AI Glossary: 100+ Terms Explained

AI Glossary: Every Term You Need to Know

This comprehensive glossary covers 100+ AI terms explained in plain English. Bookmark this page as your go-to reference for understanding AI technology.

A

AGI (Artificial General Intelligence)

What it is: Hypothetical AI that matches or exceeds human intelligence across all cognitive tasks. Unlike current AI (narrow AI), AGI would understand, learn, and apply knowledge across any domain.

Current status: Does not exist yet. Current AI systems like ChatGPT and Claude are narrow AI—excellent at specific tasks but lacking true general intelligence.

Why it matters: AGI is the long-term goal of AI research. Its development would fundamentally transform society, which is why AI safety research is crucial.

AI Agent

What it is: An AI system that can perceive its environment, make decisions, and take actions autonomously to achieve specific goals. Unlike simple chatbots, agents can use tools, execute multi-step plans, and operate with minimal human intervention.

Examples: Auto-GPT, Claude with computer use, AI sales agents that qualify leads and schedule meetings automatically.

Why it matters: AI agents represent the next evolution beyond conversational AI, enabling true automation of complex workflows.

AI Alignment

What it is: The field of research ensuring AI systems behave in accordance with human values and intentions. Alignment aims to prevent AI from pursuing goals that conflict with human wellbeing.

Why it matters: As AI becomes more powerful, ensuring it remains beneficial and controllable becomes increasingly critical. Anthropic's Constitutional AI is one approach to alignment.

API (Application Programming Interface)

What it is: A set of protocols that allows different software applications to communicate. AI APIs let developers integrate AI capabilities (like ChatGPT or Claude) into their own applications.

Example: Using OpenAI's API to add AI chat to your customer service platform.

Why it matters: APIs democratize AI access, allowing any developer to build AI-powered applications without training their own models.

Anthropic

What it is: AI safety company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei. Creator of Claude AI.

Known for: Constitutional AI approach, focus on AI safety, Claude assistant.

B

BERT (Bidirectional Encoder Representations from Transformers)

What it is: Google's 2018 language model that revolutionized NLP by reading text bidirectionally (both left-to-right and right-to-left), enabling better understanding of context.

Impact: BERT transformed Google Search and became the foundation for many modern NLP applications.

Bias (in AI)

What it is: Systematic errors in AI outputs that reflect prejudices in training data or model design. AI can perpetuate or amplify human biases around race, gender, age, and other characteristics.

Example: A hiring AI trained on historical data might discriminate against candidates from underrepresented groups.

Why it matters: Responsible AI deployment requires actively testing for and mitigating bias.

C

Chain-of-Thought (CoT) Prompting

What it is: A prompting technique where you ask the AI to explain its reasoning step-by-step before providing a final answer. This often improves accuracy on complex reasoning tasks.

How to use it: Add "Let's think step by step" or "Explain your reasoning" to your prompts.

Why it matters: CoT can dramatically improve AI performance on math, logic, and analysis tasks.

Chatbot

What it is: A software application designed to simulate conversation with human users. Modern AI chatbots use large language models to generate human-like responses.

Evolution: Rule-based chatbots → ML chatbots → LLM-powered assistants (ChatGPT, Claude)

Claude

What it is: Anthropic's AI assistant, known for strong reasoning, coding abilities, and safety-focused design. Available in multiple versions (Haiku, Sonnet, Opus) with different capability/cost tradeoffs.

Key features: 200K token context window, Constitutional AI training, strong at analysis and coding.

Constitutional AI

What it is: Anthropic's approach to AI alignment where the model is trained using a set of principles (a "constitution") rather than purely human feedback. The AI learns to evaluate and revise its own outputs.

Why it matters: Produces more predictable, principled AI behavior compared to RLHF alone.

Context Window

What it is: The maximum amount of text an AI model can process in a single conversation, measured in tokens. Larger context windows allow processing longer documents and maintaining longer conversations.

