AWS re:Invent 2025 - Customize & scale foundation models using Amazon SageMaker AI (AIM363)

Customizing and Scaling Foundation Models with Amazon SageMaker AI

Challenges with Foundation Models in Production

• Demos of foundation models look impressive, but production requirements are much more demanding
• Production needs:
  • Accurate models that can be defended
  • Cost-effective economics that work at scale
• Teams often face a false choice between accuracy and cost when using foundation models

The Power of Model Customization

• Foundation models are powerful generalists, but lack domain-specific knowledge
• Customizing models with your data and business rules can significantly improve accuracy and reduce costs
• Customization encodes your domain expertise directly into the model's weights

Overcoming the Customization Tax

• Customizing models has traditionally been complex and time-consuming
• SageMaker AI solves this with three key principles:
  1. Serverless infrastructure for training and deployment
  2. Integrated workflow from experimentation to production
  3. Production-ready customization techniques

Customization Techniques in SageMaker AI

1. Supervised Fine-Tuning (SFT): Teaches the model your domain knowledge through labeled examples
2. Direct Preference Optimization (DPO): Teaches the model your judgment and style by showing it preferred responses
3. Reinforcement Learning with AI Feedback (RLaif): Trains the model to achieve complex, subjective outcomes through feedback

Customization Workflows in SageMaker Studio

• Three customization experiences in SageMaker Studio:
  1. Visual UI-based workflows
  2. AI agent-guided experiences
  3. Code-based SDK for programmatic control
• All workflows use the same underlying service and provide a unified audit trail

New SageMaker AI Capabilities

1. Serverless Model Evaluation: Easily evaluate customized models against benchmarks or custom criteria
2. Serverless MLflow: Automatically track metrics, hyperparameters, and lineage for all model customization experiments

Compute Options

1. Serverless Training: Rapid iteration and experimentation without infrastructure management
2. SageMaker Hyperod: Persistent cluster for long-running, frontier-scale model training with fault tolerance

Key Takeaways

1. Choice: Flexibility to use open-source or proprietary models, and choose from multiple customization techniques
2. Efficiency: Serverless infrastructure and integrated workflows dramatically reduce the "customization tax"
3. Safety: Built-in governance, lineage, and security for enterprise-grade model customization

Business Impact

• Customization allows you to turn your proprietary data into a competitive advantage
• Improved accuracy and reduced costs enable new AI-powered use cases in production
• Examples:
  • Transaction categorization with 92% accuracy instead of 8% miscategorization
  • Conversational AI that sounds more human-like and aligned with your brand