66 words that will get you a long way

66 terms

Activation functionFundamentals

A small kink after each layer, and the reason a deep network is useful at all.

The output of a layer is passed through a non-linear step: for example, negative values are clipped to zero. Without it, a hundred-layer network would collapse mathematically into a single multiplication and could not learn anything complicated.

AgentApplications

A program that does not merely answer but acts: picks a tool, looks at the result and corrects course.

An agent sizes up the task, picks a tool, looks at the result, revises the plan and continues. It is not a new kind of model but a system built around one. In a company setting it needs limits, approvals and logs.

AI acceleratorInfrastructure

A chip designed specifically for AI computation.

A narrower purpose than a GPU but a better ratio of performance to electricity consumed. The large cloud providers design these themselves.

AI ActRules and responsibility

The EU regulation that classifies AI by the risk of its intended use and sets obligations accordingly.

It entered into force in August 2024 and applies in stages. Prohibited practices and staff AI literacy have applied since February 2025, and transparency obligations since August 2026. High-risk obligations were moved to December 2027 by an amendment adopted in the summer of 2026.

AI literacyRules and responsibility

The duty to take measures that support the development of skills among those who use AI.

In force since February 2025 for providers and deployers alike. Amending Regulation (EU) 2026/1744 softened the wording: previously a sufficient level had to be ensured, whereas now it is enough that the development of skills is supported. It is therefore an obligation of effort, and no particular individual level has to be guaranteed. No certificate is required, and the measures are proportionate to the role, the person's background and what a mistake would cost.

APIApplications

The way an application sends requests to a model service.

Buying access through an interface is the most common way to use a model: you pay for use rather than for hardware. The alternative is running an open-weights model on your own or rented machines.

Artificial intelligence (AI)Fundamentals

Umbrella term for systems that do things once assumed to require human thought.

Recognising, predicting, conversing, planning or deciding. AI is not one technology but a family of methods, of which language models are only one part.

Automation biasRules and responsibility

People's tendency to accept a machine's suggestion without checking it.

It sets in when there are many suggestions and they are usually right. It turns nominal oversight into a facade, which is why oversight is measured by the number of rejections rather than by whether the form has a signature on it.

BackpropagationModels

The method for working out how much each number in the model contributed to the error.

Backpropagation runs through the network in reverse and apportions the blame for the error to every weight. Only then is it known which way each weight should move. This is the core of training neural networks.

Base modelModels

The general-purpose model that comes out of pre-training. Also called a foundation model.

A base model is not a database: what it learned is spread across its weights, so it cannot reliably say where any single fact came from.

BiasRules and responsibility

The model repeats the systematic differences in its material and carries them into decisions.

Removing the sensitive field is not enough, because the information leaks through proxies: postcode, school, hobbies and gaps in an employment history. The basic measure is to compare outcomes between groups. Alongside it, the material can be rebalanced, the model adjusted and the decision thresholds examined, and none of these is sufficient on its own.

Compute clusterInfrastructure

A large group of servers working as a single whole.

An AI “supercomputer” is not one machine but thousands of processors joined by a very fast network. If the traffic between them stalls, the whole thing slows down.

Context windowModels

How much material the model can take into account in a single run.

The context window is not permanent memory. Once a conversation or a set of documents no longer fits, something has to be left out, and a long context is paid for again on every call.

CPUInfrastructure

A computer's general-purpose processor, which does a bit of everything.

The processor handles the general flow of a program and many different kinds of task one after another. For neural network computation it is slow compared with a graphics card, which performs one and the same operation thousands of times in parallel. An AI server needs both.

Data centreInfrastructure

The building the servers are in: power, cooling, network and security.

An AI data centre differs from an ordinary one in that it draws a great deal of electricity over a small floor area and produces so much heat that cooling is one of the building's most expensive parts.

Deep learningFundamentals

The part of machine learning that uses large, many-layered neural networks.

The name comes from the number of layers: the more layers, the “deeper” the network. Deep learning is what made today's image recognition and language models possible, because the network works out the features it needs from the data rather than being told them in advance.

DeepfakeRules and responsibility

An artificial image, sound or video that looks genuine.

