Skip to content

Task Statement 2.3 Study Notes

Describe AWS infrastructure and technologies for building generative AI applications

This task statement focuses on:

  1. AWS services used to build generative AI applications.
  2. Advantages of using managed AWS generative AI services.
  3. AWS infrastructure benefits such as security, compliance, safety, and resilience.
  4. Cost and architectural tradeoffs when selecting models, infrastructure, and deployment options.

1. Generative AI Application Stack

A generative AI application commonly contains three logical layers.

Layer 1: Infrastructure

This is the underlying compute, networking, storage, and security infrastructure used to train, fine-tune, and host models.

Examples include:

  • Amazon EC2
  • AWS Nitro System
  • AWS Trainium
  • AWS Inferentia
  • GPU-based EC2 instances
  • Amazon S3
  • Amazon VPC
  • AWS Regions and Availability Zones
  • AWS IAM and AWS KMS

Training and inference infrastructure

Generative AI models require significant infrastructure:

  • Training requires large amounts of compute and data processing.
  • Fine-tuning requires less compute than training from scratch, but can still require GPUs or ML accelerators.
  • Inference is the process of using a trained model to generate an output. Inference capacity must be available whenever users submit requests.

AWS provides purpose-built hardware:

Hardware Primary purpose
AWS Trainium Machine learning model training
AWS Inferentia Machine learning model inference
GPUs Training and inference for many AI workloads
AWS Nitro System Security isolation and infrastructure protection

Specialized hardware can improve price performance compared with using general-purpose CPUs or traditional GPU infrastructure.

Layer 2: Machine learning and foundation model services

This layer includes services that help developers:

  • Select models
  • Train or fine-tune models
  • Deploy models
  • Evaluate models
  • Scale inference
  • Build retrieval-augmented generation applications
  • Add agents and tools to applications

Examples include:

  • Amazon SageMaker AI
  • SageMaker JumpStart
  • Amazon Bedrock
  • Amazon Bedrock Knowledge Bases
  • Amazon Bedrock model evaluation capabilities
  • Amazon Bedrock Agents
  • Amazon Bedrock AgentCore
  • Strands Agents

Layer 3: Applications

This is the user-facing layer.

Applications can use foundation models to:

  • Generate text
  • Summarize documents
  • Create images
  • Write or debug code
  • Extract insights
  • Answer questions
  • Generate recommendations
  • Automate business processes
  • Take actions using tools and APIs

Examples include:

  • Chatbots
  • Customer service assistants
  • Employee productivity assistants
  • Document-processing applications
  • RAG applications
  • Coding assistants
  • AI agents
  • Business dashboards

2. Amazon Bedrock

What is Amazon Bedrock?

Amazon Bedrock is a fully managed service that provides API access to a selection of foundation models from Amazon and third-party model providers.

It allows developers to build generative AI applications without managing the underlying model infrastructure.

Depending on the model and capability, Bedrock can provide access to models from providers such as:

  • Amazon
  • Anthropic
  • Cohere
  • Meta
  • Mistral AI
  • Stability AI
  • Other supported providers

Amazon Titan is Amazon’s family of foundation models.

Key advantages

Amazon Bedrock provides:

  • Access to multiple foundation models through a common AWS service
  • No need to train a foundation model from scratch
  • Managed model hosting and inference
  • Model experimentation
  • Model evaluation
  • Knowledge bases and RAG capabilities
  • Agents and tool use
  • Enterprise security integration
  • Usage-based pricing for many inference modes
  • The ability to compare models for a specific use case

Important exam concept

Bedrock is generally the simpler choice when you want to consume foundation models through APIs and build applications quickly.

You do not need to manage GPU clusters, model servers, or the complete ML lifecycle.

Amazon Bedrock model selection

Different models can produce different results for:

  • Accuracy
  • Reasoning
  • Coding
  • Summarization
  • Language support
  • Image generation
  • Latency
  • Context window size
  • Cost

Bedrock provides tools such as:

  • Playgrounds for testing prompts and inference parameters
  • Model evaluation capabilities for comparing model behavior
  • Inference parameters such as temperature and maximum output tokens

Common inference parameters

Parameter Effect
Temperature Controls randomness. Higher values generally produce more varied output.
Maximum tokens Limits the length of the generated response.
Top-p or related sampling controls Controls the probability distribution used during generation.

