AWS Announces More Model Choice and Powerful New Capabilities in Amazon Bedrock to Securely Build and Scale Generative AI Applications

 At AWS re:Invent, Amazon Web Services, Inc. (AWS), an Amazon.com, Inc. company (NASDAQ: AMZN), today announced Amazon Bedrock innovations that expand model choice and deliver powerful capabilities, making it easier for customers to build and scale generative artificial intelligence (AI) applications customized to their businesses. Amazon Bedrock is a fully managed service that offers easy access to a choice of industry-leading large language models and other foundation models from AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon, along with a broad set of capabilities that customers need to build generative AI applications—simplifying development while supporting privacy and security. These announcements further democratize access to generative AI by empowering customers with even more choice of industry-leading models and new capabilities to evaluate them, simplifying how they customize models with relevant and proprietary data, supplying tools to automate the execution of complex tasks, and equipping customers with safeguards to build and deploy applications responsibly. Together, these new additions to Amazon Bedrock transform how organizations of all sizes and across all industries can use generative AI to spark innovation and reinvent customer experiences. To get started with Amazon Bedrock, visit aws.amazon.com/bedrock.

“Generative AI is poised to be the most transformational technology of our time, and we are inspired by how customers are applying it to new opportunities and tackling business challenges,” said Dr. Swami Sivasubramanian, vice president of Data and AI at AWS. “As customers incorporate generative AI into their businesses, they turn to Amazon Bedrock for its choice of leading models, customization features, agent capabilities, and enterprise-grade security and privacy in a fully managed experience. With even more tools at their fingertips, customers are using Amazon Bedrock to leverage the full potential of generative AI to reimagine user experiences, reinvent their businesses, and accelerate their generative AI journeys.”

Organizations want to use generative AI for a wide variety of use cases—like generating productivity gains, driving innovative user experiences, and reimagining work—but generative AI is evolving rapidly, with new options and innovations happening daily. With so much fluidity in this space, customers’ ability to adapt is arguably the most valuable tool of all. Organizations need to be able to experiment, deploy, iterate, and pivot using the latest and greatest models available, and be ready to immediately embrace what comes tomorrow. To address these challenges, AWS developed Amazon Bedrock to make building with—and moving between—a range of models as easy as an API call, to put the latest techniques for model customization in the hands of all developers, and to keep customers secure and their data private. This is why customers such as Alida, Automation Anywhere, Blueshift, BMW Group, Clariant, Coinbase, Cox Automotive, dentsu, Druva, Genesys, Gilead, GoDaddy, Hellmann Worldwide Logistics, INRIX, KONE, LexisNexis Legal & Professional, Lonely Planet, NatWest, Nexxiot, OfferUp, Omnicom, the PGA TOUR, Proofpoint, Salesforce, Siemens, Takenaka Corporation, and Verint have turned to Amazon Bedrock to help them harness the power of generative AI for their organizations. Today’s announcement introduces new models and capabilities that will make it even easier for customers to build and scale generative AI applications.

The latest models from Anthropic, Cohere, Meta, and Stability AI, as well as additions to the Amazon Titan family, expand model choice for customers

No single model is ideal for every use case. Models vary across capabilities, price, and performance. Customers need easy access to a variety of model choices, so they can try out different models, switch between them, and combine the best models for their needs. With Amazon Bedrock, customers can drive rapid innovation with the latest versions of models, including the newly available Anthropic Claude 2.1 and Meta Llama 2 70B, and the recently available Cohere Command Light, Cohere Embed English, Cohere Embed multilingual, Meta Llama 2 13B, and Stability AI Stable Diffusion XL 1.0—all accessible via an API. In addition to Amazon Titan Text Embeddings and Amazon Titan Text models (now generally available), AWS is introducing Amazon Titan Image Generator and Amazon Titan Multimodal Embeddings to give customers even more choice and flexibility to build generative AI applications with models. Exclusive to Amazon Bedrock, Amazon Titan models are created and pre-trained by AWS on large and diverse datasets for a variety of use cases, with built-in support for the responsible use of AI. And Amazon indemnifies customers against claims that generally available Amazon Titan models or their outputs infringe on third-party copyrights.

