AI ART Discover · Learn · Create
AI Art > AI Glossary
✨ AI INTELLIGENCE LEXICON • 200+ DEFINITIONS 224 Terms

AI Glossary & Intelligence Lexicon

Clear, plain-language definitions, architecture breakdowns, practical examples, and FAQs for the concepts shaping modern artificial intelligence.

The European Union's risk-based regulatory framework for AI systems and providers.

Practical Example

Under the EU AI Act, a system used to screen job applicants falls in a higher-risk category and must meet requirements such as risk management, documentation, and human oversight. A simple spam filter faces far lighter obligations.

A software system that can observe, reason, and take actions to achieve a goal, often using tools and memory.

Practical Example

A travel-booking agent is asked to find a flight under a set budget. It searches airline sites through tools, compares options, drops ones over budget, and asks you to confirm before buying. It keeps notes on what it has already tried so it does not repeat itself.

The work of making AI systems behave according to human intentions, values, and safety constraints.

Practical Example

A video app trains a model to maximize watch time, hoping it will recommend content people enjoy. The model may instead promote whatever keeps people scrolling, even if they later regret it. Alignment work tries to close this gap between the goal we wrote down and the goal we meant.

Policies, standards, and oversight mechanisms that guide how AI is developed and used in society.

Practical Example

A hospital wants to use an AI tool to prioritize patient scans. Before launch, an internal review board checks the tool's test results, a regulator's rules on medical devices apply, and staff are told when a human must double-check the output.

A field focused on reducing harmful behavior, failures, and misuse risks in AI systems.

Practical Example

Before a chatbot is released, testers try to get it to give dangerous instructions or leak private data. Failures they find are fixed through retraining, filters, or rules, and the system is monitored after launch for new problems.

A positional bias method that penalizes attention scores based on token distance, helping models extrapolate to longer contexts.

Practical Example

In ALiBi, a token's attention to a word far behind it is reduced by a penalty that grows with distance. Nearby words matter most, and the pattern holds for lengths beyond those seen during training.

A structured way for one software system to send requests to and receive responses from another system.

Practical Example

A weather app does not run its own forecasting. It sends a request such as a city name to a weather service's API and receives back the forecast in a structured format, which the app then displays. AI services work the same way: you send a prompt and get text back.

An iterative cycle where an AI agent observes, plans, acts, and reflects until it completes a goal or hits a stop condition.

Practical Example

A coding agent runs a test, sees a failure, edits the code, and runs the test again. It repeats until the tests pass or it reaches a step limit and asks for help.

A defined set of rules or steps that a computer follows to solve a problem or complete a task.

Practical Example

A recipe is an algorithm for cooking: the same steps in the same order give a predictable dish. In software, a sorting algorithm takes a shuffled list of names and follows fixed rules until the list is in alphabetical order.

Systematic unfairness in model outputs caused by skewed data, assumptions, or modeling choices.

Practical Example

A resume-screening model is trained on past hiring decisions at a company that mostly hired men. It learns to rate resumes with typical male patterns higher and can penalize equally qualified women. This kind of case was widely reported in coverage of automated hiring tools.

The extra cost in time, compute, or product velocity required to make systems safer and more controllable.

Practical Example

A team adds extra safety training and review steps before release. The model is safer, but the added work delays shipping and, if not done carefully, can make it refuse harmless requests.

A model component that dynamically focuses on relevant parts of an input when producing an output.

Practical Example

When a language model reads 'The trophy didn't fit in the suitcase because it was too big', attention lets it weigh 'trophy' more heavily than 'suitcase' when deciding what 'it' refers to. The model learns such weightings from data rather than following hand-written rules.

A sequence model class that generates future tokens iteratively by conditioning each new prediction on all previously generated tokens in the sequence.

Practical Example

GPT-4 and Llama generate long responses one token at a time, where each token becomes part of the prompt context for the next token.

The core training algorithm that updates model weights by propagating prediction errors backward through the network.

Practical Example

A digit-recognition network guesses '3' for an image of an '8'. Backpropagation traces the error back through the layers and calculates how each weight contributed to the mistake. Small adjustments are applied, and repeated over many examples, the guesses improve.

A standardized test or dataset used to measure and compare model performance.

Practical Example

A benchmark like MMLU gives many models the same multiple-choice questions across school and professional subjects. Because the questions and scoring are fixed, researchers can compare one model against another or a model against its own earlier version.

A consistent pattern of error or unfairness in data or model behavior.

Practical Example

A face-recognition system trained mostly on lighter-skinned faces makes more errors on darker-skinned faces. Research on commercial systems, such as the Gender Shades study, found uneven error rates across skin tone and gender.

Very large and complex datasets that require scalable storage and processing techniques.

Practical Example

A retailer records millions of purchases, clicks, and delivery events every day. No single computer can handle all of it comfortably, so the data is spread across many machines and processed in parallel to spot demand trends.

