#707: AWS News: DeepSeek R1 Models on BedRock, Improve Data Discoverability with S3 Metadata

Summary

AWS Podcast Episode #707 Summary: DeepSeek R1 Models on BedRock & Enhanced S3 Metadata

Release Date: February 10, 2025

In episode #707 of the AWS Podcast, hosted by Shruti and Jillian from Amazon Web Services, the duo delves into the latest AWS innovations, focusing on the introduction of DeepSeek R1 Models on BedRock and significant enhancements to S3 Metadata for improved data discoverability. This detailed summary captures all key points, discussions, insights, and conclusions from the episode.

Top Stories

1. DeepSeek R1 Models on Amazon BedRock

Overview of DeepSeek

The episode begins with thrilling news about DeepSeek, a cutting-edge large language model (LLM) making waves in the tech industry. Shruti emphasizes Amazon's commitment to providing diverse options for customers:

“Amazon has always said that choice matters. It gives customers the flexibility they need to choose the LLM that's right for them.”
— Shruti ([00:25])

Deployment Options

DeepSeek R1 models are now available for deployment across multiple AWS platforms:

• Amazon Bedrock
• Amazon SageMaker
• Amazon EC2 using AWS AI chips, Trainium and Inferentia

The models come in various configurations to cater to different needs:

• DeepSeek R1: The flagship model with 671 billion parameters.
• DeepSeek R10: Similar in scale to R1, offering robust performance.
• DeepSeek R1 Distill: Smaller, distilled versions ranging from 1.5 to 70 billion parameters for more resource-efficient applications.

Benefits and Flexibility

Jillian highlights the advantages of deploying DeepSeek models on AWS, particularly the flexibility and security features:

“If you’re looking at Bedrock, there’s so many already different models available... you can have additional features such as guardrails, the security features as well.”
— Jillian ([02:02])

Shruti adds that DeepSeek models are 90-95% more cost-effective than comparable models, making them an attractive option for various applications.

Getting Started with DeepSeek

There are multiple pathways to integrate DeepSeek into your projects:

• Amazon Bedrock Marketplace: Direct access to DeepSeek models.
• Bedrock Custom Model Import: For distilled versions, allowing further customization.
• Amazon SageMaker Jumpstart: Enables fine-tuning and customization.
• Amazon EC2: For those needing extensive control over infrastructure.

Shruti explains:

“You could get started in SageMaker, customize it and then pull it into Bedrock to then avail of all the added functionality such as guardrails and so on.”
— Shruti ([02:40])

Ideal Users

Jillian identifies the target audience for DeepSeek models:

“Enterprise customers that are really looking for that added layer of security that comes built in and the scalability as well.”
— Jillian ([05:02])

Organizations seeking cost optimization, security, and scalability will find DeepSeek models particularly beneficial.

2. Enhancing Data Discoverability with S3 Metadata

Introduction to S3 Metadata Enhancements

The podcast transitions to discuss enhancements in S3 Metadata, which now automatically extracts and manages rich metadata for objects stored in S3 buckets.

Jillian explains:

“It makes it really easy to discover and query your data that’s in there.”
— Jillian ([05:33])

Key Benefits

Shruti outlines the primary advantages of the updated S3 Metadata:

• Simplified Data Discovery: Reduced time and effort for data preparation without the need for complex external systems.
• Better Data Governance and Lineage Tracking: Enhanced tracking of metadata across all S3 buckets.
• Improved Query Performance: Optimized for faster analytics, benefiting use cases like AI model training and real-time inference.

She further elaborates:

“With this sort of rich metadata... real-time updates become much, much easier.”
— Shruti ([06:22])

Target Audience

Jillian details who can benefit most from these enhancements:

“Large scale, like you’ve got millions, billions, trillions of objects... Data science engineering teams... enterprise customers.”
— Jillian ([07:33])

Notable customers such as Roche, SmugMug, and PayPal are already leveraging these features to manage their extensive data repositories.