Exam tip

There is no universally “best” foundation model. The correct model depends on the application’s requirements, including:

  • Accuracy
  • Cost
  • Latency
  • Safety
  • Availability
  • Language support
  • Context length
  • Output type

Custom models in Amazon Bedrock

Amazon Bedrock can support customization options for certain models and architectures, including the ability to import custom model weights where supported.

Possible reasons to customize a model include:

  • Adapting it to a specialized domain
  • Improving performance on organization-specific tasks
  • Adjusting behavior to a particular style
  • Using proprietary model weights

Custom models usually involve additional cost and operational complexity compared with using a base model.

Amazon Bedrock pricing modes

Depending on the model and feature, Bedrock may support:

  • On-demand inference
  • Provisioned Throughput
  • Custom model deployment options
  • Token-based pricing

On-demand inference

You pay based on usage, commonly the number of input and output tokens.

Best suited for:

  • Variable workloads
  • Prototyping
  • Applications with unpredictable traffic
  • Applications that do not require guaranteed dedicated throughput

Provisioned Throughput

Provisioned Throughput reserves model capacity for a specified period or commitment.

Best suited for:

  • Predictable, high-volume workloads
  • Applications requiring more consistent throughput
  • Workloads requiring capacity guarantees
  • Production applications with strict performance requirements

Tradeoffs include:

  • Higher commitment
  • Potentially higher fixed cost
  • Risk of paying for unused capacity
  • Better predictability and performance consistency

Exam trap

Provisioned Throughput is not automatically cheaper. It can be cost-effective at sufficiently high and predictable utilization, but on-demand pricing is often better for irregular or low-volume workloads.

3. Amazon SageMaker AI

What is Amazon SageMaker AI?

Amazon SageMaker AI is a managed machine learning service used to:

  • Build ML models
  • Train models
  • Fine-tune models
  • Deploy models
  • Monitor models
  • Manage ML workflows
  • Operate models at scale

SageMaker AI provides more control over the ML lifecycle than a simple model API service.

It is appropriate when an organization needs control over:

  • Training data
  • Training algorithms
  • Model architecture
  • Fine-tuning procedures
  • Deployment configuration
  • Inference infrastructure
  • Model monitoring
  • ML pipelines

Bedrock versus SageMaker AI

Requirement More likely choice
Quickly use a managed foundation model through an API Amazon Bedrock
Compare multiple foundation models without managing infrastructure Amazon Bedrock
Train or fine-tune models with extensive control SageMaker AI
Build custom ML workflows and pipelines SageMaker AI
Deploy and manage a custom model endpoint SageMaker AI
Need detailed control over compute and model hosting SageMaker AI

Exam tip

Do not assume SageMaker is only for traditional machine learning. SageMaker AI can also be used for foundation model development, customization, training, fine-tuning, deployment, and monitoring.

4. SageMaker JumpStart

What is SageMaker JumpStart?

SageMaker JumpStart is a model hub and collection of resources that helps developers quickly begin ML and generative AI projects.

It can provide:

  • Pretrained models
  • Foundation models
  • Built-in algorithms
  • Example notebooks
  • Sample solutions
  • Datasets
  • Deployment templates
  • Fine-tuning options
  • Reference architectures and best practices

JumpStart reduces the time needed to discover and deploy suitable models.

Transfer learning

Transfer learning means using knowledge learned by a model from one task or dataset as the starting point for another related task.

For example:

  1. A foundation model is trained on a large, general-purpose dataset.
  2. The model already understands language patterns or other general features.
  3. The model is fine-tuned on a smaller, domain-specific dataset.
  4. The resulting model is adapted to the organization’s use case.

Benefits of transfer learning

  • Less training time
  • Less data required
  • Lower cost than training from scratch
  • Faster development
  • Often good accuracy for specialized tasks

Important distinction

  • Training from scratch creates a model’s knowledge from the beginning and is usually very expensive.
  • Fine-tuning adapts an existing pretrained model.
  • Prompt engineering changes the instructions sent to a model without changing its weights.
  • RAG supplies external information at inference time without necessarily changing model weights.

