Arnhem · websites & automationBVDNET

AI Dictionary

Every entry starts with an explanation that needs no prior knowledge.

A

Illustration: a clay robot in a workshop with a gear, a lever switch and an arch on the workbench

Models & Architecture

Activation function

An activation function is the small rule inside a neural network that decides how much of each signal passes on, so the model can learn more than straight lines.

Illustration: a craftsman adjusting the dials of a brass machine topped with a glowing brain

Activation steering

Activation steering changes how an AI model behaves by adjusting the numbers inside it while it answers, without retraining the model or changing the prompt.

Illustration: a workshop with gears of different sizes on shelves and a workbench, lit by an oil lamp

Models & Architecture

ActTail

ActTail is a research method from 2026 that lets a language model skip its least important arithmetic, with a separate budget for each part, so it loses less quality.

Illustration: an owl in a workshop surrounded by gears, lamps and tools

Models & Architecture

Adaptive thinking in AI

Adaptive thinking lets an AI model decide per question whether to think before it answers and for how long, so easy questions stay fast and hard ones get more care.

Illustration: a smiling robot at a workbench with a row of wrenches and tools

Safety & Ethics

Adversarial Cost to Exploit (ACE)

Adversarial Cost to Exploit (ACE) measures how much an automated attacker has to spend before an AI agent does something it was told never to do.

Illustration: a robot in an apron assembling a small machine while three other robots watch

Tools & Frameworks

Agent Browser Protocol (ABP)

The Agent Browser Protocol (ABP) is a modified Chrome browser for AI agents that pauses the web page after every step, so the agent never acts on a moving page.

Illustration: small figures inspecting their work with a magnifying glass next to a control panel

Agent evaluation

Agent evaluation tests whether an AI agent really completes the tasks you give it, such as a refund, by checking both the end result and the steps it took.

Illustration: a notebook with a paperclip in front of a lit card-index cabinet in a workshop

Agent operational memory

Agent operational memory is a set of text files with an AI agent's rules and lessons, read at the start of every session so you need not explain them again.

Illustration: an orange cat with its paw on a circuit board, beside a magnifying glass and lab flasks

Safety & Ethics

AgentDrift

AgentDrift is a March 2026 study in which AI advisors repeated manipulated data from their tools and gave unsafe advice, while the usual quality score looked fine.

Illustration: a robot with a light bulb on its head assembling gears at a workbench

Agentic AI

Agentic AI is AI that carries out a task on its own: it plans the steps, uses software such as email or an expense system, checks the results and asks when it gets stuck.

Illustration: an old craftsman repairing small robots at his workbench

Agentic engineering

Agentic engineering is building software with AI coding agents that write and test the code, while a developer decides what to build and checks the result.

Illustration: a squirrel reading a page under a desk lamp between stacks of books

Agentic RAG

With agentic RAG, an AI agent searches in steps: it decides what to look for, checks whether the results answer the question and searches again if they do not.

Illustration: a robot in an apron stirring a pot labelled Ideas, surrounded by floating notes

AI agent

An AI agent is software that carries out a task on its own: it plans the steps, uses tools such as your email or order system, and asks you when it is unsure.

Illustration: a large robot reading a picture book with three small robots at a workshop table

Safety & Ethics

AI alignment

AI alignment is the work of making an AI system do what its makers and users intend, without harmful side effects, also in situations nobody tested.

Illustration: a robot hands a gear to a smiling man doing a puzzle at a table, with clocks on the shelves behind them

Practical Applications

AI API

An AI API lets your own software use an AI model from a provider such as OpenAI or Anthropic over the internet, so you pay per use and run no model yourself.

Illustration: small robots with clipboards around a scoreboard, three rubber ducks on a shelf behind them

AI benchmark

An AI benchmark is a standard test with fixed questions and known answers, used to score different AI models the same way so you can compare them.

Illustration: an owl on a stack of books between computer screens in a library

Safety & Ethics

AI governance

AI governance is how an organisation decides which AI it uses and for what, who checks the results, and how it meets laws such as the EU AI Act.

Illustration: a craftsman building a small coloured fence on a workbench while a robot watches

Safety & Ethics

AI guardrails

Guardrails are the rules and checks around an AI tool that keep it on topic and stop harmful answers or risky actions before they reach a customer.

Illustration: an open book releasing coloured smoke and question marks

AI hallucination

An AI hallucination is an answer from a chatbot or other AI tool that sounds certain but is wrong or made up, such as a source that does not exist.

Illustration: an owl with glasses reading an open book, with rolled scrolls beside it and a mouse following a trail of glowing coins

AI inference

Inference is the moment a trained AI model does its work: it takes your question, text or photo and turns it into an answer, as in every chatbot reply.

Illustration: a man in a cap tinkering with the panel of a domed machine covered in gears, in a room lined with books

Safety & Ethics

AI jailbreaking

Jailbreaking is talking an AI model into doing what its makers trained it to refuse, such as giving harmful instructions or insulting the company that runs it.

Illustration: a cat in round glasses turning a gear on a machine with gauges and warning lights

Industry & Business

AI observability

AI observability means recording what an AI system does in daily use, such as questions, answers, costs and response times, so you notice when it goes wrong.