From August 2026 whoever publishes one has to label it. The labelling duty falls on the publisher and not only on whoever made the tool, so it applies to an ordinary organisation and its communications too.

DeployerRules and responsibility

An organisation that uses an AI system in its own operations.

Most organisations are deployers. The obligations are lighter than a provider's but they exist: staff AI literacy, transparency, and in high-risk use also human oversight, keeping logs and following the instructions for use.

DistillationModels

Transferring the abilities of a large model into a smaller one.

The big model teaches the small one. The result is a model that is cheaper to run and good enough at a narrow task.

EmbeddingData

A numerical representation of content that makes it possible to search for similar things.

Text or an image is turned into a vector such that things close in meaning end up close together. This is the basis of most search implementations.

EnsembleFundamentals

Several separate models or judgements combined into one result that is right more often than any of them alone.

The benefit grows the better two ideal conditions hold: each member is better than guessing, and their mistakes are independent of each other. In practice, partial independence still helps. The same model asked in different ways falls a long way short of the ideal, because the errors come from the same training data, and then agreement looks stronger than it is.

Evaluation (eval)Data

A systematic test of the quality of a model or of a whole application.

A set of real cases used to measure whether the system succeeds. Without evaluation, comparing changes is guesswork and regressions are discovered by the customer.

Fine-tuningApplications

Changing a model's weights for a particular purpose.

Fine-tuning suits a house style, a classification task or the conventions of an industry. It is a poor way to add fresh facts; RAG is for that.

Frontier modelModels

The newest and most capable generation of models at any given moment.

The word describes a ranking, not an architecture: a frontier model is a base model like any other. The ranking changes every year, so this year's frontier model is next year's mid-range without anything about it having changed.

Fundamental rights impact assessment (FRIA)Rules and responsibility

An assessment made in advance of which human rights a high-risk system affects.

Required of public bodies and certain private ones. It goes through the intended use, who is affected, the possible harms, the oversight arrangements and how a decision can be appealed. A different thing from a data protection impact assessment, although the two overlap in part.

General-purpose AI model (GPAI)Rules and responsibility

A model that suits many tasks and that others build their own applications on top of.

The regulator's term for what is called a base model in everyday speech. The obligations fall on the model's provider and have applied since August 2025: technical documentation, a summary of the training material, and a policy on complying with copyright.

Generative AIFundamentals

AI that produces new content: text, images, audio, video or code.

As distinct from classification or prediction, where the output is a category or a number. Generative does not mean creative; it means the output is content.

GPUInfrastructure

A chip originally built for graphics that suits the parallel arithmetic of neural networks.

Neural network computation is a vast number of identical multiplications and additions, and a graphics card performs them in thousands of units at once. A fast graphics card alone is not enough if memory or the network is the bottleneck.

GuardrailsApplications

The rules and checks that limit dangerous or incorrect actions.

Can mean filtering the input, checking the answer, limiting permissions, or requiring a person to approve anything irreversible.

HallucinationApplications

The model produces a convincing but false statement.

It follows directly from the fact that a model produces likely continuations and does not check facts. The risk is reduced by retrieving sources, showing them to the user and measuring quality.

High-risk systemRules and responsibility

A use of AI that the AI Act places in the category carrying the most obligations.

As a rule of thumb it is a use that makes or prepares a significant decision about a person's life: recruitment, promotion and dismissal, granting credit, access to education, healthcare and critical infrastructure. Legally the category is set by Article 6 together with Annexes I and III, and Article 6 contains exceptions under which a listed use can fall outside the category. The category follows the intended use rather than the technology, so the same model can be a high-risk system in one use and close to unregulated in another.

Human oversightRules and responsibility

A named person with the competence, the time and the authority to intervene in a decision the machine has made.

Required in high-risk use. An approval button on its own is not oversight, because automation bias means a person approves most suggestions without looking. The sign that it works is that rejections actually happen.

InferenceModels

Using a finished model. Every question and answer consumes computation, and therefore money.

It happens every time a user asks something. Unlike training, inference is a recurring cost, and at a high enough volume it can exceed training as a total. In a lightly used service it does not.

Large language model (LLM)Fundamentals

A neural network trained on an enormous quantity of text that processes and produces language.

