The AI Learning Hub Journal

AI Hierarchy

Diagram 1 — AI hierarchy with deep learning domains AI contains ML. ML contains three learning paradigms and deep learning. Deep learning contains three application domains: NLP, computer vision, and robotics/automation. The transformer architecture sits as the backbone. NLP leads into LLMs which contain generative AI, multimodal, agentic AI, and other app patterns. Artificial intelligence Systems that perform human-like reasoning Machine learning Models that learn from data, not explicit rules Supervised learning Labelled input → output pairs Unsupervised learning Finds patterns without labels Reinforcement learning Learns from reward signals over time Deep learning Multi-layer neural networks that learn representations automatically Transformer architecture — attention mechanism powering NLP, vision models, and more Natural language processing Chatbots, translation, summarisation Computer vision Image recognition, object detection Robotics and automation Control, navigation, perception Large language models (LLMs) Transformers trained on massive text — the dominant output of the NLP branch Generative AI Text, image, and code Multimodal Text, image, audio, video Agentic AI LLM + tools + planning loop Other app patterns Copilots, classifiers, search

The Nesting Hierarchy

The diagram shows how every AI term you encounter fits into a single hierarchy. Understanding where a product sits tells you everything about its capabilities, limitations, and failure modes.

  • Artificial Intelligence — the outermost layer: any system that simulates human reasoning
  • Machine Learning — AI that learns from data rather than explicit rules. The dominant technique in production AI today
  • Deep Learning — a subset of ML using multi-layer neural networks that automatically learn representations from raw data
  • Generative AI — Deep Learning models trained to produce new content. This is where LLMs, image generators, and code tools live

Learning Modes at a Glance

Inside Machine Learning, three training approaches dominate. We cover them in depth in the next lesson — here is just the map.

  • Supervised: labelled input/output pairs — the workhorse behind classification, scoring, and prediction
  • Unsupervised: pattern-finding without labels — clustering, anomaly detection, segmentation
  • Reinforcement: learning from reward signals — used to fine-tune LLMs and to drive agentic systems

LLMs and the Application Layer

NLP's most powerful output is the Large Language Model — a Transformer trained on massive text. From LLMs, four distinct product patterns emerge.

  • Generative AI: produces text, code, and images on demand — the foundation of copilots and content generation tools
  • Multimodal: processes and generates across text, image, audio, and video simultaneously. Today's frontier models — Claude, GPT, Gemini — are all natively multimodal
  • Agentic AI: LLM + tools + planning loop — takes sequences of actions to complete multi-step goals autonomously
  • Copilots, classifiers, search: the practical enterprise patterns that sit on top of LLMs and show up in every vendor's product pitch

Why This Matters

When someone says "we want AI," the hierarchy tells you which conversation to have — regardless of your role.

  • "AI" alone is too vague — clarify whether they mean classic ML detection, DL-based pattern recognition, or GenAI/LLM capabilities
  • Different layers carry different tradeoffs: cost, latency, explainability, data requirements, and attack surface
  • Not all ML is DL — decision trees, random forests, and SVMs still dominate many production workloads and are often the right answer

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.