Top Headlines

Following the main stories, Shruti and Jillian cover a series of updates across various AWS services. Below is a categorized breakdown of these headlines:

Marketplace

• Self-Service Seller Onboarding: Introduces support for demo and private offer requests.
• Automated Archival: Automatically archives old, unused product versions for Amazon Machine Images (AMIs), CloudFormation templates, and container products.
• Precision in Pricing: Implements eight decimal place precision for usage pricing.

Analytics

• Amazon Redshift:

  • Enhanced Query Monitoring: Improves diagnostics and performance bottleneck identification.
  • Default Security Configurations: Disables public accessibility, enables database encryption, and enforces secure connections by default.
  • History Mode for Zero ETL Integrations: Simplifies data management without the need for ETL processes.
  • New SQL Features: Introduces additional SQL functionalities for zero ETL integrations.

• Amazon EMR Serverless: Adds support for public subnets, enhancing network flexibility.

Application Integration

• Amazon SNS:

  • High Throughput Mode for FIFO Topics: Achieves up to 30,000 messages per second in US East (N. Virginia) and other regions.
  • Maintains order within message groups while reducing deduplication scope.

• Amazon EventBridge:

  • Cross-Account Event Delivery: Enables events to be delivered directly to AWS services in another account, enhancing security and reducing architectural complexity.

Jillian comments on the significance of these updates:

“I definitely love this one, especially for all the huge fans out there of event-driven architectures...”
— Jillian ([10:29])

Artificial Intelligence

• Amazon Bedrock:

  • Multimodal Support: Integrates Cohere Embed 3 multilingual and English foundation models capable of generating embeddings from both text and images.

• Amazon Q Business:

  • Image Insights: Supports uploading images directly into chat for querying related content.

• Amazon Lex:

  • Assisted Slot Resolution: Expands regions and model access, enhancing conversational AI capabilities.

• AWS Health:

  • IPv6 Support: Extends network protocol support for improved connectivity.

Jillian highlights the practical applications:

“This allows you to upload images directly to Amazon Q Business chat and ask questions related to the content of those images.”
— Jillian ([12:00])

Compute

• Zone Groups for Availability Zones:

  • Facilitates easier differentiation of local zones and Availability Zones across all regions.

• Amazon EKS:

  • New Update Strategies for Managed Node Groups: Offers control over EC2 instance updates within clusters, supporting Kubernetes version 1.32.

• AWS Elastic Beanstalk:

  • Scaling and Deployment Enhancements: Improves speeds for Windows instances and adds default support for EC2 launch templates.

Customer Engagement

• Amazon SES:

  • MailManager Enhancements: Supports defined email addresses and domain lists for better routing decisions.

• Amazon Connect:

  • Daily Headcount Projections: Provides granular staffing requirement forecasts up to 64 weeks ahead.
  • Agent Workspace Optimizations: Enhances audio performance for virtual desktops like Citrix and Amazon Workspaces.

Database

• Amazon Timestream for InfluxDB:

  • Storage Scaling: Allows dynamic scaling of storage and changing storage tiers as needed.

• Amazon Aurora PostgreSQL:

  • Support for PostgreSQL 16.6: Enhances database capabilities with the latest PostgreSQL features.

• Amazon ElastiCache:

  • One-Click Connectivity Setup: Simplifies connectivity between EC2 and caching services.

• Amazon Neptune:

  • Open Source Graph Rack Toolkit: Enhances generative AI applications with comprehensive, explainable responses using the RAG technique.

Jillian remarks on Neptune’s update:

“Having a toolkit that can help you build out your application and provide value faster is just going to help you tremendously.”
— Jillian ([17:29])

Developer Tools

• AWS CodeBuild:

  • New IAM Condition Keys: Introduces CodeBuild Project ARN and CodeBuild Build ARN for more granular access control.

• Amazon Corretto:

  • Quarterly Updates: Releases security and critical updates for long-term supported versions of OpenJDK.

