Zum Inhalt springen
AI Learning

LLM Deep Dive

iAs of: September 2026

Specific model names and benchmark scores in this module reflect the state as of September 2026. The concepts (reasoning, thinking tokens, benchmarks) are long-lived — the example models become outdated quickly.

Welcome to the first module of the Advanced path. In the Beginner course, you learned what LLMs are, how tokens work, and what embeddings mean. Now it's time to go deeper.

In this module, you'll understand how LLMs work under the hood -- not just what they do. You'll get to know the mechanisms behind modern AI systems, and afterwards you'll be able to make informed judgments about which model is right for which use case.

What to expect

  • From Prompt to Answer -- The five stations every token passes through: tokenisation, embedding, attention, MLP and sampling
  • Embeddings in Detail -- How vector spaces work and why cosine similarity is the key to semantic search
  • Context Strategies -- How LLMs handle long texts: sliding window, summarization, and the evolution from 4K to 1M tokens
  • RAG Introduction -- Why LLMs need external knowledge sources and how Retrieval-Augmented Generation works
  • Reasoning Models -- How GPT-5.6, Claude Opus 5, and Gemini 3.1 Pro natively "think" and what thinking tokens are
  • Understanding Benchmarks -- What MMLU, HumanEval, and Arena ELO actually measure -- and what they don't

Who is this module for?

You should have completed the Beginner path or already have foundational knowledge about LLMs. If terms like "token", "embedding", and "context window" sound familiar, you're in the right place.

*Learning Objective

After this module, you'll be able to analyze LLM architectures, compare models using benchmarks, and evaluate RAG as a strategy for knowledge-based applications. You'll be operating at Bloom's taxonomy levels of "Apply" and "Analyze".

Let's get started -- with the path your prompt takes through the model.

Reflect

LLMs are more than black boxes -- with an understanding of embeddings, benchmarks, reasoning, and context strategies, you can make informed model decisions. This module gives you the tools for that. Let us start with the path your prompt takes through the model.