• Anthropic’s Claude 2.1 in Amazon Bedrock: Anthropic, an AI safety and research company that builds reliable, interpretable, and steerable AI systems, has brought Claude 2.1, the latest version of their language model, to Amazon Bedrock. Claude 2.1 offers a 200K token context window and improved accuracy over long documents. Customers can now process text heavy documents like financial statements and internal datasets, and Claude 2.1 can summarize, perform Q&A, or contrast documents, and much more. Anthropic reports that Claude 2.1 has made significant gains in honesty with a 2x decrease in false statements compared to their previous model.
• Meta Llama 2 70B in Amazon Bedrock: Llama 2 is the next generation of language models by Meta. Llama 2 was trained on 40% more data than Llama 1 and has double the context length. The Llama 2 70 billion-parameter model is now available in Amazon Bedrock, in addition to the recently announced Llama 2 13 billion-parameter model. Built on top of the pre-trained Llama model, Llama 2 is optimized for dialog use cases through fine-tuning with instruction datasets and more than 1 million human annotations. The models perform competitively against multiple external benchmarks, including reasoning, coding, proficiency, and knowledge tests, and offer a compelling combination of price and performance in Amazon Bedrock.
• New Amazon Titan Image Generator available in preview: Amazon Titan Image Generator helps customers in industries like advertising, ecommerce, and media and entertainment produce studio-quality, realistic images or enhance existing images using natural language prompts, for rapid ideation and iteration on large volumes of images and at low cost. The model can understand complex prompts and generate relevant images with accurate object composition and limited distortions, reducing the generation of harmful content and mitigating the spread of misinformation. Customers can use the model in the Amazon Bedrock console either by submitting a natural language prompt to generate an image or by uploading an image for automatic editing, before configuring the dimensions and specifying the number of variations the model should generate. To edit, customers can isolate parts of an image to add or replace details (e.g., inserting a surfboard into a beach scene or replacing mountains with a forest in the background of a car advertisement), or they can extend an image’s borders with additional details in the same style as the original. Building on the commitments AWS made earlier this year at the White House, Amazon Titan applies an invisible watermark to all images it generates to help reduce the spread of misinformation by providing a discreet mechanism to identify AI-generated images and to promote the safe, secure, and transparent development of AI technology. AWS is among the first model providers to widely release built-in invisible watermarks that are integrated into the image outputs and are designed to be resistant to alterations.
• New Amazon Titan Multimodal Embeddings generally available: Amazon Titan Multimodal Embeddings helps customers power more accurate and contextually relevant multimodal search and recommendation experiences for end users. The model converts images and short text into embeddings—numerical representations that allow the model to easily understand semantic meanings and relationships among data— which are stored in a customer’s vector database. End users can submit search queries using any combination of image and text prompts. The model will generate embeddings for the search query and match them to the stored embeddings to produce more accurate and relevant search and recommendation results for end users. For example, a stock photography company with hundreds of millions of images can use the model to power its search functionality, so users can search for images using a phrase, image, or a combination of image and text (e.g., “show me images similar to the provided image, but with sunny skies”). By default, the model generates vectors that are well suited for search experiences that require a high degree of accuracy and speed. However, customers can also generate smaller dimensions to optimize for speed and performance. Amazon Titan Multimodal Embeddings joins the existing Amazon Titan Text Embeddings model, which is used to convert text input like single words, phrases, or even large documents into embeddings for use cases like search and personalization.

New capability helps customers efficiently evaluate, compare, and select the best model for their use case and business needs

Today, organizations have a wide range of model options to power their generative AI applications. To strike the right balance of accuracy and performance for their use case, organizations must efficiently compare models and find the best option based on their preferred metrics. To compare models, organizations must first spend days identifying benchmarks, setting up evaluation tools, and running assessments, all of which requires deep expertise in data science. Furthermore, these tests are not useful for evaluating subjective criteria (e.g., brand voice, relevance, and style) that requires judgment through tedious, time-intensive, human-review workflows. The time, expertise, and resources required for these comparisons—for every new use case—make it difficult for organizations to choose the optimal model for a task, limiting their use of generative AI.