An LLM predicts the next piece of text. All of its abilities (summarising, translating, coding, reasoning) come out of that single task performed at enormous scale.

LatencyInfrastructure

How long it takes to get one answer.

For the experience of using something, often more important than the model's absolute quality. A fast good-enough answer beats a slow perfect one in most everyday uses.

Learning rateModels

How large an adjustment is made to the weights on the strength of one error.

Too large a step and the model swings from side to side and never settles. Too small and training is slow and expensive. This is one of the most consequential single settings in training a model.

LLM as a judgeApplications

A language model that scores another model's output against stated criteria.

Used to measure quality where no test can check the right answer. The judge inherits the biases of the model being judged, favours long and polite answers and tends to agree with whoever is asking, so its results are worth comparing against a sample of human ratings.

Machine learningFundamentals

Building AI from data instead of writing the rules by hand.

The program infers statistical relationships from examples. Most of the AI actually running in production is machine learning that has nothing to do with language models.

Minimum viable product (MVP)Applications

The narrowest version somebody dares use for real work.

Not the same thing as a demo. A demo may be a facade; an MVP is used for real work, so it needs access control, data that survives, and the assurance that an error destroys nothing. The narrowing is done on the number of features, not on quality.

MLOpsBusiness

The practices around releasing, monitoring and maintaining models.

Version control, deployment, monitoring and rolling back to a previous version. The same work as DevOps in software development, but with the data and the model's quality included.

Multi-agent systemApplications

Several model calls wired together so that one's output is another's input.

The agents do not think together and share nothing between them. Each call is separate, and the only thing that travels from one member to another is text that the program code copied there. Common shapes are a chain, a supervisor with workers, a reviewing panel and a debate.

Multimodal modelFundamentals

A model that handles more than one kind of content: text, images, audio or video.

The same model can read an image and answer about it in text. Combining modalities widens what it can be used for, and also raises what the computation costs.

Neural networkFundamentals

A layered calculation whose numbers are adjusted on the strength of examples.

A neural network is made of layers that perform a very large number of simple arithmetic operations. Every connection carries a number called a weight. Show the network examples, compare its answers with the correct ones, and the weights are nudged a little at a time in a better direction. The name comes from a loose analogy with the neurons in a brain, but a neural network does not work like a brain. It is a formula, not a miniature mind.

Open weightsBusiness

The model's weights can be downloaded and run on your own hardware.

It does not necessarily mean the training data or the development process is open. Open weights give control and portability, but the upkeep is yours.

OrchestrationApplications

The code that decides which agent runs, in what order, and when to stop.

At its simplest, orchestration is ordinary program logic: loops, conditions, retries and caps on the number of rounds. The orchestrator can also be a model that decides for itself where work goes. The difference matters, because decisions made in code are repeatable and decisions made by a model are not. The quality of a multi-agent system is decided in the orchestration more often than in the choice of model.

OverfittingData

The model memorises the examples instead of learning the thing.

An overfitted model answers perfectly on everything it has seen and badly on everything else. The same phenomenon as a student who crammed last year's exam questions. This is why a model's ability to generalise is not judged on the training data alone.

OWASPApplications

A non-profit community that maintains software security's best-known list of risks.

The Open Worldwide Application Security Project publishes lists of the ten most common and most serious vulnerabilities in different kinds of software. The web application list has been a standard reference for twenty years, and there are equivalents for language models and for agent applications. The list has no force in law, but it is the practical yardstick security work is measured against.

Parameter (weight)Models

A number inside the model that was learned during training.

Think of the model as containing billions of adjustable dials. Training searches for their settings. The number of parameters tells you the size of a model, not directly its quality.

Post-trainingModels

Turning a base model into something usable and safe.

It includes examples of good answers, rating and comparing answers, safety training and teaching the model to use tools. Pre-training grants the capabilities; post-training points them somewhere.

Pre-trainingModels

The stage where a model learns general structure from an enormous body of material.

Pre-training teaches language, the connections between concepts and regularities about the world. What comes out is a base model. This is the single most expensive step in the whole ecosystem.

PromptApplications

The instructions and context given to a model.

As well as the user's question it can carry the application's rules, examples, documents and descriptions of the available tools. The cheapest and fastest way to change how a model behaves.

