This glossary is a living resource maintained by the AI Tuner Institute. It is designed to support students, faculty, professionals, and anyone seeking to understand the language of artificial intelligence. Terms are updated as the field evolves.
Use the alphabet navigation below to jump to a specific letter, or scroll through the full glossary.
A prediction system that generates outputs based on patterns.
The process of ensuring that an artificial intelligence system behaves in ways that match human values, goals, and expectations. Alignment focuses on safety, reliability, and preventing harmful or unintended behavior. It ensures that the system's actions remain consistent with human oversight.
The use of artificial intelligence to support human choices by providing analysis, predictions, or recommendations. The human remains the final decision maker. The AI system enhances understanding, reduces cognitive load, and helps identify patterns that may not be immediately visible.
The use of artificial intelligence to complete tasks that normally require human effort. It can include scheduling, classification, writing, or decision support.
The study and practice of guiding artificial intelligence development in ways that respect human dignity, fairness, accountability, and social responsibility. It addresses issues such as bias, transparency, privacy, and the impact of AI on individuals and communities.
The ability to understand, interpret, and work effectively with artificial intelligence systems. It includes knowing how models function, how to evaluate their output, how to communicate with them, and how to apply them responsibly in real-world contexts. AI fluency is a foundational skill for modern learners.
Moments when an artificial intelligence system produces information that is false, incorrect, or entirely invented, even though the response appears confident and believable. AI hallucinations happen when the system fills gaps, misinterprets patterns, or generates details that are not supported by data.
Coordinates multiple artificial intelligence systems, tools, or steps so they work together to complete a larger task.
The protection of personal information when artificial intelligence systems collect, process, or generate data. It includes minimizing data use, securing sensitive information, and ensuring responsible handling.
The methods used to prevent artificial intelligence systems from causing harm. It includes testing, monitoring, guardrails, and oversight practices that ensure the system behaves predictably and responsibly.
Harmful, offensive, or unsafe content produced by an artificial intelligence system. Toxicity can arise from biased training data, unclear prompts, or insufficient guardrails. Preventing toxicity requires strong alignment, filtering, and oversight.
The act of exaggerating or falsely claiming the use of artificial intelligence in a product, service, or process. AI washing misleads customers, investors, and the public by creating the appearance of advanced technology when the reality is much smaller.
An artificial intelligence system designed to perform tasks autonomously within a workflow. It can follow instructions, complete assignments, and collaborate with other systems or humans.
The structured sequence of steps required to design, train, evaluate, deploy, and maintain an artificial intelligence system. It includes data collection, preprocessing, model selection, training, testing, deployment, monitoring, and continuous improvement.
Communication that occurs when one artificial intelligence system sends information or instructions to another artificial intelligence system. Tuners must monitor these exchanges for accuracy and safety.
Uses autonomous artificial intelligence agents to complete multi-step tasks. These agents plan, act, and adapt without requiring constant human direction.
Step-by-step instructions a computer follows to solve a problem or make a decision. They are the foundation of all artificial intelligence systems.
When humans attribute human traits, emotions, or intentions to machines. Tuners must guard against this tendency.
A hypothetical system capable of understanding, learning, and performing any intellectual task that a human can perform. AGI would not be limited to specific domains. AGI does not exist today.
A theoretical form of intelligence that surpasses human capability in every domain, including reasoning, creativity, and decision making. ASI does not exist today.
Allows a model to retrieve information based on similarity or relatedness rather than exact matches. It supports reasoning, analogy, and pattern recognition.
A method that allows an artificial intelligence model to identify which parts of the input are most important at any moment. Instead of treating all words or data equally, the model assigns different levels of importance to each element.
Refers to systems that can operate or make decisions with limited human involvement.
The method a neural network uses to learn from mistakes. After making a prediction, the model measures the error and sends that information backward through the network to adjust its internal parameters.
A reasoning method where the system starts with a goal and works backward to determine the steps needed to reach it.
A method for generating text by exploring several possible response paths at the same time. Instead of choosing only the most likely next word, the model keeps a small set of promising options and selects the best overall sequence.
Occurs when an artificial intelligence system produces unfair or unequal outcomes because of flawed data or design.