Now available in preview, Model Evaluation on Amazon Bedrock helps customers evaluate, compare, and select the best models for their specific use case, using either automatic or human evaluations. In the Amazon Bedrock console, customers choose the models they want to compare for a given task, such as question-answering or content summarization. For automatic evaluations, customers select predefined evaluation criteria (e.g., accuracy, robustness, and toxicity) and upload their own testing dataset or select from built-in, publicly available datasets. For subjective criteria or nuanced content requiring sophisticated judgment, customers can easily set up human-based evaluation workflows with just a few clicks. These workflows leverage a customer’s in-house workforce, or use a managed workforce provided by AWS, to evaluate model responses. During human-based evaluations, customers define use case-specific metrics (e.g., relevance, style, and brand voice). Once customers finish the setup process, Amazon Bedrock runs evaluations and generates a report, so customers can easily understand how the model performed across key criteria and can make optimal tradeoffs and quickly select the best models for their use cases.

New model customization capabilities help customers make the most of their data, privately and securely, on AWS

Organizations want to maximize the value of their rich data sources to deliver remarkable user experiences—at scale—that are uniquely customized to reflect the company’s style, voice, and services. New, purpose-built capabilities available in Amazon Bedrock help customers personalize models privately and securely with their own data to build differentiated generative AI-powered applications.

• Knowledge Bases for Amazon Bedrock customizes model responses with contextual and relevant company data: Organizations want to supplement existing models with proprietary data to create more relevant and accurate responses. To equip the model with up-to-date information, organizations turn to retrieval augmented generation (RAG), a technique that allows customers to customize a model’s responses by augmenting prompts with data from multiple sources, including document repositories, databases, and APIs. Now generally available, Knowledge Bases for Amazon Bedrock securely connects models to proprietary data sources for RAG to deliver more accurate, context-specific responses for use cases like chatbots and question-answering systems. Knowledge bases are fully managed, so customers simply point to the location of their data. Then knowledge bases fetch the text documents and save the data to a vector database or set one up on the customer’s behalf. When a user query comes in, Amazon Bedrock orchestrates RAG by fetching text needed to augment a prompt, sending the prompt to the model, and returning the response. Knowledge Bases for Amazon Bedrock supports databases with vector capabilities, including Amazon OpenSearch, and other popular databases like Pinecone and Redis Enterprise Cloud, with Amazon Aurora and MongoDB coming soon.
• Cohere Command, Meta Llama 2, and Amazon Titan models can now be fine-tuned in Amazon Bedrock, with support for Anthropic’s Claude 2 coming soon: In addition to RAG, organizations can also leverage fine-tuning to further train the model on a specific task (e.g., text generation), using labeled datasets to adapt the model’s parameters to their business, and extending its knowledge with the lexicon and terminology used by the organization and end users. For example, a retail customer could fine-tune a model on a dataset of product descriptions to help it understand the brand style and produce more accurate descriptions for the website. Amazon Bedrock now supports fully managed fine-tuning for Cohere Command and Meta Llama 2, along with Amazon Titan Text Express, Amazon Titan Text Lite, Amazon Titan Multimodal Embeddings, and Amazon Titan Image Generator (in preview), so customers can use labeled datasets to increase model accuracy for specific tasks. Additionally, AWS customers will soon be able to fine-tune Claude 2’s performance with their data sources. To fine-tune a model, customers start by selecting the model and using Amazon Bedrock to make a copy. Customers then point to labeled examples in Amazon Simple Storage Service (Amazon S3). Amazon Bedrock incrementally trains the model (augments the copied model with the new information) on these examples, and the result is a private, more accurate fine-tuned model that delivers more relevant, customized responses. Customer data is encrypted in transit and at rest, so all valuable customer data remains secure and private. AWS and third-party model providers will not use any inputs or outputs from Amazon Bedrock to train their base models.

