AI Applications and Concepts

These flashcards cover key concepts, definitions, and mechanisms related to AI applications and systems from the lecture notes.

AI Application

A software system that embeds one or more AI models to perform tasks, translating model capabilities into practical workflows.

Large Language Model (LLM)

A neural network trained on vast amounts of text to predict the next token, generating coherent language through pattern-matching.

Model

The core engine of an AI system that maps inputs to outputs via learned statistical relationships.

Interface

The user-facing layer for interacting with the AI model, shaping the user experience and control.

Configuration

Developer-controlled settings that shape behavior within an application, acting like 'personality settings'.

Tool

A modular component that extends a model’s capabilities by adding specialized functions or access to data.

Knowledge Base

A structured, update-able collection of documents or data that supplements a model’s static training.

Extension

An add-on that connects an AI system to external services or features, broadening interoperability.

API (Application Programming Interface)

A defined set of rules for software-to-software communication that allows applications to request services or exchange data.

Plugin

A pre-built integration module that embeds new features directly into an existing software environment.

Workflow

A sequence and logic of connected actions/tools to complete a task, guiding information movement between users and AI components.

Automation

Executing routine steps without continuous user input, embedding repeatable processes into workflows.

User-Level Customization

Adjustments made in the interface to personalize outputs for specific audiences and goals.

Application-Level Customization

Developer design and configuration decisions that define an application’s capabilities and constraints.

Retrieval-Augmented Generation (RAG)

A method combining retrieval and generation to ground responses in external knowledge.

Multimodal Model

A model capable of understanding and/or generating multiple data types (text, images, audio, video).

Ecosystem

The interconnected environment of models, tools, APIs, data, and users that make AI systems functional.

Corpus / Dataset

The collection of text, images, audio, etc. used for training or evaluation.

Data Curation

The process of selecting, cleaning, deduplicating, filtering, and balancing data.

Generative Model

An AI system that creates new content resembling patterns in its training data.

Transformer

Neural network architecture using attention to model relationships across tokens, essential for modern LLMs.

Attention

A mechanism that allows the model to focus on the most relevant parts of the input when generating predictions.

Embedding

A way of converting words, sentences, or images into a vector that captures meaning and context.

Parameters / Weights

Learnable values updated during training that encode what the model has learned.

Context Window

The maximum number of tokens the model can consider at once, limiting its memory of prior text.

Diffusion Model

A process that generates images/media by starting from noise and iteratively denoising.

Multimodal Generation

Models that accept and/or produce multiple modalities in combination.

Pre-training

The initial large-scale training on diverse data to learn broad language or visual patterns.

Fine-Tuning

Further training a pre-trained model on a smaller, targeted dataset.

HRLF (Human Reinforcement Learning with Feedback)

A method where humans rank outputs, improving model responses based on usefulness and safety.

Emergent Behavior

Capabilities that appear as the scale of data or parameters increases, not evident in smaller models.

Benchmarking

Testing models on standardized tasks/datasets for comparison.

Bias Evaluation

Testing for unfair or stereotyped outputs that reflect training data imbalances.

Hallucination

Plausible-sounding but false content produced by a model.

Model Card

A transparency document that describes a model’s intended uses, performance, and risks.

Superalignment Problem

Challenges in evaluating and governing capable models when benchmarks don't capture their full abilities.

Meta prompting

A technique in prompt engineering designed to help LLMs create more precise prompts.