Intermediate Track

How tokenisation works, what a context window actually costs you in money and accuracy, and the arithmetic that makes AI spending predictable.

Zero-shot, few-shot, chain-of-thought and structured output — what each technique does, what it costs, and the failure mode that comes with it.

How retrieval-augmented generation actually works, why embeddings find meaning rather than keywords, and the failure modes nobody warns you about.

How the reason-act loop works, why tool descriptions matter more than tool names, and where the approval boundary has to sit.

The trigger-step-action pattern, why non-determinism breaks ordinary automation habits, and the four controls every AI workflow needs.

Why the same architecture handles images, audio and video, how diffusion differs from next-token prediction, and what those differences mean in practice.

How to tell whether a prompt change actually helped, why vibes-based testing misleads, and how to build a useful evaluation set in an afternoon.