DriveGrowthHQ’s AI Glossary offers plain-English definitions of key terms, from LLMs and chatbots to training, safety, governance, and emerging AI systems. Each entry breaks down what the term means, why it matters, and how it fits into the bigger AI picture.
How to use this AI glossary: This glossary is designed for both quick lookups and deeper learning. Here’s how you can use it:
- Search directly – Use ⌘/Ctrl‑F to jump to a specific term fast or click on each category grouping below
- Browse by category – Terms are grouped (Core Concepts, Training, Safety, etc.) so you can scan the section you need here
- Core Concepts — such as Artificial Intelligence (AI)
- Major AI Systems (by Company) — such as ChatGPT
- Model Types & Modalities — such as Large Language Model (LLM)
- Training & Tuning — such as Reinforcement Learning from Human Feedback (RLHF)
- Inference & Deployment — such as Latency
- Prompting, Orchestration & Agents — such as Retrieval-Augmented Generation (RAG)
- Evaluation & Quality — such as Hallucination
- Safety, Security & Privacy — such as Prompt Injection
- Data & Knowledge — such as Corpus
- Search, SEO & Content Ops (for AI‑Ready teams) — such as Google’s AI Mode
- Robotics — such as SLAM
- Legal, Ethics & Governance — such as Bias
- Practical Architecture — such as Pipeline Orchestration
- Developer Essentials — such as API
- Emerging Methods — such as Constitutional AI
- Quick Math & Hardware Terms — such as Floating Point (FP16, BF16)
- Skim the bolded term + short line – Each starts with a plain definition you can grab in seconds.
- Read the paragraph – The follow‑up sentences explain why it matters and often give a tiny example.
- Follow links or related terms – Some entries point to connected ideas so you can build understanding step by step.
Core Concepts
Artificial Intelligence (AI)
Software that can do tasks we link to human thinking, like recognizing patterns or making choices. It learns from data or rules to make better guesses over time.
Artificial General Intelligence (AGI)
AI that can understand, learn, and reason across many tasks — much like a human. Unlike today’s narrow systems, AGI would handle any intellectual job a person can do, from science to conversation, without retraining.
Artificial Superintelligence (ASI)
A hypothetical stage where AI surpasses human intelligence in nearly every field — reasoning, creativity, and strategy. ASI is still theoretical, but it shapes debates about alignment, safety, and long-term control.
Machine Learning (ML)
A way to build AI by letting computers learn from examples instead of fixed rules. The model finds patterns and uses them to predict outcomes on new data.
Generative AI
AI that creates new content — like text, images, or music — instead of just analyzing data. Models like ChatGPT or DALL·E are examples.
Large Language Model (LLM)
A deep learning model trained on massive text data to predict the next word or token. Powers chatbots, summarizers, and reasoning tools.
Prompt
The instruction or input you give an AI model. Clear prompts lead to more accurate, useful responses.
Hallucination
When an AI produces incorrect or made-up information. Happens when it fills gaps with patterns that sound right but aren’t factual.
Reinforcement Learning from Human Feedback (RLHF)
A way to make models respond more helpfully by training them using ratings from human reviewers.
Chatbot
An AI system that converses with users in natural language — answering questions, automating support, or assisting creatively.
Deep Learning
A type of ML that stacks many layers of simple math units called neurons. These layers learn features step by step, which helps with images, audio, and language.
Neural Network
A web of connected nodes (“neurons”). Each node mixes incoming numbers, applies a small function, and passes the result onward. Many small steps add up to smart behavior.
Transformer
A neural network design that uses attention to focus on the most useful parts of the input. It powers most modern language and vision models.
GPT (Generative Pre-Trained Transformer)
A type of large language model built by OpenAI. “Generative” means it creates new text, “Pre-Trained” means it learned from massive amounts of online data before fine-tuning, and “Transformer” refers to the neural network design that helps it understand relationships between words. ChatGPT and many other AI tools are based on this GPT architecture.
Token
A small chunk of text (like “cat”, “##ting”, or a space). Models read and write tokens, not full words. Token limits shape how much the model can “see” at once.
Context Window
The maximum number of tokens a model can take in and work with at one time. Bigger windows let models remember more of a conversation or document.
Embedding
A list of numbers that captures meaning. Similar things have embeddings that sit close together in this number space. Used for search, clustering, and memory.
Attention
A way for a model to weigh which tokens matter most. It lets the model “look back” over the input and pick out key parts when forming the next token.
Major AI Systems (by Company)
ChatGPT (OpenAI)
A conversational AI powered by OpenAI’s GPT models. Known for natural chat, coding help, and creative writing. Free and paid tiers exist, with Pro versions offering faster, more powerful models. Built on OpenAI’s GPT architecture — “Generative Pre-Trained Transformer”.
