The AI Dictionary
A quick cheat-sheet for all the terms you will hear when talking about prompts, models, and artificial intelligence.
The Basics
- Large Language Model (LLM)
- An AI model trained on massive amounts of text to understand and generate human language.
- Artificial Intelligence (AI)
- Computer systems designed to perform tasks that normally require human intelligence.
- Prompt
- The instruction or input given to an AI model.
- Prompt Engineering
- The practice of designing prompts to get better and more reliable AI outputs.
- Context
- Additional information that helps the model better understand the task or situation.
- Token
- A small unit of text processed by the model, such as a word or part of a word.
- Hallucination
- A false or invented answer generated by the AI that sounds believable.
- Chain of Thought
- A reasoning style where the model explains intermediate thinking steps.
- Multimodal Model
- An AI model capable of processing multiple data types such as text, images, and audio.
- Agent
- An AI system that can autonomously perform actions and use tools to complete tasks.
- Tool Calling
- The ability of an AI model to use external tools, APIs, or services during a task.
Under the hood
- System Prompt
- Hidden instructions that define the AI's behavior, rules, and personality.
- Context Window
- The maximum amount of text an AI model can process and remember at once.
- Context Injection
- Adding external information into a prompt to improve the model's response.
- Token Limit
- The maximum number of tokens a model can handle in a single interaction.
- Training Data
- The large collection of text and information used to train the model.
- Inference
- The process of generating an output using a trained AI model.
- Completion
- The text response generated by the AI model.
- Next-Token Prediction
- The core mechanism where the model predicts the most likely next token in a sequence.
- Transformer
- The neural network architecture used by modern language models like GPT.
- Attention Mechanism
- A system that helps the model focus on the most relevant parts of the input.
- Weights
- The learned internal parameters that store the model's knowledge.
- Temperature
- A setting that controls how creative or predictable the model's responses are.
- Reasoning Model
- A model optimized for deeper logical thinking and step-by-step problem solving.
- Hallucination Rate
- The frequency at which a model produces inaccurate or fabricated information.
- Grounding
- Connecting AI outputs to real data or trusted sources to improve accuracy.
- Alignment
- The process of making AI systems follow human goals, values, and safety rules.
- Guardrails
- Safety restrictions designed to prevent harmful or unwanted AI behavior.
- Jailbreak
- An attempt to bypass the AI model's built-in restrictions or safety systems.
- Few-Shot Prompting
- Providing a few examples in the prompt to guide the model's behavior.
- One-Shot Prompting
- Providing a single example in the prompt to demonstrate the desired output.
- Zero-Shot Prompting
- Asking the model to perform a task without giving examples.
- Prompt Chaining
- Breaking a complex task into multiple connected prompts.
- Structured Output
- A response generated in a fixed format such as JSON, XML, or tables.
- Few-Shot Learning
- The ability of a model to adapt to tasks using only a small number of examples.
- Embeddings
- Numerical vector representations of text used to capture semantic meaning.
- Vector Database
- A database optimized for storing and searching embeddings.
- Semantic Search
- Search based on meaning and context instead of exact keywords.
- Retrieval-Augmented Generation (RAG)
- A method where the model retrieves external information before generating a response.
- Fine-Tuning
- Additional training of a model on specialized data for a specific task or domain.