A model whose internal reasoning is difficult to interpret directly by humans.

Practical Example

A deep neural network rejects a loan application. Its prediction depends on millions of numeric weights, so nobody can point to a simple rule explaining the decision. Teams then use extra tools, such as feature-importance methods, to get partial explanations.

A multimodal model architecture that learns shared representations between text and images.

Practical Example

CLIP was trained on pairs of images and captions so that a photo of a dog and the words 'a photo of a dog' land close together in a shared space. You can then classify a new photo by comparing it to text like 'a photo of a cat' versus 'a photo of a dog', without training a new classifier.

A reasoning style where an AI model decomposes a problem into intermediate steps.

Practical Example

Asked a word problem, a model first writes out intermediate steps, such as subtracting then adding, before stating the final number. Research has shown that prompting for step-by-step reasoning can improve accuracy on math and logic problems compared with answering directly.

The processing resources required to train and run models, often measured in FLOPS or GPU hours.

Practical Example

Training a large language model can occupy thousands of GPUs for weeks, while answering one question takes a tiny fraction of that effort. Providers pay for both, and the totals show up in their hardware and electricity bills.

The maximum amount of input tokens a language model can process at once.

Practical Example

If you paste a 300-page report into a chatbot whose window is smaller than the report, the model cannot read all of it at once. Older parts may be cut off or summarized, and it can lose track of details that fell outside the window.

Training approaches that let a model keep learning from new data without forgetting prior knowledge.

Practical Example

A model trained to recognize cats is then fine-tuned only on birds. Without precautions it may become worse at cats, a problem called catastrophic forgetting. Continual learning methods, such as replaying a sample of old data, try to avoid this.

A neural network architecture that enhances text-to-image diffusion models by adding explicit spatial conditioning such as edge maps, depth contours, or human skeletal poses.

Practical Example

Using an OpenPose ControlNet to guarantee an AI-generated model strikes an exact dancing pose matching a reference photograph.

The process of assigning tags or target outputs to raw data for supervised learning.

Practical Example

To build a sentiment classifier for product reviews, people read thousands of reviews and mark each as positive, negative, or neutral. Where two labelers disagree, a reviewer decides, and the final labels become the answers the model learns from.

The surface in feature space that separates classes predicted by a classifier.

Practical Example

Imagine plotting emails by two features, such as number of links and number of exclamation marks. A classifier draws a line or curve through that plot; emails on one side are called spam and on the other not spam. Points close to the line are the uncertain ones.

A generative architecture that learns to reverse noise to synthesize images, audio, or other content.

Practical Example

An image generator is trained by taking real pictures, adding noise step by step, and learning to reverse that process. At creation time it starts from random noise and removes it gradually, steered by the text prompt, until a picture appears.

A numeric vector representation that captures semantic meaning of text, images, or other data.

Practical Example

A help center turns every article into an embedding. When a customer types 'my card was declined', the system embeds the question and finds the articles whose vectors are closest, even if they use different words like 'payment failed'.

A model specialized for converting data into vectors used for semantic search, clustering, and retrieval.

Practical Example

A company runs every product description through an embedding model once and stores the vectors. A shopper's search is embedded with the same model, and the nearest product vectors are returned.

Dense, continuous vector representations where words, documents, or images are mapped into high-dimensional space so semantic similarity translates into geometric distance.

Practical Example

Calculating the cosine similarity between the vector for "royal woman" and "queen" yields a match score close to 0.98.

Continuing training on domain-specific data to adapt a pre-trained model to a specific task.

Practical Example

A law firm starts from a general language model and continues training it on its own contract summaries. The tuned model then follows the firm's formatting and terminology more closely than the generic model.

AI systems that produce new content such as text, images, audio, video, or code.

Practical Example

You type 'write a friendly reminder email about Friday's meeting', and a model drafts one. Another tool turns the sentence 'a lighthouse at dawn' into a picture. Neither output is copied from a single source, but both reflect patterns in training data.

A graph-based index structure for fast approximate nearest-neighbor search over high-dimensional vectors.

Practical Example

A search enters at the sparse top layer, hops toward the query across long links, drops a layer, and repeats, ending in the dense bottom layer with the closest vectors. It is like using highways, then streets, then your destination's block.

When a model generates fluent but false or unsupported information.

Practical Example

Asked for court cases supporting an argument, a chatbot invents case names and citations that look real. In a widely reported 2023 incident, lawyers submitted a brief that cited such non-existent cases and were sanctioned.

A deployed API interface that receives model requests and returns predictions in production.

Practical Example

A developer deploys a model and receives a URL. The mobile app sends text to that URL and receives the model's prediction back, without the app itself containing the model.

A generative image editing process where a user isolates an unwanted or editable portion of an image with a brush mask and prompts the model to seamlessly replace it.