With Agents for Amazon Bedrock, generative AI applications can help execute multistep tasks using company systems and data sources

While models are effective at conversing and creating new content, they deliver more value if equipped to take actions, solve problems, and interact with a range of systems to complete multistep tasks (e.g., booking travel or ordering replacement parts). However, this requires custom integrations to connect models with company data sources, APIs, and internal and external systems. Developers must write code to orchestrate the interactions between models, systems, and the user, so the application can execute a series of API calls in a logical order. To connect the model with data sources, developers must implement RAG, so the model can customize its responses to the task. Finally, developers must provision and manage the requisite infrastructure, as well as establish policies for data security and privacy. These steps are time-consuming and require expertise, slowing the development of generative AI applications.

Now generally available, fully managed Agents for Amazon Bedrock enables generative AI applications to execute multistep tasks using company systems and data sources. Agents can plan and perform most business tasks, such as answering questions about product availability or taking orders. Customers can create an agent using a simple setup process, first selecting the desired model, writing a few instructions in natural language (e.g., “you are a cheerful customer service agent” and “check product availability in the inventory system”) and providing access to the company’s enterprise systems and knowledge bases. Agents automatically analyze the request and break it down into a logical sequence, using the model’s reasoning capabilities to determine the information needed. The agent then takes action by identifying the APIs to call and deciding when to call them to fulfill the request. Agents also retrieve needed information from proprietary data sources to provide accurate and relevant responses. Agents securely and privately perform this process in the background each time, relieving customers from having to engineer prompts, manage the session context, or orchestrate systems manually. With Agents for Amazon Bedrock, customers can improve the accuracy and speed of development of their generative AI applications.

With Guardrails for Amazon Bedrock, customers can implement safeguards across models based on application requirements and responsible AI policies

Organizations recognize the need to manage interactions within generative AI applications for a relevant and safe user experience. While many models use built-in controls to filter undesirable and harmful content, organizations want to further customize interactions to remain on topics relevant to their business, align with company policies, and adhere to responsible AI principles. For example, a bank might want to configure its online assistant to refrain from providing investment advice, avoid queries about competitors, and limit harmful content. As another example, after a customer service call, personally identifiable information (PII) may need to be redacted from the call summary. Organizations may need to change models, use multiple models, or replicate policies across applications, and they want a simple way to consistently deploy their preferences across all these areas simultaneously. Deep expertise is required to build custom protection systems with these kinds of safeguards and integrate them into applications, and the processes can take months. Organizations want a streamlined way to enforce key policies and rules in generative AI applications to deliver relevant user experiences and support safer use of the technology.

Now available in preview, Guardrails for Amazon Bedrock empowers customers to implement safeguards for generative AI applications that are customized to their use cases and responsible AI principles, enhancing the safety and privacy of user interactions. Guardrails drive consistency in how models in Amazon Bedrock respond to undesirable and harmful content within applications. Customers can apply guardrails to all large language models in Amazon Bedrock, as well as to fine-tuned models and in combination with Agents for Amazon Bedrock. To create a guardrail in the Amazon Bedrock console, customers start with natural language descriptions to define the denied topics within the context of their application. Customers can also configure thresholds across hate speech, insults, sexualized language, and violence to filter out harmful content to their desired level. In early 2024, customers will also be able to redact PII in models’ responses, set profanity filters, and provide a list of custom words to block interactions between users and models. Guardrails automatically evaluate both user queries and model responses to detect and help prevent content that falls into restricted categories. Customers can create multiple guardrails to support different use cases and apply the same guardrails across multiple models. Guardrails for Amazon Bedrock empowers customers to innovate safely by providing a consistent user experience and standardizing safety and privacy controls across generative AI applications.

Blueshift provides brands with marketing automation and customer data platforms to deliver personalized, customer engagement across all communication channels and devices. “Product catalogs are rapidly evolving with new content being changed every minute, and we need to continuously update our embeddings to ensure recommendations for brand audiences remain relevant,” said Manyam Mallela, co-founder and chief AI officer at Blueshift. “Amazon Titan Multimodal Embeddings in Amazon Bedrock is outperforming older models from other providers that we used, offering more nuanced and contextually relevant recommendations without complex feature engineering. Our team has seen a 10% performance improvement using Amazon Titan Multimodal Embeddings. With the robust infrastructure, security, and collaboration offered by AWS, Blueshift is poised to seamlessly integrate cutting-edge embeddings, ensuring that our recommendation solutions remain state-of-the-art and lead to improved audience engagement.”