JumpStart cost considerations

JumpStart models may require:

  • GPU-backed instances
  • Compute for fine-tuning
  • Compute for model endpoints
  • Storage for model artifacts
  • Monitoring and data transfer costs

Cost-saving practice

Delete unused SageMaker endpoints. A deployed endpoint can continue incurring charges even when it receives no requests, depending on its configuration.

Exam trap

JumpStart helps you find and deploy models, but it does not mean that the model is free or that the required compute is free.

5. Amazon Q, Amazon Quick Suite, Kiro, and Developer Tools

The exam may identify newer AWS tools used to build or consume generative AI applications.

Amazon Q

Amazon Q is a generative AI assistant from AWS.

It can support:

  • Business questions
  • Enterprise knowledge access
  • Developer assistance
  • Code generation
  • Code explanation
  • Troubleshooting
  • AWS service guidance
  • Software development tasks

Amazon Q can be connected to organizational information and business systems, subject to permissions and configuration.

Key point

Amazon Q is primarily an AI assistant and productivity solution. It is not the same as Amazon Bedrock, which is a platform for developers to build applications using foundation models.

Amazon Quick Suite

Amazon Quick Suite is an AWS business productivity and analytics experience that uses generative AI to help users work with business information and workflows.

Depending on the capability, it may help users:

  • Ask questions about business data
  • Generate insights
  • Create or interact with business content
  • Automate tasks and workflows
  • Use AI assistance across business applications

Exam guidance

Treat business-facing AI assistants and productivity tools as higher-level applications, while Bedrock and SageMaker AI are developer and ML platforms.

Kiro

Kiro is an AI-powered development environment/IDE designed to help developers build software with AI assistance.

It can support development activities such as:

  • Understanding requirements
  • Creating specifications
  • Planning implementation
  • Generating code
  • Testing
  • Debugging
  • Maintaining project context

Key point

Kiro is a developer tool. It is not a foundation model training service and is not a replacement for Bedrock or SageMaker AI.

6. Strands Agents

Strands Agents is an open-source SDK/framework for building AI agents.

An agent can:

  1. Receive a goal or request.
  2. Reason about the task.
  3. Select and invoke tools.
  4. Observe tool results.
  5. Continue working until it produces an answer or completes an action.

Tools might include:

  • APIs
  • Lambda functions
  • Databases
  • Search systems
  • Business applications
  • File systems
  • Other software services

Strands Agents can be used to create model-driven applications where the model determines which tools to use.

Agent versus chatbot

A basic chatbot generally generates responses.

An agent can:

  • Decide what steps are needed
  • Call tools
  • Access information
  • Perform actions
  • Coordinate multiple steps

Exam trap

An agent is not automatically reliable or safe merely because it uses an LLM. Tool permissions, validation, monitoring, and human approval may be required.

7. Amazon Bedrock AgentCore

Amazon Bedrock AgentCore provides capabilities for deploying and operating AI agents and agent-based applications at scale.

Agent applications typically need more than a model. They may require:

  • Runtime execution
  • Tool connectivity
  • Identity and access control
  • Session or memory management
  • Observability
  • Security controls
  • Policy enforcement
  • Integration with enterprise systems

AgentCore is designed to help organizations operationalize agents in a managed AWS environment.

Key concepts

AgentCore-related capabilities can help with:

  • Running agents
  • Connecting agents to tools and data
  • Managing agent identity
  • Monitoring agent behavior
  • Supporting enterprise-grade agent deployments
  • Controlling access to actions and resources

Exam guidance

  • Strands Agents: framework or SDK used to build agents.
  • Bedrock AgentCore: managed capabilities for deploying, running, securing, and operating agents.
  • Amazon Bedrock: managed access to foundation models and generative AI application features.
  • SageMaker AI: managed ML development and model lifecycle platform.

8. Advantages of AWS Generative AI Services

Accessibility

AWS makes generative AI accessible to organizations that do not have:

  • Large ML research teams
  • Specialized hardware
  • Expertise in distributed model training
  • Experience operating model-serving infrastructure

Developers can use APIs, SDKs, managed services, and prebuilt models.

