Fundamentals of Generative AI

Tokens are the unit of both processing and billing, roughly three-quarters of a word in English. Everything about cost and context limits is expressed in them.

Chunking is splitting long documents into pieces small enough to retrieve and feed to a model. It sounds like an implementation detail and it is not: chunks that are too small lose the context that made them meaningful, and chunks that are too large dilute relevance. Most retrieval quality problems are chunking problems.

Embeddings and vectors are the mechanism behind semantic search. Text is converted into a list of numbers positioned so that similar meanings sit near each other. This is what lets a search for "time off policy" find a document titled "annual leave", and equally, why it struggles with exact identifiers like part numbers, which carry little semantic signal.

Foundation models are large models trained broadly and adapted to many tasks. Multimodal models handle more than one kind of input or output. Diffusion models underlie most image generation.

Context engineering, an explicit objective here, is deciding what goes into the request: instructions, retrieved material, conversation history, examples. It is where most of the quality comes from and where most of the cost is incurred, which is why it deserves the name rather than being lumped in with prompting.

The foundation model lifecycle runs data selection → model selection → pre-training → fine-tuning → evaluation → deployment → feedback. Most organizations join at model selection and never touch pre-training, which is worth knowing so you can recognize when a proposal is quietly suggesting something enormous.

The genuine advantages: adaptability across tasks without retraining, conversational interaction, and the ability to generate content rather than only classify it.

The disadvantages, which AWS names directly:

• Hallucination: confident, fluent, wrong
• Limited interpretability: you cannot readily explain why a particular output appeared
• Inaccuracy: plainly wrong answers
• Non-determinism: the same input can produce different output on different runs

That last one deserves emphasis for a business audience: if a process requires an identical answer every time, a foundation model is the wrong component by design, not because of a defect awaiting a fix.

Token-based pricing and its effect on both cost and performance. You pay for input and output; input is usually the larger side because it carries instructions, retrieved context, and conversation history. Longer inputs also mean slower responses, so context length is a latency lever as well as a cost one.

Prompt caching is worth knowing by name here. Where a large part of every request is identical (a standing instruction block, a policy document), the repeated portion can be billed at a steep discount. It is an architecture decision made before you build, not an optimization applied afterward.

The AWS-specific cost levers listed in the guide: responsiveness and availability requirements, redundancy, regional coverage, token pricing, provisioned throughput (reserved capacity for predictable heavy load), and custom models. Provisioned throughput is the one that surprises people. It trades flexibility for predictable capacity and price at volume.

Model selection factors trade against each other as always: capability, performance, compliance posture, cost, latency, and complexity. There is no globally best model, only a best fit for a stated constraint.

AWS gives agentic AI real weight in this domain, and the vocabulary is worth having.

An agent uses a model plus tools to pursue a goal across multiple steps, deciding what to do next rather than following a fixed script.

The concepts named in the guide:

• Multi-agent patterns: several agents dividing work, useful when subtasks are genuinely independent and substantial
• A protocol layer connecting agents to external systems in a standardized way, so integrations are reusable rather than bespoke per application
• Memory management: what the agent retains between steps and across sessions
• Tool usage: the actions available to it
• Workflow orchestration: how the steps are coordinated

The manager-relevant translation of all of this: an agent can do exactly what its tools allow. The tool list is the permission list and therefore the risk surface. Reviewing it is the highest-value check a non-technical stakeholder can perform.

The AWS surface spans a managed service for accessing foundation models through a single interface, a broader machine learning platform for building and training, agent-building and runtime services, and developer-facing tooling. The exam tests recognition of what each is for rather than configuration detail.

Why use the managed platform rather than assembling it: lower barrier to entry, faster time to market, and, the argument that matters most in a regulated organization, a security and compliance posture you inherit rather than construct. Rebuilding that yourself is the hidden cost of the do-it-yourself route.