Extremely large collections of information that are too complex for traditional tools to process. Big data shapes how artificial intelligence learns. If the data is biased, incomplete, or messy, the model will inherit those flaws.
A metric for evaluating machine translation accuracy. BLEU compares the system's translation to one or more high quality human translations. A higher BLEU score means the artificial intelligence translation is closer to what a human would produce.
Automated systems that perform tasks without needing constant human involvement. They can answer questions, process information, or complete simple actions.
A rectangle drawn around an object in an image to help computer vision systems identify and track it.
The step-by-step explanation an artificial intelligence system produces to show how it arrived at an answer. Tuners use this to detect errors.
Chatbots designed to behave like a specific personality, role, or fictional character. They simulate style, tone, and behavior.
A system that uses artificial intelligence to hold a conversation with a person through text or voice.
An artificial intelligence system whose internal design, training data, and parameters are not publicly available. Closed source models offer stronger security and governance but limit external review and customization.
Systems designed to mimic human reasoning. These systems analyze information, learn from patterns, and support decision making.
The amount of mental effort required to understand or complete a task. Reducing cognitive load means presenting information clearly and minimizing unnecessary complexity.
Allows artificial intelligence to interpret and understand images or video. It identifies objects, people, and patterns in visual data.
The process of creating new content based on specific instructions or constraints. The model generates output that satisfies the conditions provided in the prompt.
The amount of information an artificial intelligence model can hold in its working memory at one time. If information falls outside this window, the model can no longer use it directly. The context window does not store permanent memory. It only holds information temporarily while the model is processing a request.
Teaches a model to understand similarity and difference. It learns by comparing pairs of examples and determining which ones are related and which ones are not.
The degree to which humans can guide, constrain, or shape an artificial intelligence system's behavior. High controllability ensures predictable and aligned output.
Systems designed to engage in dialogue with humans. These systems interpret language, generate responses, and maintain context across interactions.
An artificial intelligence system designed to work alongside a human, supporting tasks such as writing, analysis, planning, and decision making. A copilot enhances human capability rather than replacing it.
A large collection of text used to train or evaluate language models.
The model's ability to generate new, unexpected, or imaginative responses based on patterns it has learned. The model does not create ideas from personal experience or emotion. Creativity in artificial intelligence is a controlled form of variation rather than human creativity.
The process of discovering patterns or useful information in large datasets.
The field that studies how to collect, clean, analyze, and interpret data. It supports artificial intelligence development and evaluation.
The process of turning human behavior, actions, or experiences into data that can be stored and analyzed.
The internal processes an artificial intelligence system uses to choose an action or response. Tuners study these patterns to ensure the system behaves safely.
A model that makes predictions by following a series of branching choices. Each branch represents a question or condition, and each leaf represents an outcome.
Sometimes the transformer architecture misinterprets or mis-assembles information.
A method of training neural networks with many layers. These layers learn increasingly complex patterns, allowing the system to understand language, images, sound, and other forms of data.
Realistic but false images, videos, or audio created by artificial intelligence.
A system where all parts of the model are active for every task. It uses all its parameters each time it generates a response.
Tools that attempt to identify whether content was created by artificial intelligence. They are imperfect and must be used with caution.
A response that will always be the same when the model receives the same input under the same conditions. Deterministic behavior is useful when accuracy, consistency, and repeatability are required.
A mathematical framework that guarantees individual data privacy by introducing statistical noise into datasets or model training processes.
Generate images or other data by starting with random noise and gradually removing that noise through a learned process. Diffusion models are used in modern image generation systems.
The process of simplifying complex data by reducing the number of variables while preserving important patterns.
The process of compressing a large artificial intelligence model into a smaller, more efficient version while preserving its essential capabilities. The smaller model learns to imitate the behavior, reasoning patterns, and output quality of the larger model, even though it uses fewer parameters and requires significantly less computational power. Distillation allows organizations to deploy high-performance artificial intelligence systems on devices with limited memory, energy, or processing capacity, such as mobile phones, edge devices, or embedded systems. It also supports responsible resource use by reducing energy consumption and improving scalability. Distillation is a foundational technique for modern artificial intelligence because it ensures that advanced capabilities remain accessible, efficient, and sustainable across diverse environments.
A slow change in AI behavior over time.
Numerical representations of words, images, or other data that capture their meaning and relationships. Items that are similar in meaning are placed close together in this mathematical space.
