Thursday, 27 July 2023

Amazon expands Bedrock with conversational agents and new third-party models

by Rose White

Amazon continues to march unencumbered toward a generative AI future.

At AWS Summit New York, one of its annual conferences, Amazon announced the launch in preview of Agents for Bedrock — Bedrock being Amazon’s platform that provides a way to build generative AI-powered apps via pretrained models from startups as well as Amazon itself.

Agents allows customers to create conversational agents to deliver personalized, up-to-date answers based on their proprietary data and execute actions.

In a briefing, Swami Sivasubramanian, the VP of machine learning at AWS, told TechCrunch that a company could use Agents to create a customer service chatbot that can process orders — tapping their internal information, including customer profiles and their return policy — to customize each order.

It’s akin to the plugins system recently launched by OpenAI for its generative text AI, GPT-4 and ChatGPT, which similarly extends the models’ functionality by letting them tap into third-party APIs and databases. Indeed, there’s been a trend toward “personalized” generative models lately, with startups like LlamaIndex and Contextual AI building tools to augment models with enterprise data.

“Bedrock Agents accelerate the delivery of generative AI applications that can manage and perform tasks by making API calls to your company systems,” Antje Barth, a principal developer advocate at AWS, explains in a blog post shared with TechCrunch via email. “Agents extend AI models to understand user requests, break down complex tasks into multiple steps, carry on a conversation to collect additional information, and take actions to fulfill the request.”

To figure out which actions to take, what information to use and in which sequence to perform actions, Bedrock Agents have to be shown how to “reason” through prompts that include definitions and instructions. These prompts are generated automatically in the backend through a series of configuration steps, one of which involves telling the Agent what task it’s supposed to perform and role it’s supposed to assume (e.g. “You’re an agent designed to help with processing insurance claims and managing pending paperwork”).

The AI models powering Agents were developed internally by Amazon, a part of the company’s Titan family of AI models. But Amazon’s adding several models from third parties to the Bedrock library to coincide with the launch of Agents.

Anthropic’s Claude 2 text-generating AI and Stable Diffusion XL 1.0, the latest generative image model from AI startup Stability AI, are both available in Bedrock as of today. So are two foundation models from Nvidia-backed Cohere, specifically “Command text generation” and “Cohere Embed.” (Command is is trained to follow user commands and be useful in business applications like summarization, copywriting, dialogue, extraction and question-answering, while Embed can be applied in search or classification tasks across around 100 languages.)

The growing range of models, including the models underpinning Agents, would seem to signal Amazon’s seriousness about Bedrock — and, by extension, the generative AI field.

But Bedrock had a rocky start. Bloomberg reported in May that, six weeks after Amazon demoed the tech with an unusually vague presser and just one testimonial, most cloud customers still didn’t have access.

Given AWS’ massive reach and computational might, though, Bedright could have a fighting chance. Time will tell.