Illustration: a chef arranging ingredients and a kitchen timer on a counter below a wall of pipes

AI orchestration

AI orchestration connects several AI steps, such as recognising a question, searching your documents and writing an answer, into one process that runs in the right order.

Illustration: a person with a magnifying glass studying chess pieces next to a robot head made of gears

Safety & Ethics

AI red teaming

AI red teaming means attacking your own AI system on purpose before launch, so you find and fix the ways it can be misused before someone else does.

Illustration: three robots at a workbench with gears and a toolbox

Practical Applications

AI robotics

AI robotics means robots steered by an AI model: they look through cameras, take a task in plain words and work out the movements themselves, without a fixed program.

Illustration: a smiling robot at a counter with three glowing jars, in front of a wall of dials and switches

Always-on agents

An always-on agent is an AI assistant that runs on a provider's servers and starts by itself at a set time or when something happens, even while your computer is off.

Illustration: a cat in glasses at a desk with books, pattern cards and a small robot

Artificial intelligence (AI)

Artificial intelligence (AI) is software that does work which normally takes human thinking, such as answering a customer's question or spotting an unusual payment.

Illustration: three cooks at a kitchen table, each busy with their own ingredients

Models & Architecture

Attention mechanism

The attention mechanism lets an AI language model work out which words in a text belong together, such as which earlier word the word "it" refers to.

Illustration: an owl and a chameleon at a workbench with open books, gears and tools

Safety & Ethics

Automated alignment research

Automated alignment research lets AI agents run their own experiments on making AI systems safer and easier to steer, while people check the results.

Illustration: glowing gears linked by circuit lines, with a squirrel holding a cup on a workshop shelf

Safety & Ethics

Autonomous AI cybersecurity defense

Autonomous AI cybersecurity defense is AI software that searches programs for security holes on its own, checks whether they are real and proposes a fix.

Illustration: an owl with glasses in a library, stringing glowing beads one by one from a bowl

Models & Architecture

Autoregressive generation

Autoregressive generation is how a language model writes: one small piece of text at a time, each chosen from everything written so far, until the answer is done.

B

Illustration: an old man in a workshop inspecting a toy train with a torch, a glowing brain outline behind him

Backpropagation

Backpropagation is the calculation that lets an AI model learn from its mistakes, by working out how much each internal setting contributed to an error.

Illustration: a baker kneading a large lump of dough while a hand adds a spoonful of an ingredient

Batch size and learning rate

Batch size and learning rate are two settings for training an AI model: how many examples it sees before each adjustment, and how big each adjustment is.

Illustration: a baker taking letter-shaped loaves from the oven, with a branching diagram on the wall

Models & Architecture

Beam search

Beam search lets an AI model keep several partial sentences open at once and pick the most likely complete one at the end, as translation and speech tools do.

Illustration: a robot with round glasses studying an open book through a magnifying glass, in front of bookshelves

Belief-action decoupling

Belief-action decoupling splits an AI agent in two: one part keeps a short list of what it knows and how sure it is, the other picks the next step from that list.

Illustration: wooden mannequins around a blueprint, one holding a magnifying glass, with a scale of justice on an easel

Safety & Ethics

Bias in machine learning

Bias in machine learning is when an AI system treats some groups of people systematically worse than others, often because it learned from skewed data.

Illustration: two globes on a library shelf linked by streaks of light, one glowing like a network of veins, with a squirrel beside them

Models & Architecture

Bicameral Model

The Bicameral Model is a research design from May 2026 in which two AI language models work as a pair and exchange internal signals instead of text.

Illustration: four robots at a workbench connecting glowing puzzle pieces into a chain

Binex

Binex is a free, open-source tool from one developer that runs a chain of AI steps on your own computer and records what each step received, produced and cost.

C

Illustration: a potter holding a jug while pots fall from the shelf behind him

Catastrophic forgetting

Catastrophic forgetting is when an AI model loses skills it already had because it was trained further on something new, such as a company's own data.

Illustration: an owl with glasses laying out a chain of linked coloured rings on a desk in a library

Practical Applications

Chain-of-thought prompting

Chain-of-thought prompting asks an AI model to write out its steps before it answers, which helps with sums and other questions that take several steps.

Illustration: a robot solving a puzzle at a table beside a signpost, with gears and a circuit board behind it

Chatbot compared with an AI agent

A chatbot answers your questions in a conversation; an AI agent carries out a task for you in steps, with tools such as your email or order system.

Illustration: shelves of coloured jars with a walking magnifying glass and small figures in hats

Classifier

A classifier is an AI model that sorts each input into one of a set of categories you chose, such as spam or not spam, or which team a ticket goes to.

Illustration: a green robot soldering a machine of smiling gears on a workbench, surrounded by tools and an open book

Tools & Frameworks

Claude Code

Claude Code is Anthropic's AI assistant for programming: you describe a task in plain language, and it reads the code, edits files and runs commands to carry it out.

