Core LLM & Foundation Concepts
- Large Language Model (LLM) – A neural network trained on massive text data to predict and generate language.
- Transformer Architecture – A model design that uses attention instead of recurrence to understand context.
- Self-Attention – Mechanism that lets a model weigh which words matter most in a sentence.
- Multi-Head Attention – Multiple attention mechanisms running in parallel for richer understanding.
- Tokenization – Breaking text into smaller units (tokens) for model processing.
- Subword Tokenization – Splitting words into parts to handle rare or new words.
- Context Window – Maximum number of tokens an LLM can consider at once.
- Sliding Window Attention – Processing long text by shifting attention windows.
- Positional Encoding – Injecting word order information into the model.
- Embedding Space – Numerical space where meanings are represented as vectors.
Training & Optimization
- Pretraining – Initial training on large, general datasets.
- Fine-Tuning – Training a model further on domain-specific data.
- Instruction Tuning – Teaching models to follow human instructions.
- RLHF (Reinforcement Learning from Human Feedback) – Using human ratings to improve responses.
- Preference Modeling – Learning which outputs humans prefer.
- Loss Function – Mathematical signal that tells the model how wrong it is.
- Backpropagation – Method for updating model weights during training.
- Gradient Descent – Optimization technique to minimize error.
- Catastrophic Forgetting – When fine-tuning erases previous knowledge.
- Overfitting – Model memorizes data instead of generalizing.
Also read: Top 10 Data Modeling platforms for the AI era
Reasoning & Intelligence
- Chain-of-Thought (CoT) – Prompting the model to show step-by-step reasoning.
- Hidden Chain-of-Thought – Internal reasoning not exposed to the user.
- Tree-of-Thoughts (ToT) – Exploring multiple reasoning paths simultaneously.
- Reasoning Tokens – Internal tokens used for logical steps.
- Deliberate Reasoning – Slower, more careful inference for complex tasks.
- System 1 vs System 2 AI – Fast intuitive vs slow reasoning modes.
- Planning – Breaking tasks into ordered steps.
- Goal Decomposition – Splitting a goal into sub-goals.
- Symbolic Reasoning – Logic-based reasoning combined with neural models.
- Neuro-Symbolic AI – Hybrid of neural networks and logic systems.
Prompting & Interaction
- Prompt Engineering – Designing effective inputs for LLMs.
- Few-Shot Prompting – Providing examples inside the prompt.
- Zero-Shot Prompting – Asking without examples.
- System Prompt – Instructions that define model behavior.
- User Prompt – Input provided by the user.
- Role Prompting – Assigning a role (e.g., “act as a doctor”).
- Prompt Injection – Malicious prompts that override system instructions.
- Prompt Leakage – Exposure of system prompts unintentionally.
- Prompt Chaining – Using outputs as inputs to the next step.
- Adaptive Prompting – Dynamically changing prompts based on context.
Retrieval & Knowledge
- Retrieval-Augmented Generation (RAG) – LLM + external knowledge retrieval.
- Vector Search – Finding similar content using embeddings.
- Approximate Nearest Neighbor (ANN) – Fast similarity search method.
- Embedding Drift – When embeddings lose alignment over time.
- Chunking – Splitting documents into pieces for retrieval.
- Context Injection – Supplying retrieved data into prompts.
- Knowledge Grounding – Restricting AI to known facts.
- Hallucination – AI generating incorrect but confident answers.
- Factual Consistency – Alignment with real-world facts.
- Knowledge Cutoff – Last date of training data.
Agents & Autonomy
- AI Agent – LLM that can act, plan, and use tools.
- Agentic AI – Systems where AI makes decisions and executes actions.
- Autonomous Agent – AI operating without constant human input.
- Multi-Agent System – Multiple agents collaborating or competing.
- Tool Calling – LLM invoking APIs or software tools.
- Function Calling – Structured outputs triggering code execution.
- Agent Memory – Persisted context across interactions.
- Short-Term Memory – Context within a session.
- Long-Term Memory – Stored knowledge across sessions.
- Agent Orchestration – Coordinating multiple agents and tools.
Safety, Alignment & Trust
- AI Alignment – Ensuring AI goals match human values.
- Safety Guardrails – Constraints to prevent harmful outputs.
- Content Moderation – Filtering unsafe responses.
- Model Alignment Drift – Gradual misalignment over time.
- Red Teaming – Stress-testing AI for failures.
- Adversarial Prompting – Inputs designed to break models.
- Jailbreaking – Bypassing AI safety restrictions.
- Policy Enforcement Layer – System enforcing rules on outputs.
- AI Governance – Organizational control of AI usage.
- Human-in-the-Loop (HITL) – Humans supervising AI decisions.
Evaluation & Performance
- Benchmarking – Testing models against standard tasks.
- BLEU / ROUGE – Metrics for text quality evaluation.
- Perplexity – Measure of prediction uncertainty.
- Latency – Time to generate a response.
- Throughput – Responses per second.
- Token Cost – Cost per generated token.
- Inference Optimization – Making model outputs faster/cheaper.
- Quantization – Reducing precision to speed inference.
- Distillation – Training smaller models from larger ones.
- Model Compression – Reducing model size.
Advanced & Emerging Concepts
- Mixture of Experts (MoE) – Model activates only relevant sub-models.
- Sparse Attention – Attention applied selectively to reduce cost.
- Multimodal AI – Models handling text, images, audio, video.
- Cross-Modal Embeddings – Shared meaning across data types.
- Vision-Language Models (VLMs) – AI that understands images + text.
- Speech-to-Text (STT) – Converting speech into text.
- Text-to-Speech (TTS) – Generating speech from text.
- Synthetic Data – AI-generated training data.
- Self-Improving Models – Models that learn from usage.
- Model Routing – Choosing which model handles a task.
Enterprise AI Layer
- AI Control Plane – Central governance for AI behavior.
- Enterprise Prompt Management – Managing prompts at scale.
- LLMOps – Operating LLMs in production.
- Model Registry – Catalog of AI models.
- Inference Gateway – Central access point for AI requests.
- Observability for LLMs – Monitoring AI behavior and outputs.
- AI Audit Trail – Tracking how decisions were made.
- Semantic Guardrails – Meaning-based constraints on AI.
- Trust Layer – Systems ensuring AI reliability.
- AI Reasoning Layer – Logic layer guiding AI decisions.
As AI becomes operational, LLM terminology becomes the shared language of trust, governance, and intelligent execution.
