Published 13 April 2023

Amazon has announced the launch of Bedrock, its latest platform for generative AI models. Bedrock will host third-party models on AWS, with AWS customers able to access the models through an API. The models hosted on Bedrock are aimed at enterprise-scale AI applications and include AI21 Labs’ Jurassic-2 family, which can generate text in multiple languages, Anthropic’s Claude, which can perform a range of conversational and text-processing tasks, and Stability AI’s suite of text-to-image models. Amazon’s bespoke models include the Titan FM family, comprising a text-generating model and an embedding model.

To customize any Bedrock model, AWS customers can point the service at a few labeled examples in Amazon S3, with as few as 20 being enough. Amazon says no customer data is used to train the underlying models, but didn’t specify exactly which data was used to train its Titan FM family. The company says that the models were built to detect and remove harmful content and reject inappropriate content users input, filtering outputs containing hate speech, profanity, and violence. However, prompt injection attacks can still be used against generative AI models to write malware, identify exploits in open-source code, and generate misinformational content.

The move coincides with Amazon’s launch of CodeWhisperer, its AI-powered code-generating service. Amazon has made CodeWhisperer free of charge to developers without any usage restrictions, which suggests it hasn’t seen the uptake Amazon hoped for. CodeWhisperer launched in late June as part of the AWS IDE Toolkit and AWS Toolkit IDE extensions, and is trained on billions of lines of publicly available open-source code and Amazon’s own codebase, as well as documentation and code on public forums.

CodeWhisperer now supports several additional programming languages, including Go, Rust, PHP, Ruby, Kotlin, C, C++, Shell scripting, SQL, and Scala. CodeWhisperer Professional Tier adds single sign-on with AWS Identity and Access Management integration and higher limits on scanning for security vulnerabilities. The move is seen as an attempt to compete with GitHub’s Copilot, which had over a million users as of January, including thousands of enterprise customers. To ward off the legal challenges GitHub is facing with Copilot, CodeWhisperer highlights and optionally filters the license associated with functions it suggests that bear a resemblance to existing snippets found in its training data.