AI Terminology Explained: Your Comprehensive Dictionary
160+ AI Terms Defined A-Z
Complete Dictionary of Agentic AI & Modern Terminology
A
AGI
Artificial General Intelligence - hypothetical system capable of performing any intellectual task a human can.
Abbreviations
AI Agent
A software programme or system designed to perceive its environment (through sensors or data inputs), process that information (reasoning), make decisions, and take actions to achieve specific goals, often with a degree of autonomy.
Agentic AI
AI Alignment
Ensuring AI systems steer towards intended human goals and ethical standards, preventing harmful or unintended behaviours.
Safety
AI Safety
The field concerned with preventing accidents, misuse, and harmful consequences of AI systems, particularly as they become more autonomous.
Safety
ANI
Artificial Narrow Intelligence - AI systems designed and trained for a specific task, like recognising images, playing chess, or powering a chatbot. Most current AI is considered ANI.
Abbreviations
ANN
Artificial Neural Network - a computational model inspired by the structure of the human brain, consisting of interconnected nodes (neurones) that process and transmit information.
Abbreviations
ASI
Artificial Super Intelligence - AI that surpasses human intelligence across all fields.
Abbreviations
Action
The specific output of an agent intended to manipulate the environment (e.g., an API call, database query, or file operation).
Agentic AI
Action Module
The component that translates an AI agent's decisions into actual executions or interactions within the environment.
Architectures
Activation Function
A mathematical function within a neural network neurone that determines whether the neurone 'fires' and passes information to the next layer, introducing non-linearity.
Architectures
Actuators (Effectors)
The components or interfaces that allow an AI agent to take action within its environment (e.g., robotic arms, sending emails, updating a database, displaying information).
Agentic AI
Adaptive RAG
A retrieval strategy that dynamically selects the search method (vector, web, or none) based on query complexity and context.
Techniques
Agency
The capacity of an AI system to act independently and purposefully to achieve a goal without constant human intervention.
Agentic AI
Agent-to-Human Handoff
The seamless transfer of a task, interaction, or decision-making authority from an AI agent to a human operator, typically when the task exceeds the agent's capabilities or requires human judgement.
Safety
AgentOS (AIOS)
An operating system architecture that embeds LLMs into the kernel for agent resource management, scheduling agent threads, and optimising GPU usage.
Architectures
Agentic AI
AI systems designed to autonomously pursue complex goals and workflows with limited supervision. Unlike generative AI which focusses on content creation, agentic AI emphasises goal achievement and task execution.
Agentic AI
Agentic RAG
An evolution of Retrieval-Augmented Generation where the system actively plans search strategies, verifies cited sources, and iterates if initial retrieval is insufficient.
2025 Emerging
Agentic Workflow
A sequence of steps where an agent directs the flow of execution using logic, branching, and tool usage rather than following a linear path.
Agentic AI
Algorithm
A finite sequence of well-defined instructions followed by a computer system to perform a calculation or solve a problem.
Architectures
Artificial General Intelligence (AGI)
Hypothetical AI possessing human-level understanding and capability across all cognitive domains, able to learn and apply knowledge flexibly.
Abbreviations
AutoGPT
An early open-source framework for autonomous AI agents that can chain together LLM calls and tool usage to accomplish goals.
Architectures
Autonomous Agent
An agent operating without constant human intervention, capable of making decisions and taking actions independently to achieve objectives.
Agentic AI
B
Backpropagation
The core algorithm for training neural networks by calculating loss gradients and updating weights to minimise prediction errors.
Architectures
Bias (AI)
Systematic error in AI outputs caused by prejudiced training data, flawed assumptions, or unrepresentative datasets.
Safety
Black Box
An AI system whose internal decision-making process is opaque to humans, making it difficult to understand or audit its reasoning.
Safety
C
CV
Computer Vision - the field of AI that enables computers to 'see' and interpret visual information from images and videos.
Abbreviations
Chain of Thought (CoT)
A reasoning pattern where the model decomposes a problem into intermediate steps, making the reasoning process explicit and verifiable.
Techniques
Chatbot
A programme simulating conversation, typically reactive rather than agentic, responding to user inputs without autonomous goal pursuit.
