1.1. What is a language model?

Objectives: Understand the system that predicts the most likely next word given everything that came before it.

• Next-token prediction: Generates text one token at a time, predicting what should come next.
• The “sounds right” framing: The model predicts what sounds right, not what is right—explaining why it can invent facts.

1.2. Tokens, embeddings, and semantic space

Objectives: Deep dive into the attention mechanism: how the model decides which tokens are relevant.

• Tokenisation: Text split into word-sized chunks called tokens.
• Embeddings: Tokens represented as vectors encoding meaning in high-dimensional space.

1.3. Transformers & Training

Objectives: Understand where the model’s knowledge comes from and why it has a cutoff date.

• The Context Window: The model's fixed-size working memory.
• Pre-training: How the model acquires language, facts, and reasoning patterns.

Chain-of-thought reasoning

Think of Chain-of-thought as "asking the model to show its working, like a maths exam."

Raw LLMs often make mistakes when jumping directly to an answer for complex problems. By prompting a model to "think step-by-step," we force it to generate intermediate reasoning tokens. Because each token the model generates becomes part of its own context for the next token, this sequential processing significantly improves performance on logic and multi-stage planning tasks.

Retrieval-Augmented Generation (RAG)

How AI products "know" your company documents by connecting the model to a live knowledge base.

• The Courtroom Analogy: The LLM is the judge; RAG is like having law clerks who fetch relevant case law and precedents before a ruling is made.
• Retrieval vs Generation: Separating the act of finding information from the act of writing the response.

NVIDIA: What Is Retrieval-Augmented Generation?

The industry-standard explainer for how RAG provides context without retraining the model.

Required Industry Reading

Anthropic: Building Effective Agents

Reading Instruction: Read the first half of this post, up to and including the section on "Agents." The diagrams are especially useful. The implementation details in the second half can be skipped.

AI Agents and tool use

Moving from static prompts to autonomous decision-making systems that can use external tools.

• Workflows vs Agents: Understanding the spectrum from predefined orchestration to model-driven decision-making.
• The Augmented LLM: Combining the model with retrieval, tools, and memory.

Multimodal models

Understanding AI that processes multiple inputs—such as sight and sound—simultaneously.

• Takers, Shapers, Makers: Strategic categories for how organisations adopt multimodal technology.
• Industrial Impact: How vision-and-language models apply to maintenance and fraud detection.

McKinsey: What is multimodal AI?

A strategic overview using sensory analogies to explain the next leap in AI capability.

Practical AI judgment

Success with AI is not about technical skill—it is about knowing when to trust the model and when to step in as the human-in-the-loop.

• High-Value Tasks: Brainstorming, summarization, and format conversion.
• Risk Zones: Tasks where accuracy is critical, where you cannot spot errors, or where the "effort" itself is the point of the work.

Ethan Mollick: 15 Times to Use AI, and 5 Not To

A definitive framework for deciding which tasks to delegate and which to keep manual.

Safety Awareness

Addressing Hallucinations and Bias

Understand why AI models function like "advanced autocomplete" and learn the five practical mitigation strategies to ensure accuracy in your output.

AI safety and limitations

Hallucinations aren't bugs; they are a fundamental part of how language models generate text.

• Verification: Why you must check citations and data points against primary sources.
• Bias: Recognizing that models inherit the perspectives and flaws of their training data.

European Parliament Overview

The definitive anchor source for risk tiers (Unacceptable, High, Limited, Minimal).

SIG: Detailed Act Summary

A non-legal deep dive into compliance timelines (2025-2027) and potential fines.