Practical Example

Brushing over a pedestrian in a travel photo and prompting "empty stone cobblestone street with golden sunset reflections".

A prompt technique intended to bypass a model's safety constraints.

Practical Example

A user tells a chatbot to pretend to be a character with no rules, hoping it will answer something it would normally refuse. Developers study such prompts and retrain or add filters so similar attempts fail.

Stored key and value tensors from previous tokens that let transformers generate new tokens without recomputing past attention.

Practical Example

When a chatbot writes a long reply, it does not recompute attention inputs for every earlier token at each step. It keeps those stored values and only adds those for the newest token.

A compressed representational space where similar concepts are positioned near each other as vectors.

Practical Example

An image model compresses each picture into a short list of numbers. Pictures with similar content end up with similar lists, so moving smoothly between two points can blend one image into another.

Artificial intelligence architectures capable of understanding, interpreting, and correlating disparate modalities (text, code, image, video, audio) in unified representations.

Practical Example

An AI assistant analyzing a photograph of a handwritten whiteboard architecture diagram and outputting production-grade Terraform configuration.

A creative synthesis feature that expands an image beyond its existing canvas borders while matching style, lighting, depth of field, and compositional flow.

Practical Example

Taking a square portrait of a warrior and expanding the sides to 16:9 cinematic widescreen to reveal an army encampment behind them.

A language-model metric measuring how surprised the model is by true next tokens.

Practical Example

A language model reads a held-out news article and rates how likely each actual next word was. If it kept assigning low probability to the real words, perplexity is high, which signals a weaker fit to that kind of text.

Training steps applied after pretraining, such as instruction tuning, preference optimization, and safety tuning.

Practical Example

A base model that only continues text is taught to follow instructions, to prefer helpful answers, and to refuse harmful requests. The result feels like an assistant, though the underlying knowledge came from pretraining.

A fine-tuning technique that combines 4-bit weight quantization with LoRA adapters to reduce memory needs.

Practical Example

A researcher fine-tunes a multi-billion-parameter model on one consumer graphics card by compressing the base weights to 4-bit and training only small adapter layers.

Converting model weights to lower precision formats such as 8-bit or 4-bit.

Practical Example

A model file that is about 14 GB at 16-bit precision can shrink to roughly a quarter of that at 4-bit, so it can run on a laptop. Outputs may be slightly different from the original.

A retrieval method that generates multiple query variants, retrieves results for each, and fuses the rankings.

Practical Example

A question about 'laptop battery drain' is rewritten as 'power consumption causes' and 'battery settings tips'. Each version retrieves its own list of documents, and documents that rank well across lists rise to the top.

A prompting pattern that interleaves reasoning steps with tool-use actions to solve tasks more reliably.

Practical Example

Asked who directed a film and where they were born, the model writes a thought, searches for the film, reads the result, then searches the director and answers using what it found.

Stress-testing an AI system with adversarial prompts to reveal failures and risks.

Practical Example

Before a model launches, specialists try to coax it into giving dangerous instructions, leaking private data, or showing biased behavior. Each failure is logged and used to improve the system, then the tests are repeated.

A structured argument, supported by evidence, that an AI system is safe for a defined context of use.

Practical Example

Before deploying a drug-dosage assistant in a hospital, the team writes a safety case: the claim, the testing evidence, known failure modes, and the checks that catch them. Reviewers can then challenge the reasoning.

A deterministic integer used to initialize the pseudo-random number generator that governs the noise tensor in generative models, allowing exact output reproducibility.

Practical Example

Re-running Midjourney with seed --seed 98214 to adjust camera focal length while preserving the exact same character facial likeness.

A chunk of text processed by language models, such as a word piece or symbol.

Practical Example

The word 'unbelievable' might be split into three tokens, while a short common word like 'the' is one. Providers count tokens to limit input and to price usage.

A database optimized for storing and querying high-dimensional embedding vectors.

Practical Example

A support bot stores embeddings of every help article in a vector database. When a customer asks a question, the database returns the article vectors closest to the question's vector in milliseconds.

A dense vector representation of words capturing semantic relationships.

Practical Example

In a word-embedding model, words used in similar contexts get similar vectors. 'Doctor' ends up near 'nurse' and far from 'banana'. Classic examples such as word2vec showed that directions in the space can capture relationships.

AI TAXONOMY ARCHITECTURE

Foundational Pillars of Modern AI Systems

Master how prompt inputs, latent representations, and transformer layers interconnect.

Explore Masterclasses
1

Foundation Models

Large-scale transformer and diffusion architectures pre-trained on multi-modal datasets.

2

Prompt Conditioning

System instructions, few-shot demonstration pairs, and chain-of-thought elicitation.

3

Sampling & Steering

Temperature, Top-P stochastic controls, negative prompt penalties, and LoRA adapters.

4

Retrieval & Grounding

Vector embeddings, semantic cosine matching, and external factual store injection.