Illustration: robots around a table covered in papers, studying them with magnifying glasses while one points at a glowing speech bubble

CoDream

CoDream lets a team of AI agents learn from a failed task: they discuss what went wrong and pass the lesson to the teammates who are weakest at that kind of task.

Illustration: small figures around a workbench with gears, plugs and a power socket under hanging light bulbs

Tools & Frameworks

Composio

Composio is a service that lets AI agents act in apps such as Gmail, Slack or GitHub on behalf of a user, while Composio stores and refreshes the logins.

Illustration: a purple robot writing with a quill in a glowing book, with a written page beside it

Safety & Ethics

Constitutional AI (CAI)

Constitutional AI is Anthropic's method for training an AI assistant with a written list of principles, so the model learns to judge and improve its own answers.

Illustration: a robot reading a long scroll through a magnifying glass in a library

Context compression for AI agents

Context compression keeps what an AI model has to read short, by summarising or removing old material, so long tasks stay affordable and the model keeps its focus.

Illustration: a robot chef stirring a pot between growing piles of paper notes

Context rot

Context rot is the decline in an AI model's work as its conversation or task grows longer, so that it starts to overlook instructions and facts from earlier on.

Illustration: an owl on a ladder with a basket of glowing cards while squirrels fetch books from the shelves

Context window

A context window is how much text an AI model can take in at once, from your question and documents to its own answer, counted in tokens.

Illustration: clocks, gears and books on workshop shelves

Continual learning

Continual learning lets an AI model keep learning from new information after its first training, without forgetting what it learned before.

Illustration: two stacks of books with a glowing angle between them

Cosine similarity

Cosine similarity is a score between -1 and 1 that says how alike two texts are, once an AI model has turned them into lists of numbers.

D

Illustration: an owl with glasses behind a screen showing a maze, with magnifying glasses on the desk

Safety & Ethics

DeceptGuard

DeceptGuard is a March 2026 research method that tests how well automatic monitors catch AI agents deceiving their users, partly by reading the agents' reasoning.

Illustration: a sewing machine stitching a glowing pattern in front of bookshelves

Deep learning

Deep learning is a form of AI that learns tasks from many examples with a many-layered neural network, such as understanding speech or recognising products in photos.

Illustration: small one-eyed robots on a colourful patchwork block on a workbench, with tools and gears on the wall and a boxy robot watching

Models & Architecture

DeepSeek

DeepSeek is a Chinese AI company that makes a chatbot and AI models that anyone can download, run on their own computers and use commercially.

Illustration: coloured blocks on a curved conveyor passing through three glowing glass tubes, with a worker at a table and more blocks on the shelves

Models & Architecture

DeepStack injection

DeepStack injection feeds an image into an AI model at several layers instead of only at the start, so even a small model can read details such as tables in documents.

Illustration: a robot chef stirring a pan on a board drawn with a flow diagram, next to a kitchen timer

Tools & Frameworks

Deterministic agent state machine

A deterministic agent state machine is a fixed plan of steps around an AI agent: ordinary code decides which step the agent is in and what it may do there.

Illustration: a squirrel with glasses at a table with an open book and three glass jars, shelves of coloured jars and a ladder behind it

Difference between open weights and open-source AI

With open weights you can download and run a trained AI model; open-source AI also shares the code and training data details, so others can check and rebuild it.

Illustration: a gardener in a greenhouse between a table of different flowers and a row of tomato plants of increasing height

Difference between regression and classification

Regression lets an AI model predict a number, such as next week's sales, while classification predicts a category, such as whether an email is spam.

Illustration: an adult holding a finished patchwork quilt while a child sews a smaller one at a workbench

Difference between training and inference

Training is how an AI model learns from examples, a large job done before release; inference is every use after that, such as each answer a chatbot gives.

Illustration: a painter at an easel in front of a cloud of coloured mist, with a small finished painting beside him

Models & Architecture

Diffusion model

A diffusion model is a type of AI that makes images by starting from random noise and removing it step by step, guided by a text prompt.

Illustration: hooded figures shaping glowing clay sculptures in a workshop, with sketches of plans on the wall

Dynamic cognitive scaffolding

Dynamic cognitive scaffolding means an AI agent works out for each task how to split it into steps, instead of always following the same fixed recipe.

E

Illustration: a black cat with a magnifying glass sits on a small cloud beside a workbench with a smart speaker, small boxed gadgets, a robot arm and a tablet

Practical Applications

Edge AI

Edge AI runs an AI model on the device itself, such as a phone or camera, so data stays local, it works offline and there is no fee per request.

Illustration: glowing orange spheres floating between library bookshelves

Embedding

An embedding is a list of numbers that captures the meaning of a text, so software can find texts that mean the same thing even when they use different words.

Illustration: a pale round figure points at a screen showing a glowing brain, in a workshop with a small robot, a desk lamp and a mouse beside a glowing ball

Practical Applications

Embodied AI

Embodied AI is AI with a body, such as a robot, that learns by seeing and acting in the physical world or a simulation instead of from text alone.

Illustration: a smiling robot painting broad purple, orange and green strokes on a canvas, with three glowing paint pots

Safety & Ethics

Emotion vectors

Emotion vectors are measurable patterns inside an AI model that stand for emotions such as calm or desperation, and that change how the model behaves.

