Generative AI Core Concepts

Most explanations of generative AI stop at the chat window. This topic opens the box. It covers the transformer architecture that lets a model read a whole sentence at once, the way text becomes tokens and then numeric vectors, and the very different mechanism that turns random noise into an image. From there it moves to the decisions a builder actually makes: which model, how much context, what it costs, and when to hand the work to an agent.

The lessons build in order, and each one depends on the one before it. Tokens explain the bill. The context window explains why retrieval beats pasting everything. And the limits of a single model call explain why agents exist at all.

What This Topic Covers

• foundation models and large language models, the transformer architecture, self-attention, and what model parameters actually store
• tokens and tokenization, chunking strategies in Amazon Bedrock, and embeddings as searchable coordinates for meaning
• multimodal models, and how diffusion models generate images by reversing a noising process
• the generative AI use cases the exam names, grouped by the capability behind each one
• the seven-stage foundation model lifecycle, from data selection through deployment to feedback
• token-based pricing, and how on-demand, batch, Provisioned Throughput, and prompt caching change the cost of the same workload
• context engineering: what competes for the context window and why more context can produce worse answers
• agentic AI, the reasoning loop, multi-agent patterns, and the Model Context Protocol

Why It Matters

This topic maps directly onto Task Statement 2.1 of the AIF-C01 exam guide, including the objectives added in the 2026 revision for token-based pricing, context engineering, and agentic AI with MCP. Domain 2 carries 24% of the scored content, and the vocabulary built here is assumed by the largest domain that follows.

The value outlasts the exam. Once you can trace a request from prompt to tokens to inference to invoice, the questions that matter at work become answerable: whether this task suits a foundation model at all, which model to start with, how much context it genuinely needs, and what the whole thing will cost once real traffic arrives.