Architectures
CoT
Chain of Thought - reasoning pattern decomposing problems into intermediate steps.
Abbreviations
Cognitive Architecture
System design modelling human cognitive processes such as memory, reasoning, attention, and learning.
Architectures
Complexity Threshold
The difficulty level where a task requires agentic decomposition or System 2 reasoning rather than simple pattern matching.
Agentic AI
Computer Vision
AI field enabling computers to interpret and understand visual data from images and videos.
Architectures
Constitutional AI
Alignment method using a natural language constitution and AI feedback (RLAIF) to train models to follow ethical principles.
Safety
Context Serialisation
The process of saving and transferring an agent's state (context, memory, conversation) between sessions or agents.
Techniques
Context Window
The maximum amount of text (measured in tokens) a model can process at once in its immediate short-term memory.
Architectures
Copilot
An AI assistant that operates with a human-in-the-loop, providing support and suggestions but not full autonomy.
Agentic AI
Corrective RAG (CRAG)
RAG system that evaluates retrieved documents for relevance and correctness before generation, triggering fallback actions if needed.
Techniques
D
DL
Deep Learning - a subset of Machine Learning that uses artificial neural networks with multiple layers to analyse complex patterns in large datasets, particularly effective for tasks like image and speech recognition.
Abbreviations
DPO
Direct Preference Optimisation - more stable alternative to RLHF for aligning models to preferences.
Abbreviations
Decision-Making Mechanism
The part of an AI agent's architecture responsible for processing input and determining the appropriate action based on its goals, knowledge, and perception.
Architectures
Deep Learning (DL)
Machine learning using multi-layered neural networks to learn hierarchical representations of data.
Architectures
DeepSeek-R1
A prominent System 2 reasoning model that uses test-time compute to perform deliberate, multi-step reasoning.
2025 Emerging
Deterministic Reasoning
Logic-based reasoning that follows strict rules, ensuring predictability and consistency in outputs.
Techniques
Digital Labour
The concept of AI agents performing end-to-end job functions as autonomous workers within organisational workflows.
2025 Emerging
Direct Preference Optimisation (DPO)
Alignment method optimising directly on preference data without requiring a separate reward model, more stable than RLHF.
Techniques
Dynamic Resource Allocation
The ability of an agent or operating system to assign compute resources based on task priority and complexity in real-time.
Architectures
E
Edge AI
AI processing performed locally on devices (phones, IoT) rather than in the cloud, reducing latency and improving privacy.
Architectures
Embeddings
Dense vector representations of data (like words, images, or users) that capture their meaning or characteristics in a multi-dimensional space, allowing AI models to process them effectively.
Techniques
Embodied AI
Agents that control physical bodies (robots) and interact with the physical world through sensors and actuators.
2025 Emerging
Emergent Behaviour
Complex patterns arising from simple agent interactions, where the whole system exhibits capabilities beyond individual components.
Agentic AI
Environment
The context in which an AI agent operates. This can be a digital environment (like a database, a website, or a software system) or a physical one (like a factory floor for a robot).
Agentic AI
Episodic Memory
Storage of specific past events and interactions, allowing agents to recall 'what happened in the last meeting'.
Techniques
Explainable AI (XAI)
Methods to make AI decisions transparent and understandable to humans, addressing the black box problem.
Safety
F
Few-Shot Learning
Learning from a small number of examples, typically provided in the prompt context rather than through fine-tuning.
Techniques
Fine-tuning
Training a pre-trained model on a specific dataset to adapt it for particular tasks or domains.
Techniques
Flow Engineering
Designing the iterative workflows and logic graphs for agents, architecting how data flows between agents, tools, and memory.
2025 Emerging
Foundation Model
A broad, pre-trained model adaptable to many downstream tasks through fine-tuning or prompting.
Architectures
Frugal AI
Development philosophy focussed on creating efficient, low-cost AI systems using SLMs, quantisation, and optimisation.
2025 Emerging
Function Calling
The ability of a model to output structured data (JSON) that triggers external code, tools, or API endpoints.
