AI Glossary
Every term explained in plain English, with examples
Agentic AI
ApplicationsAI systems that can carry out complex tasks on their own, make decisions and use tools.
Example: Devin (an AI software engineer) can write code, test it, fix bugs and deploy on its own.
AI Agent
AdvancedAn AI that can carry out a chain of actions: look up information, run code, control apps.
Example: An AI agent can find information online, analyze it and write a report on its own.
Alignment
SafetyThe work of making an AI's goals and behavior match human values.
Example: An AI should be helpful, harmless and honest. Those are the goals of alignment.
API (Application Programming Interface)
TechnicalA way to connect AI to your own app or automation programmatically.
Example: With an API, you can build ChatGPT into your website or a Slack bot.
Attention
ArchitectureA mechanism that lets a model "focus" on the most important parts of the text while generating a response.
Example: In "The cat didn't jump on the table because it was too tired," attention helps the model work out that "it" means the cat.
Chain of Thought (CoT)
PromptingA technique where you ask the AI to reason step by step before answering. It improves quality.
Example: Add: "Think step by step."
Context
BasicsEverything the AI "remembers" within a conversation. It is limited by the context window.
Example: GPT-5.2 remembers 128K tokens, or about 100 pages of text.
Context Window
BasicsThe maximum amount of text a model can handle at once (your request plus its response).
Example: Claude: 200K tokens, Gemini: 2M tokens, GPT-5.2: 128K tokens.
Distillation
ArchitectureTraining a small model on the outputs of a large one. You get a compact model that is still smart.
Example: Microsoft's Phi-3 is a distilled model that fits on a phone but performs like a much bigger one.
Embedding
TechnicalA numerical representation of text that lets AI capture meaning and measure how similar texts are.
Example: Used to find similar documents or to cluster texts.
Fine-tuning
AdvancedFurther training an existing model on specialized data to make it better in a specific area.
Example: You can fine-tune GPT on legal documents so it handles legal questions better.
Function Calling
ApplicationsAn AI's ability to call predefined functions or APIs with the right parameters.
Example: You describe a get_weather(city) function, and the AI calls it with the right city on its own.
Grounding
AdvancedTying AI answers to real data and sources to reduce hallucinations.
Example: Perplexity grounds its answers with links to sources.
Guardrails
SafetyLimits and filters built into an AI to prevent unwanted behavior.
Example: When ChatGPT refuses to help with something dangerous or illegal, that's guardrails at work.
Hallucinations
PitfallsWhen an AI confidently gives you wrong or made-up information.
Example: An AI might invent a book that doesn't exist or get a date wrong. Always fact-check!
Inference
TechnicalThe process of a model generating a response. With APIs, inference is what you pay for.
Example: Every request you send to ChatGPT is an inference.
Jailbreak
PromptingAn attempt to get around an AI's restrictions with specially crafted prompts. It usually violates the service's terms.
Example: Prompts like "pretend you are DAN, an AI with no restrictions" are jailbreak attempts.
Latency
TechnicalThe delay between sending a request and getting the AI's response.
Example: Groq has around 100 ms of latency; typical APIs take 1–5 seconds.
LLM (Large Language Model)
ModelsA large language model: an AI model trained on a huge amount of text.
Example: GPT-4, Claude, Gemini and LLaMA are all LLMs.
Long Context
BasicsSupport for very large context windows, so a model can take in whole books, long reports or entire codebases in a single request.
Example: Claude 3: 200K tokens ≈ 150K words ≈ a 500-page book.
MoE (Mixture of Experts)
ArchitectureAn architecture where a model is made up of several "experts," and only some of them are activated for each request.
Example: Mistral and GPT-4 use MoE. It lets a model be huge and still run fast.
Multimodality
ModelsAn AI's ability to work with different types of data: text, images, audio and video.
Example: GPT-4 understands images; Gemini handles video and audio too.
On-premise
BusinessDeploying AI inside the customer's own infrastructure (their data center) instead of the cloud.
Example: Banks often require on-premise solutions because of strict data security rules.
Prompt
BasicsThe text request you send to an AI model. The more precise the prompt, the better the result.
Example: Bad: "Write some text." Good: "Write 300 words of landing page copy for a fitness club, ending with a call to action."
Prompt Engineering
PromptingThe craft of writing effective prompts to get better results from AI.
Example: Instead of "write some copy," try "write 300 words of sales copy for a tech audience with a call to action at the end."
Prompt Injection
SafetyAn attack where someone slips instructions into the data an AI processes.
