AI comes with a lot of jargon. This glossary explains the terms you will keep running into, in plain English, with links to tools on AIMAPER where you can see each idea in practice.
74 of 74 terms
An AI agent is a system that takes a goal, decides on steps, and uses tools (search, code, apps) to carry them out with little hand-holding. Think of it as a model that can act, not just answer. Most agents loop: plan, act, check the result, repeat.
A process where one or more AI agents handle multi-step work end to end, such as researching, drafting, and filing a report. The human sets the goal and reviews; the agents do the legwork in between.
A hypothetical AI that could learn and perform any intellectual task a person can, across domains, without being retrained for each one. No system today meets that bar, and experts disagree on how to measure it.
The work of making AI systems pursue the goals and values their builders and users actually intend. Alignment research covers everything from training methods like RLHF to studying how models behave in unusual situations.
An Application Programming Interface is a defined way for one piece of software to talk to another. Most AI models are offered through an API: you send a request with your prompt, and the provider sends back the model's output.
The technique that lets a model weigh which parts of its input matter most for each part of its output. Attention is the core idea behind transformers and the reason they handle long, connected text so well.
A model that generates output one piece at a time, each new piece based on everything produced so far. Chat-style language models are autoregressive: they predict the next token, add it, and repeat.
A standard test used to compare models, such as a set of math problems or coding tasks with known answers. Benchmarks are useful for rough comparison, but a high score does not guarantee a model will suit your specific work.
Systematic skew in a model's outputs, usually inherited from patterns in its training data. Bias can show up as unfair treatment of groups or as lopsided answers, and reducing it takes careful data choices and testing.
A prompting approach where the model works through a problem step by step before giving an answer. Spelling out intermediate reasoning often improves results on math, logic, and planning tasks.
Software you talk to in a conversation, by text or voice. Modern chatbots are usually built on large language models, which lets them handle open-ended questions instead of following a fixed script.
Sorting inputs into predefined groups, such as marking emails as spam or not, or tagging support tickets by topic. It is one of the oldest and most widely used machine learning tasks.
The field of AI that helps machines interpret images and video: recognising objects, reading text, detecting faces, or measuring movement. It powers everything from photo search to quality checks on factory lines.
The maximum amount of text (measured in tokens) a model can consider at once, including your prompt, any documents, and its own reply. Anything beyond the window is simply not seen by the model.
An AI assistant embedded inside a tool you already use, such as a code editor or office suite, that suggests or completes work as you go. You stay in control; the copilot speeds up the routine parts.
Adding tags or answers to raw data so a model can learn from it, for example drawing boxes around cars in photos. Labeled data is the backbone of supervised learning.
A branch of machine learning that uses neural networks with many layers to learn patterns from large amounts of data. Most of today's headline AI, from image generators to chatbots, is built with deep learning.
Synthetic media, usually video or audio, made with AI to convincingly show a real person saying or doing something they never did. The same technology has legitimate uses in film and dubbing, which is why consent and labeling matter.
A type of generative model that learns to turn random noise into a clear image (or audio, or video) step by step. Many popular image generators, including Stable Diffusion, work this way.
Training a smaller "student" model to imitate a larger "teacher" model. The result is cheaper and faster to run while keeping much of the original's ability.
A list of numbers that represents the meaning of a piece of text, an image, or other data. Items with similar meanings get similar embeddings, which makes them the foundation for semantic search and recommendations.
Getting a model to do a task by showing it a handful of examples in the prompt, rather than retraining it. Two or three good examples often shape the format and tone of the output far better than instructions alone.
Taking a model that has already been trained and training it a bit more on your own examples, so it gets better at a specific task, style, or domain. It is cheaper than training from scratch and often more reliable than prompting alone.
A large model trained on broad data that can be adapted to many tasks. GPT, Claude, Gemini, and Llama are examples; most AI products are built on top of one.
A feature that lets a language model ask your software to run a specific function, such as checking the weather or looking up an order, and then use the result in its answer. It is how chatbots connect to real systems.
A generative approach where two networks compete: one creates fake samples and the other tries to spot them. Over time the creator gets very good. GANs were the leading image-generation method before diffusion models took over.
AI that creates new content, such as text, images, music, code, or video, rather than only analysing existing data. It learns patterns from training data and produces fresh output that follows them.
Short for Generative Pre-trained Transformer, the family of language models from OpenAI. The term is also used loosely for any chat-style language model, though strictly it names OpenAI's models.
A Graphics Processing Unit is a chip built to do many calculations in parallel. That makes GPUs ideal for training and running neural networks, and they are the main hardware behind modern AI.
Tying a model's answers to verifiable sources, like your documents or a live web search, instead of relying only on what it memorised in training. Grounded answers are easier to check and less prone to made-up facts.
Rules and filters placed around an AI system to keep its behaviour within safe, on-topic, or on-brand limits, for example blocking certain requests or checking outputs before they reach users.
When an AI model states something false or invented with full confidence, such as a fake citation or a wrong date. It happens because models predict plausible text, not verified facts, so important answers should be checked.
Running a trained model to get an output, such as asking a chatbot a question. Training builds the model; inference is using it. Most AI pricing is based on inference volume.
A prompt crafted to trick a model into ignoring its safety rules. Providers continually patch known jailbreaks, and using them usually breaks terms of service.
The date after which a model has no training data. Ask about later events and the model either does not know or guesses, unless it is connected to search or other live sources.