Systems that have a physical presence in the world and can sense, move, and act within their environment. Embodied AI supports robotics, autonomous vehicles, and systems that require real-time perception.
The internal mathematical environment where an artificial intelligence model organizes its learned representations of words, images, or other data. Items that share similar meaning or contextual relationships are placed close together in this space, while unrelated items are positioned farther apart. Embedding space allows the model to understand nuance, similarity, analogy, and conceptual relationships without relying on explicit rules. It is the foundation for reasoning, classification, semantic search, and retrieval-augmented generation. Embedding space transforms human language and perception into structured numerical patterns that artificial intelligence systems can interpret, compare, and manipulate with precision.
Refers to unexpected behaviors that appear in large artificial intelligence systems even though they were not directly programmed. Tuners must watch for these behaviors.
Combine multiple models to produce a single, stronger prediction. Each model contributes its own perspective, and the ensemble merges these perspectives to improve accuracy.
A measure of uncertainty in the model's predictions. High entropy means the model sees many possible next words with similar likelihood. Low entropy means the model strongly favors one option.
The process of identifying important pieces of information in text, such as names, dates, locations, or key terms.
The process of transferring an AI responsibility to a higher authority.
Methods that help humans understand why an artificial intelligence system made a particular decision or produced a specific output.
The ability of a system to grow, adapt, or add new capabilities without being redesigned.
Information created by artificial intelligence that is false, invented, or not grounded in real data.
A research area focused on understanding and addressing the societal implications of artificial intelligence. FATE examines how systems can unintentionally reinforce stereotypes, discrimination, or inequity, and develops methods to ensure that artificial intelligence supports fairness, responsibility, and human dignity.
Occurs when the output of an artificial intelligence system influences the future data it receives. Managing feedback loops is essential for responsible deployment.
Allows a model to learn a new task from only a small number of examples. This approach is useful when data is limited or expensive to collect.
The process of taking a pre-trained model and training it further on a smaller, specialized dataset. This allows the model to adapt to a specific task, domain, or style.
Artificial intelligence systems designed to operate efficiently with limited data, computation, or energy. It emphasizes accessibility, sustainability, and responsible resource use.
Systems made of two models that learn together. One model creates new data, and the other evaluates whether the data looks real. GANs are used to generate images, audio, and other synthetic content.
Creates new content such as text, images, audio, or code by learning patterns from large amounts of data. It predicts what should come next based on context.
A type of artificial intelligence model that learns from large amounts of text and then generates new text based on patterns it has learned.
A model or component that creates new data, such as text, images, or audio, based on patterns it has learned.
A retrieval-augmented generation method that uses a knowledge graph to organize information. It retrieves structured relationships rather than isolated text, improving reasoning and factual accuracy.
A type of computer chip that can perform many calculations at the same time. Artificial intelligence systems use GPUs to train and run models quickly.
The connection between a model's internal representations and real-world meaning. A grounded system understands concepts in ways that correspond to actual objects, actions, or experiences.
Rules and safety systems that prevent artificial intelligence from producing harmful, false, or inappropriate content. Tuners help design and enforce these guardrails.
The shared work between humans and artificial intelligence systems. It focuses on combining human judgment, creativity, and ethics with the model's speed, pattern recognition, and analytical capabilities.
The design and development of AI systems that prioritize human needs, values, and well-being. It ensures that technology supports human judgment and respects human dignity.
The study of how people engage with digital systems. It examines how interfaces and tools can be designed to be intuitive, efficient, and supportive of human goals.
A system design principle in which humans are intentionally embedded within the operation of artificial intelligence. HITL ensures that automated processes remain accountable, interpretable, and aligned with ethical or contextual standards. It is the point where human reasoning and machine precision meet to preserve trust, fairness, and adaptability.
A setting chosen before training an artificial intelligence model. It controls how the model learns. Tuners do not adjust hyperparameters, but they must understand that these settings influence model behavior.
The ability of an artificial intelligence system to identify objects, patterns, or features within an image.
The process by which an artificial intelligence model generates an output after it has been trained. Inference is distinct from training, which is the process of learning those patterns.