Current sizes:

• Claude 3.5: 200,000 tokens (~150,000 words)
• GPT-4 Turbo: 128,000 tokens (~96,000 words)
• Gemini 1.5: 1,000,000+ tokens

Copilot

What it is: Microsoft's AI assistant brand, used across products. GitHub Copilot is an AI pair programmer; Microsoft 365 Copilot assists with Office tasks.

Note: "Copilot" has become a generic term for AI assistants that work alongside humans.

D

Deep Learning

What it is: A subset of machine learning using neural networks with many layers (hence "deep"). Deep learning powers most modern AI breakthroughs, including image recognition, language models, and game-playing AI.

Relationship: AI > Machine Learning > Deep Learning > Neural Networks

Diffusion Models

What it is: AI models that generate images by starting with random noise and gradually refining it into coherent images. Powers tools like DALL-E, Midjourney, and Stable Diffusion.

How it works: Trained by adding noise to images, then learning to reverse the process.

E

Embedding

What it is: A numerical representation of text (or images, audio) as a vector of numbers. Embeddings capture semantic meaning, allowing AI to understand that "dog" and "puppy" are related.

Use cases: Semantic search, recommendations, clustering similar content, RAG systems.

Emergent Behavior

What it is: Capabilities that appear in large AI models without being explicitly trained for. As models scale, they sometimes develop unexpected abilities.

Examples: GPT-3 demonstrated few-shot learning; larger models show improved reasoning.

F

Few-Shot Learning

What it is: An AI's ability to perform a task after seeing only a few examples, rather than requiring thousands of training samples.

How to use it: Include 2-5 examples in your prompt before asking the AI to perform the task.

Example prompt:

Classify sentiment:
"Great product!" → Positive
"Terrible experience" → Negative
"It arrived on time" → Neutral

Now classify: "Best purchase ever!"

Fine-Tuning

What it is: The process of further training a pre-trained AI model on specific data to customize its behavior for particular tasks or domains.

When to use: When you need consistent, specialized outputs that prompting alone can't achieve.

Cost: Requires significant data and compute resources; often thousands of examples needed.

Foundation Model

What it is: A large AI model trained on broad data that can be adapted to many different tasks. GPT-4, Claude, and Gemini are foundation models.

Why it matters: Foundation models enable "build once, use everywhere" AI development.

G

Generative AI

What it is: AI systems that create new content—text, images, audio, video, or code—rather than just analyzing existing content.

Examples: ChatGPT (text), DALL-E (images), Suno (music), Runway (video)

GPT (Generative Pre-trained Transformer)

What it is: OpenAI's family of language models. GPT models are pre-trained on vast text data, then fine-tuned for specific applications.

Versions: GPT-1 (2018) → GPT-2 (2019) → GPT-3 (2020) → GPT-4 (2023) → GPT-4 Turbo (2024)

Guardrails

What it is: Safety mechanisms built into AI systems to prevent harmful, inappropriate, or off-topic outputs. Guardrails filter inputs and outputs to keep AI behavior within acceptable bounds.

Examples: Content filters, topic restrictions, output validators.

H

Hallucination

What it is: When an AI generates false information presented as fact. The AI confidently produces plausible-sounding but incorrect or fabricated content.

Examples: Citing non-existent research papers, inventing historical events, making up statistics.

Mitigation: Use retrieval-augmented generation (RAG), fact-checking, and asking AI to cite sources.

Hugging Face

What it is: A platform and community for sharing machine learning models, datasets, and applications. Often called the "GitHub of machine learning."

Why it matters: Democratizes access to AI models; hosts thousands of open-source models.

I

Inference

What it is: The process of using a trained AI model to generate outputs from new inputs. When you chat with ChatGPT, the model is performing inference.

Cost consideration: Inference costs (per-token pricing) are ongoing, unlike one-time training costs.

In-Context Learning

What it is: An AI model's ability to learn from examples provided within the prompt itself, without any fine-tuning or parameter updates.

