📖 AI Glossary

Agent (AI Agent)

An AI system that can take actions — search the web, write code, call APIs — to achieve a goal, not just answer questions.

Algorithm

A set of mathematical steps a computer follows to solve a problem. Machine learning algorithms find patterns in data.

Attention Mechanism

The part of a transformer that lets the model weigh which words in a sentence matter most for understanding each other word. The breakthrough that enabled LLMs.

Backpropagation

The algorithm that trains neural networks by calculating errors and adjusting connection weights to reduce them.

Bias (AI)

Systematic errors in AI output caused by patterns in training data that reflect historical or cultural biases.

ChatGPT

OpenAI's conversational AI product, built on GPT-4. One of the most widely used LLM-based assistants globally.

Claude

Anthropic's AI assistant, used in SumQ. Known for longer context windows and strong reasoning.

Context Window

The maximum amount of text an LLM can process at once — its working memory. GPT-4: ~128K tokens. Claude 3: ~200K tokens.

Deep Learning

Machine learning using neural networks with many layers. Enabled modern AI breakthroughs in image, speech and language understanding.

Diffusion Model

The AI technique behind image generators like DALL-E and Midjourney. Learns to remove noise from images to generate new ones.

Embedding

A numerical representation of text that captures its meaning. Similar meanings have similar numbers. Powers semantic search and RAG.

Few-shot Prompting

Giving the AI 2-3 examples in your prompt before asking it to continue the pattern.

Fine-tuning

Further training a pre-trained model on domain-specific data to specialise its behaviour and style.

Foundation Model

A large model trained on broad data that can be adapted for many tasks — GPT-4, Claude, Gemini are all foundation models.

Gemini

Google's large language model family, used in SumQ. Available in multiple sizes for different use cases.

Generative AI

AI that creates new content — text, images, code, audio, video — based on patterns learned from existing content.

GPT (Generative Pre-trained Transformer)

OpenAI's model architecture. "Pre-trained" on vast internet text. "Transformer" is the neural network architecture.

GPU

Graphics Processing Unit — the chip that powers AI training and inference. Originally for gaming, now essential for AI.

Hallucination

When an AI generates confident-sounding but false information. A fundamental limitation of current LLMs.

Human-in-the-loop

Keeping a human in the decision process when AI is used — especially for high-stakes or irreversible actions.

Inference

Using a trained AI model to generate responses. Training teaches the model; inference is the model doing its job.

Knowledge Cutoff

The date after which a model has no training data. Events after this date are unknown to the model.

LLM (Large Language Model)

An AI model trained on massive text datasets that can generate and understand language. ChatGPT, Claude, Gemini are LLMs.

LoRA

Low-Rank Adaptation — an efficient technique to fine-tune AI models by only updating a small fraction of their weights.

Machine Learning

Teaching computers to find patterns from data automatically, instead of manually writing rules.

Model

The output of training — a file containing learned parameters that can make predictions or generate content.

Multimodal

An AI that can process multiple types of input — text, images, audio and video — not just one.

Neural Network

A mathematical model loosely inspired by the brain — layers of connected nodes that transform input into output.

Overfitting

When a model memorises training data instead of learning general patterns, making it perform poorly on new data.

Parameter

A numerical value in a neural network that stores what the model learned. GPT-4 has an estimated 1 trillion parameters.

Perplexity

Both a measure of how well a language model predicts text, and a popular AI search tool that uses web results.

Prompt

The input you give to an AI — your question, instruction or context. Better prompts give better answers.

Prompt Engineering

The skill of writing effective AI instructions to get better, more accurate and more useful outputs.

RAG (Retrieval-Augmented Generation)

An approach where AI retrieves relevant documents first, then generates an answer grounded in them — reducing hallucination.

RLHF

Reinforcement Learning from Human Feedback — training AI using human preferences to align its behaviour.

Semantic Search

Search that finds results by meaning, not just exact keyword matching. Powered by embeddings.

SumQ

A multi-model AI platform that queries multiple LLMs in parallel and synthesises a consensus answer.

Supervised Learning

Training with labelled examples — "this is spam, this is not spam." The model learns to predict labels for new data.

Token

The basic unit an LLM processes — roughly ¾ of a word. "ChatGPT" ≈ 2 tokens.

Training

The process where a model learns from data by adjusting its parameters to minimise prediction errors.

Transformer

The neural network architecture behind all modern LLMs, introduced in the 2017 paper "Attention Is All You Need".

Vector Database

A database optimised for storing and searching embeddings — finding semantically similar content quickly.

Zero-shot Prompting

Asking an AI to do something without giving any examples — relying on its training alone.