Calda’s AI Glossary

Abstraction in AI refers to simplifying complex systems by focusing on high-level concepts while hiding lower-level implementation details.

Adaptive AI can adjust its behavior and improve performance over time based on new data and environmental feedback.

Agent-based AI workflows use autonomous agents to perform tasks by reasoning, acting, and coordinating across systems or tools.

AGI refers to a theoretical AI that can understand, learn, and apply knowledge across any domain like a human being.

An AI adoption framework outlines the strategic, technical, and organizational steps a company should follow to implement AI solutions effectively.

An AI agent is an autonomous system that uses AI to perceive its environment, reason, and take actions toward a defined goal.

AI agent architecture is the structural design that defines how the components of an AI agent (like reasoning engine, memory, and tools) interact to perform tasks.

AI agent frameworks are toolkits or platforms that simplify the development of agent-based systems by providing components like memory, reasoning, and tool integration.

AI auditing frameworks provide guidelines and processes for evaluating the fairness, transparency, and accountability of AI systems.

The AI Bill of Rights is a set of principles proposed to protect individuals from harm caused by automated systems, covering fairness, privacy, transparency, and accountability.

AI compliance frameworks are structured guidelines that ensure AI systems meet legal, ethical, and industry standards.

AI context awareness is the ability of AI systems to understand and respond based on situational information, such as user intent, location, or past interactions.

AI copilot tools are AI-powered assistants embedded in workflows to help users write code, generate content, or make decisions in real time.

AI data anonymization involves masking or altering personal identifiers in datasets to preserve privacy while enabling data use.

AI data scraping is the automated extraction of information from websites or digital sources to create training datasets or support AI functions.

AI diffusion models are generative models that create images or other outputs by iteratively denoising data from random noise.

AI embeddings are numerical representations of data (like words or images) in a vector space that preserve relationships and meaning.

AI evaluation techniques are methods used to measure an AI system’s performance, including accuracy, bias, robustness, and generalization.

AI fine-tuning is the process of retraining a pre-trained model on a smaller, domain-specific dataset to specialize its behavior.

AI hallucinations refer to instances where AI models generate outputs that are factually incorrect or entirely fabricated.

AI observability tools provide visibility into how AI models perform, predict, and fail.

AI overfitting occurs when a model learns training data too well—including noise—resulting in poor generalization to new data.

AI prompt frameworks are structured approaches to designing effective prompts that guide AI model behavior and outputs.

An AI token is a unit of text (often a word or subword) used as input or output by language models for processing language.

AI tokenization is the process of breaking down text into smaller units (tokens) that an AI model can understand and process.

AI TTS is a technology that converts written text into human-like spoken audio using machine learning.

AI workflows are orchestrated sequences of tasks involving AI models, data, and tools to automate and solve business or technical problems.

Artificial Intelligence (AI) is the field of creating systems that simulate human intelligence to perform tasks like reasoning, learning, and decision-making.

Bayesian networks are probabilistic graphical models that represent relationships among variables using conditional dependencies.

Bias in AI refers to unfair or prejudiced outcomes in AI systems, often stemming from imbalanced or flawed training data.

Cognitive architectures are computational frameworks that model human-like reasoning, memory, and learning in AI systems.

CoT prompting is a technique that encourages language models to explain reasoning steps explicitly before answering, improving performance on complex tasks.

Custom GPTs are personalized versions of OpenAI’s GPT models, tailored using specific instructions, data, or functions for specialized use cases.

Deep learning is a subset of machine learning that uses multi-layered neural networks to model complex patterns in data.

Deterministic models always produce the same output for a given input, with no randomness in behavior or results.

Discriminative AI models focus on distinguishing between classes by learning decision boundaries from labeled data.

Embeddings are vector representations of text, images, or other data that capture semantic relationships in a high-dimensional space.

Ethical AI principles are guidelines ensuring AI systems are developed and deployed in ways that are fair, transparent, accountable, and aligned with human values.

Explainable AI aims to make the decision-making processes of AI systems transparent and understandable to humans.

Foundation models are large pre-trained neural networks that can be adapted to a wide variety of downstream tasks via prompting or fine-tuning.

