Has been updated On Thursday, Amazon Web Services announced an API platform called Bedrock. The platform hosts generative AI models built by top startups AI21 Labs, Anthropic, and Stability AI.
Generative AI is exploding in popularity due to the development of models that can generate text and images. Commercial tools developed by hot startups such as OpenAI and Midjourney have amassed tens of millions of users, and Big Tech is scrambling to catch up.
While Microsoft and Google are competing to bring generative AI chatbots to their search and productivity suites, Amazon’s strategy has been to remain fairly neutral and, like a sort of machine learning Swiss, to cloud platforms. provides access to the latest models. This is a win-win for any startup that agrees to work with the e-commerce giant. A developer pays to use his API to access the startup’s model. AWS provides all the underlying infrastructure that fully manages and delivers those services.
Swami Sivasubramanian, Machine Learning Specialist at AWS, said in a blog post:
“First, they need an easy way to find and access high-performing content. [foundational models] It gives excellent results and is perfect for its purpose. Second, customers want seamless integration into their applications without managing huge clusters of infrastructure or incurring large costs. “
Amazon Bedrock currently offers large-scale language models (Jurassic-2 from AI21 Labs and Claude from Anthropic) that can process and generate text, as well as the text-to-image model Stable Diffusion from Stability AI. Bedrock also offers two of his foundation models, exclusive to Amazon under the Titan brand. Not to be confused with Google’s Titan brand.
Developers can build their own generative AI-powered products and services on top of these Bedrock-maintained APIs, providing their own labeled training examples to fine-tune their models for specific tasks. Adjustable. According to Amazon, this customization process allows organizations to tailor their training to their specific application without worrying about private training data being leaked, misplaced, or used to train other large-scale language models. You can adjust the neutral network.
Amazon also promotes the Trainium and Inferentia families of custom-designed machine learning acceleration chips. You can use these chips to train and run this kind of generative model in your own cloud. The new EC2 instance type, named Trn1, is powered by Trainium silicon, which a developer reportedly saves up to 50% on training costs compared to his other EC2 instances. .
Trn1 instances are optimized to distribute training across multiple servers and have up to 1600 Gbps of network bandwidth. Developers can also spin up “ultraclusters” that scale up to 30,000 Trainium chips, providing more than 6 exaflops of computing.
Also, for inference work, AWS says the Inf2 EC2 instance type is now generally available and is accelerated by the cloud giant’s Inferentia2 silicon.
Finally, Amazon is making its AI pair programming tool, CodeWhisperer, publicly available and free for qualified developers to use. Expanded to support 10 new languages including Go, Kotlin, Rust, PHP, SQL and more, in addition to Python, Java, JavaScript, TypeScript and C#.
“Today, we believe CodeWhisperer is the most accurate, fastest, and safest way to generate code for AWS services such as Amazon EC2, AWS Lambda, and Amazon S3,” said Sivasubramanian. ®
update
I noticed that Amazon’s announcement states that “customers simply point Bedrock at a few labeled examples from Amazon S3, and the service can fine-tune its model for specific tasks.” So I wondered if Bedrock should be used with his other AWS resources. Both must be paid by the customer.
AWS representatives have been communicating that Bedrock is a standalone, fully managed API service that does not require other AWS resources such as S3, but customers can mix Bedrock APIs and AWS services together as needed. and will be billed by Amazon. as needed. Alternatively, the customer can use her Bedrock alone and pay only for it.
