This domain is 24% of the scored exam. It tests vocabulary (token, embedding, chunking, multimodal), what generative AI is good and bad at, and which AWS service to reach for: Bedrock, SageMaker JumpStart, Amazon Q or PartyRock. Questions are conceptual and scenario-based; none ask you to build anything.
Key concepts
| Term |
Meaning |
| Foundation model (FM) |
A large model pre-trained on broad data (often unlabeled) that can be adapted to many tasks without training from scratch |
| LLM |
A foundation model for language; generates text by predicting the next token |
| Token |
The unit a model reads and writes: a word, part of a word, or punctuation. Limits and prices are measured in tokens (roughly 3 to 4 English characters each) |
| Context window |
The maximum tokens (input plus output) a model handles in one request |
| Embedding |
A list of numbers (a vector) that captures the meaning of text, an image or other data. Similar meanings land close together |
| Vector |
The numeric array itself; vectors are compared with distance measures such as cosine similarity |
| Chunking |
Splitting long documents into smaller pieces before embedding them, so retrieval returns focused passages |
| Prompt |
The input you give the model; prompt engineering is crafting it for better output |
| Transformer |
The neural-network architecture behind modern LLMs; its self-attention mechanism weighs how relevant each token is to every other token, and it processes tokens in parallel |
| Multimodal model |
Takes or produces more than one type of data (text plus image plus audio) |
| Diffusion model |
Generates images (and video, audio) by starting from random noise and removing it step by step, guided by a prompt. Stable Diffusion, Titan Image Generator and Nova Canvas are examples |
| Inference parameters |
Settings like temperature and max tokens that shape output (see the next domain) |
How the pieces fit
- Text is split into tokens.
- Tokens become embeddings.
- A transformer processes them and predicts the next token, repeated until the answer is done.
- For images, a diffusion model denoises toward a picture.
Embedding models are different from text-generation models: they output vectors for search, clustering and classification, not prose.
Generative AI use cases
| Use case |
Example |
| Text generation / summarisation |
Summarise a long contract or call transcript |
| Chatbots and virtual assistants |
Customer support agents grounded in company data |
| Translation |
Multi-language content |
| Code generation |
Completion, explanation, tests, upgrades (Amazon Q Developer) |
| Image, video and audio generation |
Marketing images, music, voice |
| Search |
Semantic search and Q&A over documents |
| Recommendation engines |
Personalised suggestions and explanations |
| Customer service agents |
Agents that look up orders and take actions |
| Data extraction and classification |
Pull fields out of messy text |
Foundation model lifecycle
| Stage |
What happens |
| Data selection |
Collect and curate large, diverse, high-quality data |
| Model selection |
Choose an architecture or an existing FM |
| Pre-training |
Train on massive unlabeled data using self-supervised learning. By far the most expensive step |
| Fine-tuning |
Adapt to a task or domain with a smaller labeled dataset |
| Evaluation |
Benchmarks, automatic metrics, human review |
| Deployment |
Host as an API or endpoint |
| Feedback |
Collect user feedback and monitoring data to improve the model or prompts |
Most organisations skip pre-training and start from an existing FM.
Advantages of generative AI
- Adaptability: one model handles many tasks, and different prompts redirect it with no retraining.
- Responsiveness: instant, natural-language answers at scale.
- Simplicity: natural-language interface; low barrier to entry; fast prototypes.
- Productivity, personalisation and creativity at lower unit cost.
Disadvantages and risks
| Limitation |
What it means |
Common mitigation |
| Hallucination |
Confident but false output |
RAG, grounding checks, citations, human review |
| Inaccuracy |
Wrong or outdated facts (knowledge stops at the training cutoff) |
RAG with current data |
| Nondeterminism |
The same prompt can return different answers |
Lower temperature; don’t use for tasks needing exact repeatable outputs |
| Interpretability |
Hard to explain why it produced something |
Transparency docs, source citations |
| Bias and toxicity |
Reflects biases in training data |
Guardrails, evaluation, diverse data |
| Privacy and IP risks |
Sensitive data or copyrighted material in prompts or outputs |
Data controls, guardrails |
| Cost and latency |
Large models are slow and expensive |
Smaller models, caching, batching |
Choosing a generative AI model
Factors: model type and modality (text, image, embedding), performance (quality on your task), capabilities (reasoning, languages, tool use), constraints (latency, cost, context window), compliance (data residency, licensing, provider terms) and customisation options. Always evaluate candidates on your own data; leaderboards are only a starting point.
