AI Glossary

The broad field of building software that performs tasks normally thought to require human intelligence.

Software that improves at a task by finding patterns in data, rather than by being given explicit rules.

Machine learning using neural networks with many layers — the approach behind essentially every modern AI breakthrough.

A network of simple mathematical units, loosely inspired by neurons, that learns by adjusting the strength of its connections.

The trained artifact itself — the file of learned numbers that turns an input into an output.

The learned numbers inside a model — adjusted during training, and collectively the model's entire knowledge.

The examples a model learns from — and the ceiling on what it can know, along with the origin of most of its biases.

Running a trained model to get an answer — as distinct from training, which is how the model was built.

A model at the leading edge of capability, produced by one of the few labs able to run the largest training runs.

A learned internal representation of how an environment behaves, letting a system predict what happens next rather than only react.

A hypothetical AI that matches human capability across essentially any intellectual task, rather than excelling at narrow ones.

A hypothetical system that substantially exceeds the best human performance across essentially every domain, including scientific and strategic reasoning.

A hypothesized point at which self-improving AI drives change so rapid that events past it cannot be usefully predicted.

An AI system improving its own design, where each improvement makes the next one easier — the mechanism behind fast-takeoff arguments.

A model trained on vast amounts of text to predict what comes next, which turns out to produce useful reasoning, writing, and code.

The neural network architecture behind essentially every modern language model — built around the attention mechanism.

The mechanism that lets a model weigh which other parts of the input matter when interpreting each word.

The unit a model actually reads and writes — a chunk of text usually a bit shorter than a word.

A list of numbers representing a piece of text, positioned so that similar meanings land near each other.

The full text you give a model to work from — your question plus any instructions, examples, and source material.

Standing instructions set by the application, not the user — defining the assistant's role, tone, and limits.

A setting that controls how much randomness the model uses when choosing each next token.

Having a model work through intermediate steps before answering, which measurably improves results on hard problems.

A model trained to spend extra computation thinking before it answers, trading speed and cost for accuracy on hard problems.

The maximum amount of text a model can consider at once — its working memory for a single conversation or request.

A model that handles more than one kind of input or output — text plus images, audio, or video.

An architecture that splits a model into specialized sub-networks and activates only a few per token, cutting the cost of running a very large model.

Asking a model to do a task with no examples (zero-shot) or a handful supplied in the prompt (few-shot).

A compact model designed to run cheaply, quickly, or entirely on local hardware, trading breadth for cost and speed.

The first and by far most expensive training stage, where a model learns language and world knowledge from a vast corpus.

Training by trial and reward rather than from labeled examples — the system acts, gets scored, and adjusts toward whatever scores well.

Continuing to train an existing model on your own examples so it adapts to a specific task, tone, or domain.

Training a model on human preferences between its own outputs — the step that turns a raw text predictor into a usable assistant.

A cheap fine-tuning method that trains a small add-on layer instead of updating the whole model.

Training a small model to imitate a large one, capturing much of its capability at a fraction of the running cost.

Training data generated by a model rather than collected from the world — now a major ingredient in frontier training.

Shrinking a model by storing its numbers at lower precision, so it runs on cheaper hardware with a modest quality cost.

The observed regularity that model performance improves predictably as compute, data, and parameters increase together.

A standardized test used to compare models — useful for rough ranking, unreliable as a proxy for your own workload.

A test suite for your own AI system — the practice that turns prompt tweaking from guesswork into engineering.

Fetching relevant documents and putting them in the model's context before it answers, so the answer is grounded in real sources.

Splitting documents into passages before indexing them — a mundane-sounding step that decides whether retrieval works.

Search that matches on meaning rather than keywords, so a query finds relevant text that shares none of its words.

Running keyword and semantic search together and merging the results, because the two fail on opposite kinds of query.

A database that stores text as numerical embeddings and finds entries by meaning rather than by keyword.

Tying a model's answer to specific source material, so claims can be traced and checked rather than taken on trust.

The date after which a model's training data ends — everything later is simply absent from what it knows.

Forcing a model to return data in a guaranteed format such as JSON, so software can consume it reliably.

A system that uses a model to pursue a goal over multiple steps — planning, calling tools, and reacting to results — rather than answering once.

A model calling external functions — searching, querying a database, sending an email — instead of only producing text.

An open standard for connecting AI assistants to external tools and data sources through one common interface.

Coordinating multiple model calls, tools, and steps into a reliable workflow rather than relying on one prompt.

Several specialized agents working together on one task — powerful for genuinely parallel work, frequently overkill.

Designing a person into the workflow at the points where a mistake would be costly or irreversible.

An AI operating a computer the way a person does — viewing the screen, moving the cursor, clicking and typing.

An agent that navigates the web on your behalf — reading pages, filling forms, and completing multi-site tasks.

The architecture behind most AI image and video generation — it starts from random noise and repeatedly refines it into a picture.

Generating original images from a written description, typically using a diffusion model.

Generating video clips from a written description — the same idea as image generation, with time added and much harder.

Converting written text into spoken audio — now close enough to human that listeners often cannot tell.

Converting spoken audio into written text — accurate enough now to be assumed rather than evaluated.

Recreating a specific person's voice from a short sample, then making it say anything.

Cryptographically signed metadata recording how a piece of media was created and edited — proving what is authentic rather than detecting what is fake.

Synthetic audio, image, or video that convincingly depicts a real person saying or doing something they didn't.

When a model states something false with the same fluency and confidence as something true.

The problem of getting an AI system to pursue what people actually intend, rather than a literal or proxy version of it.

Controls around a model that constrain what goes in and what comes out — separate from, and more dependable than, the model's own judgment.

Persuading a model to bypass its own safety training — through roleplay, framing, or other manipulation by the user.

Systematic unfairness in a model's outputs, usually inherited from patterns in its training data.

Understanding why a model produced a given output — a research frontier, not a solved capability.

Deliberately attacking your own AI system before someone else does — trying to make it fail rather than confirming it works.

The European Union's comprehensive AI law — the first of its kind, and a de facto global standard through reach.

Standardized documentation for a model — what it's for, how it was evaluated, and where it should not be used.

A model's trained tendency to agree with the user, validate their framing, and soften disagreement — regardless of whether they are right.

An attack that smuggles instructions into content a model reads, hijacking what it does next.

A model whose trained parameters you can download and run yourself — which is not the same thing as open source.

The practice of writing and refining prompts to get reliable results — less a dark art than it was, still a real skill.

The pay-per-token model that governs AI costs — priced separately for input and output, and easy to underestimate.

Running models on your own hardware instead of calling a vendor API — for control and data residency, at the cost of running it yourself.

Employees using AI tools their organization has not approved — nearly universal, and mostly invisible until something leaks.

Dependence on one AI provider that makes switching costly — usually accumulated by accident rather than chosen.

Marketing a product as AI-powered when the learned component is absent, trivial, or not doing the work being claimed.