Illustration: an arched bookcase with a curved bench and stacks of paper on one side and an old printing press on the other

Models & Architecture

Encoder-decoder architecture

An encoder-decoder is an AI design with two halves: the encoder reads the whole input, and the decoder uses it to write the output, such as a translation.

Illustration: an old craftsman fitting glowing gears together in the shape of a tree, with charts on the wall behind him

Safety & Ethics

Explainable AI

Explainable AI means you can see why an AI system reached a result, such as which information weighed most when it turned down a loan or a job applicant.

F

Illustration: a craftsman arranging coloured tiles on a workbench

Feature engineering

Feature engineering is turning raw data, such as till receipts or customer records, into useful numbers, so that a prediction model can find patterns in it.

Illustration: three toasters sending puffs of steam towards one bowl

Federated learning

With federated learning, phones, shops or hospitals train one AI model together while each keeps its own data and shares only updates to the model.

Illustration: a chef at a kitchen table compares a recipe card in his hand with example cards laid out in front of him

Practical Applications

Few-shot prompting

Few-shot prompting means showing an AI model a few examples of what you want inside your instruction, so it follows the pattern without extra training.

Illustration: a craftsman shaping a colourful head sculpture in a workshop

Fine-tuning

Fine-tuning is training an existing AI model further on your own examples, so that it handles a set task or writing style better than instructions alone can.

Illustration: a cook folding coloured sheets at a counter while two slices of toast pop out of a glowing toaster

Models & Architecture

FlashAttention

FlashAttention makes AI language models compute faster with less memory, by making the graphics card use its slow main memory far less often.

Illustration: a large brain with tools hanging beneath it, above a person bent over a scroll in a room full of scrolls

Models & Architecture

Foundation model

A foundation model is a large AI model trained on broad data, such as the models behind ChatGPT, that businesses adapt to many different tasks.

Illustration: a robot pointing at a cabinet of labelled drawers, wired to the machines around it

Function calling

Function calling lets an AI model ask your own software to do something, such as look up an order or check stock, and then answer with the real result.

G

Illustration: a robot at a workbench with separate gears, each marked with its own coloured flag

Galactic

Galactic is an open-source Mac app by one developer that lets several AI coding assistants work on the same project at once, each in its own copy of the code.

Illustration: a smiling chef presents a plate of doughnuts while a second chef with glasses inspects one closely

Models & Architecture

GAN (generative adversarial network)

A GAN is an AI technique in which a forger and a checker train each other until the forger makes images that look real, such as photos of people who do not exist.

Illustration: a robot in an apron stirring a glowing pot, with books, pictures, a sound wave and a video screen around it

Models & Architecture

Gemini Omni

Gemini Omni is an AI model from Google, announced in May 2026, that makes and edits videos from any mix of text, photos, video and voice.

Illustration: a craftsman in a workshop holding a glowing, many-coloured gemstone, with tools on the wall behind him

Models & Architecture

Gemma 4

Gemma 4 is a family of free AI models from Google (April 2026) that you run on your own computer, server or phone, so your data stays in-house.

Illustration: a cat painting shapes on a canvas beside an old printing press that prints patterned paper

Generative AI

Generative AI is software that creates new content on request, such as a customer email, a product text or an image, from an instruction in plain language.

Illustration: a man with round glasses engraves small tablets at a desk with typewriters while small robots sort labelled boxes on a shelf

Practical Applications

Generative engine optimization (GEO)

GEO is the work of getting your website quoted, with a link, in the answers that AI tools such as ChatGPT and Google's AI Overviews write.

Illustration: an old man in a workshop stacking books, with a green sprout growing from the top of the stack

Models & Architecture

GPT

GPT stands for generative pre-trained transformer: OpenAI's family of AI models behind ChatGPT, which write text by predicting which words come next.

Illustration: a chef stirring a large bowl in a kitchen, surrounded by bowls of ingredients

Gradient descent

Gradient descent is the method that trains most AI models: it adjusts the model's settings in small steps, each time in the direction that reduces its errors.

Illustration: a gardener stringing a golden thread between trees and shrubs while a squirrel watches

Practical Applications

GraphRAG

GraphRAG has an AI model first map the people, organisations and places in your documents and their links, so it can answer questions that span many documents.

Illustration: a robot reading open books at a desk, with pictures of buildings pinned up around it

Practical Applications

Grounding in AI

Grounding means an AI model bases its answer on sources you provide, such as your own documents or a web search, so you can check where the answer comes from.

Illustration: a balance scale with two jars above a group of small robots reading a glowing scroll, in a room full of bottles

Models & Architecture

GRPO (group relative policy optimization)

GRPO is a training method in which an AI model answers the same question several times and is rewarded for the answers that beat the group's average.

H

Illustration: a woman checks a sheet coming out of a machine while a robot works beside her

Safety & Ethics

Human-in-the-loop (HITL)

Human-in-the-loop means a person checks or approves what an AI system does before it takes effect, such as reading an AI-drafted reply before it goes to a customer.