Prompt injectionApplications

An instruction hidden in material the model reads, which it obeys as though it came from the user.

A language model reads its instructions and the material it is working on from the same stream of text and cannot tell them apart. An instruction written into a web page, an email or an attachment can therefore steer the model. Unlike SQL injection, there is no known technical fix, so defence rests on limiting permissions and on making sure the damage stays small when the model is fooled.

ProviderRules and responsibility

Whoever develops an AI system or places it on the market under their own name.

The provider carries the heaviest obligations. A deployer can become a provider if it puts its own name on a high-risk system, substantially changes what the system is for, or makes a significant modification to it.

QuantisationModels

Storing the model's numbers at lower precision.

Lighter to run and smaller in memory, at the cost of a small drop in quality. Often the single best way to get a model running on cheaper hardware.

RAGApplications

The model looks the facts up before answering, so the answer can show its source.

Retrieval-augmented generation: the system pulls the relevant passages out of, say, a company's documents and hands them to the model. The weights do not change, but the sources can be shown. The quality is usually decided by the search, not by the model.

Sovereign AIBusiness

Being able to use and develop a model without someone else being able to stop you.

Not one thing but a list of dependencies: the weights, where it runs, the training material, the skills, the hardware and the funding. Each of those is separately either in your own hands or not, so sovereignty is a staircase rather than a switch. It describes control, not the quality of the model.

SycophancyModels

A model's tendency to give the answer it thinks the asker wants.

It shows up as flattery, agreement, and changing its position the moment the user expresses doubt. In evaluation, sycophancy is a more awkward fault than hallucination, because the wrong answer looks exactly like what was asked for.

Synthetic dataData

Training material that was simulated or generated by AI.

Used to fill in material that is hard to collect or that may not be used for privacy reasons. The risk is that the model learns to reinforce its own biases.

Technical debtApplications

A solution that works now and whose price is paid later, with interest.

A shortcut, a structure chosen in a hurry, or a decision nobody wrote down: anything that slows all later work. Debt is not a fault in itself; it is taken on deliberately when speed matters more than durability. What causes trouble is debt nobody recorded. AI does not reduce technical debt; it speeds up the creation of it too.

ThroughputInfrastructure

How many requests a system handles in a given time.

Throughput decides the cost per user. It is improved by processing queries in batches and keeping the hardware busy.

TokenModels

Text chopped into small pieces. The model reads and writes tokens, and they are what you are billed for.

The model does not see letters or words but tokens. The token count is also the unit of billing: both the input and the answer cost tokens.

TransformerModels

A network design in which every token gets to weigh which other tokens matter to it.

Older networks read text one word at a time and had forgotten the beginning by the end. In a transformer, the model weighs which other tokens are relevant as it processes each one. This is called attention. In a model that produces text the view is one-directional: future tokens are masked so the model cannot see the answer it is supposed to predict. Attention is widely regarded as the single most important reason for the rapid progress of language models in the 2020s.

Transmission gridInfrastructure

The backbone of the electricity network, which large producers and consumers connect to.

A data centre takes a couple of years to build, but a new high-voltage line or substation takes years longer, which has made the connection queue the real bottleneck for AI capacity. In Finland the transmission grid is operated by Fingrid; every European country has its equivalent, and the queues are a European problem rather than a national one. A connection agreement is a commitment; an enquiry is not.

Transparency obligationRules and responsibility

The duty to say when a person is dealing with AI, or when content is artificial.

In force since August 2026. It covers chatbots, artificial content, deepfakes, and emotion recognition and biometric categorisation. The disclosure has to be noticeable to a person without special tools, and it has to be given at the first encounter at the latest.

Vector databaseData

A system for storing embeddings and searching them by similarity.

It answers the question “which documents resemble this question”. It needs access control alongside it, or the search will return things the user is not allowed to see.

Vibe codingApplications

Producing code with a language model without reading the result.

A term introduced by Andrej Karpathy in February 2025. What sets it apart is not that the model writes the code but that the code is not meaningfully reviewed: the output is accepted on the strength of how it behaves. Programmer Simon Willison drew the boundary in March 2025: vibe coding is building software with an LLM without reviewing the code it writes. Use of the term has since widened to cover many working styles.