Agentic AI
G
GAIA
General AI Assistants Benchmark - tests fundamental abilities like reasoning, tool use, and multimodality on real-world tasks.
Techniques
GenAI
Generative AI - a category of AI models capable of creating new content, such as text, images, music, code, or synthetic data, based on the data they were trained on.
Abbreviations
Generative AI
AI capable of creating new content (text, images, code) based on learned patterns from training data.
Architectures
Goal Decomposition
Breaking high-level, abstract goals into executable sub-tasks that can be tackled sequentially or in parallel.
Agentic AI
Goal-Based Agent
An AI agent that operates with explicit goals and plans a sequence of actions to achieve those goals.
Agentic AI
Graph of Thoughts (GoT)
Non-linear reasoning structure where reasoning steps can be combined and aggregated from arbitrary points in the thought process.
Techniques
GraphRAG
RAG using Knowledge Graphs for structured retrieval, enabling multi-hop reasoning by traversing entity relationships.
2025 Emerging
Grounding
Linking AI outputs to verifiable sources or real-world data to reduce hallucinations and increase accuracy.
Techniques
H
Hallucination
Generation of incorrect, fabricated, or nonsensical information with high confidence, a key challenge in LLMs.
Safety
Headless AI Agent
Backend agent operating without a user interface, performing tasks by directly manipulating software and APIs.
Agentic AI
Hierarchical Multi-Agent System (HMAS)
Layered agent organisation with supervisors managing workers, creating clear command structures for complex tasks.
Architectures
Human-in-the-Loop (HITL)
Interaction pattern requiring human validation and approval before critical actions are executed.
Safety
I
Inference
The process of using a trained model to generate outputs or predictions on new inputs.
Architectures
Inference-Time Compute
The paradigm where models are allocated extended computational time during generation to explore multiple reasoning paths.
2025 Emerging
Intent Recognition
Understanding the underlying goal behind a user input, essential for routing and task planning.
Techniques
Interruptibility
The ability to safely stop an autonomous agent mid-execution, a critical safety feature ('The Red Button').
Safety
J
JEPA/V-JEPA
Joint-Embedding Predictive Architecture - a model architecture for predicting relationships in data, particularly video.
Architectures
K
Kahneman-Tversky Optimisation (KTO)
Alignment method using unpaired binary feedback based on psychological principles of decision-making.
Techniques
Knowledge Distillation
Transferring knowledge from a large 'teacher' model to a smaller 'student' model whilst preserving performance.
Techniques
Knowledge Graph
Structured data network of entities (nodes) and relationships (edges) representing semantic information.
Architectures
L
LAM
Large Action Model - model trained to execute UI actions and interact with software.
Abbreviations
LLM
Large Language Model - model trained on text to generate language.
Abbreviations
Large Action Model (LAM)
Model trained to execute UI actions, understanding interfaces and software logic to navigate and interact autonomously.
2025 Emerging
Large Language Model (LLM)
Model trained on vast text data to generate human-like language and understand linguistic patterns.
Architectures
Learning Agent
An AI agent that improves its performance over time based on its experiences and interactions, often using machine learning techniques.
Agentic AI
Liquid Neural Networks (LNNs)
Adaptive networks with time-continuous states that adjust after training, ideal for dynamic environments and robotics.
2025 Emerging
LoRA (Low-Rank Adaptation)
Efficient fine-tuning technique that adds small trainable matrices to frozen model weights, reducing computational cost.
Techniques
M
MAS
Multi-Agent System - system of interacting independent agents.
Abbreviations
MCP
Model Context Protocol - universal standard for connecting AI to tools and data.
Abbreviations
ML
Machine Learning - a subset of AI that focusses on enabling systems to learn from data, identify patterns, and make decisions without being explicitly programmed for every scenario.
Abbreviations
Machine Learning (ML)
Systems that learn patterns from data without explicit programming, improving performance through experience.
Architectures
Mixture of Agents (MoA)
Architecture combining multiple LLMs where proposer models generate responses and an aggregator synthesises the best output.
2025 Emerging
Mixture of Experts (MoE)
Model architecture activating only subsets of parameters (experts) for each token, improving efficiency and capability.