Claude (Anthropic)
A chatbot from Anthropic focused on safety and helpfulness. Claude is tuned with “Constitutional AI,” a training method that follows written principles. Often praised for balanced, cautious answers.
Grok (xAI)
Developed by Elon Musk’s xAI, integrated with X (formerly Twitter). Marketed as witty and connected to real‑time platform data.
Perplexity AI (Perplexity)
A search‑focused AI assistant. Answers with citations and sources, blending chatbot and search engine. Competes with Google AI Overviews and ChatGPT browsing.
Gemini (Google DeepMind)
Google’s flagship AI family, built on the Gemini architecture. Used in Google Search (AI Overviews, AI Mode), Workspace tools, and APIs. Multimodal, combining text, code, and images.
Copilot (Microsoft)
A branding for Microsoft’s integrations of OpenAI’s models into products like Office, GitHub, and Windows. GitHub Copilot is a developer tool that suggests code in real time.
Meta AI (Meta / formerly Facebook)
Meta’s assistant powered by LLaMA models, integrated into Facebook, Instagram, and WhatsApp.
Llama (Meta)
An open‑weight family of models released by Meta. Popular with developers because it can be fine‑tuned for private or custom use. Llama powers many open‑source AI apps.
Mistral (Mistral AI)
European startup offering compact but powerful models, including “Mixtral” (a mixture‑of‑experts model). Known for efficiency and open‑weight releases.
DeepSeek (DeepSeek AI)
A fast-growing Chinese AI company known for open-weight large language models like DeepSeek-R1 and V3. It focuses on efficient training and cost-effective performance using mixture-of-experts designs, positioning it as an emerging global rival to ChatGPT and Gemini.
Command R (Cohere)
A family of models tuned for retrieval‑augmented generation (RAG). Geared toward enterprise search and knowledge management tasks.
Pi (Inflection AI)
A personal AI designed to be supportive and empathetic in tone. Focused on conversation rather than technical depth.
Model Types & Modalities
Large Language Model (LLM)
A model trained on huge text sets to predict the next token. It can chat, write, summarize, translate, and reason to a degree.
Chatbot
A program powered by rules or AI that talks with users in natural language. Modern chatbots often use LLMs to answer questions, give support, or act like an assistant.
Synthetic Agents / Digital Humans
AI avatars or characters that simulate human tone, emotion, and interaction in real time.
Multimodal Model
A model that can work across text, images, audio, or video. For example, it can read a photo and answer questions in text.
Vision Model
A model that understands images or video. It handles tasks like object detection, image captions, and scene understanding.
Embodied AI
AI connected to a physical body — robots, drones, or wearables — that perceives, decides, and acts in the real world.
Speech‑to‑Text (ASR)
Tech that turns spoken words into written text. Useful for captions, notes, and voice commands.
Text‑to‑Speech (TTS)
Tech that turns written text into natural‑sounding speech. Used in voice assistants and accessibility tools.
Diffusion Model
A generative model that starts with random noise and removes it step by step to form an image or sound. Popular in image generation.
Mixture‑of‑Experts (MoE)
A model with many “experts,” where only a few activate for each input. This saves compute while keeping strong results.
Training & Tuning
Pretraining
The first big learning phase on vast data, often with a simple goal like “predict the next token.” It gives the model broad knowledge.
Supervised Fine‑Tuning (SFT)
Training on labeled examples to teach the model task‑specific behavior (e.g., “given a question, write a short, polite answer”).
RLHF (Reinforcement Learning from Human Feedback)
A step where humans rate model outputs. The model is then nudged to produce responses people prefer.
DPO / ORPO (Preference Optimization)
Training methods that align a model to chosen answers without a full reinforcement loop. They are simpler than classic RLHF.
PEFT (Parameter‑Efficient Fine‑Tuning)
Ways to adapt a big model by training only a small part of it. Saves time and money.
LoRA / QLoRA
PEFT methods that add small “adapters” (LoRA) and can use lower‑precision numbers (QLoRA). They make fine‑tuning cheaper.
Distillation
Teaching a smaller model to mimic a larger one. You keep most of the skill with less compute.
Quantization (8‑bit, 4‑bit)
Storing numbers with fewer bits to shrink models and speed them up. Some accuracy can drop, so you balance size and quality.
Curriculum Learning
Training on easier examples first, then harder ones. Like how people learn math step by step.
Frontier / Reasoning Models
Advanced models built for logic and reasoning, aiming to go beyond simple prediction or pattern matching.
