AI Glossary
54+ AI terms explained in plain English. No jargon required.
A
AGI (Artificial General Intelligence)
AI that can do any intellectual task a human can — learning, reasoning, problem-solving across all domains. Doesn't exist yet. What we have now is "narrow AI" that's good at specific tasks.
AI (Artificial Intelligence)
Computer systems that can do tasks that normally require human intelligence — like understanding language, recognising images, or making decisions.
AI Alignment
The challenge of making sure AI systems do what humans actually want, not just what they're technically programmed to do. A major focus of AI safety research.
AI Bias
When AI systems produce unfair results because of problems in their training data or design. For example, facial recognition working worse for certain ethnicities.
Algorithm
A set of rules or steps a computer follows to solve a problem. Like a recipe, but for machines.
Anthropic
The company that makes Claude. Founded by former OpenAI researchers focused on AI safety.
API
Application Programming Interface — a way for different software to talk to each other. How apps connect to AI services like ChatGPT behind the scenes.
B
Backpropagation
How neural networks learn from mistakes — errors are sent backwards through the network to adjust its settings. Technical but fundamental.
Benchmark
A standardised test to measure AI performance. Like an exam that lets researchers compare different AI systems.
BERT
An influential AI model from Google (2018) that improved how computers understand language. Led to better search results.
Big Data
Extremely large datasets that require special tools to process. AI needs big data to learn patterns.
Black Box
When an AI system makes decisions but we can't see how. The inputs and outputs are visible, but the reasoning inside is hidden.
C
ChatGPT
OpenAI's conversational AI assistant, launched November 2022. The tool that brought AI into mainstream awareness.
Claude
Anthropic's AI assistant. Known for longer context windows and more nuanced responses.
Cloud Computing
Running software on remote servers accessed via the internet. Most AI tools run in the cloud, not on your device.
Computer Vision
AI that understands images and video — recognising faces, reading text, identifying objects.
Context Window
How much text an AI can consider at once. Like the AI's working memory. Larger windows mean it can handle longer documents.
D
DALL-E
OpenAI's AI image generator. Type a description, get an image.
Deep Learning
AI using neural networks with many layers. The "deep" refers to depth of layers. Powers most modern AI breakthroughs.
Deepfake
AI-generated fake video or audio of real people. Can make someone appear to say things they never said.
DeepMind
Google's AI research lab. Created AlphaGo and AlphaFold. Based in London.
Diffusion Model
The technology behind AI image generators. Learns to create images by reversing a process of adding noise.
F
Fine-Tuning
Taking a pre-trained AI and training it further on specific data. How companies customize general AI for their needs.
Foundation Model
A large AI model trained on broad data that can be adapted for many tasks. GPT-4 and Claude are foundation models.
G
GAN
Generative Adversarial Network — two AI systems competing. One creates fake data, one tries to detect fakes. Led to realistic image generation.
Gemini
Google's latest AI model family, competing with GPT-4. Powers Google's AI products.
Generative AI
AI that creates new content — text, images, music, code. The current AI boom is largely about generative AI.
GPT
Generative Pre-trained Transformer. The architecture behind ChatGPT. Pre-trained on vast text, then fine-tuned for conversation.
GPU
Graphics Processing Unit — computer chips originally for gaming graphics, now essential for AI. NVIDIA dominates this market.
H
Hallucination
When AI confidently generates false information. A major limitation — AI doesn't know what it doesn't know.
Hugging Face
A platform for sharing AI models and datasets. Like GitHub but for AI.
L
Language Model
AI that predicts and generates text. The technology behind ChatGPT, Claude, and similar tools.
Large Language Model (LLM)
Language models with billions of parameters, trained on massive text datasets. GPT-4, Claude, Gemini, Llama.
Llama
Meta's open-source large language model. Made powerful AI more accessible to researchers and developers.
M
Machine Learning
AI that learns from data instead of being explicitly programmed. The dominant approach in modern AI.
Midjourney
Popular AI image generator known for artistic quality. Accessed through Discord.
Multimodal AI
AI that can work with multiple types of input — text, images, audio, video. GPT-4 with vision is multimodal.
N
Natural Language Processing (NLP)
AI that understands and generates human language. The field that gave us chatbots, translation, text analysis.
Neural Network
AI systems loosely inspired by brains — layers of connected nodes processing information. Foundation of deep learning.
NVIDIA
The company dominating AI chips. Their GPUs power most AI training.
O
OpenAI
The company behind ChatGPT, GPT-4, and DALL-E. Started as nonprofit, now mostly for-profit.
Overfitting
When AI memorises training data instead of learning general patterns. Does great on training data, poorly on new data.
P
Parameter
A setting in a neural network adjusted during training. GPT-4 has over a trillion parameters.
Prompt
The instruction or question you give to AI. Better prompts get better results.
Prompt Engineering
The skill of crafting effective prompts. Part art, part science.
R
RAG
Retrieval-Augmented Generation — AI that looks up information before responding. Reduces hallucinations by grounding in real data.
Reinforcement Learning
AI that learns through trial and error, getting rewards for good outcomes. How AI learned to beat humans at games.
RLHF
Reinforcement Learning from Human Feedback — using human preferences to train AI. Key to making ChatGPT helpful and safe.
T
Temperature
A setting controlling AI randomness. Low = more predictable, high = more creative/random.
Token
The units AI processes text in. Roughly 3/4 of a word. "Tokenization" splits text into these pieces.
Training
The process of teaching AI from data. Can take days to months for large models.
Transformer
The architecture behind modern language models. Introduced in 2017, revolutionised NLP. The "T" in GPT.
Turing Test
Alan Turing's proposed test for machine intelligence — can it fool a human in conversation? Passed by modern chatbots in limited contexts.
Z
Zero-Shot Learning
AI performing tasks it wasn't specifically trained for. Modern LLMs are surprisingly good at this.