Architectures
MoA
Mixture of Agents - architecture combining multiple LLMs for superior outputs.
Abbreviations
MoE
Mixture of Experts - model activating subsets of parameters for efficiency.
Abbreviations
Model
The output of a machine learning algorithm after it has been trained on data. It represents what the algorithm has learned and is used to make predictions or decisions on new data.
Architectures
Model Context Protocol (MCP)
Universal standard for connecting AI systems to data and tools, solving the N×M integration problem ('USB-C for AI').
2025 Emerging
Model-Based Reflex Agent
An AI agent that maintains an internal model or representation of the environment to help it make decisions, even if the environment is not fully observable. It uses memory to track the state of the world.
Architectures
Multi-Agent System (MAS)
System of multiple interacting independent agents collaborating to solve complex problems beyond single-agent capability.
Agentic AI
Multimodal AI
AI processing multiple data types - text, images, audio, video - enabling richer understanding and generation.
Architectures
N
NLP
Natural Language Processing - field enabling computers to understand human language.
Abbreviations
Natural Language Processing (NLP)
Field concerned with interaction between computers and human language, enabling understanding and generation of text.
Architectures
Neuro-symbolic AI
Combining neural networks (pattern recognition) with symbolic logic (reasoning) for more robust and explainable systems.
2025 Emerging
O
Ontology
Formal representation of concepts and relationships within a domain, providing structured knowledge representation.
Architectures
Orchestrator
Central agent managing task delegation in multi-agent systems, coordinating workers and synthesising outputs.
Agentic AI
Overfitting
When an AI model performs exceptionally well on the data it was trained on but poorly on new, unseen data because it has essentially just memorised the training data rather than learning general patterns.
Safety
P
Parameter-Efficient Fine-Tuning (PEFT)
Training methods that update only a small subset of model parameters, reducing computational requirements.
Techniques
Perception
An AI agent's ability to receive and interpret information from its environment using sensors (in physical agents) or data inputs/APIs (in software agents).
Agentic AI
Plan-and-Solve
Prompting strategy where agents explicitly plan their approach before executing actions, improving success rates.
Techniques
Planning Module
A component that enables an AI agent to develop a sequence of steps or a strategy to achieve a specific goal.
Architectures
Polyphonic AI
System where agents operate in parallel providing simultaneous multi-perspective analysis rather than sequential processing.
2025 Emerging
Procedural Memory
Memory of skills and how to execute tasks, encoded as tool definitions and prompt libraries.
Techniques
Prompt
The input or instruction given to an AI model, especially a generative AI or LLM, to guide its output. It can be a question, a command, or a piece of text to be continued.
Techniques
Prompt Engineering
Optimising inputs and instructions to guide AI outputs towards desired behaviours and formats.
Techniques
Pruning
Removing unnecessary parameters from a model to compress it whilst maintaining performance.
Techniques
Q
Quantisation
Reducing model precision (e.g., from 32-bit to 4-bit) to decrease memory usage and increase inference speed.
Techniques
R
RAG
Retrieval-Augmented Generation - combining a generative model with retrieval systems to ground answers in specific data sources.
Abbreviations
RLHF
Reinforcement Learning from Human Feedback - fine-tuning models based on human preference rankings.
Abbreviations
RPA
Robotic Process Automation - software robots designed to mimic human actions when interacting with digital systems to perform repetitive, rule-based tasks. Often enhanced with AI.
Abbreviations
ReAct (Reason + Act)
Prompting paradigm where models generate explicit 'Thought' (reasoning), 'Action' (tool call), and 'Observation' (result) loops.
Techniques
ReWOO
Reasoning WithOut Observation - decoupling planning from execution by generating complete plans upfront before tool use.
Techniques
Reasoning Engine
The core model or system driving agent decisions, typically an LLM or specialised reasoning model.
Architectures
Reflection
Agent pattern of critiquing its own output to improve future performance through self-assessment.
Techniques
Reflexion
Framework for reinforcement learning via linguistic feedback, where agents learn from error messages and critiques.
Techniques
Reinforcement Learning (RL)
Learning paradigm where agents learn through rewards and penalties based on action outcomes.