Dentsu is one of the world’s largest providers of integrated marketing and technology services. “We work at the convergence of marketing, technology, and consulting to drive people-centered transformations for brands that want to shape society for the better, and generative AI is changing our ability to deliver at scale and speed for clients, augmenting, not replacing, our 72,000-strong team around the world,” said Brian Klochkoff, executive vice president of Innovation & Emerging Technologies at dentsu. “Specifically, Amazon Bedrock gives us the enterprise control and ease-of-use to deploy third-party models for decentralized usage across our product and engineering teams. This allows our teams to innovate with the latest and greatest generative AI advancements in a safe and responsible space, while inventing cutting-edge opportunities for clients.”

Druva enables cyber, data, and operational resilience for every organization with their leading, at-scale software-as-a-service (SaaS) solution. “It takes time and effort to manually build architecture that incorporates real-time and regularly changing data into applications, so we wanted a way to automate this process,” said David Gildea, vice president of Product, Generative AI at Druva. “We built our new service Dru—an AI co-pilot that IT teams can use to access company information and perform actions in natural language—in Amazon Bedrock because it provides fully managed and secure access to an array of foundation models. Next, we plan to integrate agents and knowledge bases into Dru to simply implement and automate RAG and improve end-user experiences, which we expect will lead to 70% more accurate responses, 50% faster response times, and 50% lower costs compared to similar models on other platforms. With Agents and Knowledge Bases for Amazon Bedrock, we can add new capabilities, like support ticket creation, with increased velocity and without having to re-engineer Dru—enabling us to deliver the latest generative AI-powered solutions to our customers.”

GoDaddy is a leading domain registrar, commerce, and web hosting company serving more than 20 million customers. “Everyday entrepreneurs—from bakers to plumbers—use GoDaddy to establish their businesses online, and we are heavily leaning into generative AI as it can help these entrepreneurs compete with and become the next big brand,” said Travis Muhlestein, chief data and analytics officer at GoDaddy. “We have long worked with AWS because of their dedication to customers, history of democratizing cloud services, and current mission of making generative AI accessible to businesses of all sizes. Amazon Bedrock plays a critical role in our AI platform, and we plan to use Amazon Titan Text Express to accelerate internal development patterns. We are especially excited to test Amazon Titan Image Generator because it has the potential to help small businesses create professional marketing assets that solidify their brand identities and can propel their businesses exponentially.”

INRIX is a global leader in connected car services and transportation analytics. “The insights we provide our customers are informed by a vast amount of data, so it is critical that we pursue the most efficient and effective generative AI solutions to drive data analysis through our new causality platform,” said Bryan Mistele, chief executive officer at INRIX. “Amazon Bedrock’s model evaluation capability simplifies the process of identifying the right foundation model for our needs by aggregating key metrics, expediting our selection from days to minutes. For content needing sophisticated judgement, we can use our own employees at scale to evaluate the model’s responses and ensure they accurately reflect our domain and the brand voice of our chatbot application. With a visual and comprehensive resulting report, not only does the capability boost our confidence that we have selected the right model for the task, but it helps us focus instead on delivering innovative services to move people, cities, and businesses forward.”

MongoDB’s mission is to empower innovators to create, transform, and disrupt industries by unleashing the power of software and data. “More and more customers across industries want to take advantage of generative AI to build next-generation applications, but many of them are concerned about data privacy and ensuring the accuracy of the outputs from AI-powered systems,” said Sahir Azam, chief product officer at MongoDB. “To meet customers where they are, we made MongoDB Atlas available as a knowledge base for Amazon Bedrock, so our joint customers can securely build generative AI applications with their operational data to create personalized experiences with the trust and accuracy their end users expect. Through this integration, customers can access industry-leading foundation models and use their data processed by MongoDB Atlas Vector Search to create applications that deliver more relevant outputs with the right context. Leveraging the data privacy best practices built into Knowledge Bases for Amazon Bedrock, customers can spend less time on the operational overhead of generative AI and focus more on deploying the technology to provide highly engaging end-user experiences on AWS.”

