On the second day of AWS re:Invent 2024, Dr. Swami Sivasubramanian, vice president of AI and data at AWS, announced a series of new advancements for Amazon Bedrock, the company’s platform that enables businesses to create generative AI applications. On day one, we saw AWS CEO Matt Garman and Amazon CEO Andy Jassy demonstrate new models and features. On Thursday, December 5, Sivasubramanian introduced new model capabilities with an overview of how AWS is affecting change in their respective industries through generative AI.
The new upgrades aim to provide greater flexibility and control to build and deploy generative AI applications faster and more efficiently. All of the announcements show AWS’s commitment to model choice and optimizing how inference is performed at scale. In his insightful speech, Sivasubramanian asserted that AWS, with its revolutionary technology, was not only shaping the present but also setting the stage for future innovations to take flight.
Here’s a look at key moments from the opening speech:
What’s new with Amazon Bedrock?
During his keynote speech, Dr. Sivasubramanian introduced some new features for Amazon Bedrock. The latest updates include expanded template options, access to more than 100 specialized templates through the Amazon Bedrock Marketplace, improved prompt management tools, and new features for Knowledge Bases (an online library of information self-service) and data automation. Dr Sivasubramanian said these features aim to provide flexibility, extend inference and maximize data usage. While other features are in preview, Amazon Bedrock Marketplace is live. The AWS official also said that models of Luma AI, Poolside and Stability AI will be added to Amazon Bedrock soon.
New AI features in Amazon SageMaker
The AWS AI model creation and deployment service, Amazon SageMaker, has benefited from four innovations aimed at making the development of generative AI and machine learning cost-effective, faster, and easier to scale. The new innovations aim to help businesses get started instantly with popular models, optimize their training processes, and integrate seamlessly with partners’ AI tools. Features include curated training recipes, flexible training plans, task governance, and integrated partner AI applications.
These advances mean customers will benefit from faster, more affordable AI solutions. Businesses can now expect more personalized, efficient and innovative experiences, like smarter chatbots, faster recommendations and improved automation of everyday tasks.
Bedrock Amazon Marketplace
One of the big announcements on day two was the new Amazon Bedrock Marketplace. This is a site that provides access to over 100 popular and specialized AI models, including Mistral NeMo, Falcon RW, and more. Users can choose templates tailored to their needs, deploy them on scalable AWS infrastructure via fully managed endpoints, and integrate them securely. using Bedrock APIs. It also includes guardrails, agents, and strong security and privacy protections. According to AWS, this marketplace simplifies model discovery, deployment, and integration.
SageMaker HyperPod gets new features
To meet the growing demands of AI, AWS announced new features for SageMaker HyperPod. These included flexible training plans to streamline capacity bookings, saving weeks of training time and enabling working within budgets and deadlines. On the other hand, task governance in SageMaker HyperPod automates the management and prioritization of compute resources, maximizing utility and completing high-priority tasks efficiently. SageMaker also integrates AI applications from partners such as Comet and Fiddler, reducing the time spent configuring third-party tools and accelerating the model development lifecycle. These innovations aim to improve resource efficiency, reduce development complexity, and improve the speed of AI deployment for customers.
Advanced AI Generation Improvements on Bedrock
Dr. Sivasubramanian, in his keynote speech, also presented a series of innovations that simplify and optimize the development of generative AI. While prompt caching facilitates context reuse in API calls, intelligent prompt routing improves response quality and cost efficiency by directing requests to the best-fit AI model. To address the challenges of augmented recovery generation (RAG), AWS introduced Kendra Gen AI Index, which integrates with more than 40 enterprise data sources for accurate and engaging outputs.
On the other hand, Bedrock Knowledge Bases now enable structured data retrieval and use of knowledge graphs through GraphRAG support, enabling richer, more accurate answers in Gen AI applications. AWS also introduced Bedrock Data Automation, which processes unstructured multimodal data to improve AI generation insights.
For security reasons and to ensure ethical use, AWS introduced Bedrock Guardrails which provide customizable protections and automated reasoning checks. Meanwhile, the new multimodal toxicity detection filters harmful image content.
The author is at AWS re:Invent 2024 in Las Vegas at the invitation of AWS.