AI Glossary

AI Agent

An LLM equipped with tools (web search, file writing, code execution) that can complete tasks autonomously, making decisions and using tools sequentially without human intervention at every step.

Foundation model

A large LLM trained on general data, then used as a base for specific applications. Examples: GPT-4, Claude 3, Gemini Ultra.

LLM

Large Language Model. An AI model trained on massive amounts of text that learns to predict and generate language. It's the technology behind ChatGPT, Claude and Gemini.

Multimodal

A model capable of processing multiple input types: text, images, audio, video. GPT-4o and Gemini 1.5 are examples of multimodal models.

SLM

Small Language Model. Compact LLMs (under 10 billion parameters) that run on consumer devices like laptops and phones. Examples: Phi-3 Mini, Gemma 2B.

Context window

The maximum amount of text a model can "keep in mind" during a conversation. The larger it is, the more the model remembers. Measured in tokens.

Embedding

A numerical representation of the meaning of text as a vector. It allows models to compare sentences by "semantic similarity" — the foundation of AI search systems.

Fine-tuning

The process of further training an existing model on specific data to adapt it to a precise task. Like "specialising" a general practitioner in surgery.

Inference

The moment when the model "thinks" and generates a response, as opposed to training. It's the phase that happens every time you send a message to ChatGPT.

MCP

Model Context Protocol. Anthropic's open standard for connecting LLMs to external tools (databases, APIs, file systems) in a standardised way. It's becoming the "USB for AI agents".

RAG

Retrieval-Augmented Generation. A technique that allows an LLM to search for updated information in external documents before responding, overcoming the training data cutoff limitation.

RLHF

Reinforcement Learning from Human Feedback. An alignment technique that uses human ratings to teach the model to prefer helpful, safe and honest responses.

Token

The unit of text measurement for LLMs. Roughly ¾ of a word in English. Model costs and limits are measured in tokens (e.g. "128k token context window").

Open source / Open weights

Models with publicly available weights that anyone can download and run. Examples: Llama 3 (Meta), Mistral, Gemma. Opposite: closed models like GPT-4.

Hallucination

When an LLM generates false information presenting it as true, with great confidence. Still an open problem: models "invent" facts, citations or non-existent links.

Prompt

The instruction or question you send to an AI model. Writing effective prompts has become a genuine skill (prompt engineering).