Innovations in Amazon Bedrock at AWS re:Invent
Amazon Web Services, Inc. (AWS) introduced new advancements in Amazon Bedrock at AWS re:Invent, expanding model options and enhancing capabilities for building generative artificial intelligence (AI) applications tailored to individual businesses. Amazon Bedrock, a fully managed service, now offers a wider selection of industry-leading large language models and foundation models from AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon. These enhancements aim to simplify the development process while prioritizing privacy and security. By providing customers with a diverse range of models and tools, Amazon Bedrock empowers organizations to leverage generative AI for innovation and improved customer experiences.
Expanded Model Choices and Capabilities
The latest models from Anthropic, Cohere, Meta, and Stability AI, along with additions to the Amazon Titan family, offer customers a broader selection to cater to diverse use cases. Customers can now access models like Anthropic Claude 2.1, Meta Llama 2 70B, Cohere Command Light, Cohere Embed English, Cohere Embed multilingual, Meta Llama 2 13B, and Stability AI Stable Diffusion XL 1.0 through an API. Additionally, Amazon introduces Amazon Titan Image Generator and Amazon Titan Multimodal Embeddings to provide more flexibility in building generative AI applications with models that are pre-trained on extensive datasets.
- Anthropic Claude 2.1: This model offers improved accuracy over long documents, making it suitable for processing text-heavy content like financial statements and internal datasets. Claude 2.1 excels in tasks such as summarization, Q&A, and document contrasting, with enhanced honesty and decreased false statements.
- Meta Llama 2 70B: Trained on a vast dataset with double the context length of its predecessor, this model is optimized for dialog use cases through fine-tuning. It competes effectively in various benchmarks, offering a compelling mix of price and performance.
- Amazon Titan Image Generator: This model assists industries like advertising and ecommerce by generating realistic images or enhancing existing ones based on natural language prompts. It ensures accurate object composition and reduces harmful content, supporting rapid ideation and iteration.
- Amazon Titan Multimodal Embeddings: This model enables more accurate and contextually relevant search and recommendation experiences by converting images and text into embeddings stored in a vector database.
Efficient Model Evaluation
To help customers select the most suitable model for their requirements, Amazon Bedrock now offers Model Evaluation, available in preview. This feature allows customers to compare models based on predefined criteria or subjective evaluations. Automatic evaluations involve selecting criteria like accuracy and toxicity, while human-based evaluations focus on nuanced content requiring human judgment. By streamlining the evaluation process, customers can quickly identify the best models for their specific use cases.
Enhanced Model Customization
Amazon Bedrock introduces new customization capabilities to enable customers to personalize models with their proprietary data securely. With Knowledge Bases, organizations can augment model responses with contextual and relevant data from multiple sources, enhancing the accuracy and specificity of AI applications. Additionally, fine-tuning support for models like Cohere Command, Meta Llama 2, and Amazon Titan allows customers to adapt models to their business needs using labeled datasets, thereby improving model accuracy for specific tasks.
Streamlined Task Execution with Agents
The introduction of Agents for Amazon Bedrock enables generative AI applications to perform multistep tasks using company systems and data sources efficiently. By simplifying the setup process and automating interactions between models and systems, Agents enhance the accuracy and speed of developing generative AI applications, enabling organizations to deliver more sophisticated user experiences.
Implementing Safeguards with Guardrails
Guardrails for Amazon Bedrock empower customers to implement safeguards tailored to their applications and responsible AI policies. By defining denied topics and setting thresholds for harmful content, organizations can ensure that models respond appropriately and adhere to company guidelines. Guardrails provide consistency in managing interactions within generative AI applications, supporting a safe and relevant user experience.
In conclusion, the latest innovations in Amazon Bedrock at AWS re:Invent aim to democratize access to generative AI, empowering organizations to leverage cutting-edge models and capabilities to drive innovation and enhance customer experiences.
