Amazon Bedrock vs SageMaker AI: which one, and why

A company wants to quickly deploy a popular open-source pre-trained model into its own AWS account, where it can control the instances and fine-tune the model with its own data. Which AWS option BEST meets this requirement?

• AAmazon SageMaker Ground Truth
  Ground Truth creates labeled datasets with human annotators. It does not deploy or fine-tune open-source models.
• BAmazon SageMaker JumpStart
  Correct. SageMaker JumpStart is a model hub with pre-trained and open-source models that can be deployed to SageMaker AI endpoints in your account and fine-tuned.
• CAWS Glue DataBrew
  DataBrew prepares data. It does not host models.
• DAmazon Bedrock on-demand model access
  Bedrock offers many popular models, so it is tempting. Its models are served as a managed API; the company would not control the instances hosting the model, which this requirement demands.

A mid-sized insurer wants to add text summarization to its claims portal using several foundation models from different providers. It has no ML team and does not want to manage servers or GPUs. Which AWS service BEST meets these needs?

• AAmazon Bedrock
  Correct. Bedrock offers many providers' FMs through one API with no servers or GPUs to manage.
• BAmazon EC2 with GPU instances
  The insurer would have to deploy, patch, and scale models itself, which it wants to avoid.
• CAmazon SageMaker JumpStart
  JumpStart offers many pre-trained models, so it is tempting. It deploys models to endpoints on instances the customer chooses and manages, which the insurer wants to avoid; Bedrock is serverless.
• DAmazon Comprehend
  Comprehend offers pre-trained NLP such as entity and sentiment detection, not a choice of FMs for summarization.

A research lab has fine-tuned an open-source model and must control the exact instance types, use its own inference container, and keep the model weights in its own account. Which production approach BEST meets these requirements?

• ACall a general FM through Amazon Bedrock on-demand pricing
  On-demand Bedrock models are managed by AWS and the provider, so the lab cannot choose instances or bring a custom inference container.
• BImport the fine-tuned weights with Amazon Bedrock Custom Model Import
  Custom Model Import does host your own weights, so it is tempting. It is serverless: the lab cannot choose instance types or run its own inference container, which are explicit requirements.
• CSelf-host the model on a SageMaker AI endpoint the lab configures
  Correct. Self-hosting on SageMaker AI lets the lab pick instance types, bring its own container, and deploy its own weights while AWS manages the underlying hosting.
• DBuy Amazon Bedrock Provisioned Throughput for a base model
  Provisioned Throughput reserves dedicated capacity, so it sounds like control. The lab still would not pick instance types or containers, and a base model is not the lab's own fine-tuned open-source model.

A research company needs full control over training a custom model on its own data, including choosing algorithms, managing training jobs, and hosting on dedicated endpoints. Which AWS service is designed for this?

• AAmazon SageMaker Canvas
  Canvas builds custom models from company data, so it is tempting. It is a no-code tool that automates algorithm choice, so it does not give full control over algorithms and training jobs.
• BAmazon Bedrock model customization
  Bedrock can fine-tune models on company data, so it is tempting. It adapts existing foundation models in a managed way; it does not let the company choose algorithms or manage training jobs and endpoints.
• CAmazon Comprehend custom classification
  Comprehend trains custom models on your data, so it is tempting. It only supports specific text tasks and hides the algorithm and infrastructure, the opposite of full control.
• DAmazon SageMaker AI
  Correct. SageMaker AI supports the full ML lifecycle with control over training jobs and endpoints.