AWS unveils new AI chatbot, chips, Nvidia partnership

Amazon Q

Amazon Q is AWS' attempt at creating an enterprise-grade generative AI chat application.

"What the early providers in the space have done is exciting, and it's genuinely super useful for consumers," Selipsky said during his keynote, referring to ChatGPT from OpenAI, partner of AWS rival Microsoft, and Bard from Google.

But Selipsky claimed that other chat applications are not work-friendly and fail to address security and data privacy concerns.

Amazon Q is a generative AI-powered assistant designed to "work for you at work," according to AWS.

Amazon customers can use Q to get answers to questions, generate content and streamline tasks, according to the cloud vendor.

The new feature uses AWS applications for business intelligence, contact centers and supply chain management.

It's a way of showcasing that they are helping customers benefit from generative as part of the products and services that AWS is offering.

Developers can access Amazon Q through a conversational interface from the AWS Management Console, documentation pages and on Slack and other third-party services and applications.

"It's a way of showcasing that they are helping customers benefit from generative AI as part of the products and services that AWS is offering," said Gartner analyst Jim Hare.

While Amazon Q is similar to the Microsoft Copilot and Google Duet AI digital assistants, the capability addresses a need within AWS' customer community for a chatbot that can be embedded and infused with other AWS products and services, Hare added.

For example, Q is integrated into Amazon's cloud contact center service, Amazon Connect, and its business intelligence service, Amazon QuickSight. It will soon be incorporated into AWS Supply Chain.

"The fact that this kind of conversational system has been baked into a variety of different parties and services that helps [different] individuals improve their productivity and get information to generate content, I think is an important step," Hare said.

Moreover, using the capability with third-party applications like Slack and Salesforce and having information from those applications indexed or pulled in to create findable and usable information will be essential for customers looking for alternatives to Microsoft and Google, he said.

Meanwhile, some enterprises using a beta version of using Amazon Q are improving their productivity, according to founder and analyst R "Ray" Wang.

For example, developers from Amazon and IT company Persistent Systems found that Amazon Q helped Persistent Systems speed up its coding by up to a third, according to Wang.

"That's huge," Wang said. "That's like you need one less developer or two less developers per project."

Moreover, having a chat capability incorporated into the development environment, with the ability to not only provide answers to queries but also troubleshoot errors with language models, is a plus for AWS users, Wang added.

It's also helpful that Q is not running on OpenAI's large language model (LLM) GPT-4, especially after the recent leadership shakeup at the AI startup, Hare said.

"They're using other models that are more use case-specific," he said. "The fact that AWS offers a choice of different models gives them that flexibility, versus relying on one single model."

Updates to Bedrock

Through AWS Bedrock, its managed foundation model service, AWS offers access to models from startups including Anthropic, AI21 Labs, Cohere and Stability AI, as well as Amazon models.

On Tuesday, Amazon revealed that customers can customize foundation models with private data.

Bedrock now supports fine-tuning for Meta Llama 2, Cohere Command Light and Amazon Titan models.

Fine-tuning for Anthropic's foundation model, Claude, will come soon, according to Amazon.

While fine-tuning with private data was just introduced, one customer of both AWS and Anthropic is customizing and fine-tuning its models using reinforcement learning from human feedback.

In October, LexisNexis launched a new product, Lexis+AI, which uses several versions of Claude through AWS Bedrock.

"One of the reasons we wanted to work with AWS and Anthropic is that they would allow us to fine tune one of these foundational models with legal use cases," LexisNexis CTO Jeff Reihl said in an interview.

With the Anthropic foundation model from Bedrock, LexisNexis fine-tuned the model using reinforcement learning with human feedback to create a model optimized for various legal use cases. Those include conversational search, document drafting, summarization and document upload.

LexisNexis also uses some of its own internal legal professionals to help augment the model for specific applications, Reihl said.

Amazon also revealed that Guardrails for Amazon Bedrock is available in preview. Customers can apply safety guardrails to all LLMs in Bedrock, including fine-tuned models. Using the feature, customers can define policies, set up a set of topics that can't be broached and configure thresholds to filter harmful content.

As a company with customers concerned with privacy and security, LexisNexis chose AWS because of its ability to protect customers' data, Reihl said.

"In the years we've worked with AWS, we felt very confident in their security model, and I think ... the announcement today is just another example of that," he said.

Amazon also revealed that Agents for Amazon Bedrock, a new capability that helps developers create agents that speed up the delivery of generative AI applications, is now generally available.

New chips and partnerships

Beyond a choice of models, AWS provides hardware choices.

On Tuesday, the cloud provider introduced two new generations of AI chips: AWS Trainium2 and Graviton4.

Trainium2 is designed to deliver faster training than the first generation of Trainium chips and will make it possible for customers to train foundation models and LLMs faster.

Graviton4 provides better compute than the current Graviton3 processors, according to the vendor.

AWS also expanded its partnership with AI hardware and software vendor provider Nvidia. Both vendors will bring Nvidia GH200 Grace Hopper Superchips with new multi-node NVLink technology to the cloud. The platform will be available on Amazon Elastic Compute Cloud.

Nvidia and AWS will also collaborate to host Nvidia's AI-training-as-a-service system, DGX Cloud, on AWS. This will provide developers with the largest shared memory single memory in a single instance, Nvidia claimed.

In addition, Nvidia and AWS are collaborating to build a giant AI cloud supercomputer with the Nvidia Grace Hopper Superchip, hosted by AWS.

"With this announcement, AWS is showing that it's going to continue to offer silicon diversity to its customers," said Futurum Research analyst Dan Newman. "They need to be flexible as a hyperscale cloud provider with a massive diversification of services, so they don't alienate themselves from certain customers by not having the products and solutions they want."

While Nvidia partnering with yet another cloud provider is not surprising -- Nvidia has already teamed with Google and Microsoft -- "having the DGX Cloud available on AWS is significant because that's where a lot of AWS customers' data is," said Forrester analyst Mike Gualtieri.

All these product moves dispel the idea that AWS is behind in generative AI, according to industry analysts.

"The series of announcements, including the Bedrock announcements, are just a big step that AWS has made," Gualtieri said. "A lot of people said they were late. They can't say that anymore."

Esther Ajao is a TechTarget Editorial news writer covering artificial intelligence software and systems.