Business value and metrics
Judge success with business measures, not only model scores:
- Efficiency (time saved, tickets handled per agent, productivity)
- Conversion rate and average revenue per user (ARPU)
- Customer lifetime value and satisfaction
- Accuracy of answers
- Cross-domain performance: works across several tasks and subject areas
- Cost per interaction and ROI
AWS services for generative AI
| Service |
What it is |
Pick it when |
| Amazon Bedrock |
Fully managed service giving API access to foundation models from Amazon and other providers (Anthropic, Meta, Mistral, Cohere, AI21, Stability AI and others) with a unified API, plus Knowledge Bases, Agents, Guardrails, evaluation and customisation |
You want to build generative AI apps without managing infrastructure; your data isn’t used to train the base models |
| Amazon SageMaker JumpStart |
Hub of pre-trained and open-source models deployed to your own SageMaker endpoints; fine-tune with your code |
You need full control of the model, instance and hosting, or a model Bedrock doesn’t offer |
| PartyRock |
A no-code Bedrock playground for building small generative AI apps by describing them. For learning and experimenting; not a production platform |
Experiment, demo, learn |
| Amazon Q Business |
Assistant that answers questions over your company data with connectors and respects existing permissions |
Employees need answers from internal systems |
| Amazon Q Developer |
Coding assistant in IDEs, CLI and console: code generation, explanation, transformation, troubleshooting |
Developer productivity |
| Amazon Q in QuickSight |
Natural-language BI questions and data stories |
Business analysts |
| Amazon Q in Connect |
Real-time agent assist for contact centres |
Customer service |
| AWS Trainium / Inferentia |
AWS-designed chips for lower-cost training and inference |
Cost-efficient large-scale training or inference |
| Amazon Nova / Titan |
Amazon’s own foundation models available in Bedrock |
Many tasks within AWS |
Bedrock versus SageMaker JumpStart
|
Bedrock |
JumpStart |
| Infrastructure |
Serverless; none to manage |
You choose instances and endpoints |
| Model access |
Many leading FMs via one API |
Hundreds of open and proprietary models |
| Control |
Less (managed) |
More (full) |
| Effort |
Lowest |
Higher |
| Pricing |
Per token, or provisioned throughput |
Instance hours |
Advantages of AWS generative AI services
- Accessibility and lower barrier to entry: APIs and no-code tools.
- Speed to market: skip building and hosting models.
- Efficiency and cost-effectiveness: pay per use; no large GPU cluster.
- Security, compliance and safety: data encrypted, private connectivity, no sharing of your prompts with model providers, compliance programs.
- Responsible AI features built in, such as Guardrails.
- Choice of models, so you can swap as better ones appear.
Cost tradeoffs
| Pricing choice |
How it works |
Best for |
| On-demand |
Pay per input and output token (or per image) |
Unpredictable or low traffic; experiments |
| Batch inference |
Submit many prompts for asynchronous processing at a discount |
Large offline jobs |
| Provisioned throughput |
Commit to model units for a term and get guaranteed throughput; required to use custom (fine-tuned) models in Bedrock |
Steady high-volume traffic needing consistent performance |
| Self-hosted (SageMaker) |
Pay for instance hours whether or not busy |
Full control, steady heavy load |
Other cost drivers: model size (larger costs more and is slower), number of tokens (long prompts and long answers cost more), customisation (fine-tuning and continued pre-training add training and storage cost), and Region (not every model is in every Region; cross-Region inference helps availability). Think responsiveness, availability, redundancy and performance against cost.
Common traps
- “No infrastructure to manage, multiple models, one API” is Bedrock; “choose the instance type and deploy an open model” is JumpStart.
- PartyRock is for experimenting, not production.
- Bigger is not always better: a smaller model is cheaper and faster when it is good enough.
- Hallucination is fixed by grounding, not by a bigger model alone.
- Provisioned throughput costs money whether or not you use it.
Exam checklist
- Define token, embedding, vector, chunking, transformer, diffusion, multimodal, foundation model.
- List the foundation-model lifecycle in order.
- Name three advantages and four disadvantages of generative AI with a mitigation each.
- Pick Bedrock, JumpStart, PartyRock or Amazon Q for a scenario.
- Compare on-demand, batch and provisioned throughput.
- Name business metrics (efficiency, conversion, ARPU, lifetime value) for judging a generative AI project.