Learn AI in the right order

A Large Language Model (LLM) is an AI system trained on massive amounts of text that can understand, generate, and reason about language. It predicts what comes next in a sequence of text, and through that mechanism becomes capable of writing, summarising, translating, and answering questions.

An LLM is trained by processing billions of documents and learning to predict the next word (or token) given what came before. Through this training process, the model builds a compressed representation of how language works. At inference time, it generates text one token at a time, sampling from a probability distribution over possible next tokens.

LLMs are the foundation of almost every practical AI product people use today — chatbots, coding assistants, writing tools, customer service systems, and search. Understanding what they are is the prerequisite for understanding everything built on top of them.

Because the way LLMs work explains their failure modes. They hallucinate because they are predicting probable text, not retrieving verified facts. They perform better with clearer prompts because the prompt is the only context they have. Understanding the mechanism helps you use them intelligently and spot when they're unreliable.

Step 6 — Generative AI and LLMs. By the time you reach this step, you'll understand training (Step 3) and the Transformer architecture (Step 4), which makes LLMs much easier to grasp as a system built on those foundations.

RAG — how to give an LLM access to your own data. Context Window — what limits how much an LLM can consider at once. Hallucination — understanding when and why LLMs produce incorrect outputs.

An embedding is a way of representing a word, sentence, document, or image as a list of numbers (a vector) in a high-dimensional space. Items that are semantically similar end up numerically close together in that space, even if they use completely different words.

A model trained on language learns to place words into a numerical space where their position reflects their meaning and relationships. "Paris" ends up near "France" and near "city." You can even do arithmetic: king − man + woman ≈ queen. More advanced embedding models encode entire sentences or documents into single vectors that capture their full meaning.

Embeddings are how AI systems find meaning rather than just matching text. A search that uses embeddings finds documents about the topic you meant, not just documents that contain your exact words. They are the foundation of semantic search, RAG systems, recommendation engines, and much of modern NLP.

Because embeddings behave differently from keyword search in ways that matter for debugging and system design. If a RAG system retrieves the wrong documents, it's usually an embedding quality problem. Knowing how embeddings work helps you diagnose failures and choose the right embedding model for your use case.

Step 4 — Deep Learning, and practically essential for Step 5 (Projects) when you build any retrieval or search system. Vector databases (Step 5) only make sense once you understand what they're storing.

Vector Database — where embeddings are stored and searched at scale. RAG — the retrieval pattern that uses embeddings to give LLMs access to external knowledge.

An AI agent is a system where a language model is given tools, memory, and the ability to take actions in the world to complete goals autonomously. Instead of just responding to a single message, an agent can plan a sequence of steps, use tools like web search or code execution, observe results, and continue until the goal is complete.

An agent works by giving an LLM a goal and a set of tools it can call. The LLM reasons about what step to take next, calls a tool (e.g. search the web, run code, read a file), gets back a result, incorporates that result into its reasoning, and continues this loop until the goal is achieved or it gets stuck. The quality of the agent depends on the capability of the underlying LLM and the design of the tools and feedback loop.

Agents represent the shift from AI that answers to AI that acts. They can complete tasks that previously required a human to execute each step — research workflows, coding tasks, data pipelines, document processing. They are one of the fastest-developing areas in practical AI.

Because agents can make mistakes quickly and at scale. An agent that misunderstands a goal might send emails, delete files, or make purchases before a human notices. Understanding how the tool-call loop works — and where it can go wrong — is essential for deploying agents responsibly.

Step 7 — Specialisation, though the concept appears in Step 6 (LLMs). You need to understand LLMs well before agents make full sense, because an agent is ultimately an LLM orchestrating tools via a feedback loop.

Agentic AI — the broader concept of autonomous AI systems. Workflow — how agents are structured as sequences of steps. RAG — a pattern agents use to access external knowledge.