The AI Learning Hub Journal

LLM Landscape

Diagram 2a — LLM Core LLM box centre with generative AI, multimodal, agentic AI and other app patterns inside. Model quality left. LLM techniques right with RAG flagged as advanced. LLM mechanics band below with temperature and sampling flagged as advanced. Guardrails referenced as advanced topic. Model quality Pre-training Next-token prediction at scale Inference Running the model on new inputs relates to Large language models (LLMs) Transformers trained on massive text — predict, generate, and reason Generative AI Creates text, image, and code Multimodal Text, image, audio, and video Agentic AI LLM + tools + planning loop Other app patterns Copilots, classifiers, search Guardrails Runtime constraints — safety, scope, output filters, hallucination controls ▲ Advanced topic — explored in depth in the advanced section LLM techniques applied externally Prompt engineering Shapes model input at runtime RAG Injects external knowledge ▲ Advanced — 8 architectures explored further Fine-tuning Retrains on domain-specific data RLHF Aligns outputs via human feedback applied to LLM mechanics — how the model works under the hood Context window All text the model sees at once Tokenisation Text split into model units Embeddings Semantic vector representations Temperature and sampling Controls output randomness ▲ Advanced — 7 parameters explored further

One Diagram, the Whole Picture

LLMs come with a lot of vocabulary — and most of it gets used loosely. This diagram organises every key concept into a single map so you always know what something is, where it fits, and how it relates to everything else. The lessons in this module each go deep on one area. This lesson is the overview that ties them together.

  • LLM mechanics: pre-training, inference, tokens, context windows, embeddings, and sampling — covered in lesson 2
  • Guardrails and hallucinations: the safety layer around every deployment — covered in lesson 3
  • LLM techniques: prompt engineering, RAG, fine-tuning, and RLHF — covered in lesson 4
  • Infrastructure: how models reach users safely at enterprise scale — covered in lesson 5

Pre-Training vs Inference: Two Distinct Phases

Every LLM goes through two distinct phases. Pre-training is a one-time process where the model reads enormous amounts of text and learns to predict what comes next — that is how it acquires general knowledge. It costs tens of millions of dollars for frontier models and is entirely vendor territory. Inference is what happens every time someone sends a message — the model uses what it learned to generate a response. This is where enterprise cost, latency, and deployment decisions live.

  • Pre-training: next-token prediction at scale — the model learns patterns from billions of examples; never a customer responsibility
  • Inference: running the model on new inputs — happens on every query; the cost compounds at enterprise scale
  • Fine-tuning sits between the two: you continue training on your data, but starting from a pre-trained checkpoint — much cheaper than training from scratch, but still requires ML infrastructure
◆ See it for yourself
Open Next-Word Prediction in the library →

Watch the model choose the next word, one at a time.

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