Amazon is entering the field of generative AI by introducing Amazon Bedrock, a platform that allows third-party startups to host pre-trained models on AWS. Bedrock offers access to pre-trained models from startups like AI21 Labs, Anthropic, and Stability AI, as well as in-house models called Titan FMs. This move follows AWS’ recent partnerships with generative AI startups and its investments in technology for building productive AI apps. With Bedrock, AWS aims to enable businesses to apply machine learning at scale to solve real-world problems using generative AI.
Bedrock and custom models
Amazon’s Bedrock is a significant move by the company to capture a share of the generative AI market, projected to be worth around $110 billion by 2030. Bedrock allows AWS customers to access AI models from various providers, including AWS itself, through an API. While pricing details are unclear, Amazon has stated that Bedrock is aimed at large customers building “enterprise-scale” AI applications, differentiating it from other model hostings services like Replicate, Google Cloud, and Azure.
The third-party models hosted on Bedrock include AI21 Labs’ Jurassic-2 family, which are multilingual and capable of generating text in multiple languages. Claude, Anthropic’s model on Bedrock, can perform various conversational and text-processing tasks. Stability AI’s suite of text-to-image models, including Stable Diffusion, can generate images, art, logos, and graphic designs.
The terms of the model licensing and hosting agreements between Amazon and the third-party vendors on Bedrock have been kept private. Still, it is presumed that vendors were incentivized by AWS’ extensive reach and potential revenue-sharing opportunities.
Amazon’s Bedrock is a new entry into the generative AI market, aiming to provide AWS customers with large-scale, enterprise-level AI capabilities. Bedrock hosts third-party models, such as Jurassic-2 for multilingual text generation, Claude for everyday tasks, and Stable Diffusion for text-to-image age.
In addition to Bedrock, Amazon also offers custom generative AI models under the Titan FM family, which currently includes text-generating and embedding models. The text-generating model can perform tasks like writing blog posts, summarizing documents, and extracting information from databases. In contrast, the embedding model translates text inputs into numerical representations for semantic meaning.
AWS customers can customize any Bedrock model by providing labelled examples from their data, with as few as 20 examples being sufficient. Amazon emphasizes that no customer data is used to train the underlying models, prioritizing data privacy.
While Microsoft has seen success with its Azure OpenAI Service, there are legal concerns surrounding generative AI, including copyright infringement allegations and attribution issues. However, Amazon asserts that its Titan FM models are designed to detect and remove harmful content and reject inappropriate inputs, focusing on responsible use of the technology.
Despite concerns about potential biases and misuse of generative AI, Amazon is committed to monitoring the regulatory landscape and ensuring the responsible use of its models. While brands may have liability concerns, individual customers may find value in the capabilities Bedrock and the Titan FM models offer, particularly considering there is currently no charge for their use.
CodeWhisperer, Trainium and Inferentia2 launch in GA
As part of its generative AI push, Amazon has made CodeWhisperer, its AI-powered code-generating service, free for developers without any usage restrictions. This move comes as a response to the growing popularity of GitHub’s Copilot, which already has over a million users, including enterprise customers. Amazon has also launched CodeWhisperer Professional Tier to catch up, targeting the corporate market. The Professional Tier offers additional features such as single sign-on with AWS Identity and Access Management integration and higher limits on scanning for security vulnerabilities.
CodeWhisperer, launched in late June as part of the AWS IDE Toolkit and AWS Toolkit IDE extensions, is trained on billions of publicly available open-source code, Amazon’s codebase, and documentation and code from public forums. It can autocomplete functions in Java, JavaScript, and Python based on comments or keystrokes. Recently, CodeWhisperer expanded its language support to include Go, Rust, PHP, Ruby, Kotlin, C, C++, Shell scripting, SQL, and Scala. To address legal challenges, CodeWhisperer now highlights and optionally filters the license associated with suggested functions that resemble existing snippets in its training data.
In addition to CodeWhisperer, Amazon has announced the general availability of Elastic Cloud Compute (EC2) Inf2 instances powered by AWS Inferentia2 chips, previewed last year at Amazon’s re: Invent conference. These instances are designed to accelerate AI runtimes, offering improved throughput and lower latency for better overall inference price performance. Furthermore, Amazon EC2 Trn1n instances powered by AWS Trainium, Amazon’s custom-designed chip for AI training, are also now generally available.
These instances provide up to 1600 Gbps of network bandwidth and are designed to deliver up to 20% higher performance over Trn1 for large, network-intensive models.
With these launches, Amazon is confident in offering adequate cloud infrastructure for generative AI and addressing the cost concerns of customers when dealing with these models. However, the competition in the fertile AI space, particularly from rivals like Google and Microsoft, remains strong, and the impact of these new offerings on the market is yet to be determined.
Leave a Reply