Architectures
Reinforcement Learning from AI Feedback (RLAIF)
Using AI-generated ratings and feedback to train models, scaling beyond human annotation capacity.
Techniques
Reinforcement Learning from Human Feedback (RLHF)
Training method using human ratings of model outputs to align behaviour with human preferences.
Techniques
Retrieval-Augmented Generation (RAG)
Enhancing model outputs by retrieving relevant information from external knowledge sources before generation.
Techniques
S
Self-RAG
Model self-assessing relevance and quality during generation using reflection tokens to verify evidence support.
Techniques
Semantic Memory
Storage of general facts and world knowledge (e.g., 'What is the company's refund policy?').
Techniques
Sensors
The components or interfaces that allow an AI agent to perceive its environment (e.g., cameras, microphones, data feeds, APIs).
Agentic AI
Sentiment Analysis
Determining emotional tone and attitude expressed in text, from positive to negative.
Techniques
Simple Reflex Agent
The most basic type of AI agent that makes decisions based only on the current perception of the environment, using a set of predefined condition-action rules. It has no memory.
Architectures
Small Language Model (SLM)
Compact, efficient models optimised for specific tasks or edge deployment with reduced computational requirements.
Architectures
Superintelligence
Hypothetical AI surpassing human intelligence across all domains by orders of magnitude.
Abbreviations
Supervised Learning
A type of machine learning where the algorithm is trained on labelled data (input-output pairs), learning to map inputs to correct outputs.
Techniques
Supervisor Pattern
Hierarchical agent orchestration where a supervisor coordinates and delegates to specialised worker agents.
Architectures
Swarm Intelligence
Decentralised collective agent behaviour where complex global patterns emerge from local interactions.
2025 Emerging
System 1 Thinking
Fast, intuitive generation based on pattern matching - the default mode of standard LLMs.
Techniques
System 2 Thinking
Slow, deliberate reasoning involving explicit problem decomposition and verification steps.
2025 Emerging
T
Tensor
A mathematical object used to represent data in deep learning, similar to a multi-dimensional array.
Architectures
Test-Time Compute
Computing resources allocated during inference for reasoning and exploration, the key scaling law of 2025.
2025 Emerging
Token
Basic unit of text processing in LLMs, typically representing words, subwords, or characters.
Architectures
Tokenizer
The component in NLP models that breaks down raw text into tokens for processing.
Techniques
Tool Use
Agent ability to invoke external software, APIs, calculators, or code execution environments.
Agentic AI
Training Data
The dataset used to teach an AI model. The quality and quantity of this data significantly impact the model's performance.
Techniques
Transformer
Foundational architecture of modern LLMs based on self-attention mechanisms, introduced in 2017.
Architectures
Tree of Thoughts (ToT)
Branching reasoning structure exploring multiple solution paths before selecting the optimal approach.
Techniques
Trusted Agent
Agent certified through governance checks for high-stakes autonomous operation without immediate oversight.
2025 Emerging
U
Unsupervised Learning
A type of machine learning where the algorithm learns from unlabelled data, finding patterns, structures, or relationships within the data on its own (e.g., clustering data).
Techniques
Utility-Based Agent
A sophisticated agent that considers not just achieving a goal, but also the 'utility' or desirability of the outcome. It evaluates different possible actions and chooses the one expected to maximise its utility function.
Agentic AI
V
Vector Database
Database optimised for semantic similarity search using high-dimensional embeddings, essential for RAG systems.
Architectures
Vector Space
The multi-dimensional space where embeddings exist, allowing mathematical operations to represent relationships between data points.
Techniques
Vibe Coding
Development style where humans provide high-level intent ('vibes') and AI handles implementation details.
2025 Emerging
W
World Model
Internal simulation of the environment allowing agents to predict action consequences before execution.
Agentic AI
Z
Zero-Shot Learning
Performing tasks without any training examples, relying solely on pre-trained knowledge and instructions.
Techniques
Contact Us
This web site is operated by Hello Leads Ltd.
Hello Leads Ltd does not provide any finance advice and only acts as an introducer to regulated companies.
Please fill in our no obligation form and we will get back to you regarding your enquiry.