The structured sequence of steps an artificial intelligence system follows when generating an output after training is complete. It includes receiving the prompt, converting the input into tokens, retrieving relevant internal representations, applying model parameters, generating probabilities for possible next tokens, selecting the final output, and delivering the response to the user. The inference pipeline must be optimized for speed, accuracy, and reliability because it determines how quickly and effectively the model responds in real-world environments. Tuners monitor the inference pipeline to ensure that each stage behaves predictably, remains aligned with safety rules, and produces output that reflects the model’s training and ethical constraints.
Trains a model to follow human-written instructions more accurately. It improves clarity, alignment, and responsiveness.
An attempt to force an artificial intelligence system to ignore its safety rules. Tuners must detect and prevent jailbreak attempts.
A joint embedding predictive architecture that learns by predicting relationships between different parts of data rather than predicting the data itself. V-JEPA applies this approach to visual information.
The process of transferring knowledge from a large, complex model to a smaller, more efficient model. The smaller model learns to mimic the larger one's behavior while using fewer resources.
A structured network of facts, concepts, and relationships. It represents information as connected nodes and links, allowing artificial intelligence systems to reason about how ideas relate to one another.
A type of artificial intelligence system trained on massive amounts of text so it can understand and generate human language.
The amount of time it takes for an artificial intelligence system to respond after receiving a prompt.
The internal mathematical space where a model organizes its learned representations. Similar ideas, images, or patterns are placed near one another in this space.
The practice of managing the full lifecycle of large language models, including deployment, monitoring, evaluation, safety, and continuous improvement.
The raw numerical scores a model produces before converting them into probabilities. They represent the model's unprocessed assessment of how likely each possible next word or token might be.
Low-rank adaptation is a method that adapts a large model by training only a small number of additional parameters. This approach reduces cost and preserves the original model while enabling customization.
The mathematical measure of how far a model's predictions are from the correct answers during training. The model uses this measure to adjust its weights and reduce errors.
Platforms that allow users to build applications with minimal programming by using visual tools and prebuilt components.
A method where a computer system improves its performance by learning from data instead of being given step-by-step instructions. It finds patterns and uses them to make predictions or decisions.
A mathematical system that moves from one state to another based only on the current state, not on the full history.
A model that uses many specialized parts called experts. Only the experts needed for a specific task are activated, making the system faster and more efficient.
The discipline of managing machine learning systems from development through deployment and maintenance.
The trained artificial intelligence system that makes predictions, generates content, or performs tasks.
Connects multiple artificial intelligence models so that the output of one becomes the input of another. This supports complex workflows and multi-step reasoning.
The model may see patterns that are not real, similar to humans seeing shapes in clouds.
How strongly an artificial intelligence model believes that a particular response is the most appropriate choice. Model confidence does not guarantee correctness.
The collection of techniques used to reduce the size, memory requirements, and computational demands of an artificial intelligence model without significantly reducing its performance. Compression may include quantization, pruning, distillation, or other methods that simplify the model’s internal structure. These techniques allow organizations to deploy artificial intelligence systems on devices with limited resources while maintaining accuracy and reliability. Model compression supports sustainability, accessibility, and responsible deployment by reducing energy consumption and enabling artificial intelligence to operate efficiently across diverse environments. It is essential for scaling artificial intelligence beyond high-power servers and into everyday tools and applications.
Occurs when an artificial intelligence system becomes less accurate over time because the real-world data it encounters changes.
A model that can understand and generate more than one type of data, such as text, images, audio, or video.
The ability to connect information across multiple steps or sources to reach a conclusion.
Performs one specific task or a small group of related tasks. It does not understand anything outside its assigned purpose.
Refers to the multiple possible meanings that a sentence or phrase can have. Artificial intelligence systems must resolve ambiguity to interpret user intent correctly.
The ability of an artificial intelligence system to interpret meaning, intent, and context from human language.
A computer system made of many connected layers that learn patterns from data. Neural networks recognize relationships, make predictions, and support most modern artificial intelligence systems.
Platforms that allow users to build applications without writing any programming code.
Selects from the smallest group of possible next words whose combined probabilities reach a chosen threshold. Also called top-p sampling.
An artificial intelligence system whose learned parameters are publicly released. Open-weight models support transparency and innovation, but they also require careful oversight to prevent misuse.