If a developer can write code, call an API, and use the AWS SDK, they can often integrate generative AI capabilities without being an ML expert.

Lower barrier to entry

Managed services reduce the need to:

  • Acquire GPUs
  • Build model-serving infrastructure
  • Manage operating systems
  • Configure distributed training
  • Create model deployment systems
  • Maintain complex AI platforms

This allows teams to focus on:

  • Business requirements
  • Application logic
  • User experience
  • Data quality
  • Evaluation
  • Responsible AI controls

Efficiency

AWS managed AI services can improve development efficiency by providing:

  • Pretrained models
  • Ready-to-use APIs
  • Model hubs
  • Example notebooks
  • Built-in integrations
  • Deployment automation
  • Evaluation tools
  • Scaling capabilities

Transfer learning and fine-tuning can also reduce the time and data needed to create a specialized solution.

Cost-effectiveness

Using managed services can be more cost-effective than owning and operating an AI infrastructure platform because the organization avoids or reduces:

  • Up-front hardware purchases
  • Data center costs
  • Hardware maintenance
  • Infrastructure administration
  • Idle capacity
  • Specialized operational staffing

However, managed services are not always the lowest-cost option. High-volume, predictable workloads may benefit from dedicated infrastructure or provisioned capacity.

Speed to market

Pretrained models, JumpStart resources, Bedrock APIs, and managed deployment capabilities help organizations move from idea to prototype and production faster.

This supports:

  • Rapid experimentation
  • Faster proof-of-concept development
  • Model comparison
  • Faster integration with business applications
  • Quicker iteration based on user feedback

Ability to meet business objectives

Generative AI can support business goals such as:

  • Improving customer service
  • Increasing employee productivity
  • Automating repetitive tasks
  • Reducing operational costs
  • Creating new products and services
  • Improving decision support
  • Generating content
  • Extracting insights from documents and data
  • Enabling new revenue streams

The correct solution should be selected based on the business objective rather than simply choosing the newest or largest model.

9. AWS Global Infrastructure for GenAI Applications

AWS infrastructure is organized around:

  • Regions
  • Availability Zones
  • Edge locations
  • Regional and global services

AWS Regions

A Region is a separate geographic area containing multiple Availability Zones.

Regions support:

  • Data residency requirements
  • Disaster recovery
  • Geographic proximity to users
  • Regional service availability
  • Compliance requirements

Availability Zones

An Availability Zone is an isolated location within a Region.

Using multiple Availability Zones can improve:

  • Availability
  • Fault tolerance
  • Resilience
  • Recovery from infrastructure failure

Edge locations

Edge locations support services such as Amazon CloudFront and help deliver content closer to users.

They can reduce latency for application content, although model inference itself may still occur in a specific AWS Region.

10. Security, Compliance, Responsibility, and Safety

Security is particularly important for generative AI because applications may process:

  • Personal information
  • Financial information
  • Operational data
  • Proprietary business data
  • Model weights
  • Prompt and response data
  • Enterprise documents

Three critical AI system components

An AI system can be analyzed as:

  1. Input
  2. Model
  3. Output

Security and governance controls should apply to all three.

Input protection

Protect inputs against:

  • Sensitive data exposure
  • Unauthorized access
  • Prompt injection
  • Malicious instructions
  • Data poisoning
  • Untrusted documents

Model protection

Protect:

  • Model weights
  • Fine-tuned artifacts
  • Training data
  • Configuration
  • Inference endpoints
  • Access credentials

Output protection

Validate outputs for:

  • Confidential information leakage
  • Hallucinations
  • Toxic or unsafe content
  • Incorrect recommendations
  • Unauthorized actions
  • Regulatory violations

AWS security controls

Important security mechanisms include:

AWS Identity and Access Management

IAM controls:

  • Who can access AI services
  • Which resources they can access
  • Which actions they can perform

Use least privilege whenever possible.

Encryption

Use encryption:

  • At rest, such as with AWS KMS
  • In transit, using secure protocols such as TLS

Encryption helps protect prompts, responses, training data, model artifacts, and stored documents.

