are helpful, honest, and safe. It is a major reason modern chat assistants feel genuinely useful rather than just technically capable. Most of the well-known consumer AI models went through this process.
Self-supervised Learning
A training method where the AI learns from data that has not been labelled by humans, by setting itself tasks like predicting the next word in a sentence. This matters because labelled data is expensive and limited, while unlabelled text and images are practically infinite. It is how the biggest language models were trained.
Semantic Search
A search approach that understands the meaning of what you type rather than just matching keywords. So if you search for “how to fix a leaking tap,” it can find an article titled “repairing a dripping faucet” because the meaning is the same. It is powered by embeddings and is far more useful than old-fashioned keyword matching.
Sentiment Analysis
An AI technique that figures out the emotional tone behind a piece of text — whether it is positive, negative, or neutral. Businesses use it to understand customer reviews, social media posts, and support tickets at scale. It turns thousands of comments into a simple picture of how people feel.
System Prompt
An instruction given to an AI before you ever speak to it that sets its behaviour, role, and boundaries. For example, a system prompt might tell a customer service bot to be polite, stay on topic, and never give medical advice. You usually do not see it, but it shapes every answer you get.
Token
A small chunk of text — roughly a word or part of a word — that an AI uses as its basic unit when reading and writing. A model does not see whole words the way you do; it sees sequences of tokens. When you hear about a model's “context length” or the cost of using an API, it is usually measured in tokens.
Tokenizer
The tool that splits text into tokens before an AI processes it. The same word might be one token or several depending on the tokenizer and the language involved. It sits between you and the model, quietly breaking your words into pieces the AI can handle.
Temperature
A setting that controls how random or predictable an AI's responses are. A low temperature makes the model cautious and consistent, giving similar answers each time. A high temperature makes it more creative and varied, but also less reliable — useful for brainstorming, risky for factual answers.
Training Data
The examples an AI learns from during its training phase. It can be text, images, audio, numbers, or anything else the model is built to understand. The old saying “garbage in, garbage out” applies perfectly — a model is only as good as the data it learned from.
Transformer
The specific type of neural network architecture behind virtually all modern language models. Its key innovation is the “attention” mechanism, which lets it weigh the importance of different words in a sentence all at once rather than one by one. It is the reason AI went from clumsy to fluent so quickly.
Transfer Learning
When a model takes what it learned from one task and applies it to a different but related task, rather than starting from scratch. This is why a model trained on general text can be adapted to summarise legal documents with relatively little extra training. It makes AI far more practical and affordable to customise.
Vector Database
A specialised storage system designed to hold embeddings — the numerical representations of meaning — and find similar ones quickly. It is what lets an AI search through thousands of documents and pull out the most relevant ones based on meaning rather than exact keywords. Vector databases are a key part of many modern AI applications, including RAG systems.
Vision Model
An AI model specifically trained to understand and interpret images. It can describe what is in a photo, read text within an image, or identify objects and people. Vision models power features like visual search, medical image analysis, and the image-understanding abilities of multimodal assistants.
Weights
The specific numerical values inside a model that represent what it has learned. Each connection between neurons has a weight, and together all the weights encode the model's knowledge. When you download a model file, what you are really downloading is a large collection of weights.
Workflow Automation
Using AI to handle repetitive, multi-step tasks that would otherwise need a person — things like sorting emails, extracting data from invoices, or routing customer queries. It combines AI with ordinary software to free humans from routine work. For small businesses, it is often where AI delivers the most obvious value.
Still Confused by a Term?
AI vocabulary keeps growing, and no glossary can capture every word the moment it appears. If you came across a term that is not here, or if an explanation still does not quite click, that is completely normal — even people who work in AI full time look things up every day.
The best approach is to keep this page handy, return whenever you need it, and not worry about memorising anything. Understanding comes from repeated exposure, not from cramming. Over time, the jargon starts to feel less like a foreign language and more like familiar shorthand.
Remember: You do not need to know every technical term to use AI well. You just need enough understanding to ask good questions and recognise when something sounds off. That is a skill anyone can build, one term at a time.
Want Help Making AI Work for You?
Avondale.AI can help you cut through the jargon and find the practical, genuinely useful ways AI can help your business or project — explained in plain English, with no overwhelm.
Get in Touch with Steve