AI glossary

LLM, token, RAG, agent, AGI, superintelligence … The key terms of modern AI, explained in plain language.

AGI (artificial general intelligence)

AI that can do essentially any intellectual task a skilled human can, across domains, without being specially built for it. Definitions differ; it has not been reached.See also: Superintelligence (ASI / SI)

AI agent

A system that pursues a goal over many steps: it plans, uses tools (browser, code, files, apps), checks results and decides what to do next.See also: Tool use / function calling, Model Context Protocol (MCP)

Alignment

Making AI systems reliably pursue the goals and values their developers and users intend — and not harmful shortcuts.See also: Reward hacking

Artificial intelligence (AI)

Computer systems that perform tasks we associate with human intelligence — understanding language, recognising images, reasoning, planning.

Chain of thought

Asking a model to reason step by step before answering. It improves maths, logic and planning; reasoning models do it internally.See also: Reasoning model

Context window

How much text (in tokens) a model can consider at once — the prompt, documents and conversation so far. Anything outside it is invisible to the model.

Deepfake

AI-generated or altered image, video or voice that realistically imitates a real person. Under the EU AI Act such content must be labelled.

Diffusion model

The technique behind most image and video generators: start from noise and remove it step by step until an image matching the prompt appears.

Edge AI / on-device AI

AI that runs on your own phone, laptop or browser instead of in the cloud. Private, works offline, no per-use cost.See also: WebGPU / WebAssembly, Quantization

Embedding

A list of numbers that captures the meaning of text or an image. Similar meanings get similar numbers — the basis of semantic search and RAG.

EU AI Act

The EU’s law on artificial intelligence, regulating AI by risk level: banned practices, high-risk systems, transparency duties and rules for general-purpose models.

Fine-tuning

Further training an existing model on your own examples so it adopts a style, format or domain knowledge.

Hallucination

When a model states something false with full confidence — an invented fact, quote or source. Always verify important facts.See also: Retrieval-augmented generation (RAG)

Inference

Running a trained model to get an answer — as opposed to training it. Most everyday AI cost is inference.

Large language model (LLM)

A transformer trained on huge amounts of text to predict the next token. From that simple objective emerge writing, translation, coding and reasoning skills.See also: Token, Transformer

Machine learning

Instead of writing rules by hand, a system learns patterns from examples (data) and uses them to make predictions.See also: Neural network

Model Context Protocol (MCP)

An open standard for connecting AI assistants to tools and data sources — like a USB port for AI. One connector works with many assistants.

Multimodal

A model that works with several kinds of input or output — text, images, audio, video — in one system.

Neural network

A model made of layers of simple connected units whose connection strengths (weights) are adjusted during training. “Deep learning” means many layers.

Open weights

A model whose trained parameters are published, so anyone can download, run and adapt it. Not always fully open source (data and training code may be closed).

Prompt

The instruction and context you give a model. Clear goals, context, examples and a required output format produce much better results.See also: Chain of thought

Quantization

Storing model weights with fewer bits (e.g. 4 instead of 16) so it needs less memory and runs faster — with a small loss in quality. It makes local and in-browser AI possible.

Reasoning model

A model trained to “think” — spend extra compute on an internal chain of reasoning — before answering. Slower, but much stronger on hard problems.

Red teaming

Deliberately attacking a model to find harmful behaviour, jailbreaks and weaknesses before bad actors do.

Retrieval-augmented generation (RAG)

The model first searches your documents or the web, then answers using what it found. Reduces hallucinations and keeps answers current.See also: Embedding

Reward hacking

When an AI finds an unintended shortcut to score well on its objective instead of doing the real task — for example, looking up test answers.

RLHF

Reinforcement learning from human feedback: people rate model answers and the model is trained to prefer the better ones. Key to making assistants helpful and safe.

Superintelligence (ASI / SI)

Intelligence that greatly exceeds the best humans in practically every field. The stated long-term goal of several frontier labs — and the “.si” in Hypervision’s story.See also: Alignment

Token

The unit a language model reads and writes — usually a word piece. In English one token is about ¾ of a word; Slovenian needs more tokens per word. Pricing and limits are counted in tokens.See also: Context window

Tool use / function calling

The ability of a model to call external functions — search, calculators, databases, APIs — and use the results in its answer.

Transformer

The neural-network architecture (2017) behind almost every modern language model. Its “attention” mechanism lets each word look at every other word in the context.

WebGPU / WebAssembly

Browser technologies that let websites use your graphics card (WebGPU) and run fast compiled code (WebAssembly) — the engines behind Hypervision’s edge tools.