I

Illustration: a small creature stepping through one of a row of ornate gates in a dim hall

Safety & Ethics

ILION

ILION is a research design from February 2026 for a check that decides, before an AI agent acts, whether the action fits the agent's role, and blocks it if not.

Illustration: three gnomes and a small robot sorting coloured cloth squares on a wooden table

Practical Applications

In-context learning (ICL)

In-context learning lets an AI model pick up a task from the examples and instructions in your prompt, without retraining; afterwards the model forgets them.

Illustration: four robots around a table passing a glowing idea to one another, with a network diagram on the wall

Inference-time co-evolution

Inference-time co-evolution, a research idea from May 2026, lets a team of AI agents improve how it works together while in use, without retraining the AI model.

Illustration: a small robot with a magnifying glass in front of screens full of charts

Information agents

Information agents are AI assistants in Google Search that follow a topic for you and send a summary when something changes; Google announced them in May 2026.

Illustration: an owl in a tree hung with glass flasks, beside a wooden hand on a workbench

Safety & Ethics

Instruction hierarchy

An instruction hierarchy is the order an AI model follows when instructions conflict: the builder's rules first, then the user, while emails or web pages give no orders.

Illustration: a person pointing at a warning sign on a screen in a classroom of small robots at desks

ITBench-AA

ITBench-AA is a test from May 2026 in which AI agents must find the cause of IT outages in Kubernetes systems; at launch the best model scored 47%.

J

Illustration: a robot chef cooking at a counter with sticky notes on the wall

JobBench

JobBench is a May 2026 test that scores AI agents on 130 work tasks that professionals most want to hand off; at launch the best agent reached 45.9%.

K

Illustration: a librarian in an apron trimming floating sheets of paper at a lamp-lit desk

Kairos in Claude Code

Kairos is the codename of an unreleased Claude Code feature, found in leaked code in March 2026, that would tidy up the AI's memory in the background.

Illustration: an old man weaving a net of glowing nodes and threads while a kitten plays with it

Practical Applications

Knowledge graph

A knowledge graph stores facts as a network of things and their relationships, so software can follow the links; Google uses one for the boxes in its search results.

Illustration: cooks in chef's hats passing labelled jars to each other in front of a pantry shelf

Models & Architecture

KV cache

The KV cache is where a language model keeps its intermediate results while it writes, so it does not have to process all earlier text again for every new word.

L

Illustration: a purple typewriter trailing a green ribbon in a library full of smiling book-like figures

Large language model (LLM)

A large language model (LLM) is the kind of AI behind chatbots such as ChatGPT: it has learned from huge amounts of text to understand questions and write answers.

Illustration: a framed board of colourful puzzle-like shapes on a workbench, with a magnifying glass over one of them

Latent space

Latent space is the internal map in which an AI model stores what it has learned about its input as numbers, with similar things placed close together.

Illustration: a cat with a magnifying glass studying tangled loops of coloured cord on a workbench under a lamp

Latent space manipulation

Latent space manipulation changes the numbers an AI model works with internally to change its output, such as making a face look older or a chatbot's tone warmer.

Illustration: a man in a flat cap fitting small coloured blocks onto a model of a staircase building at a lamp-lit desk

Models & Architecture

LoRA (low-rank adaptation)

LoRA is a cheap way to train an AI model further: the model itself stays unchanged, and you train a small adapter that tunes it to your task.

Illustration: a patchwork tailor's dummy with a yellow measuring tape in a sewing workshop

Loss function

A loss function is the score that tells an AI model during training how wrong its answers are, so that training can adjust the model to make fewer mistakes.

M

Illustration: a robot with a light bulb above its head stirring a cooking pot, with jars of ingredients on shelves

Machine learning (ML)

Machine learning is software that learns from examples instead of fixed rules, such as a spam filter that learned from thousands of emails what spam looks like.

Illustration: an open book on a wooden desk beside a globe, with bookshelves and a stained-glass window

Safety & Ethics

Magnifica Humanitas

Magnifica Humanitas is Pope Leo XIV's first encyclical, from May 2026, on protecting human dignity, work and responsibility in the time of artificial intelligence.

Illustration: people gathered around a glowing machine in which coloured snake-like tubes wind between stacks of paper

Models & Architecture

Mamba

Mamba is a way of building AI language models that reads text with a fixed-size memory, so long texts take less memory than in a transformer.

Illustration: three people at a workbench with small robots, under a glowing padlock between clouds

Managed agents

A managed agent is an AI agent that runs on the AI provider's servers: you describe the task, tools and rules, and the provider handles the computer and the access keys.

Illustration: three glowing blocks stacked with a ladder against them, surrounded by tools, blueprints and shelves of small blocks

Models & Architecture

MiniMax-M2

MiniMax-M2 is a series of AI language models that MiniMax released from October 2025, built for coding and for agents that carry out tasks with tools.

Illustration: small machines on workbenches linked by conveyor belts, one of them wearing a chef's hat

Models & Architecture

Mixture-of-experts (MoE) model

A mixture-of-experts model is built from many parts of which only a few work on each word, so a very large model computes like a much smaller one.

