Generative AI Glossary

This glossary covers fundamental generative AI concepts to help you communicate technical ideas clearly. Don't memorize definitions. Focus on understanding how concepts connect so you can reason about AI capabilities, limitations, and design decisions effectively.

Core concepts

AGI (Artificial General Intelligence): Hypothetical AI that can understand and perform any intellectual task a human can, across all domains. Distinct from current narrow AI systems built for specific tasks. Current AI systems, including advanced LLMs, are not AGI.

AI Agent: Autonomous system that perceives its environment, makes decisions, and takes actions to achieve goals without direct human control (e.g. Siri). Can use tools, execute multi-step workflows, and adapt based on feedback.

Agentic System: AI architecture where models can independently plan, use tools, and execute complex tasks. Goes beyond simple question-answering to actively solving problems through iterative reasoning and action.

Artificial Intelligence (AI): Broad field of creating machines capable of performing tasks that typically require human intelligence, including perception, reasoning, learning, and language understanding.

Foundation Model: Large-scale AI model trained on broad data that serves as a base for multiple downstream applications. Can be adapted for various tasks through fine-tuning. Examples: GPT-4, Claude, BERT.

Generative AI: AI systems that create new content (text, images, audio, video) by learning patterns from training data. Distinct from discriminative AI that classifies or predicts from existing data.

LLM (Large Language Model): Neural network trained on vast text data to understand and generate human-like language. Size typically measured in billions of parameters. Examples: GPT-4, Claude, Llama, PaLM.

Modality: Type of data an AI model can process or generate. Unimodal handles one type (text-only), multimodal handles multiple types (text, images, audio, video). GPT-4 with vision is multimodal.

Neural Network: Computing system inspired by biological neural networks, consisting of interconnected nodes (neurons) organized in layers. Learns by adjusting connection weights based on training data.

Transformer: Neural network architecture using attention mechanisms to process sequential data. Foundation of modern LLMs. Enables parallel processing and captures long-range dependencies better than earlier architectures.

Training & development

Annotation: Process of labeling data (text, images, audio) used to train AI models. Labeled datasets teach models to recognize patterns, categories, and relationships. Quality and quantity of annotations directly impact model performance.

Fine-tuning: Adapting a pre-trained model to specific tasks by training on smaller, specialized datasets. More efficient than training from scratch. Can modify all parameters (full fine-tuning) or only some (parameter-efficient fine-tuning).

Instruction Tuning: Fine-tuning technique where models learn to follow instructions and user intent. Trained on datasets of instruction-response pairs. Makes models more helpful and aligned with user expectations.

Pre-training: Initial training phase where models learn from massive unlabeled datasets. Models learn general patterns, language structure, and world knowledge before being fine-tuned for specific tasks.

RLHF (Reinforcement Learning from Human Feedback): Training technique where human evaluators rank model outputs, and the model learns to generate responses humans prefer. Used to align models with human values and preferences.

Training Data: Dataset used to teach AI models. Quality, diversity, and size significantly impact model capabilities and biases. LLMs train on trillions of tokens from books, websites, and other text sources.

Transfer Learning: Technique where knowledge gained from training on one task is applied to different but related tasks. Foundation models leverage transfer learning (trained once, adapted many times).

Model interaction

Chain of Thought (CoT): Prompting technique encouraging models to show reasoning steps before final answers. Improves performance on complex problems requiring multi-step reasoning. Example: "Let's think step by step..."

Context Window: Maximum amount of text (measured in tokens) a model can process at once. Includes both input prompt and generated output. Larger context windows enable longer conversations and document processing.

Few-shot Learning: Providing models with a few examples in the prompt to demonstrate desired behavior. More guidance than zero-shot, less than fine-tuning. Typically 2-10 examples.

One-shot Learning: Providing a single example in the prompt to guide model behavior. Middle ground between zero-shot and few-shot approaches.

Prompt: Input text given to an AI model to generate a response. Can include instructions, context, examples, and the actual query. Prompt design significantly affects output quality.

Prompt Engineering: Art and science of crafting effective prompts to elicit desired outputs from AI models. Involves techniques like few-shot learning, chain of thought, and role-playing to improve results.

RAG (Retrieval-Augmented Generation): Technique combining information retrieval with text generation. Model searches external knowledge base for relevant information before generating response. Reduces hallucinations and enables access to current information.

Zero-shot Learning: Model's ability to perform tasks without explicit training or examples for that specific task. Relies on knowledge gained during pre-training and instruction tuning.

Model behavior & limitations

Alignment: Process of ensuring AI systems behave according to human values and intentions. Includes techniques like RLHF, red teaming, and constitutional AI to make models helpful, harmless, and honest.

Bias: Systematic skew in model outputs reflecting prejudices in training data or design choices. Can manifest as gender, racial, cultural, or other stereotypes. Addressing bias is critical for fair AI systems.

Hallucination: When models generate plausible-sounding but factually incorrect or nonsensical information. Occurs because models predict likely text rather than retrieve facts. Major challenge for reliability in factual domains.

Temperature: Parameter controlling randomness in model outputs. Low temperature (0-0.3) makes outputs more focused and deterministic. High temperature (0.7-1.0) increases creativity and variation. Used to balance creativity with coherence.

Token: Basic unit of text models process. Can be words, subwords, or characters. English text averages ~4 characters per token. Models have token limits for context windows and output generation.