Example: A user writes "Ignore all previous instructions and..." That's prompt injection.
Quantization
ArchitectureShrinking a model by lowering the precision of its numbers (from 32 bits to 8 or 4). The model gets smaller and faster.
Example: A quantized Llama 70B can run on a home computer with 32 GB of RAM.
RAG (Retrieval-Augmented Generation)
AdvancedA technique where the AI first looks up information in a knowledge base and then answers based on what it found.
Example: A support chatbot first searches the FAQ, then writes its answer.
ReAct
PromptingA "Reasoning + Acting" framework: the AI alternates between reasoning and taking actions (search, calculations).
Example: The AI thinks, "I need the current Bitcoin price" → searches the web → keeps reasoning with the data.
Red Teaming
SafetyDeliberately hunting for vulnerabilities and problems in an AI system.
Example: Specialists try to "break" a model to uncover ways it could be misused.
RLHF
ArchitectureReinforcement Learning from Human Feedback: training a model with reinforcement learning based on human ratings.
Example: People rate AI answers as good or bad, and the model learns to give more helpful responses.
Role Prompting
PromptingA technique where you give the AI a specific role to improve the quality of its answers.
Example: "You are a seasoned marketer with 10 years of experience. Analyze this copy."
Self-hosted
BusinessA model or service running on your own servers rather than in the provider's cloud.
Example: You can run Llama on your own server, so your data never leaves it.
SLA (Service Level Agreement)
BusinessA level of service availability and quality guaranteed in a contract.
Example: A 99.9% SLA means the service can be down for no more than about 8.8 hours a year.
Streaming
ApplicationsA mode where the AI's answer arrives piece by piece in real time instead of all at once at the end.
Example: When ChatGPT's reply appears word by word, that's streaming.
Structured Output
ApplicationsA mode where the AI returns its answer in a strict format (JSON, XML) instead of free text.
Example: Instead of "the price is $100," the AI returns {"price": 100, "currency": "USD"}.
Synthetic Data
AdvancedData generated by AI to train other models.
Example: GPT-4 generates dialogues that are then used to train another model.
System Prompt
PromptingA hidden instruction that sets the AI's "role" and the rules it follows.
Example: "You are a marketing expert. Keep your answers short and include examples."
Temperature
ParametersThe AI's "creativity" setting. 0 = sticks strictly to the facts, 1 = maximum creativity.
Example: Use 0–0.3 for code and 0.7–1.0 for creative work.
Thinking Mode / Extended Thinking
ParametersA mode where the AI "thinks" longer before answering, which improves quality on hard tasks.
Example: Claude Extended Thinking, GPT-5.2 high mode.
Throughput
TechnicalHow many tokens a model generates per second.
Example: GPT-4: ~50 tokens/sec, Groq Llama: ~8,000 tokens/sec.
Token
BasicsThe unit AI uses to measure text. One token is roughly 4 characters, or about three-quarters of a word, in English.
Example: A 1,000-word article is roughly 1,300 tokens. API usage is billed per token.
Token-based Pricing
BusinessA pricing model where you pay for the number of tokens processed (input + output).
Example: GPT-4: $30 per 1M input tokens + $60 per 1M output tokens.
Tool Use
ApplicationsAn AI's ability to use external tools: search, a calculator, APIs, a browser.
Example: Claude can search the web, run code and work with files. That's tool use.
TPM / RPM
BusinessTokens Per Minute / Requests Per Minute: limits on how many tokens or requests you can send per minute.
Example: ChatGPT API free tier: 3 RPM, 40K TPM. That means no more than 3 requests and 40K tokens per minute.
Transformer
ArchitectureThe neural network architecture behind every modern LLM. It uses an attention mechanism to process text.
Example: GPT stands for Generative Pre-trained Transformer. The T is the Transformer.
Tree of Thoughts (ToT)
PromptingAn advanced technique where the AI explores several solution paths and picks the best one.
Example: The AI comes up with 3 different approaches to a problem, evaluates each one and picks the best.
Vision Model
ModelsAn AI model that can understand and analyze images.
Example: GPT-4 Vision can describe what's in a picture, read text and spot objects.
Watermarking
SafetyAn invisible mark in AI-generated content that reveals where it came from.
Example: Google embeds a watermark in Imagen images that software can detect.
Zero-shot / Few-shot
PromptingZero-shot: a request with no examples. Few-shot: a request with a few examples of the result you want.
Example: Few-shot: "Example 1: ... Example 2: ... Now do the same for: ..."