Discover how companies are responsibly integrating AI in production. This invite-only event in SF will explore the intersection of technology and business. Find out how you can attend here.
A year after Amazon Web Services (AWS) first announced the Amazon Bedrock generative AI effort, the platform is getting a major update that brings new models and improved security.
Amazon Bedrock is a platform that aims to make it easier for developers to build and deploy gen AI models. The service was first previewed in April 2023 and officially became generally available in September 2023. Amazon Bedrock provides users with easy access to a number of foundation models from leading vendors including Anthropic, Stability AI, Cohere, Meta and A121, as well as providing its own Amazon Titan models.
With today’s update, AWS is expanding the model lineup by enabling developers to bring their own models. There is also a new Amazon Titan embeddings model update that offers the promise of smaller embeddings that could help to accelerate vector database and retrieval augmented generation (RAG) use cases. Security is also a focus of the update, with the debut of a new guardrails capability. The new updates come as AWS continues to face pressure from cloud rivals including Google and Microsoft that are also expanding gen AI services.
“We’re looking at Amazon Bedrock as a place where people come and build applications using gen AI,” Vasi Philomin, VP and GM of Generative AI at AWS told VentureBeat.
VB Event
The AI Impact Tour – San Francisco
Join us as we navigate the complexities of responsibly integrating AI in business at the next stop of VB’s AI Impact Tour in San Francisco. Don’t miss out on the chance to gain insights from industry experts, network with like-minded innovators, and explore the future of GenAI with customer experiences and optimize business processes.
Bring your own model to Amazon Bedrock
While Amazon Bedrock already provides access to many of the most popular foundation models, the platform to date has not made it particularly easy for users to bring their own models.
The new Custom Model Import feature will now allow users to bring their own model to Amazon Bedrock. Philomin said that organizations in various industries including FinTech and healthcare might have started out with their own models for AI. Or they might have developed and customized models using the Amazon SageMaker platform which is focussed on data scientists.
“We allow them to bring those models into Amazon Bedrock and they can stand right next to all the models we already support in a completely serverless fashion,” Philomin explained. “They can compare the model and they can use it across multiple workflows on Amazon Bedrock to build their applications.”
As part of the Amazon Bedrock platform update, the Model Evaluation service is now also generally available. The service enables users to compare the performance of different models for various use cases. The model evaluation capability on Amazon Bedrock allows customers to evaluate models both automatically and through human evaluations. For automated evaluation, Philomin said that customers select models to compare and provide prompts reflective of their application. The system then provides various scores to compare model performance. For human evaluation, humans can rank model outputs side-by-side to determine which performs best for an application. Both types of evaluation help customers determine the best models for their specific needs.
Amazon Titan debuts Text Embeddings v2 and Titan Image Generator
A critical part of AWS’ overall gen AI strategy is the company’s own Amazon Titan family of models.
As part of today’s update, Amazon Titan Text Embedding V2 is making its debut. An embedding model is a foundational element for enabling a vector database search. Vector embeddings help to convert documents and other data into vectors that can then be stored and searched for RAG-based gen AI deployments. Among the most common text embedding libraries in use today are the OpenAI ada models. Amazon Titan offers an alternative that AWS claims can reduce the overall storage needed for vector embedding by up to four times while still retaining high accuracy.
Beyond embeddings, the Amazon Titan Image Generator which was previewed by AWS in November 2023, is now generally available. Philomin explained that, unlike a consumer-grade general-purpose image generator, Amazon Titan Image Generator was built to solve business use cases. The first core use cases are for the advertisement, media and entertainment sectors. The tool enables organizations to generate realistic studio-quality images.
AWS looking to keep gen AI on track, with guardrails
As part of its continuing efforts to help better secure gen AI for enterprises, AWS is also launching Guardrails for Amazon Bedrock.
The basic concepts behind guardrails are being deployed by multiple vendors in the industry, including Nvidia with its NeMo Guardrails technology. Philomin said that the Guardrails for Amazon Bedrock feature goes beyond what Nvidia is doing as it combines protections for both safety and privacy.
The Guardrails for Amazon Bedrock functionality allows customers to configure natural language policies to help ensure model safety across their generative AI applications. On the safety side, customers can set rules to prevent models from discussing harmful topics or generating toxic, hateful, or inappropriate content. Guardrails also help detect and handle personally identifiable information for privacy. By applying guardrails policies, AWS claims that customers can block up to 85% more harmful or objectionable content compared to relying on the models’ inherent safeguards alone.
Looking forward, Philomin said that there is still lots left for AWS to do to further improve and expand Amazon Bedrock.
“We have to continue to add more workflows that help people actually build applications and there’s a lot more to do there,” he said.
