Concepts
Core
- LLM: A model that generates text by predicting the next token.
- Foundation Model: broad AI model trained on vast amounts of data that can be adapted to many tasks.
- Base Model: pre-trained model that predicts text but lacks instruction-following tuning.
- Parameter: Internal variables that store the model's learned patterns.
- Token: piece of text (word or symbol) used as the basic unit of processing.
- Tokenization: process of converting raw text into tokens.
- Embedding: Numerical vectors that represent the semantic meaning of tokens.
- Latent Space: mathematical "map" where embeddings are organized by meaning.
- Multimodality / Multimodal AI: ability to process multiple input types (text, images, audio).
- MCP: open standard that enables AI models to connect to data sources and tools seamlessly.
Prompting Layer
- Prompt: The complete input (instructions + context) sent to the model.
- System Prompt: High-level instructions defining the model's persona/limits.
- User Prompt: specific question or request from the human.
- Prompt Engineering: craft of refining prompts to get better, more accurate outputs.
- Context Window: maximum number of tokens a model can "remember" at one time.
- Context Engineering: Managing and optimizing the information provided within the context window.
- Zero-Shot Learning: Task performance without any provided examples.
- Few-Shot Learning: Including a few examples in the prompt to guide the output.
- Chain of Thought: Prompting the model to show its step-by-step reasoning.
RAG
- RAG (Retrieval-Augmented Generation): Adding external data to a prompt before generation.
- Semantic Search: Searching for information based on meaning rather than just keywords.
- Chunk: smaller segment of larger document used for indexing in RAG.
- Vector Database: specialized database designed to store and search embeddings efficiently.
- Grounding: Linking model responses to verifiable, external facts to prevent errors.
Training
- Pre-training: Initial training on massive datasets to learn language patterns.
- Fine-tuning: Training a pre-trained model on a specific dataset to shape behavior.
- Instruct Model: model fine-tuned specifically to follow directions.
- Alignment: Ensuring the AI's goals match human values (helpful, honest, harmless).
- RLHF (Reinforcement Learning from Human Feedback): Using human ranking to guide model behavior.
- Distillation: Creating a smaller, faster "student" model from a large "teacher" model.
- Train-time Compute: The total computational power used during the training phase.
- Inference-time Compute (Reasoning Models): The extra processing a model does while thinking before it answers.
Execution
- Inference: The process of the model actually generating an output from an input.
- Latency: The delay between sending a prompt and receiving the response.
- Temperature: A setting that controls the randomness/creativity of the output.
- Evals (Evaluations): Systematic tests used to measure a model's performance.
- Benchmarks: Standardized tests (like MMLU) used to compare different models.
- Workflow: fixed, predefined sequence of steps for an LLM to follow.
Agents
- AI Agent: system where the LLM plans and uses tools to achieve a goal.
- Agentic AI: AI designed with high degree of autonomy and goal-oriented behavior.
41. Agentic RAG: RAG system where agent decides which data to retrieve and how to use it.
- Multi-agent System: Multiple AI agents working together to solve complex tasks.
- Agent2Agent: Communication and collaboration between different AI agents.
- Reasoning Models: Models (OpenAI's o1) designed to "think" longer to solve complex logic problems.
Safety
- Hallucination: When a model confidently generates false or fabricated info.
- Guardrails: Rules or systems that block unsafe, biased, or inappropriate content.
- Prompt Injection: security vulnerability where a user tries to trick the AI into ignoring its instructions.