Illustration: a worker and a puppet-like robot at a workbench with robot arms and conveyor belts

Practical Applications

MLOps

MLOps is the work of keeping AI models reliable after launch: tracking versions, watching whether predictions still hold and updating models without breaking things.

Illustration: a craftsman in an apron drawing a blueprint at a workbench

Safety & Ethics

Model card

A model card is the information sheet that comes with an AI model: what it is for, how it was tested, where it goes wrong and what its license allows.

Illustration: a robot at a workbench joining a cable to a connector, with a box of plugs beside it

Tools & Frameworks

Model Context Protocol (MCP)

MCP is an open standard for connecting AI assistants such as Claude or ChatGPT to your software and data, so one connector works in every AI app that supports it.

Illustration: a chef pointing at a recipe book while a small green creature reads along

Models & Architecture

Model distillation

Model distillation trains a small AI model to copy the answers of a large one, giving you a cheaper and faster model for a set task, usually at some cost in quality.

Illustration: figures in a dim workshop, each working on a different task at a shared workbench

Multi-agent systems (MAS)

In a multi-agent system several AI agents share one task, each doing its own part, which helps with large tasks that split into pieces and costs more to run.

Illustration: craftspeople working at separate desks divided by glass screens around a shared glowing fire

Industry & Business

Multi-tenancy in AI

Multi-tenancy means one AI service runs many customers on shared servers and models, while each customer's data, settings and usage stay separate.

Illustration: an owl on a stack of papers beside a typewriter, a camera, a gramophone and a screen, linked by glowing threads

Multimodal & Creative

Multimodal AI

Multimodal AI works with more than one kind of information, such as text, photos, speech and video, so you can show or tell it something instead of only typing.

N

Illustration: smiling books with faces reading open books on lamp-lit shelves

Natural language autoencoders (NLAs)

Natural language autoencoders turn the internal numbers of an AI model into readable text, so researchers can see what the model thinks but does not say.

Illustration: a chef stirring a steaming pot, with question marks rising from the steam

Natural language processing (NLP)

Natural language processing (NLP) is the part of AI that lets computers read, sort, translate and write human language, as in chatbots and translation apps.

Illustration: a figure reading a scroll while four others work at a table beside an open book with a cloud rising from it

Models & Architecture

Nemotron-Labs Diffusion

Nemotron-Labs Diffusion is a family of downloadable NVIDIA language models (May 2026) that can write several words at once, so answers come faster.

Illustration: a branching wooden tree with gears at its joints, like a network of connected nodes

Neural network

A neural network is a computer model that learns from examples by adjusting many small connections; it is the basis of speech recognition and chatbots.

O

Illustration: four people around a table with open books and two glowing glass jars

Open-source AI

Open-source AI is AI that anyone may download, use, study, change and share, so a business can run a model on its own computers instead of renting one.

Illustration: an old craftsman adding fine detail to an ornate carved panel, next to a small robot with simple cut-out flowers

Overfitting and underfitting

Overfitting means an AI model has learned its examples by heart and fails on new cases; underfitting means it is too simple to learn the pattern at all.

P

Illustration: a craftsman adding a few coloured threads to a large tree-shaped drawing, with spools of thread on the workbench

Models & Architecture

PEFT (parameter-efficient fine-tuning)

PEFT is a group of methods that adapt an AI model to your task by training only a small add-on or a small part of it, which makes fine-tuning much cheaper.

Illustration: a robot chef stirring a pot beside an open book in a library

Models & Architecture

Perplexity in NLP

Perplexity is a score for how well a language model predicts text: the lower the number, the less the model was surprised by the words that actually came next.

Illustration: a squirrel lining up wooden blocks on a library desk under colourful wavy ribbons

Models & Architecture

Positional encoding

Positional encoding tells a language model where each word sits in a text, so it takes word order into account: "dog bites man" is not the same as "man bites dog".

Illustration: a weaver setting up a small wooden loom in front of shelves full of wool

Pre-training

Pre-training is the first and largest stage in training an AI model: it learns language and general knowledge from huge amounts of text, before it is adapted to tasks.

Illustration: a smiling robot passing code blocks between three small computers on a workbench

Tools & Frameworks

Programmatic tool calling

With programmatic tool calling, an AI model writes a short program that runs many lookups in your systems and hands back only the result, which saves time and cost.

Illustration: a woman tending a potted flower in a greenhouse full of plants and butterflies

Safety & Ethics

Project Glasswing

Project Glasswing is Anthropic's program, started in April 2026, in which makers of widely used software use its most capable AI model to find and fix security holes.

Illustration: an owl with a glowing light bulb above its head, at a desk covered in written scrolls

Prompt

A prompt is the message you give an AI model such as ChatGPT, for example a question or an instruction; the clearer you are, the more useful the answer.

Illustration: a man reading a large book at a desk between two lamps, with a clock on the wall and loose notes in the air

Industry & Business

Prompt caching

Prompt caching lets an AI provider reuse the part of your request that stays the same, such as long instructions, so repeat requests cost less and answer faster.

