Glossary
AI terms, defined in one breath
45 terms. Each definition is short enough to read in a meeting and precise enough to use in one.
A
AgentAn AI system that pursues a goal by taking actions with tools, in a loop, rather than answering a single question.AlignmentThe broad effort to make models behave as their developers and users intend, including being honest and refusing harmful requests.APIApplication programming interface: the way software calls a model, billed per token, as opposed to a chat app used by a person.AttentionThe mechanism that lets a transformer weigh which earlier tokens matter when producing the next one.
B
Batch APIA way to submit many requests to be processed asynchronously, usually within hours, at about half price.BenchmarkA standard test used to compare models, such as a set of coding tasks or exam questions.
C
Closed modelA model reachable only through its vendor's API; the weights are not published.ComputeThe processing power, usually GPU or TPU hours, used to train and run models. The scarce input of the AI industry.Context windowThe maximum text a model can consider in one request: prompt, documents, conversation and answer combined.
D
DeepfakeSynthetic audio, image or video that convincingly impersonates a real person.DistillationTraining a smaller model to imitate a larger one's outputs, capturing much of its capability at lower cost.
E
EmbeddingA list of numbers that represents the meaning of a piece of text, so similar texts have similar numbers.Evaluation (eval)A test set with known-good answers used to measure a model or system on your own tasks.
F
Fine-tuningFurther training a model on your own examples to change its behaviour or specialise it.Frontier modelA model at or near the state of the art at the time, typically the newest top tier from a major lab.
G
GPUGraphics processing unit: the chip type, dominated by Nvidia, that performs the parallel arithmetic AI training and inference require.GroundingSupplying a model with trusted sources (documents, search results) so its answer is tied to evidence rather than memory.GuardrailsChecks placed around a model to block unwanted inputs or outputs, such as filters for sensitive data or policy violations.
H
I
J
L
Large language model (LLM)A neural network trained on vast amounts of text to predict the next token, which turns out to be enough to write, summarise, reason and code.LatencyHow long a model takes to respond, usually measured as time to first token and tokens per second.
M
Mixture of experts (MoE)A model design where only a subset of parameters activates for each token, giving large capacity at lower inference cost.Model Context Protocol (MCP)An open standard for connecting AI applications to tools and data sources through a common interface.MultimodalA model that accepts or produces more than one kind of input or output: text, images, audio, video.
N
O
P
ParametersThe numbers inside a model that training adjusts; loosely, its size. Frontier models have hundreds of billions to trillions.PromptThe text sent to a model: instructions, context and the question.Prompt cachingCharging a reduced rate for the repeated prefix of a prompt, such as a system prompt or a document reused across requests.Prompt injectionAn attack in which text the model reads, such as a web page or email, contains instructions that hijack its behaviour.
R
Rate limitThe maximum requests or tokens per minute a vendor allows an account, often tiered by spending history.Reasoning budgetA setting that controls how many thinking tokens a reasoning model may use before answering.Reasoning modelA model trained to work through a problem step by step, generating intermediate 'thinking' tokens, before answering.Retrieval-augmented generation (RAG)Finding the relevant passages from a document store and giving them to the model with the question, so answers are grounded in your data.RLHFReinforcement learning from human feedback: training a model to prefer outputs that human raters rate highly.
S
Synthetic dataTraining data generated by a model rather than collected from humans.System promptStanding instructions a developer gives a model that apply to every conversation, such as persona, rules and output format.
T
TokenThe unit models read and write; roughly three-quarters of an English word. Pricing and context limits are counted in tokens.Tool use (function calling)A model's ability to request that the surrounding software run a function, such as a search or a database query, and return the result.TrainingThe process of adjusting a model's parameters on data. Pre-training builds general ability; post-training shapes behaviour.TransformerThe neural network architecture behind nearly all current language models, built around a mechanism called attention.
V
W
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.