AWS Re:Invent Keynote Announcements (2024)
After watching AWS’s 2.5-hour re:Invent keynote, it powerfully reinforces what I’ve been telling my customers and peers: AI success requires so much more than just LLMs and prompt engineering.
A key theme was the concept of building blocks—the idea that complex applications can be broken down into simpler components. This approach has allowed AWS to develop hundreds of services that customers can combine in innovative ways, and this really fits well in the era of generative AI which is revolutionizing industries and where startups are at the forefront of this disruption.
The announcements demonstrate that organizations need a complete ecosystem – from specialized hardware to robust data management to development tools – to truly leverage AI at scale. AWS has been building the full stack of technology required to make AI work in real business environments, and the latest improvements do not disappoint at all!
Sure they introduced Amazon’s new Nova family of foundational models (LLMS), but I think a jewel was AWS’s Automated Reasoning capability in Bedrock, which addresses the critical challenge of hallucinations in AI models. It prevents factual errors and hallucinations in model outputs by mathematically verifying statements against source documentation. Crazy and much needed!
As someone who works with different organizations implementing AI, I’ve seen firsthand how companies often fixate on LLMs while overlooking critical infrastructure needs. This keynote showcases exactly what I’ve been advocating – you need the entire foundation to make AI work effectively.
Announcement list
Amazon Bedrock Updates:
- Model Distillation: New capability to create smaller, faster, cheaper models that run 500% faster and cost 75% less
- Automated Reasoning Checks: New feature to prevent hallucinations and ensure factual accuracy
- Multi-Agent Collaboration: Enhanced support for complex workflows using multiple autonomous agents
Amazon Q Updates:
- Q Developer: New autonomous agents for generating unit tests, documentation, and code reviews
- Q Business: Integration with QuickSight and support for third-party ISV access to Q index
- Q Business Automate: New capability to automate complex workflows across multiple applications
Amazon Nova Models:
- Text Models: Nova Micro (text-only)
- Multimodal Models: Nova Lite, Pro, and Premier (coming Q1 2024)
- Nova Canvas: New image generation model
- Nova Reel: New video generation model for 6-second clips (expanding to 2 minutes)
- Coming in 2024: Speech-to-speech model (Q1) and multimodal-to-multimodal model (mid-year)
Amazon SageMaker Updates:
- Next-generation SageMaker with integrated analytics capabilities
- SageMaker Unified Studio: New single development environment for data and AI
- SageMaker Lakehouse: New Apache Iceberg compatible data lake
- Zero-ETL for Applications: New capability for analyzing data across third-party SaaS applications
Training Infrastructure:
- Trainium 2: General availability of AWS’s custom AI chip
- Trainium 3: Announced for late 2024, offering 2x more compute and 40% better efficiency
- Project Rainier: Massive Trainium 2 cluster being built with Anthropic
Content based on category with timestamps:
| Category | Product/Service | Timestamp |
|---|---|---|
| Compute & Infrastructure | ||
| EC2 | 13:20 | |
| Graviton4 | 17:50 | |
| P6 instances | 20:38 | |
| Trainium2 GA | 23:23 | |
| Trainium2 Ultra | 25:50 | |
| Trainium3 | 34:38 | |
| Elastic VMware Service | 2:08:06 | |
| Storage | ||
| S3 Table Buckets | 42:33 | |
| S3 Metadata | 45:51 | |
| Databases | ||
| Aurora D SQL | 59:56 | |
| DynamoDB Global Tables Updates | 1:01:11 | |
| AI/ML Services | ||
| Nova Models | 1:48:20 | |
| Nova Canvas | 1:53:45 | |
| Nova Reel | 1:55:12 | |
| Q Developer autonomous agents | 2:01:46 | |
| Q Business with QuickSight | 2:25:18 | |
| Q Business APIs | 2:27:34 | |
| Q Business Automate | 2:31:47 | |
| SageMaker Next Generation | 2:39:04 | |
| SageMaker Unified Studio | 2:39:39 | |
| SageMaker Lakehouse | 2:42:02 | |
| Data Management | ||
| Zero-ETL for Applications | 2:41:16 |
