AI Glossary
46 AI terms explained in plain English. No jargon, no PhD required. Click any term for the full definition with practical context.
A
AI Agent
An AI that can actually do things, not just talk. Instead of just answering "here's how to book a flight," an AI agent can actually search flights, compare prices, and book one for you — handling multiple steps on its own.
AI Alignment
Making sure AI does what humans actually want, not just what it was technically told to do. Like the story of the genie who grants wishes too literally — alignment is about making sure the AI understands the spirit of your request, not just the letter.
API (Application Programming Interface)
The "behind the scenes" way to use AI. Instead of chatting in a browser window, developers use APIs to plug AI into their own apps — like how a food delivery app uses a maps API to show you where your driver is.
Attention Mechanism
How AI decides what to pay attention to. When reading "The cat sat on the mat because it was tired," the attention mechanism helps the AI understand that "it" refers to "the cat" — by paying attention to the relationship between those words.
C
Chain-of-Thought (CoT)
Asking the AI to "show its working" like a maths teacher would. When the AI thinks through problems step by step instead of jumping to an answer, it makes fewer mistakes — especially with tricky questions.
ChatGPT
The AI chatbot that started it all — made by OpenAI. You type questions or instructions, and it writes back. Think of it as a very knowledgeable assistant you can have a conversation with. The free version uses a good model; the paid version ($20/month) gets you the best models.
Claude
Anthropic's AI assistant — a major competitor to ChatGPT. Claude is known for being thoughtful, following complex instructions well, and having a very large "memory" (context window). Available through claude.ai or the API.
Context Window
The AI's short-term memory. A bigger context window means the AI can read longer documents or remember more of your conversation. If you paste a 100-page document, you need a model with a large enough context window to read it all.
F
Few-Shot Learning
Teaching the AI by showing examples. Instead of explaining what you want, you show it: "Here's example input 1 and its ideal output. Here's example 2 and its output. Now do the same for this new input."
Fine-tuning
Teaching an already-smart AI to be an expert in your specific area. Like how a general doctor can specialise in cardiology through additional training, fine-tuning makes a general AI model better at one particular job.
Foundation Model
A general-purpose AI brain that can be specialised for different jobs. Like how a university degree gives you broad knowledge that you can then apply to a specific career, a foundation model provides general intelligence that can be tailored to specific tasks.
Function Calling (Tool Use)
When AI can use tools, not just talk. Function calling lets AI actually do things — check the weather, search a database, send an email — by generating specific "commands" that your software then carries out.
G
Gemini
Google's answer to ChatGPT. Gemini is built into Google products (Search, Gmail, Docs) and is available as a standalone chatbot. It comes in different sizes — from the tiny Nano (runs on phones) to the powerful Ultra (for complex tasks).
Generative AI
AI that creates new things. Unlike older AI that just sorted emails into spam/not-spam, generative AI writes essays, creates artwork, composes music, and writes code. If it can create something new, it's generative AI.
GPT (Generative Pre-trained Transformer)
The name for OpenAI's AI models — the ones behind ChatGPT. The letters stand for "Generative Pre-trained Transformer." Each new version (GPT-4, GPT-5) is smarter and more capable than the last.
GPU (Graphics Processing Unit)
The powerful computer chips that AI runs on. Originally made for video games, GPUs turned out to be perfect for AI because they can do thousands of calculations at the same time. NVIDIA makes the most popular ones (like the H100 and B200).
Grounding
Keeping AI honest by connecting it to real facts. Instead of letting the AI just make things up from memory, grounding forces it to check actual sources — like making a student cite their sources in an essay.
Guardrails
Safety bumpers for AI. Like guardrails on a road that stop you going off a cliff, AI guardrails stop the model from generating harmful content, sharing dangerous information, or behaving in ways its creators didn't intend.
I
Inference
Actually using the AI to get answers. Training is like studying for an exam; inference is like taking the exam. Every time you ask ChatGPT a question, it's doing inference — and that costs computing power.