Retrieval Transformer (RETRO / Hyena)
New architectures that blend retrieval and transformer design to scale efficiently on long-context tasks.
Synthetic Data
Data created by models or rules. Used to fill gaps, balance classes, or protect privacy.
Contrastive / Retrieval-Augmented Contrastive Models
Models trained by comparing correct vs. incorrect answers to improve accuracy in retrieval-based generation.
Inference & Deployment
Inference
Running a trained model to get outputs (like a chat reply or a label). This is the “serving” part after training is done.
Latency
How long it takes to get a response. Users notice delays, so lower latency is better.
Throughput
How many requests a system can handle per second or minute. Higher is better for busy apps.
KV Cache
A memory that stores past attention keys and values so the model can respond faster during long chats.
Speculative Decoding
A speed‑up trick. A small model guesses several tokens. The big model checks them. If they look fine, the system accepts them and moves on.
Beam Search / Sampling (Temperature, Top‑p)
Ways to pick the next token. Beam search looks for high‑probability paths. Sampling adds randomness. Temperature and top‑p control how “creative” outputs feel.
Temperature
A model setting that controls how random or creative the output is. Lower values (like 0.2) make responses more focused and factual, which can reduce hallucinations. Higher values (like 0.8 or 1.0) make the model explore more options, leading to varied or creative answers.
Serverless / Edge Inference
Running models on demand or near users. Cuts cost and lag for global apps.
Custom / On-Device AI (TinyML)
AI models optimized to run directly on phones or local devices for speed, privacy, and offline use.
GPU / TPU / NPU
Chips made for heavy math. GPUs are common. TPUs come from Google. NPUs power on‑device AI.
Vector Database
A database for embeddings. It finds “nearest neighbors” in vector space. Great for search, RAG, and memory.
Prompting, Orchestration & Agents
Prompt
The instructions you give a model. Clear prompts get better results. Structure helps.
System Prompt
Hidden or first message that sets the role and rules. It guides tone, safety, and style.
Few‑Shot Prompting
You show examples in the prompt so the model learns the pattern for this task.
Tool Use / Function Calling
Letting a model call tools (like a calculator or database). The model decides when to call and how to use the results.
Custom GPT
A personalized version of ChatGPT that follows specific instructions, data, or style. Created through OpenAI’s GPT builder, these tailored models specialize in topics, workflows, or brand tone — like your own mini-assistant built on GPT technology.
RAG (Retrieval‑Augmented Generation)
The model looks up facts from your data, then writes an answer. This cuts hallucinations and keeps answers current.
Memory
A way to store key facts from past chats so the model can personalize later replies.
Memory Architecture
Persistent memory systems that let AI recall context, preferences, or prior sessions.
Agent
A loop where the model thinks, decides on an action, calls tools, reads results, and repeats until it reaches a goal.
Agentic AI
Represents AI systems that act on goals without constant input — autonomous agents completing multi-step tasks.
Multi-Agent Systems (Multiagentic AI)
Multiple AI agents working together (or competing) to solve problems, plan workflows, or simulate scenarios.
Planner / Executor
A pattern where one model plans the steps and another carries them out. Helps with longer tasks.
Agent Loop / ReAct Pattern
A reasoning-and-action cycle where the model thinks, takes an action, observes the result, and repeats.
Guardrails
Rules and checks that shape what the model can say or do. Used for safety, policy, and brand tone.
Evaluation & Quality
Benchmark
A test set that measures model skill (e.g., reasoning, coding, or safety). Scores help compare models.
Evals
Your own tests for your use case. Often a mix of automated checks and human review.
Hallucination
When a model states something that is not true or not in the source. RAG, citations, and stricter prompts help.
Grounding
Tying answers to trusted sources. Grounded answers can be checked and cited.
Explainable AI (XAI)
AI that’s transparent — humans can understand why it made a decision; vital for trust and regulation.
Self‑Consistency
A decoding trick: ask the model to solve a problem several times and pick the most common answer.
Verifier Model
A second model that checks the first model’s work. It can score truth, style, or safety.
Safety, Security & Privacy
Prompt Injection
A malicious input that tries to override your system prompt or make the model leak secrets. Sanitization and allow‑lists help.
Jailbreak
A method users try to bypass safety rules. Good guardrails and monitoring reduce risk.
Data Exfiltration
When a prompt tricks the model into revealing private data. Use strict tool scopes and redaction.
PII (Personally Identifiable Information)
Data that can identify a person. Handle it with care and follow laws.
Differential Privacy
A way to train or use models while adding noise so no single person’s data stands out.
Federated Learning
Training across many devices without moving raw data to a central server. Only updates are shared.
Content Credentials / C2PA
A standard to label content with how it was made, including if AI helped. Builds trust.
