AI Glossary
A comprehensive guide to AI and LLM terminology
A
Agent Memory
The capacity of an AI agent to store and retrieve past data in order to enhance how it performs tasks.
Agentic AI
AI system that operates on its own, making decisions and interacting with tools as needed.
Agentic RAG
An advanced RAG setup where the AI agent(s) chooses on its own when to search for external information and how to use it.
Agents
AI systems that make decisions on the fly, choosing actions and tools without being manually instructed each step.
Alignment
The final training phase to align model's actions with human ethics and safety requirements. QA for values and safety.
Attention Mechanism
A component in neural networks that helps the model focus on relevant parts of input data, similar to how humans pay attention to specific aspects of information.
B
Bias
Systematic errors in AI models that can lead to unfair or prejudiced outcomes, often reflecting biases present in training data.
C
Chain of Thought (CoT)
A reasoning strategy where an AI agent/model explains its thinking step-by-step, making it easier to understand and improving performance.
Chaining
A method where an AI agent completes a task by breaking it into ordered steps and handling them one at a time.
Cognitive Architecture
The overall structure that manages how an AI system processes input, decides what to do, and generates output.
Context Window
The maximum amount of text or tokens that an AI model can process at once, determining how much context it can consider.
Continual Pretraining
A training method where AI model improves by learning from large sets of new, unlabeled data.
D
Design Patterns
Reusable blueprints or strategies for designing effective AI agents.
Distillation
Compressing a large AI's knowledge into a smaller model by teaching it to mimic predictions.
E
Embedding
A numerical representation of text, images, or other data that captures semantic meaning and relationships in a vector space.
F
Few-Shot Learning
A technique where an AI model learns to perform a task with only a few examples, rather than requiring extensive training data.
Fine-Tuning
The process of adapting a pre-trained model to a specific task by training it on a smaller, task-specific dataset.
Function Calling
Allows AI agents to dynamically call external functions based on the requirements of a specific task.
H
Hallucination
When an AI model generates false or made-up information that appears to be factual, often occurring in text generation.
Human-In-The-Loop (HITL)
A design where humans stay involved in decision-making to guide the AI's choices.
Hybrid Models
Models that blend fast and deep thinking. Adapting their reasoning depth depending on how hard the problem is.
I
Inference
The process of using a trained AI model to make predictions or generate outputs based on new input data.
K
Knowledge Graph
A structured representation of information that shows relationships between different entities, helping AI systems understand connections.
L
Latent Space
A compressed representation of data where similar items are closer together, allowing AI models to understand and generate new content.
LLM
AI model that creates content like text or images, often used in generative tasks.
Long-Term Memory
Memory that enables an AI to keep and access information across multiple sessions or tasks. What we see in ChatGPT, Claude, etc.
LRM
Large Reasoning Models: built for complex, logical problem-solving beyond simple generation.
M
MCP (Model Context Protocol)
A protocol that allows AI agents to connect with external tools and data sources using a defined standard, like how USB-C lets devices plug into any compatible port.
Mixture of Experts (MoE)
A neural network model setup that directs tasks to the most suitable sub-models for better speed and accuracy.
Model Parameters
The internal variables that an AI model learns during training, determining how it processes and responds to input data.
Multi-Agents
A setup where several AI agents work together, each handling part of a task to achieve a shared goal more effectively.
N
Neural Network
A computing system inspired by biological neural networks, consisting of interconnected nodes that process information.
O
One-Shot Learning
A machine learning approach where a model learns to recognize new objects or concepts from a single example.
Orchestrator
A coordinator that oversees multiple AI agents, assigning tasks and deciding who does what and when. Think about it as a manager of agents.
Overthinking
When an AI agent spends too much time or uses excessive tokens to solve a task often fixed by limiting how deeply it reasons.
P
Post-Training
Further training of a model after its initial build to improve alignment or performance. Pretty same what's Alignment.
Procedural Memory
AI's ability to remember how to perform repeated tasks, like following a specific process or workflow it learned earlier.
Prompt Engineering
The practice of carefully crafting input prompts to get desired outputs from AI models, especially in language models.
Q
Quantization
The process of reducing the precision of model parameters to make AI models more efficient and faster to run.
R
ReAct
An approach where an AI agent combines reasoning and acting. First thinking through a problem, then deciding what to do.
Reasoning
AI models that evaluate situations, pick tools, and plan multi-step actions based on context.
Reflection
A method where an AI agent looks back at its previous choices to improve how it handles similar tasks in the future.
Reinforcement Learning
Method of training where AI learns by trial and error, receiving rewards or penalties.
Reinforcement Learning from Human Feedback (RLHF)
Uses human feedback to shape the model's behavior through rewards and punishments.
Retrieval-Augmented Generation (RAG)
A technique where model pulls in external data to enrich its response and improve accuracy or get a domain expertise.
Routing
A strategy where an AI system sends parts of a task to the most suitable agent or model based on what's needed.
S
Self-Healing
When an AI agent identifies its own errors and fixes them automatically. No human involvement or help needed.
Self-Supervised Learning
A training approach where models learn from unlabeled data by creating their own learning objectives.
Short-Term Memory
A form of memory in AI that holds information briefly during one interaction or session.
Structured Outputs
A method where AI agents or models are required to return responses in a specific format, like JSON or XML, so their outputs can be reliably used by other systems, tools or can be just copy/pasted elsewhere.
Supervised Fine-Tuning
Training AI model with labeled data to specialize in specific tasks and improve performance.
T
Temperature
A parameter that controls the randomness of AI model outputs, with higher values leading to more creative but potentially less focused responses.
Test-Time Scaling
A technique that lets an AI agent adjust how deeply it thinks at runtime, depending on how complex the task is.
Token
The basic unit of text that an AI model processes, which can be a word, part of a word, or a character.
Tools
External services or utilities that an AI agent can use to carry out specific tasks it can't handle on its own. Like web search, API calls, or querying databases.
Transfer Learning
Using knowledge gained from solving one problem to help solve a different but related problem.
Transformer
A type of neural network architecture that uses attention mechanisms to process sequential data, forming the basis of many modern LLMs.
V
Vector Database
A specialized database that stores and retrieves vector embeddings, enabling efficient similarity search for AI applications.
Vertical Agents
Agents built for a specific field like legal, healthcare, or finance, so they perform better in those domains.
W
Weight
A parameter in a neural network that determines the strength of connections between neurons.
Workflows
Predefined logic or code paths that guide how AI system, models and tools interact to complete tasks.
Z
Zero-Shot Learning
The ability of an AI model to perform tasks it hasn't been explicitly trained on, using its general understanding.