Why it matters: Enables customization of AI behavior without expensive retraining.

J

Jailbreaking

What it is: Techniques used to bypass an AI's safety guardrails and make it produce outputs it would normally refuse.

Note: Attempting to jailbreak AI systems often violates terms of service and can have legal implications.

K

Knowledge Cutoff

What it is: The date after which an AI model has no training data. Events after the cutoff are unknown to the model unless provided via web search or documents.

Current cutoffs (approximate):

• GPT-4: April 2024
• Claude 3.5: April 2024
• Gemini: Varies with web access

L

LangChain

What it is: An open-source framework for building applications powered by language models. Provides tools for chaining prompts, connecting to data sources, and building AI agents.

Use cases: RAG systems, AI agents, complex AI workflows.

Large Language Model (LLM)

What it is: AI models trained on massive text datasets to understand and generate human language. "Large" refers to billions of parameters.

Examples: GPT-4 (~1.7T parameters), Claude, Gemini, Llama

Why it matters: LLMs are the foundation of modern AI assistants and generative AI.

Llama

What it is: Meta's family of open-source large language models. Llama models can be downloaded, modified, and deployed freely (with some restrictions).

Significance: Democratized access to capable LLMs; sparked open-source AI development.

LoRA (Low-Rank Adaptation)

What it is: An efficient fine-tuning technique that trains only a small number of additional parameters rather than the entire model. Makes fine-tuning faster and cheaper.

Why it matters: Enables fine-tuning of large models on consumer hardware.

M

Machine Learning (ML)

What it is: A subset of AI where systems learn from data rather than following explicit programming. ML algorithms improve through experience.

Types: Supervised learning, unsupervised learning, reinforcement learning.

Midjourney

What it is: An AI image generation tool known for artistic, high-quality outputs. Accessed through Discord.

Known for: Aesthetic image quality, particularly for artistic and photorealistic styles.

Model

What it is: In AI, a mathematical representation learned from data that can make predictions or generate outputs. "The model" refers to the trained AI system.

Components: Architecture (structure) + Parameters (learned values)

Multimodal AI

What it is: AI systems that can understand and generate multiple types of content—text, images, audio, video—not just one modality.

Examples: GPT-4V (text + images), Gemini (text + images + audio), Claude (text + images)

N

Natural Language Processing (NLP)

What it is: The field of AI focused on enabling computers to understand, interpret, and generate human language.

Applications: Translation, sentiment analysis, chatbots, summarization, search.

Neural Network

What it is: A computing system inspired by biological brains, consisting of interconnected nodes (neurons) organized in layers. The foundation of deep learning.

How it works: Input layer → Hidden layers (processing) → Output layer

O

OpenAI

What it is: AI research company founded in 2015, creator of GPT models, ChatGPT, DALL-E, and Whisper. Originally non-profit, now has a for-profit arm.

Products: ChatGPT, GPT-4, DALL-E, Whisper, Sora

Overfitting

What it is: When an AI model performs well on training data but poorly on new data. The model has "memorized" rather than "learned."

Analogy: A student who memorizes test answers but can't apply concepts to new problems.

P

Parameter

What it is: The numerical values in a neural network that are learned during training. More parameters generally mean more capability but also more compute cost.

Scale:

• GPT-3: 175 billion parameters
• GPT-4: ~1.7 trillion parameters (estimated)
• Llama 2: 7B, 13B, 70B versions

Perplexity

What it is: Both a metric (measuring how well a language model predicts text) and an AI search engine company that combines search with LLM-powered answers.

Perplexity AI: Search engine providing cited, conversational answers.

Prompt

What it is: The input text you provide to an AI model. The prompt instructs the AI on what task to perform and how.

Components: System prompt (behavior instructions) + User prompt (specific request)

Prompt Engineering

What it is: The practice of crafting effective prompts to get desired outputs from AI models. A crucial skill for maximizing AI utility.

Techniques: Chain-of-thought, few-shot examples, role-playing, structured outputs.

