3.1.3. Foundation Models: Bedrock and JumpStart
💡 First Principle: Foundation models encode broad knowledge from massive datasets. Instead of training from scratch (expensive, data-hungry), you start with a foundation model and adapt it to your specific task—like hiring an experienced chef and teaching them your restaurant's recipes instead of training someone from culinary school.
Amazon Bedrock provides API access to foundation models from multiple providers (Anthropic, Meta, Amazon, Cohere, etc.) without managing infrastructure. You can use models directly via API for tasks like text generation, summarization, and code completion. For domain-specific needs, Bedrock supports fine-tuning with your data and continued pre-training to extend the model's knowledge.
SageMaker JumpStart provides pre-trained models that you can deploy or fine-tune within SageMaker Studio. It includes open-source models (Hugging Face, Stability AI) and solution templates for common ML tasks. Unlike Bedrock, JumpStart models run on SageMaker infrastructure that you manage.
| Feature | Bedrock | JumpStart |
|---|---|---|
| Hosting | Fully managed by AWS | You manage SageMaker endpoint |
| Models | Curated from multiple providers | Open-source + AWS models |
| Customization | Fine-tuning, continued pre-training | Fine-tuning, transfer learning |
| Cost model | Pay per API call (tokens) | Pay for compute (instances) |
| Best for | Quick API access, text/gen AI tasks | Full control, custom hosting, non-text models |
Choosing between Bedrock and JumpStart:
If the question says "quickly deploy a text generation model with no infrastructure management," the answer is Bedrock. If it says "fine-tune an open-source model and deploy with full control over the endpoint," the answer is JumpStart. The distinguishing signals are infrastructure control and model provider diversity.
⚠️ Exam Trap: Bedrock and JumpStart are not interchangeable. Bedrock is for generative AI workloads (text, conversation, code generation) with a serverless API model. JumpStart is for broader ML tasks (vision, tabular, NLP) where you need SageMaker's training and hosting infrastructure. A question about classifying medical images would point to JumpStart, not Bedrock.
Reflection Question: A company wants to add a customer support chatbot that understands their product documentation. Should they use Bedrock or JumpStart? What if they also need to classify support ticket images?
