Generative AI Explained

A method of using AI agents to break down complex tasks into smaller steps and coordinate execution across tools or systems.

AI with human-level intelligence across all intellectual tasks, capable of understanding, learning, and applying intelligence broadly.

Autonomous AI systems that perceive their environment, process information, make decisions, and take actions to achieve specific goals, often interacting with humans or other systems.

The broad field of computer science dedicated to creating machines that can perform tasks typically requiring human intelligence.

AI far surpassing human intelligence in all fields, including creativity, wisdom, and social skills.

An autonomous AI system that can generate its own prompts and execute tasks toward a goal without constant user input.

An AI task management agent that prioritizes and creates new tasks based on a goal, simulating a mini project manager.

A systematic error in an AI's output that reflects the flawed assumptions or prejudices present in its training data.

A prompting technique that asks a model to generate its intermediate reasoning steps before giving a final answer.

When an AI model loses track of the full conversation or task context due to token limits or poor memory handling.

The maximum amount of information (tokens) an AI model can consider at once.

A powerful Generative AI model that creates new data (especially images and more recently with text) by gradually reversing a noise-adding process, "denoising" from pure static to a clear image.

Numerical representations capturing the meaning or characteristics of data (words, sentences, images); similar meanings have similar numerical representations.

Further training an already developed Generative AI model on a smaller, specific dataset to adapt it for a particular task or domain.

A tool for exploring the internal neurons of a GAN, showing how specific units correspond to interpretable concepts like "tree," "door," or "tower."

An interactive in-browser tool that helps users visualize and train simple GANs (Generator & Discriminator) step-by-step on toy 2D distributions.

A Generative AI model with two competing neural networks: a "Generator" creating fakes, and a "Discriminator" identifying them, improving through competition.

AI that creates content by predicting the next token (text, images, music, code), rather than just analyzing existing data.

An optimization algorithm used to adjust a model’s parameters by minimizing the error in predictions.

Rules or controls that limit what an AI system can do or say to ensure safe, ethical, and aligned behavior.

When an AI model confidently generates incorrect, fabricated, or nonsensical information and presents it as fact.

The process where a trained AI model takes a new input (like a Prompt) and generates an output or makes a prediction. It's the "runtime" phase of an AI model.

The process of fine-tuning a pre-trained model on paired examples of instructions and ideal responses to make it better at following user commands.

A powerful AI trained on vast text data to understand, generate, and respond to human language naturally.

A high-dimensional space where complex data (like images, text, or audio) is encoded as numerical vectors capturing its essential features.

A parameter-efficient fine-tuning method that injects small, trainable low-rank matrices into a frozen pre-trained model.

A subset of Artificial Intelligence (AI) that focuses on enabling systems to learn from data, identify patterns, and make decisions with minimal human intervention.

AI models enhanced with the ability to store and retrieve information from memory across sessions or tasks.

A model architecture that routes different inputs to different specialized sub-models (“experts”) to improve efficiency and scalability.

A standardized set of rules or an interface that allows an AI model (especially LLMs and AI agents) to effectively connect with and use external tools, functions, or APIs to accomplish tasks.

Internal values or "knobs" within an AI model adjusted during Training Process; more parameters allow learning more complex patterns.

AI systems capable of processing and generating content across multiple types of data simultaneously, such as text, images, audio, and video.

A computational model inspired by the structure and function of the human brain, consisting of interconnected "nodes" or "neurons" organized in layers, designed to recognize patterns and learn from data.

Overfitting happens when a model learns too much from the training data and performs poorly on new data; underfitting is when it doesn’t learn enough.

The input or instruction given to an AI model to generate a response.

A technique where the output of one prompt is used as the input to another, allowing for more complex reasoning or multi-step tasks.

The skill of crafting effective Prompts to achieve desired AI results by understanding how the AI "thinks".

A type of attack where a user inserts unexpected instructions into a prompt to override or alter the AI’s intended behavior.

AI models designed to "think" step-by-step, performing logical inference, planning, and problem-solving, rather than simply predicting the next token based on probability.

The process of rigorously stress-testing an AI model by acting as an adversary to find its flaws, vulnerabilities, and potential for harmful behavior before release.

Humans rate Generative AI outputs, and this feedback further trains the AI to produce more helpful, accurate, and aligned responses.

A technique that enhances a Large Language Model (LLM) by allowing it to retrieve relevant information from an external, trusted knowledge base (like a database or documents) before generating a response.

A network component that lets a model weigh the importance of each input element relative to others when producing an output.

A learning approach that combines a small amount of labeled data with a large amount of unlabeled data to improve training.

A machine learning approach where the model is trained on labeled data, meaning the correct answers are provided during training.

Artificially generated data that mimics real-world data, used for training or testing AI models.

A special instruction given to an AI model before the user's prompt to set behavior, tone, or rules for how it should respond.

Hyperparameters controlling the randomness of a model's next-token selection—temperature adjusts distribution "sharpness," top-p limits sampling to the most probable tokens.

The fundamental "building block" of text processed and generated by a Large Language Model (LLM), typically part of a word, a whole word, or punctuation.

A language model that learns to decide when and how to call external tools—like calculators or search engines—during generation.

The vast information (text, images, code) an AI model learns from during initial development, forming its knowledge and abilities.

A neural network design enabling AI to process and generate sequences (like text) by efficiently "paying attention" to relevant parts of the input, regardless of distance.

An emerging approach that explores multiple parallel chains of thought (branches), evaluates them, and backtracks to the most promising paths.

A machine learning approach where the model learns patterns in data without any labeled answers.

A specialized database designed to store and search high-dimensional vectors (Embeddings), often used for similarity search in AI applications.

Techniques where a model performs a task with zero, one, or only a handful of examples provided at inference time.