Multi-factor authentication

MFA helps protect administrator and user accounts from credential compromise.

VPC and private connectivity

Network controls can help restrict access to:

  • Model endpoints
  • Data stores
  • Internal applications
  • Enterprise systems

Logging and monitoring

Use AWS monitoring and logging capabilities to detect:

  • Unusual access
  • Failed requests
  • Data exfiltration
  • Unexpected model behavior
  • Excessive usage
  • Security incidents

AWS Nitro System

The AWS Nitro System uses specialized hardware and firmware to improve isolation and enforce security restrictions for EC2 instances.

Nitro-based instances include instances using:

  • GPUs
  • AWS Trainium
  • AWS Inferentia
  • Other specialized accelerators

The Nitro System helps protect customer workloads and data from unauthorized access by reducing the ability of the underlying host system to access customer instance memory and data.

Exam tip

Nitro is an infrastructure security and isolation technology. It is not a generative AI model, model evaluation service, or guardrail service.

Shared responsibility

AWS is responsible for security of the cloud, including:

  • Physical data centers
  • Underlying hardware
  • Core infrastructure
  • Managed service infrastructure

The customer is responsible for security in the cloud, including, depending on the service:

  • IAM permissions
  • Application code
  • Data classification
  • Prompt design
  • Network configuration
  • Encryption choices
  • Model and endpoint configuration
  • Output validation
  • User access controls

The exact division depends on whether the service is:

  • Infrastructure
  • Platform
  • Managed AI service
  • Fully managed application

Exam trap

Using a managed AWS AI service does not remove the customer’s responsibility for data governance, permissions, application security, and responsible use.

11. Generative AI-Specific Risks

Prompt injection

An attacker attempts to manipulate the model through malicious instructions in a prompt or retrieved document.

Potential consequences include:

  • Ignoring system instructions
  • Revealing confidential information
  • Calling unauthorized tools
  • Producing unsafe outputs

Controls include:

  • Input validation
  • Prompt and instruction separation
  • Retrieval filtering
  • Least-privilege tool permissions
  • Output validation
  • Human approval for sensitive actions

Data poisoning

Data poisoning occurs when malicious or incorrect data is added to training, fine-tuning, or retrieval sources.

Potential consequences include:

  • Biased results
  • Incorrect recommendations
  • Backdoors
  • Unreliable responses

Controls include:

  • Data provenance
  • Dataset review
  • Quality checks
  • Access control
  • Versioning
  • Monitoring
  • Validation before training or indexing

Model inversion

Model inversion attempts to infer sensitive information from model behavior or outputs.

Controls include:

  • Limiting sensitive data in training
  • Access control
  • Output filtering
  • Privacy-preserving techniques
  • Monitoring
  • Restricting overly detailed responses

Other common risks

  • Hallucinations
  • Bias
  • Toxic content
  • Intellectual property concerns
  • Sensitive data leakage
  • Excessive agency
  • Unauthorized actions
  • Model theft
  • Insecure tool use
  • Overreliance on model output

12. Cost Tradeoffs

Selecting a generative AI architecture involves more than comparing model prices.

Important factors include:

  • Responsiveness
  • Availability
  • Redundancy
  • Performance
  • Regional coverage
  • Token consumption
  • Provisioned capacity
  • Custom model requirements
  • Data transfer
  • Storage
  • Monitoring
  • Development and operational effort

Hosting your own model

When hosting a model yourself, you pay for:

  • Compute instances
  • GPUs or accelerators
  • Storage
  • Networking
  • Load balancing
  • Monitoring
  • Scaling
  • Security
  • Patching
  • Operations
  • Potential model license fees

Advantages

  • More control
  • Potentially lower unit cost at very high utilization
  • Ability to customize infrastructure
  • Greater control over model serving and deployment

Disadvantages

  • High up-front and ongoing infrastructure costs
  • Operational complexity
  • Capacity planning requirements
  • Hardware maintenance
  • Need for ML and infrastructure expertise
  • Risk of idle capacity

Managed model APIs

With services such as Amazon Bedrock, the customer can use a managed model through an API.