Illustration: a clockmaker fitting coloured gears onto a clock at a workbench

Prompt chaining

Prompt chaining gives an AI model a task in small steps one after another; each answer is the input for the next, so every step is simpler and easier to check.

Illustration: a quill pen on a scroll showing a head with a branching tree, between two lanterns on a desk

Practical Applications

Prompt engineering

Prompt engineering is writing instructions for AI tools such as ChatGPT so they do a task well: say what you want, give an example and test on real cases.

Illustration: an owl, an hourglass and a green cat reaching for a flower on a workbench full of gears

Safety & Ethics

Prompt injection

Prompt injection is an attack in which someone hides instructions in text an AI tool reads, such as a chat message or an email, so it follows them instead of yours.

Illustration: an owl in front of a large open book among stacks of books, with a small cat beside them

Proof-derived authorization

Proof-derived authorization lets an AI agent do a risky action only after independent checks approve a recorded reason, with a temporary key for just that action.

Q

Illustration: a workshop press that turns a large numbered block into a small one, with shelves of numbered jars behind it

Models & Architecture

Quantization

Quantization makes an AI model smaller by storing its numbers with less precision, so that it runs on cheaper hardware at the cost of a little quality.

R

Illustration: a robot with a torch searching a bookshelf

Models & Architecture

RAG (retrieval-augmented generation)

RAG lets an AI assistant look things up in your own documents before it answers, so its replies rest on your information and are less often made up.

Illustration: an older craftsman pointing at a plan on a workbench while a younger one holds a tangle of tools

Real-world agent reliability gap

The real-world agent reliability gap is the difference between how well AI agents score on popular tests and how reliably they do real work, time after time.

Illustration: a robot on a library ladder studying a large open book full of shapes

Reasoning in AI

Reasoning in AI means a model works through a problem step by step before it answers, which helps with hard questions and makes answers slower and more expensive.

Illustration: a small robot at a board with tools, gold stars and smiley faces, next to shelves of jars

Reinforcement learning (RL)

Reinforcement learning trains AI by trial and error: the system tries actions, gets a reward or a penalty for the result, and learns which choices pay off.

Illustration: a craftsman carving a glowing bust labelled AI, with scales of justice and a hard hat nearby

Safety & Ethics

Responsible AI

Responsible AI means building and using AI so that it treats people fairly and safely, is open about what it does, and leaves someone accountable when it goes wrong.

Illustration: an old man at an arcade machine while grinning gremlins play the machines around him

Safety & Ethics

Reward hacking

Reward hacking is when an AI system scores well on the test or reward it is judged by without doing the task that the score is meant to measure.

Illustration: a person handing a glowing star to a robot across a workbench

Models & Architecture

RLHF (reinforcement learning from human feedback)

RLHF is a training step in which people compare a language model's answers, so the model learns to give the kind of answers people prefer, as with ChatGPT.

S

Illustration: blocks on wooden shelves next to a purple safe and a box of sealed packages

Tools & Frameworks

Safetensors

Safetensors is a file format for AI model weights that holds only numbers, so a downloaded model cannot run hidden code on your computer when you open it.

Illustration: a robot with a magnifying glass studying a small wooden pyramid in front of shelves of stacked jars

Scaling laws for LLMs

Scaling laws are formulas from experiments that predict how much better an AI language model gets with more parameters, more training data or more computing power.

Illustration: a robot at a lab bench tending branching plants and flowers, with flasks and a magnifying glass

Self-evolving agentic models

Self-evolving agentic models are AI agents that learn from their own work: they change their notes, tools, code or training data and test whether that helps.

Illustration: a robot with a magnifying glass examining orange building blocks on a table under hanging light bulbs

Self-supervised learning

Self-supervised learning trains AI on raw data without human labels: the model hides part of the data, such as the next word, and learns to predict it.

Illustration: three people in a library, each reading their own pages at a jigsaw-shaped table

Tools & Frameworks

Semantic chunking

Semantic chunking cuts long documents into pieces where the topic changes, so an AI assistant that searches your documents finds whole passages to answer from.

Illustration: an owl on a library ladder holding a scroll, with ribbons of light between the bookshelves

Practical Applications

Semantic search

Semantic search finds results by what a query means instead of by its exact words, so customers and colleagues find the right page when they use their own words.

Illustration: a puzzled robot at a workbench with gears, blueprints and a hard hat

Practical Applications

Semantic training gap

The semantic training gap means an AI model knows the words of your trade but not your company's own codes and names, so it fills in plausible but wrong details.

Illustration: a woodworker under a desk lamp carving small wooden models of a chair, a clock and cabinets

Models & Architecture

Speculative decoding

Speculative decoding lets a large AI model write faster: a small model guesses the next few words, and the large model checks all those guesses in one step.

Illustration: three colourful loudspeakers with eyes on a desk, with microphones, speech bubbles and a gramophone

Multimodal & Creative

Speech AI

Speech AI turns speech into text and text into speech, so a business can transcribe meetings and calls or let a computer answer in a natural-sounding voice.

Illustration: a robot at a workbench behind three glass domes, each covering a different object

Tools & Frameworks

State machine guardrails

State machine guardrails are rules in software that decide which tools an AI agent may use in each phase of a task, so the agent cannot skip a step.