A neural network trained on huge amounts of text to predict the next word (token). That simple objective, at large scale, produces models that can write, summarise, translate, answer questions, and write code.
The delay between sending a request to a model and getting a response. Low latency matters for voice assistants and live chat; for batch jobs it matters much less than cost.
A lightweight fine-tuning method that trains a small set of extra weights instead of the whole model. LoRA files are small and easy to share, which is why they are popular for customising image models with a style or character.
The broad practice of building systems that learn patterns from data rather than following hand-written rules. Deep learning and LLMs are subsets of machine learning.
An open standard, introduced by Anthropic in 2024, for connecting AI assistants to external tools and data sources in a consistent way. Instead of a custom integration for every app, a tool exposes an MCP server and any compatible assistant can use it.
The learned numbers inside a neural network that determine its behaviour. Releasing the weights lets anyone run or modify the model; keeping them private means you can only use it through the provider.
A model that can work with more than one type of input or output, such as text, images, audio, and video, within a single system. You can show it a photo and ask questions about it, for example.
The field of AI focused on understanding and producing human language. Translation, sentiment analysis, summarisation, and chatbots all fall under NLP.
A computing system loosely inspired by the brain, made of layers of connected units that adjust their connections as they learn from data. It is the basic building block of deep learning.
Technology that reads text from images or scanned documents and turns it into editable text. Modern AI-based OCR also handles handwriting, tables, and messy layouts.
A model whose weights (and ideally code and training details) are publicly released so anyone can run, study, or adapt it. Licences vary, so "open" does not always mean free for every commercial use.
When a model learns its training data too closely, including the noise, and then performs poorly on new data. It is like memorising answers to past exams instead of understanding the subject.
The adjustable values inside a model that are set during training. Parameter count (for example 7 billion or 70 billion) is a rough indicator of a model's size, though not a direct measure of quality.
The instruction or input you give an AI model. A prompt can be a single question or a long brief with context, examples, and formatting rules; the clearer it is, the better the output tends to be.
The craft of writing and refining prompts to get reliable, high-quality results from a model. It includes giving context, examples, constraints, and a clear output format.
An attack where hidden or malicious instructions in content a model reads (a web page, email, or file) try to override what the user or developer intended. It is a key security concern for agents that browse or read documents.
Shrinking a model by storing its weights with fewer bits, for example 4-bit instead of 16-bit. Quantized models use less memory and run faster, often on consumer hardware, with a small loss in quality.
A setup where the system first searches a knowledge source, like your documents or a database, then passes the relevant snippets to the model so it can answer from them. RAG keeps answers current and specific without retraining the model.
A language model trained to spend extra computation thinking through a problem before answering. These models tend to do better on math, coding, and multi-step logic, at the cost of slower and pricier responses.
A training approach where a model learns by trial and error, getting rewards for good actions and penalties for bad ones. It is how game-playing AIs improve and is part of how chat models are tuned.
Reinforcement Learning from Human Feedback: people rate or rank model answers, and those preferences are used to train the model to respond in ways humans find more helpful and safe. It is a key step in turning a raw language model into a useful assistant.
Search that matches on meaning rather than exact keywords. Ask for "cheap places to stay" and it can find results about "budget hotels", usually by comparing embeddings.
Detecting the emotional tone of text, such as positive, negative, or neutral. Businesses use it to scan reviews, social posts, and support tickets at scale.
Converting spoken audio into written text, also called automatic speech recognition (ASR). It powers meeting transcripts, captions, and voice commands.
Data generated artificially, often by AI, rather than collected from the real world. It helps fill gaps, protect privacy, or create rare examples for training and testing.
Background instructions given to a model before the conversation starts, setting its role, tone, and rules. Users usually do not see it, but it shapes every answer.
A setting that controls how random a model's output is. Low temperature gives focused, predictable answers; high temperature gives more varied and creative ones, with more risk of going off track.
Generating a picture from a written description. You type "a red bicycle leaning on a blue wall at sunset" and the model produces an image to match.
Turning written text into spoken audio. Modern AI voices sound natural, can carry emotion, and in some tools can be cloned from a short voice sample with consent.
Generating a short video clip from a written description or a still image. It is one of the newest generative AI areas and is improving quickly in length and realism.
The unit a language model reads and writes: often a word, part of a word, or a punctuation mark. In English, a token averages about three-quarters of a word. Pricing and context limits are measured in tokens.
The examples a model learns from. Its quality, size, and variety largely decide what the model knows, how accurate it is, and what biases it carries.
Reusing what a model learned on one task as a head start for another. Fine-tuning a pre-trained model is the most common form.
The neural network design, introduced by Google researchers in 2017 in the paper "Attention Is All You Need", that underpins nearly all modern language models. It processes whole sequences in parallel using attention.
Using AI to increase an image or video's resolution while adding plausible detail, rather than just stretching pixels. Handy for restoring old photos or preparing images for print.
A database built to store embeddings and quickly find the ones most similar to a query. It is the usual storage layer behind semantic search and RAG systems.
Building software mainly by describing what you want to an AI in plain language and accepting its code, with little manual editing. The term was coined by Andrej Karpathy in early 2025; it works well for prototypes but still needs review for anything serious.
Creating a synthetic copy of a specific person's voice from recordings. Reputable tools require the speaker's consent, since the same technique can be misused for fraud.
Asking a model to do a task it was not specifically shown examples of, relying only on instructions. Large models handle many tasks zero-shot, which is why simple prompts often work.