Advantages

  • Fast implementation
  • No need to manage model servers
  • Elastic scalability
  • Usage-based pricing
  • Access to multiple model providers
  • Reduced operational burden

Disadvantages

  • Less infrastructure control
  • Token costs can become significant at scale
  • Model and feature availability can vary by Region
  • API latency and quotas may affect performance
  • Some advanced customization options may not be available

Token-based pricing

A token is a unit of information processed by a model.

For text models, tokens may represent:

  • Parts of words
  • Words
  • Punctuation
  • Special symbols

Billing may distinguish between:

  • Input tokens
  • Output tokens

Some other generative AI modalities may use different units, such as:

  • Image inputs
  • Image outputs
  • Audio duration
  • Characters
  • Requests

Cost drivers

Token usage increases when:

  • Prompts are long
  • Conversation history is repeatedly included
  • Retrieved documents are large
  • Outputs are long
  • System instructions are verbose
  • Many agents or tools are called
  • Multiple model calls are made for one user request

Cost optimization techniques

  • Limit unnecessary prompt content
  • Reduce the context window
  • Retrieve only relevant documents
  • Limit maximum output tokens
  • Use smaller models for simpler tasks
  • Cache repeated results where appropriate
  • Route requests to different models based on complexity
  • Monitor usage with budgets and alerts
  • Avoid unnecessary agent loops

Exam trap

Token pricing is generally based on tokens processed, not simply the number of API calls. A single request with a large prompt and response can cost more than many small requests.

13. Responsiveness and Latency Tradeoffs

A larger or more capable model may provide better reasoning or quality, but it can have:

  • Higher latency
  • Higher cost
  • Greater compute requirements

A smaller model may provide:

  • Lower cost
  • Faster responses
  • Lower resource usage

But it may provide lower accuracy for complex tasks.

Possible design choices

  • Use a small model for classification or simple extraction.
  • Use a larger model for complex reasoning.
  • Use streaming responses to improve perceived responsiveness.
  • Use asynchronous processing for long-running jobs.
  • Place applications close to users and dependent services.
  • Use caching for repeated requests.

Exam tip

“Fastest” and “most accurate” are not always the same choice. Model selection is a tradeoff.

14. Availability, Redundancy, and Regional Coverage

Higher availability often requires additional architecture and cost.

Options may include:

  • Multiple Availability Zones
  • Multiple Regions
  • Failover architectures
  • Replicated data stores
  • Backup model providers
  • Queues and retry mechanisms
  • Cross-Region disaster recovery

Tradeoff

Multi-Region deployment can improve:

  • Disaster recovery
  • Regional resilience
  • User latency
  • Business continuity

But it may also increase:

  • Compute costs
  • Data replication costs
  • Operational complexity
  • Compliance concerns
  • Model availability challenges

Important point

Not every foundation model or Bedrock feature is available in every AWS Region. Regional coverage should be checked before designing the application.

15. Performance and Provisioned Throughput

On-demand capacity

Best when:

  • Traffic is unpredictable
  • The application is in development
  • Usage is low or intermittent
  • Cost flexibility is more important than guaranteed capacity

Provisioned Throughput

Best when:

  • Traffic is high and predictable
  • Consistent performance is required
  • Capacity needs to be reserved
  • The application has production-level throughput requirements

Provisioned capacity can reduce uncertainty around latency and throttling, but unused provisioned capacity may still incur charges.

16. Custom Models and Fine-Tuning Tradeoffs

Customization can improve performance for a specific use case but adds:

  • Training or fine-tuning cost
  • Data preparation cost
  • Evaluation requirements
  • Model storage costs
  • Deployment costs
  • Monitoring requirements
  • Governance complexity

Before fine-tuning, consider whether the same result can be achieved using:

  • Prompt engineering
  • Few-shot examples
  • RAG
  • Knowledge Bases
  • A different foundation model
  • A smaller specialized model

General decision sequence

  1. Start with prompt engineering.
  2. Add RAG if the model needs current or private knowledge.
  3. Evaluate different models.
  4. Fine-tune only if the use case justifies the extra cost and complexity.
  5. Train from scratch only for exceptional requirements.