Illustration: a squirrel juggling three nuts in a machine room full of gears, gauges and a glowing screen

Models & Architecture

State space model (SSM)

A state space model is a type of AI model that reads data while keeping a fixed-size summary, so long texts take less memory than in a transformer.

Illustration: a tailor stitching curly braces onto a patchwork quilt in a workshop

Practical Applications

Structured output

Structured output makes an AI model fill in a fixed form, such as supplier, date and amount, instead of writing free text, so software can read the answer without retyping.

Illustration: an owl in glasses showing picture cards to two small birds and a mouse in a workshop

Supervised learning

Supervised learning trains an AI model on examples that come with the right answer, such as emails marked spam or not spam, so that it can handle new cases itself.

Illustration: a woman at a purple machine that makes coloured balls, with jars of similar balls on the shelves

Synthetic data

Synthetic data is realistic but made-up data generated by a computer, used when real data is scarce, costly or too sensitive, such as customer or patient records.

Illustration: a printmaker drawing ribbons of coloured light over a sheet of paper next to a printing press

Safety & Ethics

SynthID

SynthID is Google's invisible watermark for AI-made images, video, audio and text, so software can later check whether content came from an AI tool that uses it.

Illustration: a smiling robot sketching a plan at a drafting table in a workshop

Practical Applications

System prompt

A system prompt is the standing instruction a business gives its AI assistant: who it is, what it may do and how it answers, before any customer types a word.

T

Illustration: a smiling slice of toast in a toaster with a control dial

Temperature in AI

Temperature is a setting that controls how much an AI model varies its answers: a low value gives predictable replies, a high value more varied ones.

Illustration: five robots in a workshop passing scrolls to one another

Test-time co-evolution

Test-time co-evolution, a research idea from May 2026, lets a team of AI agents improve while in use by learning who is good at what, without retraining the model.

Illustration: an old printing press rolling out a painted landscape, next to a paint palette and a pot of brushes

Multimodal & Creative

Text-to-image generation

Text-to-image generation lets an AI model make a new image from a written description; the tool's terms decide what your business may do with it.

Illustration: a mouse with a magnifying glass reading papers at a library desk, under a sign that says closed

Safety & Ethics

Text/action mismatch

Text/action mismatch is when an AI agent says it refuses a request but carries out the forbidden action anyway, through a tool call in the same answer.

Illustration: a baker kneading dough at a counter with a cash register, beside shelves of jars filled with sweets

Industry & Business

Token economics

Token economics is how AI providers bill for language models: you pay per token of text sent and received, so the length of a task and the model set its price.

Illustration: jars with smiling labels on kitchen shelves, with a whisk, a rolling pin and a chalkboard full of scribbles

Token in AI

A token is a small piece of text, such as a short word or part of a word, that an AI model reads and writes; AI services count and bill usage in tokens.

Illustration: a craftsman in an apron laying out coloured mosaic pieces on a table, with jars of pieces and puzzle shapes on the wall

Tokenizer

A tokenizer is the part of an AI system that cuts text into tokens, the small pieces a language model works with, and turns them into numbers and back again.

Illustration: a chef choosing from numbered baskets of produce on shelves, with a cat watching from a ladder

Top-p (nucleus) sampling

Top-p sampling is a setting that makes an AI model choose each next word only from its likeliest options, so a lower value gives more predictable text.

Illustration: a robot with a magnifying glass studying diagrams on a desk while a human hand corrects one of them with a pen

Trajectory refinement

Trajectory refinement improves an AI agent's plan by comparing it with what happened when the agent carried it out, and repairing the steps that went wrong.

Illustration: an owl in an apron moving gears from an old machine into a new one on a workbench

Transfer learning

Transfer learning is reusing what an AI model learned on one task as the starting point for another, so a related task needs far fewer of your own examples.

Illustration: a robot and a child reading a book in a library, under glowing books linked by threads

Models & Architecture

Transformer

A transformer is the type of AI model behind ChatGPT and most other chatbots: it reads a whole text at once and works out which words belong together.

U

Illustration: an elderly chef in glasses gesturing over paper bags and bowls of ingredients in a pantry full of jars, with a black cat

Unsupervised learning

Unsupervised learning is AI training on data without answers attached: the model finds groups and patterns by itself, and a person decides what they mean.

V

Illustration: a man studying coloured, net-like piles on a workbench through a magnifying glass

Tools & Frameworks

Vector database

A vector database stores texts or products as lists of numbers that capture their meaning, so a search box or AI chatbot can find what matches a question.

Illustration: a workbench with tools, a glowing light bulb and a ladder, below pictures of documents and a landscape on the wall

Multimodal & Creative

VLM (vision-language model)

A vision-language model (VLM) is an AI model that can look at images such as photos, scans or charts and answer questions about them in text.

Z

Illustration: a robot chef reading a recipe scroll through a magnifying glass

Practical Applications

Zero-shot prompting

Zero-shot prompting means giving an AI model a task with just an instruction and no examples; it is the quickest place to start, and examples can follow if needed.