During the AWS Summit New York event on July 26, 2023, AWS made several exciting announcements concerning Generative AI (GenAI) which should be relevant for anyone in an engineering or leadership role within the AI space. AWS is most certainly aware of both the increased cultural awareness of GenAI as well as the intense interest in the technology expressed by companies across nearly every industry; the primary idea expressed during the AWS Summit event is that AWS seeks to commit fully to these technologies and continue to offer a broad range of AI-related services that enable businesses to host their AI infrastructure in the cloud at any scale. More specifically, AWS aims to provide a best-in-class experience for developing and deploying the kinds of GenAI applications that have been getting so much public attention in recent months.
Broadly speaking, the most notable AI-related products/services, such as ChatGPT and DALL-E, involve using a Foundational Model (FM) and then fine-tuning this model for a more specific use case. Because of this, if AWS wants to continue to maintain relevance in the rapidly developing AI ecosystem, it is critical that they offer a wide variety of high-quality FMs on their platform. Luckily, July’s announcement includes a variety of updates to the available FMs on AWS’ Amazon Bedrock service, which is their fully-managed service offering for building and hosting GenAI applications. New FMs available on Bedrock are as follows:
Now, regardless of how many models are available on Bedrock, or which model a business may choose, it is typically difficult to integrate GenAI applications with a company’s existing systems. An example of this provided in the AWS Summit was a chatbot that needs to apply edits to some customer’s order of a pair of shoes. To help solve the challenge of integrating GenAI apps, such as a chatbot, with business data (such as customer orders), AWS announced Agents for Amazon Bedrock. This new capability to the existing Bedrock service will allow for GenAI apps hosted on Bedrock to more easily perform actual tasks “such as managing retail orders or processing insurance claims” or, in the case of the shoe-related example, change a user’s order to involve a pair of brown shoes instead of black ones. Agents for Amazon Bedrock will help to provide GenAI apps with capabilities that are highly relevant for end-users, whether that be changing the color of some shoes or processing someone’s car insurance claim automatically.
With regard to providing relevant capabilities to end-users, AWS also announced the vector engine for Amazon OpenSearch Serverless. Generally speaking, vector embedding (which is what this feature pertains to) allows for GenAI apps to provide more appropriate experiences to customers rather than the application misunderstanding the intent of the user. To once again reference the shoe-related example, this feature will help GenAI apps to understand that it is the color of the shoes themselves that need to be changed, not the color of the box they will arrive in, or the color of the laces attached to them. While the storage of vectors has already been possible in AWS’ standard/serverful Opensearch offering as well as in AWS RDS Aurora Postgres and AWS RDS Postgres, this announcement helps to provide the serverless variant of OpenSearch with the full range of features necessary for proper usage with GenAI workloads.
Generative AI on AWS
Interested in GenAI on AWS? Check out the link to learn more!
Given that there are so many options available for storage of data on AWS, it is often difficult to match and link records across a variety of services. To help solve this problem, AWS also announced AWS Entity Resolution, which is an AI-powered service to match/link/prepare your data for analysis by software applications such as GenAI systems. This can be really helpful for people who work across services because even though the service is new, best case scenario: you can treat all the services like a large data lake you can query information out of, and worst case scenario: you can improve data integrity by matching related records.
While it may seem that AWS is unveiling a variety of different solutions across numerous isolated AWS services, Amazon in fact seeks to increase the amount of native integrations available between AWS services moving forward. Specifically, one new integration that is now available is AWS Glue Studio support for Amazon CodeWhisperer. This integration will allow for code generation within Glue Studio notebooks. Moreover, Amazon seeks to leverage its increased focus on GenAI to enhance existing services. An example of this that was provided during the conference was the announcement that QuickSight Q would now support natural language processing for both the generation of charts as well as verbal descriptions of your data. These newly-added GenAI features within QuickSight Q should tremendously increase the usefulness of the service for anyone who needs to share data insights with their team. In particular, the GenAI features of QuickSight Q allow for anyone to get started with the service very easily because you can just verbally describe the kind of graph you want and have it automatically generated from the data you already have stored.
The final part of the conference does in fact reflect the announcement of a service that does not necessarily have an obvious relation to other AWS services: AWS HealthScribe. This service allows for recordkeeping of clinical documentation in a HIPAA-compliant manner on AWS. HealthScribe leverages GenAI to generate insights from stored patient data as well as transcribes audio using existing AI technologies. The insights generated from this patient data should definitely make this service noteworthy for any and all companies in the healthcare industry, as these insights will allow for large amounts of patient data to be condensed into more manageable pieces of information for healthcare professionals to act upon. The biggest advantage of this would be better patient outcomes and decreased time-per-patient for doctors. Better patient outcomes should be expected here because AI would be able to identify older/interfering diseases/conditions that could affect a diagnosis that would have otherwise been missed if a doctor was manually reading notes. Doctors will also be very quick in diagnosis because an AI would be able to summarize any relevant information.
The last noteworthy announcement made during this talk was that EC2 P5 instances are now available. If you are someone using a previous generation of this instance family, then the importance of this announcement is clear: usage of this new instance type will provide better cost/performance for your AI-related workloads. Otherwise, if you are not familiar with this family of instances, then this announcement may safely be disregarded.
One final point I would like to emphasize [that is not an explicit service announcement] is that not only does AWS desire to commit fully to Generative AI, but they also aim to integrate this technology into existing services. The previously-mentioned QuickSight Q is a perfect example of this; Amazon is now bringing GenAI capabilities directly to this service and allowing for the generation of visual reports automatically. We should expect that in the future AWS will continue to announce further GenAI integrations into existing services. It is reasonable to expect that services such as CloudWatch may someday allow for the querying of logs using natural language instead of using Amazon’s custom query language seen in CloudWatch Logs Insights.
Learn how Claude3 Opus, now available on Amazon Bedrock, outperforms its peers on common evaluation benchmarks for AI systems.