Fuzzy logic is a form of logic that allows reasoning with degrees of truth rather than strict true/false values, useful for handling uncertainty in AI systems.

GANs are AI models that use two neural networks—a generator and a discriminator—that compete to create increasingly realistic synthetic data.

Generative AI guardrails are safety mechanisms and policies put in place to prevent harmful, biased, or inappropriate outputs from AI models.

GPT-4o is a multimodal version of OpenAI’s GPT-4 model, capable of processing and reasoning across text, images, and audio in real time.

GPTs are large language models trained on massive datasets to generate human-like text using transformer-based neural network architectures.

Grounding in AI refers to linking abstract concepts or outputs from a model to real-world data, context, or actions to ensure relevance and accuracy.

HITL is an approach where human feedback or intervention is incorporated into the AI training or decision-making process to improve accuracy and safety.

This involves connecting AI agents to external tools or services via APIs to enable them to perform actions, retrieve data, or automate workflows.

LangChain is an open-source framework for building AI agents and applications using large language models, with support for chaining tools, memory, and logic.

LCEL is a declarative language within LangChain for defining complex chains of prompts, tools, and models in a structured and readable way.

LLMOps refers to the set of tools and practices for deploying, monitoring, and managing large language models in production environments.

MMLU is a benchmark that evaluates a language model’s multitask accuracy across diverse academic and professional subjects.

Model drift occurs when an AI model's performance degrades over time due to changes in input data distributions or external conditions.

Multimodal AI refers to models that can process and integrate information from multiple data types—such as text, images, and audio—simultaneously.

Multimodal LLMs are large language models that accept and generate inputs and outputs across various modalities, including text, images, and sound.

NAS is a technique that uses algorithms to automatically discover the optimal neural network architecture for a given task.

This is a framework developed by the U.S. National Institute of Standards and Technology to guide organizations in managing risks related to AI systems.

NLG is a subfield of AI focused on generating human-like text from structured or unstructured data.

NLI is a task in NLP that involves determining whether a given hypothesis logically follows from, contradicts, or is neutral toward a given premise.

One-shot learning is an AI technique where a model learns to recognize patterns or make decisions from just one labeled example.

Open-source LLMs are large language models whose code and/or training data are made publicly available for use, study, and modification.

Predictive AI refers to systems that analyze historical data to forecast future outcomes or behaviors.

Prompt design is the practice of crafting inputs to guide large language models toward generating specific or accurate responses.

Prompt engineering involves strategically designing and refining prompts to improve the performance and reliability of LLM outputs.

Prompt injection is a security vulnerability where users manipulate inputs to influence or override the behavior of an AI model.

Prompt tuning is a fine-tuning technique where a small set of learnable parameters is optimized to improve model performance on specific tasks.

RAG is an approach that combines LLMs with external search or knowledge retrieval to improve the relevance and accuracy of generated responses.

Reinforcement learning is a training method where an AI learns to make decisions by receiving rewards or penalties for its actions over time.

RNNs are neural networks designed to handle sequential data by maintaining memory of previous inputs through internal loops.

Semi-supervised learning uses a combination of labeled and unlabeled data to train models, reducing the need for large annotated datasets.

SLMs are lightweight versions of large language models designed for lower computational cost while retaining task-specific capabilities.

Synthetic data is artificially generated data that mimics real-world data, used to train or test AI models without privacy or scarcity issues.

Task-specific language models are AI models fine-tuned to perform a single, well-defined language task with high accuracy.

The EU AI Act is a comprehensive regulatory framework from the European Union aimed at ensuring safe and ethical AI development and deployment.

Topic modeling is an NLP technique that identifies themes or topics in large collections of text data using unsupervised learning methods.

Unsupervised learning is a machine learning method where algorithms discover hidden patterns in data without using labeled outputs.

Vector databases are optimized to store and retrieve high-dimensional vector embeddings used for similarity search in AI systems.

Vector search is a retrieval method that finds items most similar to a query based on mathematical proximity in a high-dimensional space.

Zero-shot learning allows AI models to perform tasks they haven’t been explicitly trained on by leveraging generalized knowledge from other domains.