📖 AI Glossary

A type of AI trained on massive amounts of text data that can understand and generate human-like text. Examples: GPT-4, Claude, Gemini, Llama. LLMs power most modern AI tools.

The art and science of crafting effective inputs (prompts) for AI models to get the desired outputs. Includes techniques like chain-of-thought, few-shot examples, and role-playing.

The basic unit of text that AI models process. Roughly 4 characters or 0.75 words in English. LLM pricing is based on tokens — both input (your prompt) and output (the AI response) cost tokens.

The maximum amount of text an AI can process at once. Measured in tokens. Claude 3.5 has 200K tokens (~150,000 words), allowing it to read entire books in one conversation.

When an AI generates information that sounds confident but is factually incorrect or made up. A key limitation of LLMs. RAG and grounding techniques help reduce hallucinations.

AI that can process and generate multiple types of data — text, images, audio, video, and code. GPT-4o, Claude 3.5, and Gemini are multimodal. The future of AI is multimodal.

The process of running a trained AI model to generate outputs. Contrasts with training (teaching the model). When you chat with ChatGPT, you're using inference. Inference costs are what you pay per API call.

A way for software to communicate. AI APIs let developers add AI capabilities to their apps. You pay per use (tokens or calls). OpenAI API, Anthropic API, and Google Gemini API are the most popular.

AI models whose weights are publicly available. Can be downloaded and run locally. Examples: Llama 4, Mistral, Gemma, Phi. No per-call costs but requires compute infrastructure.

A technique that gives AI access to external information at query time. Instead of relying only on training data, the AI retrieves relevant documents and uses them to generate more accurate, up-to-date answers.

An AI system that can take actions autonomously to complete tasks. Unlike simple chatbots, agents can use tools, browse the web, write and run code, and take multi-step actions to achieve goals.

A numerical representation of text, images, or other data as vectors in high-dimensional space. Similar concepts end up close together in this space, enabling semantic search and RAG systems.

The neural network architecture behind all modern LLMs. Introduced by Google in 2017, it uses "attention mechanisms" to understand relationships between words across long sequences. The "T" in GPT.

Connecting AI responses to verified, real-world information sources to reduce hallucinations. Techniques include RAG, web search access, and connecting to databases. Google's AI uses grounding to cite sources.

AI systems that autonomously plan and execute multi-step tasks with minimal human supervision. They use tools, make decisions, and iterate toward goals. Claude Code, Devin, and AutoGPT are examples.

The architecture behind image generation AI like Stable Diffusion, DALL-E, and Midjourney. Learns to remove noise from images, enabling generation of realistic images from text descriptions.

A database optimized for storing and searching vector embeddings. Essential for RAG systems and semantic search. Popular options: Pinecone, Qdrant, pgvector, Chroma.

The process of further training a pre-trained model on a specific dataset to specialize it for a particular task or domain. Like teaching a general expert to become a specialist.

A training technique where humans rate AI outputs and those ratings are used to improve the model. Key to making models helpful, harmless, and honest. Used by OpenAI, Anthropic, and Google.

Asking an AI to perform a task without giving any examples. The AI relies on its training to understand what's needed. Contrast with few-shot learning, where you provide 2-5 examples.

Providing 2-5 examples in your prompt to show the AI what you want. Dramatically improves output quality for complex or unusual tasks. Each example is called a "shot".

A prompting technique that asks the AI to "think step by step" before answering. Dramatically improves reasoning on complex problems. Adding "let's think step by step" can increase accuracy by 40%.

Instructions given to an AI before the conversation starts to set its behavior, personality, and constraints. Invisible to users but shapes every response. Used by companies to customize AI for their product.

A parameter controlling how random or creative an AI's outputs are. Temperature 0 = deterministic and focused. Temperature 1+ = creative and varied. Use low temp for factual tasks, high for creative.

Standardized tests used to measure and compare AI model capabilities. Common benchmarks: MMLU (knowledge), HumanEval (coding), MATH, BigBench. Often gamed by training on test data.