In the ever-evolving landscape of generative AI, Amazon has thrown its hat into the ring with the introduction of Amazon Bedrock, a foundation model service that promises to simplify the development and scaling of generative AI applications. As part of a limited preview, we had the opportunity to explore Bedrock and share our first impressions with you.
The Bedrock Concept
Amazon Bedrock is positioned as a part of the broader initiative to enable Generative AI on AWS. It functions as a service that provides access to various foundation models through a straightforward API, eliminating the need for users to manage the underlying infrastructure and maintenance tasks associated with running these models. In essence, Bedrock is Foundation-Models-as-a-Service, designed to streamline the process of building and scaling generative AI-based applications.
Bedrock vs. Sagemaker Jumpstart
You might be wondering how Bedrock compares to Amazon’s Sagemaker Jumpstart. While Sagemaker Jumpstart offers a range of pretrained models that can be fine-tuned or deployed on demand, Bedrock takes a step further by offering larger models as a managed service. This simplifies the utilization of foundation models to a simple API request.
Supported Models on Bedrock
As of now, Amazon Bedrock supports NLP and Vision Foundation models. The initial lineup includes models from notable players in the generative AI space:
Anthropic’s Claude Family
- Claude: With a context size of 12K tokens, suitable for handling sophisticated dialogue and creative content generation tasks.
- Claude Instant: With 9K tokens for faster performance and more straightforward language tasks.
Anthropic’s Claude models emphasize excellent performance, steerability, and safety through the Constitutional AI approach.
AI21’s Jurassic Family
Two models from AI21 Labs’ Jurassic family are accessible through Bedrock:
- J2 Grande Instruct
- J2 Jumbo Instruct
The Jurassic models offer multilingual capabilities, supporting languages such as English, Spanish, French, German, Italian, Portuguese, and Dutch.
Amazon’s Titan Family
Amazon’s foundation model family, Titan, is also part of Bedrock:
- Titan Text Large: A generative language model.
- Titan Text Embeddings: A model designed for generating text embeddings, suitable for applications like retrieval and recommendation.
Titan models come with content filtering capabilities to detect and remove harmful content.
Stable Diffusion for Image Generation
Bedrock extends its support beyond text to include image generation through Stable Diffusion models. Currently, Stable Diffusion XL is available, enabling text-to-image prompting, image-to-image prompting, and image inpainting.
First Impressions
As privileged early users of Amazon Bedrock, here are our initial observations:
Model Variety
One of the standout features of Bedrock is the extensive variety of models available. Unlike some competitors limited to proprietary offerings, Bedrock provides a broader selection, empowering users to make informed trade-offs in terms of quality, performance, and cost.
Focus on Safety
Amazon Titan models showcased a robust filtering system, effectively blocking unsafe or harmful content and guiding users to adhere to the terms of service. Such safety measures are crucial for user-facing applications involving these models.
Playgrounds
Similar to other generative AI offerings, Amazon Bedrock features a user-friendly playground. This allows users to test various prompts and settings for each model before integrating them into their applications, facilitating quick testing and prototyping.
Ease of Access
True to Amazon’s promise, Bedrock lives up to being the easiest way to build and scale generative AI applications with foundation models. Accessing a model is as simple as sending a request with the appropriate parameters and input, and the addition of Bedrock support to the official AWS Python SDK, boto3, further simplifies the process.
In conclusion, Amazon Bedrock emerges as a versatile and user-friendly addition to the generative AI space, offering a diverse range of models and emphasizing ease of use, safety, and rapid development. As the generative AI landscape continues to evolve, Bedrock positions itself as a formidable contender, providing developers with powerful tools to unlock the full potential of foundation models.
Author: Shariq Rizvi