AWS re:Invent 2024 GENAI Updates
1. Generative AI & Bedrock Enchancements
Automated Reasoning Checks:
Ensures factual accuracy by verifying model outputs mathematically—critical for high-stakes use cases like insurance claims.
Bedrock Agents with Multi-Agent Collaboration:
This new feature allows agents to work together on complex workflows, sharing insights and coordinating tasks seamlessly.
Supervisor Agents:
Enables Management of dozens (or possibly hundreds) of task-specific agents, deciding if tasks run sequentially or in parallel and resolving conflicts. For example: A global coffee chain analyzing new store locations. One agent analyzes economic factors, another local market dynamics, and a third financial projections. The supervisor agent ties everything together, ensuring optimal collaboration.
Intelligent Prompt Routing:
Optimizes model selection within a family of models based on the complexity of the prompt for quality and cost, potentially reducing costs by up to 30% without compromising accuracy.
Prompt Caching:
Allows reuse of frequent contexts, reducing costs by up to 90% and latency by up to 85% for supported models.
New API data ingestion
Introduces support for reranking models, custom connectors, and streaming data ingestion and , direct API for efficient data ingestion improving efficiency and real-time data availability.
RAG evaluation
Enables automatic assessment based on LLM-as-a-judge and optimization of Retrieval Augmented Generation applications using Amazon Bedrock Knowledge Bases.
2. Nova
AWS unveiled Nova, a new family of multimodal generative AI models designed for diverse applications in text, image, and video generation. Here’s what’s new:
a. Nova Text-Generating Models
Four Models:
- Micro: Text-only, low latency, fast response.
- Lite: Handles text, images, and video; reasonably quick.
- Pro: Balances speed, accuracy, and cost for multi-modal tasks.
- Premier(in training): Most advanced; ideal for complex workloads and custom model training. Capabilities: Context windows of up to 300,000 tokens (225,000 words); expanding to 2 million tokens in early 2025. Fine-tunable on AWS Bedrock for enterprise-specific needs. Use Cases: Summarizing documents, analyzing charts, and generating insights across text, image, and video.
b. Generative Media Models
Nova Canvas: Creates and edits images using text prompts. Offers control over styles, color schemes, and layouts.
Nova Reel: Generates six-second videos from prompts or reference images, with customizable camera motions like pans and 360° rotations. A two-minute video generation feature is coming soon.
3. Upcoming Features
Speech-to-Speech Model (Q1 2025):
- Transforms speech with natural human-like voice outputs.
- Interprets verbal and nonverbal cues like tone and cadence.
Any-to-Any Model (Mid-2025):
- Processes text, speech, images, or video inputs and generates outputs in any of these formats.
- Applications include translation, content editing, and AI assistants.