Input Tokens
The words you send to the AI. When you type a question, those words get converted into tokens, and you're charged for each one. Input tokens are usually cheaper than output tokens — it costs less for the AI to read than to write.
L
Large Language Model (LLM)
A computer program that has read billions of pages of text and learned to write and understand language like a human. When you chat with ChatGPT or Claude, you are talking to an LLM.
Latency
How long you wait before the AI starts answering. When you press Send and there's a pause before text appears — that's latency. Smaller, faster models have lower latency. Bigger, smarter models often have higher latency.
LoRA (Low-Rank Adaptation)
A clever shortcut for customising AI models without needing a supercomputer. Instead of retraining the whole model (expensive), LoRA adds small "adapter" layers that steer the model's behaviour — like adding a lens filter to a camera instead of buying a new camera.
M
Model Collapse
What happens when AI trains on other AI's homework. Like a game of telephone where the message gets worse each time, if AI models keep learning from each other's outputs instead of real human-written text, the quality gradually degrades.
Multimodal
An AI that can see, hear, and read — not just one of those. A multimodal AI can look at a photo of your broken dishwasher, listen to the weird noise it makes, and tell you what's wrong with it, all in one conversation.
O
Open Source (in AI)
Free AI models that anyone can download and use. Like how Linux is a free operating system anyone can use, open-source AI models like Llama let you run AI on your own computer without paying a subscription.
Output Tokens
The words the AI writes back to you. These are usually more expensive than input tokens because generating new text takes more computing power than reading text. Think of it like how writing an essay takes more effort than reading one.
P
Prompt
The message or instruction you type to the AI. Just like how you phrase a question to a person matters, how you phrase your prompt to an AI matters. A clearer prompt gets a better answer.
Prompt Engineering
The art of talking to AI effectively. Like learning to ask the right questions, prompt engineering is about figuring out how to phrase your requests so the AI gives you exactly what you want.
R
RAG (Retrieval-Augmented Generation)
Giving the AI a reference book before asking it questions. Instead of relying on memory alone, RAG lets the AI look up real information first — like how a student performs better on an open-book exam than a closed-book one.
Reasoning Model
AI models that "think harder" before answering. Instead of blurting out the first answer, reasoning models take extra time to work through problems step by step — like how you'd solve a hard maths problem on paper rather than in your head.
RLHF (Reinforcement Learning from Human Feedback)
Training the AI by having humans rate its answers as good or bad. The AI learns to give more answers like the ones humans liked and fewer answers like the ones they didn't. It's like training a dog with treats — rewarding good behaviour.
S
Self-Hosting
Running AI on your own computer or server instead of using someone else's service. Like cooking at home vs. ordering takeaway — you have more control and potentially lower costs, but you need the equipment and skills to do it.
System Prompt
The "secret instructions" that tell the AI how to behave. When you use a customer service chatbot, it has a system prompt saying something like "You are a helpful customer service agent for Company X. Only answer questions about our products."
T
Temperature
A dial that controls how creative or predictable the AI is. Turn it down for factual answers (like maths or coding). Turn it up for creative writing or brainstorming. Think of it like the difference between a careful accountant (low) and a jazz musician (high).
Token
Think of tokens as the "syllables" that AI reads. The word "hamburger" might be split into "ham", "bur", "ger" — three tokens. AI companies charge based on how many tokens you use, like paying per word.
Tokenisation
How AI chops up your text into pieces it can understand. Like breaking a sentence into individual building blocks. The word "unhappiness" might become "un", "happiness" — two tokens. Different AI models chop text up differently.
Top-p (Nucleus Sampling)
Another way to control how creative or predictable the AI is, alongside temperature. At low top-p, the AI only picks from the most likely next words. At high top-p, it considers more unusual word choices too.
Transformer
The breakthrough invention that made modern AI possible. Before transformers, AI read text word by word. Transformers can look at all the words at once and understand how they relate to each other — like reading a whole page instead of one word at a time.