OfferUp is one of the largest mobile marketplaces for local buyers and sellers in the U.S. and is changing the way people transact in their communities by providing a uniquely simple and trusted experience. “With millions of listings posted daily, it is important that we continually improve our personalized search and recommendation experiences for our users,” said Melissa Binde, chief technology officer at OfferUp. “To achieve this goal, we are experimenting with Amazon Titan Multimodal Embeddings, with the aim of revolutionizing local commerce through cutting-edge semantic search capabilities. During an initial evaluation with the new multimodal model, we have observed substantial improvement in relevance recall for keyword searches. This advancement will significantly expedite successful matches, benefiting both our buyers and sellers.”

Salesforce is a leading AI customer relationship management (CRM) platform, driving productivity and trusted customer experiences powered by AI, CRM, and data. “AI is an integral part of our commitment to help companies connect with their customers in new and personalized ways,” said Kaushal Kurapati, senior vice president of Product at Salesforce. “Amazon Bedrock is a key component of our open ecosystem approach to models, seamlessly grounding models with customer data, and integrating them into workflows across Salesforce. And with the addition of new evaluation capabilities, comparing foundation models across multiple criteria—including qualitative aspects such as friendliness, style, and brand relevance—will make operationalizing models easier and faster than ever.”

SmartBots AI helps enterprises build their own conversational chatbots to deliver elevated customer support and sales experiences. “It is critical that our customers can trust the responses of the conversational chatbots they are building with SmartBots AI, as these responses represent their brands and will be used by their employees and clients,” said Jaya Prakash Kommu, co-founder and chief technology officer at SmartBots AI. “Because our chatbot development service is powered by Amazon Bedrock, customers’ chatbots inherit AWS data safety and privacy best practices. Their chatbots will also have access to Guardrails for Amazon Bedrock to set, control, and avoid inappropriate and unwanted content in user prompts and responses that their chatbots generate, so users have safe and brand-adhering experiences. Our customers are excited for this safety feature to build and use generative AI responsibly.”

TTEC Digital uses AI to improve customer experiences through cutting-edge contact center technologies. “Our clients are always looking for ways to shorten contact center response times for their customer interactions, so we are building a generative AI-powered solution using Amazon Bedrock to partially automate information gathering and response generation,” said Dave Seybold, chief executive officer at TTEC Digital. “We will soon be deploying Amazon Titan Text Express within our clients’ enterprise contact center environments to enable rapid retrieval of customer data through simple prompting, helping their contact center representatives provide fast, more informed responses. We predict this solution will increase customer satisfaction, decrease average call times, reduce post-call workload for representatives, and lower labor costs. With Amazon Titan Text Express, our clients’ contact center representatives will be empowered to address a broader range of customer issues and extend the transformative advantage of generative AI to their customers.”

About Amazon Web Services

Since 2006, Amazon Web Services has been the world’s most comprehensive and broadly adopted cloud. AWS has been continually expanding its services to support virtually any workload, and it now has more than 240 fully featured services for compute, storage, databases, networking, analytics, machine learning and artificial intelligence (AI), Internet of Things (IoT), mobile, security, hybrid, virtual and augmented reality (VR and AR), media, and application development, deployment, and management from 102 Availability Zones within 32 geographic regions, with announced plans for 15 more Availability Zones and five more AWS Regions in Canada, Germany, Malaysia, New Zealand, and Thailand. Millions of customers—including the fastest-growing startups, largest enterprises, and leading government agencies—trust AWS to power their infrastructure, become more agile, and lower costs. To learn more about AWS, visit aws.amazon.com.

About Amazon

Amazon is guided by four principles: customer obsession rather than competitor focus, passion for invention, commitment to operational excellence, and long-term thinking. Amazon strives to be Earth’s Most Customer-Centric Company, Earth’s Best Employer, and Earth’s Safest Place to Work. Customer reviews, 1-Click shopping, personalized recommendations, Prime, Fulfillment by Amazon, AWS, Kindle Direct Publishing, Kindle, Career Choice, Fire tablets, Fire TV, Amazon Echo, Alexa, Just Walk Out technology, Amazon Studios, and The Climate Pledge are some of the things pioneered by Amazon. For more information, visit amazon.com/about and follow @AmazonNews.