The process of adjusting a model's parameters to improve performance. It ensures that the model learns efficiently and produces accurate results.
Occurs when a model learns the training data too closely and performs poorly on new data. It means the model memorized instead of generalizing.
The process of dividing a task into smaller parts so they can be completed at the same time. This speeds up artificial intelligence training and response generation.
The internal values that a model learns during training. These values determine how the model recognizes patterns, interprets information, and produces responses.
The ability of artificial intelligence to identify regularities in data, such as repeated behaviors, structures, or relationships.
Adapts a model by training only a small subset of parameters. It reduces cost while preserving performance.
Extensions that allow an artificial intelligence system to interact with external tools, services, or data sources.
AI models guess the next word or token based on patterns, not truth. When the pattern is weak or ambiguous, the model fills the gap incorrectly.
Uses data and patterns to forecast future events or outcomes.
Recommends actions based on predictions. It answers the question: what should we do next?
Represents uncertainty by assigning probabilities to different outcomes. It supports reasoning, prediction, and decision making under uncertainty.
A mathematical description of how likely different outcomes are. The model selects from this distribution based on patterns it learned during training.
The input a user provides to an artificial intelligence system to guide its behavior. A prompt does not program the model. Instead, it activates patterns the model learned during training. Clear and well-structured prompts help the model produce accurate, relevant, and meaningful output.
The skill of writing clear instructions that guide an artificial intelligence system to produce accurate and useful responses.
A security risk in which a user intentionally inserts hidden or misleading instructions into a prompt to manipulate an artificial intelligence system.
Quantization reduces the precision of a model's numerical values to make it smaller and faster. Pruning removes unnecessary weights or connections. Both techniques improve efficiency without significantly reducing performance.
Uses principles of quantum mechanics to perform calculations that are difficult or impossible for classical computers.
The controlled introduction of unpredictability into the model's decision-making process. Randomness does not mean chaos. It is a mathematical technique that widens the range of possible outputs.
The ability of an artificial intelligence system to draw conclusions, make inferences, and follow logical steps.
An artificial intelligence system designed to perform structured thinking tasks such as analysis, inference, planning, and multi-step problem solving.
The technique of using a model's output as the input for the next step. It allows the system to refine ideas, plan tasks, or build multi-step reasoning chains.
A training method where an artificial intelligence system learns by receiving rewards or penalties for its actions.
The commitment to design, build, and deploy artificial intelligence systems that are ethical, transparent, fair, and aligned with human well-being.
A method where an artificial intelligence system searches for information in external sources and uses that information to produce more accurate answers.
The process an artificial intelligence system uses to locate relevant information from external sources before generating a response. It includes converting the user’s query into embeddings, searching through databases or knowledge stores, ranking the most relevant results, and delivering those results to the model for final generation. Retrieval pipelines strengthen accuracy, reduce hallucinations, and ensure that artificial intelligence systems remain grounded in verifiable information. They are essential for retrieval-augmented generation, GraphRAG, and other advanced reasoning methods that rely on external knowledge rather than internal prediction alone. A well-designed retrieval pipeline transforms artificial intelligence from a pattern-matching system into a research-supported reasoning tool.
The field that designs, builds, and operates machines capable of sensing, acting, and interacting with the physical world.
The structured rules, constraints, and oversight mechanisms that prevent artificial intelligence systems from producing harmful, biased, or inappropriate content. They include filtering systems, alignment techniques, monitoring tools, and human-in-the-loop checkpoints that ensure the model behaves predictably and responsibly. Safety guardrails protect individuals and communities by preventing toxic output, reducing hallucinations, and enforcing ethical standards. They also support organizational integrity by ensuring that artificial intelligence systems remain aligned with legal requirements, cultural expectations, and human values. Safety guardrails are essential for responsible artificial intelligence deployment because they transform raw capability into trustworthy, safe, and accountable behavior.
The process an artificial intelligence model uses to select one response from the many possibilities in its probability distribution.
A supervised environment where AI performs tasks and the learner observes.
Describe how the performance of an artificial intelligence model improves as the amount of data, computation, or parameters increases.
A method where a model learns patterns from unlabeled data by predicting missing parts of the data.
Identifies the emotional tone in text, such as positive, negative, or neutral.