Shruti expresses enthusiasm for the IAM improvements:

“Anything that can help me write better IAM, I am totally in favor of...”
— Shruti ([18:00])

Front End Web and Mobile

• AWS Amplify:
  • Data Client in AWS Lambda: Enables the use of Amplify data client within Lambda functions, simplifying data operations without raw GraphQL queries.

Internet of Things

• AWS IoT SiteWise:
  • Support for Null and NaN Data: Enhances data quality handling from industrial data sources, making SiteWise more versatile for various applications.

Management and Governance

• Amazon CloudWatch:
  • Database Insights: Analyzes historical snapshots of OS processes to correlate database load spikes with OS metrics.
  • IPv6 Support for Synthetics: Extends network protocol support for synthetic monitoring.
  • Observability Add-On: Simplifies onboarding for EKS workloads.
  • AWS Managed Notifications: Enhances the management of AWS Health notifications directly from the console.

Media Services

• AWS Elemental MediaConnect:
  • Diagnostic Metrics: Adds metrics to monitor video and audio stream quality, detecting issues like black frames and audio silence for proactive troubleshooting.

Migration and Modernization

• AWS DataSync:

  • Kerberos Authentication: Supports authentication for self-managed SMB file servers, enhancing security for data transfers.

• AWS Transfer Family:

  • Custom Directory Locations: Allows storage of AS2 files and metadata in specified directories, improving file management flexibility.

Networking and Content Delivery

• Concurrent VPN Connections for AWS Client VPN:
  • General Availability: Expands VPN connectivity options, enhancing network resilience and accessibility.

Conclusion

Shruti and Jillian wrap up the episode by inviting listeners to connect with them on LinkedIn and Twitter for more insights and updates. They encourage the audience to stay engaged and continue building innovative solutions using AWS services.

---

This comprehensive summary encapsulates the key discussions and insights from AWS Podcast Episode #707, providing a valuable resource for developers and IT professionals seeking to stay updated with the latest AWS advancements.

Transcript

[A]

This is episode 707 of the AWS podcast, released on February 10th, 2025. Hello, everyone. Welcome to another update show with AWS Podcasts. My name is Shruti. I will be one of your hosts for today and joining me is Jillian. Hey there, Jillian.

[B]

Hello, Shruti. So excited to be here today.

[A]

I'm. I know we have some exciting updates and with our new format top stories, let's get started with some fun ones. So to kick off today's show, we have exciting news about deepseek, the new large language model that everyone is talking about in the industry. And we've always said from the very beginning, Amazon has always said that choice matters. It gives customers the flexibility they need to choose the LLM that's right for them. And so now these deep seq R1 models are available for deployment in Amazon Bedrock and Amazon SageMaker. And if you really want, you can also get started on Amazon EC2 using our AWS AI chips, Trainium and Inferentia. Now, these models have a few different flavors. There's DeepSeek R1, which is the main model, so to speak, or it has 671 billion parameters. And then there's deep seq R10, which also has the same amount of parameters. These are two giant whopping models. But then we also have deep seq R1 distill, which are smaller models, which are distilled versions of the larger ones, which range anywhere from 1.5 to 70 billion parameters. And so it's really exciting to be able to have all of these available on AWS and in a few different ways. Julian, what are some of the benefits of having these models available on aws?

[B]

I think it really goes back to that choice. Right. If you're looking at Bedrock, for example, there's so many already different models that are available. And now you've got deepseek as another one where you can be able to easily, within one ui, compare the different outputs, be able to have additional features such as guardrails, the security features as well that come with using Bedrock. And then of course, if you want to go on the other side, I love that when we were talking about SageMaker as well, where you can customize it, really, if you have additional requirements that you're currently not able to do, if you need that customization.