17. RAG and Embeddings

Retrieval-Augmented Generation combines:

  1. A user query
  2. A retrieval system
  3. Relevant external information
  4. A foundation model that generates the answer

Documents are converted into embeddings.

Embeddings

An embedding is a numerical vector representation of data, such as:

  • Text
  • Images
  • Documents

Semantically similar content has similar vector representations.

Embeddings can be:

  • Stored
  • Indexed
  • Searched
  • Used to retrieve relevant context for an LLM

Benefits of RAG

  • Uses organization-specific information
  • Can use more current data
  • Avoids retraining for every data change
  • Can reduce hallucinations when retrieval is accurate
  • Supports citations or source references

Costs and tradeoffs

RAG introduces additional costs and complexity:

  • Embedding generation
  • Vector database storage
  • Indexing
  • Retrieval latency
  • Document processing
  • Context token costs
  • Data access controls
  • Security of retrieved content

Exam trap

RAG does not retrain or change the foundation model. It provides relevant information to the model during inference.

18. Exam Tips and Common Traps

High-value distinctions

  • Amazon Bedrock: Managed access to foundation models and generative AI capabilities through APIs.
  • SageMaker AI: Build, train, customize, deploy, and monitor ML models with greater control.
  • SageMaker JumpStart: Model hub and quick-start resources for pretrained models, algorithms, and solutions.
  • Amazon Q: Generative AI assistant for developers, employees, and business users.
  • Kiro: AI-assisted development environment.
  • Strands Agents: SDK/framework for building agents.
  • Bedrock AgentCore: Capabilities for running, securing, and operating agents at scale.
  • AWS Trainium: Training accelerator.
  • AWS Inferentia: Inference accelerator.
  • Nitro System: EC2 infrastructure security and isolation.
  • RAG: Retrieves external data at inference time; it is not model retraining.
  • Fine-tuning: Changes model behavior or weights using task-specific data.
  • Prompt engineering: Changes instructions without changing model weights.

Common incorrect assumptions

  1. “Bedrock requires customers to manage GPUs.”
    Generally false. Bedrock is managed and abstracts much of the underlying infrastructure.

  2. “SageMaker JumpStart models are free.”
    False. The model may be available through JumpStart, but compute, storage, and endpoints can incur costs.

  3. “Provisioned Throughput is always cheaper.”
    False. It is most useful for predictable, sustained workloads.

  4. “A managed AWS service removes all security responsibilities.”
    False. Customers still manage permissions, data, application security, and responsible use.

  5. “A larger model is always the best model.”
    False. Smaller models may be cheaper and faster and may be sufficient.

  6. “RAG fine-tunes the foundation model.”
    False. RAG provides retrieved context during inference.

  7. “Training from scratch is the normal approach.”
    False. Most organizations use pretrained models, prompting, RAG, or fine-tuning.

  8. “Multi-Region deployment is free.”
    False. It improves resilience but adds compute, replication, data transfer, and operational costs.

  9. “Token cost only depends on output length.”
    False. Input tokens and output tokens may both contribute to cost.

  10. “Agents only generate text.”
    False. Agents can invoke tools, access systems, and take actions, which introduces additional security and governance requirements.

Final Review Summary

For the exam, remember:

  • AWS offers multiple ways to build generative AI applications, from high-level assistants to highly customizable ML platforms.
  • Amazon Bedrock is the common choice for quickly accessing and experimenting with foundation models through APIs.
  • SageMaker AI is the choice when you need greater control over training, fine-tuning, deployment, and the ML lifecycle.
  • SageMaker JumpStart accelerates model discovery, experimentation, fine-tuning, and deployment.
  • Managed AWS services reduce the barrier to entry, improve speed to market, and reduce infrastructure management.
  • AWS infrastructure provides security, isolation, resilience, compliance support, and specialized AI hardware.
  • AI security must protect the input, model, and output.
  • Cost depends on model choice, token usage, latency, availability, redundancy, regional design, provisioned capacity, and customization.
  • The cheapest architecture is not always the fastest, most available, or most accurate architecture.
  • Start with the simplest solution—prompting or a managed model—and add RAG, fine-tuning, custom models, agents, or provisioned capacity only when justified by business requirements.