Tokenization: Process of breaking text into tokens the model can process. Different models use different tokenization strategies. Affects how models understand text structure and multi-language content.

Advanced concepts

Attention Mechanism: Component allowing models to focus on relevant parts of input when generating output. Transformers use self-attention to capture relationships between all words in a sequence, regardless of distance.

Embedding: Numerical representation (vector) of text, images, or other data that captures semantic meaning. Similar concepts have similar embeddings. Used for semantic search, clustering, and similarity comparison.

Inference: Process of using a trained model to generate outputs. Contrast with training. Inference is typically much faster and cheaper than training but still requires significant computation for large models.

Parameter: Numerical value learned during training that determines model behavior. More parameters generally mean more capability but also more computational cost. Modern LLMs have billions to trillions of parameters.

Perplexity: Metric measuring how well a model predicts text. Lower perplexity indicates better prediction. Used to evaluate language model performance. A confused model has high perplexity.

Applications & use cases

Chatbot: Conversational AI system designed to interact with users through natural language. Can range from simple rule-based systems to sophisticated LLM-powered assistants capable of complex reasoning.

Code Generation: AI capability to write, complete, or debug code based on natural language descriptions or partial code. Examples: GitHub Copilot, GPT-4 code interpreter.

Image Generation: AI systems creating images from text descriptions (text-to-image) or other images. Examples: DALL-E, Midjourney, Stable Diffusion. Use diffusion models or GANs.

Multimodal AI: Models processing and generating multiple data types (text, images, audio, video). Enables richer interactions like describing images, generating images from text, or answering questions about videos.

Sentiment Analysis: NLP task determining emotional tone in text (positive, negative, neutral). Used for customer feedback analysis, social media monitoring, and content moderation.

Summarization: Generating concise versions of longer texts. Extractive summarization selects key sentences. Abstractive summarization generates new text capturing main ideas.

Model architectures

Autoregressive Model: Generates output sequentially, predicting one token at a time based on previous tokens. Most LLMs are autoregressive. Enables flexible generation but slower than parallel approaches.

Diffusion Model: Generates images by learning to reverse a noise-adding process. Starts with random noise and iteratively refines into coherent images. Used in Stable Diffusion, DALL-E 2.

GAN (Generative Adversarial Network): Two neural networks: generator creates fake data, discriminator tries to detect fakes. They compete, improving both. Early approach for image generation, now less common than diffusion models.

Evaluation & safety

Benchmark: Standardized test measuring AI model performance on specific tasks. Examples: MMLU (general knowledge), HumanEval (coding), HELM (holistic evaluation). Enables comparison across models.

BERTScore: Evaluation metric comparing generated text to reference text using contextual embeddings. Measures semantic similarity rather than exact word matches. More robust than traditional metrics for paraphrased or varied outputs.

BLEU (Bilingual Evaluation Understudy): Metric measuring overlap between generated text and reference translations. Originally designed for machine translation. Scores 0-100 based on n-gram precision, with higher scores indicating better match.

Embedding Cosine Similarity: Metric measuring similarity between two text embeddings by calculating the cosine of the angle between their vectors. Ranges from -1 (opposite) to 1 (identical). Used to evaluate semantic similarity of generated text.

Evals: Systematic tests measuring specific AI model capabilities or behaviors. Can be automated (using metrics) or human-evaluated. Help track model improvements and catch regressions across versions.

Gates: Quality thresholds that must be met before deploying model changes. Examples: minimum accuracy score, maximum toxicity rate, pass rate on critical test cases. Prevent shipping models that regress on important metrics.

Live-traffic Scoring: Evaluating model performance on real user interactions in production. Uses metrics like user satisfaction, task completion, and engagement. Provides ground truth for how models perform in actual use.

METEOR (Metric for Evaluation of Translation with Explicit ORdering): Text evaluation metric considering synonyms, stemming, and word order. Originally for machine translation. More nuanced than BLEU by accounting for semantic similarity beyond exact matches.

Red Teaming: Process of deliberately testing AI systems for vulnerabilities, harmful outputs, or unintended behaviors. Helps identify and fix safety issues before deployment.

Regression Suite: Collection of tests ensuring new model versions don't perform worse than previous versions on established benchmarks. Catches unintended degradation when making improvements. Critical for maintaining quality standards.

Responsible AI: Framework for developing and deploying AI ethically. Includes considerations of fairness, transparency, privacy, accountability, and societal impact.

ROUGE (Recall-Oriented Understudy for Gisting Evaluation): Metric measuring overlap between generated summaries and reference summaries. Focuses on recall (how much reference content appears in generation). Commonly used for evaluating summarization tasks.

Synthetic Sweeps: Large-scale automated testing using generated or synthetic data to evaluate model behavior across many scenarios. Enables testing edge cases and rare situations without waiting for real-world examples. Complements live-traffic evaluation.

Toxicity: Harmful, offensive, or inappropriate content generated by AI systems. Can result from biased training data or adversarial inputs. Requires filtering and safety mechanisms to mitigate.

Industry terms

API (Application Programming Interface): Interface allowing applications to access AI model capabilities. Most commercial LLMs accessed via APIs. Usage typically priced per token.

Open Source: Models with publicly available code and often weights, allowing community use and modification. Examples: Llama, Mistral. Contrast with closed-source models like GPT-4.

Proprietary Model: Commercial AI model where architecture, weights, or training data are not publicly available. Examples: GPT-4, Claude. Often accessed via paid APIs.