[A]

Yeah, yeah. And I mean, it's reported to be 90 to 95% more cost effective than comparable models. Again, it all depends on what your application and use cases, but it's something to try out for sure. Guardrails for sure. For that added layer of security is really interesting. And so within Amazon Bedrock, there's actually two ways that you can get started. There is Amazon Bedrock Marketplace where these models are available, or for the distilled versions, you could also do Amazon Bedrock Custom Model Import. So combining all these approaches that we talked about, Gillian, is like you could get started in SageMaker, customize it and then pull it into Bedrock to then avail of all the added functionality such as guardrails and so on. So that's really exciting. So that's Amazon Bedrock. I kind of inadvertently already talked about SageMaker Jumpstart, which is where you can also take the model and start playing with it, fine tune it, customize it as you need, and then as I mentioned, if you really want a lot of control on your infrastructure and just want to do it your own way, you can also get started on Amazon EC2, either using our GPU based instances, but our AWS AI chips, Trainium and Inferentia also support these models out the gate. And all of these options are supported both via API as well as console based deployment and they integrate really well with existing AWS services and tools. So who are these sort of models ideal for? Like, who should be looking at this?

[B]

Yeah, there's a lot of different types of people. I definitely see teams that want like quick API based integration. Maybe with those pre trained foundation models, there could be some organizations that maybe need advanced customizations and that's where the SageMaker or you need like just on straight up EC2 instances. Of course, on Bedrock you've got the enterprise customers that are really looking for that added layer of security that comes built in and the scalability as well. And of course cost conscious, which who isn't these days, right? Everyone's always looking to make things that are just as cost optimized as possible.

[A]

Absolutely, yeah. So again, this is our commitment to providing a diverse set of models, both in terms of helping you achieve your budget and cost efficiency goals, but also just your accuracy and performance goals while maintaining the enterprise grade security and scalability. So that is deep seek on AWS. The next top story we have is S3 metadata. Gillian, what can you tell us about this?

[B]

This one's really cool. I mean, it automatically extracts and manages rich metadata for objects that are already uploaded into an S3 bucket. So why would you want to do this? Well, it makes it really easy to discover and query your data that's in there. The metadata tables that are stored in Amazon S3 tables, they're going to be storage optimized for tabular data. So it then makes, it's going to be the makes it much faster for analytics. So use case would be like maybe you've got AI model training, maybe you have real time inference. There could be business analytics, those are some common use cases. And of course if you're a lot of customers are looking for iceberg compatible tools and so that's definitely supported as well.

[A]

Right, right. And so with this sort of rich metadata there's some key benefits. First is the simplified data discovery. Because of the fact that you have this metadata, there's reduced time and effort for data preparation. You don't need complex external custom metadata management systems. It now comes built in with S3 and with that you also get better data governance and lineage tracking because you are tracking the metadata for all of these sort of S3 buckets. And then it becomes much more queryable using a wide range of tools. Not to mention performance also obviously improves because of all this. So real time updates are become much, much more easier. And as you mentioned, some of these faster or rather workloads that have requirements that are very stringent like AI model training or real time inference or just faster analytics becomes much, much easier. Okay, who should be looking at this?

[B]

There's a number of different types of organizations. I would say definitely the large scale. Like you've got millions, billions trillions of objects which I'm sure there's some people who are out there that do tell us. Data science engineering teams. Let's say there's companies which a lot of companies these days got multimodal data. Another fancy word for saying you've got text, images, video, audio, so data of different types, enterprise customers. So those that want to be able to discover their data really easily. It's so much data in S3 it can be feel maybe overwhelming to discover it. Of course AI ML pipelines and there's already some customers that are using it these types of scale. Roche, SmugMug, PayPal. That's awesome.

[A]

Yeah, that's awesome. We are very customer obsessed and so it's good to know that this sort of new launch is already helping customers. All right, well now that we have covered our top stories, we are going to get into the top headlines and cover the rest of the update show.

[B]

Awesome. So of course we're going to start off with the marketplace. The AWS Marketplace adds self service seller onboarding support for demo and private offer requests. The AWS Marketplace also announces the availability of automated archival of old unused product versions that are no longer available publicly for subscription. This feature is available for Amazon machine images with cloudformation templates and container products. The AWS Marketplace introduces eight decimal place precision for usage pricing.

