Glossary
AI terms from our issues, explained in plain English. A new term may appear each day.
A
Agent
An agent is a program that uses artificial intelligence to carry out an assigned task step by step. It can choose its next steps and use tools, such as running calculations or working with files, so it needs clear instructions and limits on what it is allowed to do.
ExampleYou assign an agent to process files overnight and set a fixed spending limit for the paid services it uses in its work.
Source: simonwillison.net
AI slop
AI slop is a term for low-quality content created by artificial intelligence in large quantities, such as text, images, or music. This content floods the available selection and makes it harder to find something useful or interesting.
ExampleWhen choosing music, AI slop among the recommended tracks can get in the way, making it harder to find something you want to listen to.
Source: The Verge AI
AI training
AI training is the process in which artificial intelligence learns from a large number of examples, such as texts or images. It looks for recurring patterns in them and uses these patterns to adjust how it generates responses or carries out tasks.
ExampleYour work emails and documents that a company sells from its archive can also serve as learning materials for AI training.
Source: MIT Technology Review AI
B
Benchmark
A benchmark is a comparative test that measures how well artificial intelligence handles a particular set of tasks. When different models solve the same tasks under the same conditions, their results can be compared. However, a good score alone does not guarantee that AI will handle your work equally well.
ExampleWhen choosing AI to review program code, a benchmark helps you compare how well different models find errors in the same set of samples.
Source: github.blog
C
classifier
A classifier is an artificial intelligence model that assigns content to predefined categories based on what it has learned from examples. For instance, it can distinguish between spam and a regular email, or determine whether a text contains sensitive information.
ExampleWhen your email inbox moves a suspicious message to the spam folder, a classifier may have identified it as advertising.
Source: github.blog
computer use
Computer use with AI means that artificial intelligence interacts directly with application interfaces: for example, clicking buttons or filling in forms. Based on your instructions, it can carry out individual steps of a task instead of simply describing how to do them.
ExampleWhen filling in a web form, you can ask AI to enter the information into the appropriate fields using computer use.
Source: The Decoder
context
Context is the information available to artificial intelligence when it is working on a task, helping it understand what you want it to do. This can include previous messages, reference materials, or details about what you are currently working on. When important information is missing, AI may misunderstand the request.
ExampleWhen you ask an AI assistant in a work application to summarize a document in progress, the context includes information about which document you have open and what it contains.
Source: aws.amazon.com
context window
A context window is the amount of content AI can work with at once when responding. It determines how much of the prompt, previous messages, and other reference materials can fit into the conversation. When a conversation is too long, a summary can help by taking up less space, but it may not preserve every detail.
ExampleWhen you switch to a model with a smaller context window while editing code, a summary of the conversation so far helps carry over important information so you can continue working.
Source: Claude Code changelog
custom instructions
Custom instructions are saved instructions that tell AI how to tailor its responses to your needs. You can use them to specify, for example, the writing style or response format, so you do not have to repeat the same requirements with every prompt.
ExampleYou specify in your custom instructions that responses should be concise and organized into bullet points to make them easier to read while you work.
Source: Google Workspace Updates
G
generative model
A generative model is an artificial intelligence system that creates new content, such as text, an image, or music, based on a prompt. It uses patterns it has learned from earlier examples.
ExampleWhen you give a generative model a request for a cheerful birthday song, it can create both music and vocals based on that request.
Source: The Verge AI
L
local AI
Local AI means that artificial intelligence processes tasks directly on your computer or another device you own. You do not need to send the data for such a task to an external server for processing, but the device must have sufficient computing power and memory.
ExampleWhen summarizing a work document, local AI can allow its contents to be processed directly on your computer.
Source: NVIDIA AI Blog
M
model context protocol
MCP is a common way to connect an AI assistant to other programs and data sources. It allows the assistant to use connected tools, for example to request reference materials for an answer.
ExampleWhen preparing a work report, an AI assistant can use MCP to obtain the necessary data from a connected tool, provided it has permission to access that data.
Source: zive.cz
model parameter
A model parameter is an internal numerical value that artificial intelligence adjusts during learning and that influences its results. The number of parameters describes the size of a model, but on its own does not indicate how good its results will be.
ExampleWhen choosing AI to create environments, knowing that the Odyssey-3 model has 14 billion parameters tells you its size, but does not in itself guarantee a good result.
Source: The Decoder
N
neural network
A neural network is a computational system made up of many interconnected simple units that process information together. Depending on how these connections are configured, it transforms inputs, such as an image or text, into a result, such as a recognized letter or a response.
ExampleWhen AI converts a photo of a receipt into text, it may use a neural network to recognize the letters.
Source: JetBrains AI blog
O
open-weight model
An open-weight model is an artificial intelligence system whose learned settings are available to download. This allows you to run it on your own computer instead of sending prompts to an external service; however, you need a sufficiently powerful computer and must comply with the license terms.
ExampleWhen you run an open-weight model on your own computer while programming, it can help you edit code without sending the code to an external service.
Source: blog.jetbrains.com
R
reference image
A reference image is a visual example you give artificial intelligence to guide it when creating a new image. It helps show the appearance, composition, or style you want, instead of having to describe everything in words alone.
ExampleWhen creating an invitation, you provide an example of an artistic style as a reference image so that AI can follow it in the design.
Source: The Decoder
S
self-evaluation
Self-evaluation means that AI assesses the quality of its own response or completed work. This can help it find errors and adjust its approach, but during review it may overlook the same shortcomings it missed while carrying out the task.
ExampleWhen AI drafts an email to a customer, you ask it to check whether it has answered all of the customer’s questions; this prompts it to perform self-evaluation.
Source: The Decoder
speech model
A speech model is an artificial intelligence system that processes or generates spoken language. Depending on its purpose, it can understand what a person is saying or generate speech, including its tone and delivery.
ExampleWhen you listen to an audiobook, a speech model can distinguish between the voices of different characters and deliver their lines as nervous whispers or angry shouts.
Source: Amazon Science
T
text-to-speech
Text-to-speech is the automatic generation of spoken audio from written text. The system reads the text aloud; when AI is used, the delivery can imitate natural human speech, including pronunciation and intonation.
ExampleWhen you have an article read aloud by an artificial voice, you are using text-to-speech.
Source: AWS ML Blog
Token
A token is a small piece of text that artificial intelligence works with when reading a prompt or generating a response. It can be a whole word, part of a word, or even a punctuation mark. The number of tokens is often used to calculate the cost of processing a prompt and generating a response.
ExampleWhen you ask AI to summarize a long document, it uses tokens both when reading the prompt and when creating the summary.
Source: aws.amazon.com
training data
Training data are examples that artificial intelligence learns from to recognize patterns and perform tasks. They can be texts, images, or audio; their content and quality influence what AI learns.
ExampleIf you are developing AI for product inspection, images of defects in metal or glass can serve as training data so that it learns to recognize those defects.
Source: NVIDIA AI Blog
training environment
A training environment is a set of situations and tools in which AI learns to perform tasks. For example, it can simulate apps on a phone so that AI can practice using them.
ExampleIf an AI assistant is to operate apps on your phone in the future, it can first learn to use them in a training environment with simulated apps, as is the case with OpenMobile-2.
Source: The Decoder