AWS expands AI strategy with support for different models

Amazon Web Services Inc. is ready for the next generation of workers — and by that, the company means a host of different models and artificial intelligence agents.

The cloud giant is preparing for the upcoming AWS re:Invent later this year, with a series of product updates based around intelligent automation and agentic AI.

TheCUBE’s John Furrier talks with AWS’ Atul Deo about the development of AI agents.

“We talk about agentic capabilities, but particularly models being able to follow instructions and being good at it … make a huge difference in the entire workflow,” said Atul Deo (pictured), director of product, AWS Bedrock, at AWS. “We are seeing employability for some of these cloud models. We also have been busy launching some of our first party models with Nova.”

Deo spoke with theCUBE’s John Furrier at the AWS Mid-Year Leadership Summit event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how Amazon is responding to a rapidly changing AI ecosystem.

The combined power of different models

Since announcing a new generation of foundation models at the last re:Invent, AWS has made intelligent prompt routing generally available. This tool enables users to combine the advantages of cheaper and larger, more capable models. Another product that offers the best of both worlds is Bedrock’s model distillation feature, which transfers intelligence from a larger model to a smaller, more specialized model.

“We’ll generate additional data for the distillation process based on the prompts that a customer provides,” Deo explained. “It can give a few indicator 30, 40 prompts of what it wants kind of generally for the distillation purpose. Then behind the scenes we can generate 10 times more data, which is basically synthetic data, then that synthetic data response of that larger model then gets used to essentially kind of make the smaller model more targeted and focused.”

Two of the hottest areas for generative AI have been code generation and sales and marketing. Part of making AI a good assistant for customer service, code or even real estate is providing agents standardized access to relevant context through Model Context Protocol, according to Deo.

“The two things that I like about MCP, number one, is … [it] can discover different sources and it doesn’t have to be hardwired beforehand,” he said. “And second one is unlike an API, which is typically stateless, this one can maintain state and can be secure.”

Even though AWS wants to offer up the best models for potential customers, Deo acknowledged that there are a number of different companies competing in the “AI petri dish,” as Furrier calls it. Amazon’s goal is to support customers for any models they choose to employ.

“You’re probably going to see the rise of different models even being used in the same workflow more and more often,” Deo said. “We already had some of that with RAG, where everybody had to use an embeddings model, and so everybody had to use at least one kind of generative model so that that kind of combination existed in the past. But you’re going to see increasing combinations of multiple models being used by customers even in the same workflow.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the  AWS Mid-Year Leadership Summit:

Photo: SiliconANGLE