R

RAG (Retrieval-Augmented Generation)

What it is: A technique combining AI generation with information retrieval. The system retrieves relevant documents, then uses them as context for generating responses.

Benefits: Reduces hallucinations, enables up-to-date information, grounds responses in source material.

Use cases: Enterprise search, customer support, knowledge bases.

Reinforcement Learning (RL)

What it is: A type of machine learning where an agent learns by interacting with an environment and receiving rewards or penalties.

AI examples: Game-playing AI (AlphaGo), robotics, RLHF for language models.

RLHF (Reinforcement Learning from Human Feedback)

What it is: A training technique where human preferences guide AI learning. Humans rate AI outputs, and the model learns to produce preferred responses.

Significance: Key technique behind ChatGPT's helpful, harmless behavior.

S

Semantic Search

What it is: Search that understands meaning and intent rather than just matching keywords. Uses embeddings to find conceptually related content.

Example: Searching "how to fix slow computer" also returns results about "speed up PC performance."

Stable Diffusion

What it is: An open-source image generation model developed by Stability AI. Can be run locally on consumer hardware.

Significance: Democratized image generation by being fully open-source.

System Prompt

What it is: Hidden instructions that define an AI's behavior, personality, and constraints. Set by developers, not visible to end users.

Example: "You are a helpful customer service agent for Acme Corp. Be friendly and concise. Never discuss competitor products."

T

Temperature

What it is: A parameter controlling AI output randomness. Higher temperature = more creative/random; lower temperature = more focused/deterministic.

Typical range: 0.0 (deterministic) to 2.0 (very random)

Use cases: Low for factual tasks, higher for creative writing.

Token

What it is: The basic unit of text that AI models process. Roughly 4 characters or 0.75 words in English.

Why it matters: AI pricing and context limits are measured in tokens.

Example: "Hello, world!" ≈ 4 tokens

Training

What it is: The process of teaching an AI model by exposing it to data and adjusting its parameters to minimize errors.

Phases: Pre-training (general knowledge) → Fine-tuning (specific tasks) → RLHF (behavior alignment)

Transformer

What it is: The neural network architecture behind modern LLMs, introduced in the 2017 paper "Attention Is All You Need." Uses self-attention to process sequences.

Impact: Revolutionized NLP; foundation of GPT, BERT, Claude, and virtually all modern language models.

V

Vector Database

What it is: A database optimized for storing and searching embeddings (vectors). Essential for RAG systems and semantic search.

Examples: Pinecone, Weaviate, Chroma, Milvus

Vision Language Model (VLM)

What it is: AI models that can understand both images and text, enabling tasks like image description, visual question answering, and document analysis.

Examples: GPT-4V, Claude (with vision), Gemini

W

Weights

What it is: The numerical values in neural network connections that are adjusted during training. Synonymous with parameters in most contexts.

"Loading weights": The process of initializing a model with pre-trained parameter values.

Z

Zero-Shot Learning

What it is: An AI's ability to perform a task without any task-specific examples, relying only on general training and instruction understanding.

Example: Asking an AI to translate to a language it wasn't specifically trained to translate to, based on its general language understanding.

Quick Reference Card

| Term | One-Line Definition | |------|---------------------| | AGI | Hypothetical human-level AI across all domains | | API | Interface for integrating AI into applications | | Context Window | Max text an AI can process at once | | Embedding | Numerical representation of text meaning | | Fine-Tuning | Customizing a model with specific data | | Hallucination | AI generating false information | | LLM | Large language model like GPT-4 or Claude | | Prompt | Input instructions given to AI | | RAG | Retrieval + generation for grounded responses | | RLHF | Training AI using human preferences | | Temperature | Controls output randomness | | Token | Basic unit of text (~4 characters) | | Transformer | Architecture powering modern AI |

Need help implementing AI in your business? Our team can help you navigate the AI landscape and choose the right solutions. Contact us for a consultation.