Amazon Web Services expanded the variety of large language models (LLMs) and other foundation models available on Amazon Bedrock in a move to help customers find the right foundation model for each unique artificial intelligence (AI) use case.

Bedrock is a serverless API-based service designed to help developers without technical expertise regarding the intricacies of machine learning (ML) or the management of complex AI systems build AI applications based on a choice of foundation models. Bedrock also "provides a set of tools that make it easy for developers to build generative AI (genAI)-powered apps," GM of Amazon Bedrock Atul Deo told SDxCentral.

Providing quality choices for customers drove the decision to partner with a number of AI startups on Bedrock. Foundation models available through Amazon Bedrock include those created by AI21 Labs, Anthropic, Cohere, Meta and Stability AI. Bedrock also includes access to the Titan foundation models created by AWS.

To realize the full value of AI, most enterprises need to train and fine-tune their own models rather than running something right out of the box. In the past year and a half, Deo has seen a new type of tooling set spring up that "allows companies to build applications that make specific use of their data and their software systems," he said during an interview at AWS re:Invent. "Otherwise, the models are generic. They're pre-trained, and they understand the world but they don't understand the company and their use cases."

A company with a specific AI use case in mind should consider three aspects: accuracy, latency and cost. And the needs of the applications will determine which foundation model is right for the job.

Take a high-volume, end-user-facing use case that summarizes millions of books, for example. "I may not want to invest in the most capable, most powerful model because usually the most powerful model is going to be slower. It's also going to be more expensive," Deo said. "Do you really need your biggest weapon for a particular task? Probably not." But "you don't need one generic model for everything," either.

If a customer feels their best fit is to use a foundation model that isn't available with Bedrock — like one from OpenAI, for example — "we just want to hear the feedback and understand why that is the case," Deo said. "Because the portfolio that we have — we believe that it addresses every possible use case based on the customers that we've worked with." If at some point "there is a gap, of course we want to address that," he added.

AWS' approach to the AI market is "to create, to innovate with speed and agility and to surround themselves with partners like Deloitte," Deloitte Principal JB McGinnis told SDxCentral during AWS re:Invent. With Bedrock, AWS created an ecosystem of players that McGinnis feels is a "very inclusive approach." Of course, there's also the competitive aspect. "We just saw the sphere," he said, referring to Google Cloud's conveniently timed ad takeover of the Las Vegas sphere during AWS' annual conference.

"I respect the way that they've approached the market with genAI," McGinnis said.

To help genAI applications execute multistep tasks, AWS added Bedrock Agents, which help solve problems for any specific scenario by making use of the company's software systems and unique data.

"Think about a chatbot agent that is geared for customer service. You can point it to different APIs and different resources and within a few steps — a relatively low effort, no-code workflow — you can get a chatbot up and running and you can answer questions for your end users. That is the kind of experience that Agents for Amazon Bedrock unlocks," Deo said.

Agents don't stop at chatbots, however. "It is about more than just being able to converse," Deo said. It's about being able to execute tasks and invoke APIs in the background to complete multistep processes with ease. "Agents powered by foundation models are now capable of dealing with a lot more ambiguity than ever before," Deo said.

The cloud hyperscaler also introduced a new set of guardrails for Amazon Bedrock meant to install safeguards across models based on application requirements and responsible AI policies.

Many businesses don't want a certain set of topics to be discussed in a chatbot because "they only want their end users to talk about what matters to their business," Deo said. Bedrock's denied topics and content moderation features, however, are important new pieces of guardrails that address this specific customer concern. To that point, Bedrock guardrails fuel consistency in terms of how Amazon Bedrock foundation models respond to unwanted or harmful content input to genAI apps.

"Think about it as pre-processing of the inputs that get fed in and post-processing of the results that come from the model," and "we are giving those controls to the developer," Deo said.

In the future, AWS has plans to launch additional guardrails for personally identifiable information (PII) input reduction and a blocked words list.