A system's ability to understand, interpret, and reason about physical space. It includes recognizing shapes, distances, directions, and spatial relationships.
A type of digital asset designed to maintain a stable value by linking its price to a reference asset such as a fiat currency. Stablecoins are used for payments, trading, and digital settlement because they reduce volatility compared to other cryptoassets. Their stability depends on the issuer’s ability to maintain reserves, honor redemption claims, and operate within a regulated framework. When stablecoins become widely used for payments, they may pose risks to financial stability and therefore require stronger oversight, including prudential regulation, consumer protection rules, and operational safeguards.
A method used to generate images by gradually transforming random noise into a coherent picture.
A method where multiple models are combined to improve performance.
The degree to which a user can guide or shape a model's behavior through prompts, settings, or constraints.
A term used to describe how language models repeat patterns from their training data without true understanding.
Can take actions toward goals, make decisions, and complete tasks with limited human direction. It can plan, reason, and operate across multiple steps.
Information organized in a clear format such as tables, spreadsheets, or databases.
A training method where the model learns from labeled examples provided by humans.
Artificially generated information created by models rather than collected from real-world sources. Synthetic data must be carefully evaluated to ensure it reflects real-world patterns without introducing bias.
A setting that controls how predictable or creative a model's responses will be. A lower temperature produces consistent and highly predictable output. A higher temperature increases creativity, variation, and risk. Temperature does not measure heat. It is a mathematical value that influences how widely the model searches through its learned patterns.
A structured numerical object used to store data in artificial intelligence systems. It can represent a single number, a list, a matrix, or more complex multi-dimensional structures.
A Google-developed chip designed specifically to accelerate neural network machine learning workloads.
A small unit of text, such as a word or part of a word, that artificial intelligence uses to process language.
Tthe process of converting text into smaller units called tokens so an artificial intelligence system can interpret and generate language. These tokens may represent whole words, sub-words, or even individual characters, depending on the model’s design. Tokenization allows the model to break complex language into structured components that can be analyzed mathematically. It also determines how the model handles spelling, punctuation, compound words, and multilingual text. Effective tokenization reduces ambiguity, improves accuracy, and ensures that the model can process language consistently across different contexts. Without tokenization, artificial intelligence systems would be unable to understand or generate human language in a structured and predictable way.
A method that limits the model's choices to the k most likely next words or tokens. A smaller k produces more predictable output, while a larger k increases diversity.
Also called nucleus sampling. Selects from the smallest group of possible next words whose combined probabilities reach a chosen threshold.
The collection of text, images, audio, or other information used to teach an artificial intelligence model. The quality, diversity, and accuracy of the training data directly influence the model's performance and reliability.
If the data is biased, incomplete, or contradictory, the model inherits those flaws.
A type of artificial intelligence model that processes information by examining relationships between words, sentences, or data elements all at once rather than one piece at a time. Transformers are the foundation of many modern AI systems, including large language models.
A human supervisor who oversees AI behavior.
Measures whether a machine can produce responses that seem human to a person.
How people feel, think, and behave when interacting with AI systems. A strong user experience ensures that AI systems are supportive, respectful, and easy to use.
The process of evaluating a model's performance on data it has not seen during training. It ensures that the model generalizes well and does not simply memorize patterns.
A numerical representation of text, images, or other data. Vectors allow artificial intelligence systems to compare meaning, find similarities, and retrieve information.
Stores numerical representations of text, images, or other data known as embeddings. It allows artificial intelligence systems to search for meaning, similarity, and relationships rather than exact matches.
A model that applies transformer techniques to images instead of text. It breaks an image into small patches and uses attention to understand how the patches relate to one another.
The internal numerical values that determine how strongly different inputs influence a model's predictions. Weights do not store facts. They represent mathematical relationships learned from data.
Describes a future work environment where humans and artificial intelligence systems collaborate seamlessly. It emphasizes augmentation, creativity, well-being, and shared decision making.
An artificial intelligence system that builds an internal representation of its environment. It predicts how actions lead to outcomes and uses this understanding to plan, reason, and adapt.
Refers to existential risk, meaning threats that could cause irreversible harm to humanity. Tuners do not manage x-risk, but they should understand the term.
Allows a model to perform a task it has never been explicitly trained on. The model uses its general understanding of language and concepts to infer the correct behavior.