Red Teaming
Acting like an attacker to find model weaknesses before real users do.
AI Safety Research
Scientific work to reduce model risks and prevent harmful or uncontrolled AI behavior.
Model Card / System Card
A public document that explains a model’s data, limits, and risks. Supports responsible use.
Shadow AI
Unofficial AI use inside companies; a governance and compliance risk.
Data & Knowledge
Corpus
A large body of text, images, or audio used to train or test models.
Data Pipeline
Steps to gather, clean, label, split, and store data for training and evals.
Deduplication
Removing near‑duplicate items so the model does not “memorize” repeats.
Metadata
Data about your data (source, date, author). Useful for filtering and trust.
Ground Truth
The correct answer you use to check model outputs. Labeled by humans or gold‑standard tools.
Search, SEO & Content Ops (for AI‑Ready teams)
AI Overviews (AIO)
Google’s AI summaries on some searches. They pull facts from the web and show them above results.
AI Mode (Google)
A chat‑style experience in Google Search that lets users ask follow‑ups and compare options.
Zero-Click / Zero-Prompt AI
Systems that surface content automatically — before users even ask — often in search or recommendation engines.
LLM Visibility
How well your brand or content gets used and cited by AI systems. Clean structure, sources, and expertise help.
Entity
A real‑world thing (person, place, brand). Clear entities make it easier for models to understand and link your content.
Structured Data / Schema
Machine‑readable markup that labels things on a page (like products or FAQs). It helps both search engines and LLMs.
Content Chunking
Breaking long content into sections with headers and summaries. It boosts retrieval and RAG.
Citations
Links or references that show where facts came from. They build trust for both users and models.
Robotics
Robot Operating System (ROS)
A set of tools and libraries that help robots move, sense, and plan.
SLAM (Simultaneous Localization and Mapping)
A way for a robot to build a map and track its location at the same time.
Sensor Fusion
Combining data from cameras, lidar, GPS, and more to get a clearer picture of the world.
Actuator
A part that moves something in the real world (motors, servos). The model decides; actuators do.
Path Planning
Picking a safe, efficient route from A to B while avoiding obstacles.
Legal, Ethics & Governance
Bias
When a model favors certain groups unfairly. It can come from the data or design. Testing helps find and fix it.
Copyright
Legal rules about who owns content. Matters for training data and generated outputs.
Attribution
Giving credit to sources. It supports trust and legal safety.
Policy / Usage Guidelines
Rules for what the model can and cannot do. Good policies protect users and your brand.
AI Governance
Company-level frameworks for managing AI responsibly — covering data use, fairness, and accountability.
Transparency
Clear notes about how the system works and where the data came from. Users deserve to know.
Practical Architecture
Pipeline Orchestration
Coordinating steps like retrieval, prompt building, tool calls, and post‑processing.
Observability
Tracking prompts, outputs, errors, and cost. Helps debug and improve quality.
Rate Limiting
Controls how many requests a client can make. Protects stability and cost.
Caching
Saving results so repeated requests are faster and cheaper.
Canary Release
Rolling out changes to a small slice of users first. You watch metrics and expand if all is well.
Developer Essentials
API
A set of rules for how code talks to services. Models often live behind APIs.
JSON
A simple text format for structured data. Many tools expect JSON outputs from models.
Schema‑Constrained Output
Forcing the model to reply in a strict format (like a JSON schema). It makes apps more reliable.
Vibe Coding
Developers describe what they want in natural language, and the AI writes or edits the code automatically.
Retries / Backoff
Trying a failed request again after waiting a bit. Helps with flaky networks and rate limits.
Cost per Token
The price to process tokens in and out. Tracking it keeps your app affordable.
Emerging Methods (secondary terms)
Self‑Reflection / Self‑Critique
Having a model review its own answer and fix weak spots.
Tree‑of‑Thought / Graph‑of‑Thought
Letting a model branch into many solution paths, then pick the best one.
Program‑Aided Language (PAL)
Combining code with language models so the model plans and the code runs precise steps.
Neural Symbolic AI
Combines neural networks with rule-based logic for more structured, interpretable reasoning.
Constitutional AI
A model follows a written set of rules (“a constitution”) during training to keep answers within policy.
Watermarking
Marking AI‑made content so tools can spot it later. Helps with misuse detection.
Quick Math & Hardware Terms
Floating Point (FP16, BF16)
Number formats used on GPUs and TPUs. Lower precision speeds things up with small accuracy trade‑offs.
Throughput vs. Latency
Throughput is volume over time. Latency is time per request. You usually tune for both.
Batching
Handling many requests at once to use hardware better. It lowers cost but can add wait time.