[A]

Next up we have a few updates under Analytics. Amazon Redshift launches enhanced Query Monitoring to improve query monitoring and diagnostics. This enables you to efficiently identify and isolate performance bottlenecks. Amazon Redshift announces enhanced default security configurations for for new warehouses. These changes include disabling public accessibility, enabling database encryption and enforcing secure connections by default when creating a new data warehouse. Amazon Redshift announces support for History mode for zero ETL integrations. Amazon Redshift introduces new SQL features for zero ETL integrations. Amazon EMR Serverless adds support for public subnets.

[B]

Next up is application integration. Amazon SNS now supports high throughput mode for SNS FIFO topics with default throughput matching SNS standard topics across all regions. When you enable high throughput mode, SNS FIFO topics will maintain order within message group while reducing the deduplication scope to the message group level. So with this change you can leverage up to 30,000 messages per second per account by default in the US East 1 region and Nikkei messages per second in the US west to Oregon region and Europe Ireland region. Amazon EventBridge now allows you to deliver events directly to AWS services in another account. This feature enables you to use multiple accounts to improve security and streamline business processes while reducing the overall cost and complexity of your architecture. I definitely love this, especially for all the huge fans out there of event driven architectures. I think this one of being able to implement best practices. Doing so when you're managing multiple accounts is a really good one to bookmark.

[A]

Excellent. All right, moving on to our updates under Artificial Intelligence. In addition to what we just covered earlier in the episode Deepseek, Amazon Bedrock now offers multimodal Support for Cohere Embed 3 multilingual and Embed 3 English foundation models that generate embeddings from both text and images. This addition can enable enterprises to unlock significant value from the vast amounts of data, including visual content. So just as Jillian mentioned, multimodal is basically a different term for different types of data, different modalities of data. And now Bedrock offers multimodal support with Cohere Embed 3 multilingual as well as the Embed 3 English. Amazon bedrock flows Announces preview of multi turn conversation Support Amazon Q Business now supports insights from images uploaded in chat that is really awesome. This will allow you to upload images directly to Amazon Q Business chat and ask questions related to the content of those images. So if you are already using Amazon Q Business to tap into your data, get insights from your data, and let's say you have some image or a snapshot of a report or something like that, you could just upload that, ask questions and get going. Amazon Lex expands Assisted Slot resolution regions and model access Amazon Lex Global Resiliency now supports cloud formation and existing alias replication. AWS Health now supports Internet Protocol version 6 IPv6.

[B]

Now we've got Compute AWS now supports the Zone groups for Availability zones across all AWS regions, making it easier for you to differentiate groups of local zones and Availability zones. Amazon EKS now offers new update strategies for managed node groups, which give you control over how Amazon EC2 instances in your clusters are updated with new configurations or for new kubernetes versions. This feature provides flexibility to make changes to your EKS cluster nodes in a way that best suits your use case while reducing operational overhead and compute costs. Amazon eks and Amazon Eksgistro now supports Kubernetes version 1.3.2. Kubernetes version 1.32 introduces several improvements, including stable support for custom resource field selectors and auto removal of persistent volume claims created by Statefulsets. This release removes version 1 beta 3 API version of Flow Schema and priority level configuration. AWS Elastic Beanstalk improves scaling and deployment speeds for Windows instances with EC2 fast launch AWS Elastic Beanstalk adds default support for EC2 launch templates when creating new environments. Next we've got customer engagement. Amazon SES announces that MailManager now supports defined email addresses and domain lists, which are used as part of the MailManager rules engine to distinguish between known and unknown addresses. This functionality adds both the mechanisms to upload and manage email addresses and domain lists, and the rules engine controls to make routing decisions based on whether a given address in a message envelope is on such a list or not. Amazon Connect now provides daily headcount projections in capacity plan downloads, enhancing your ability to review staffing requirements with greater precision. While capacity plans already provided weekly and monthly projections, this launch allows you to access details day by day account requirements for up to 64 weeks into the future. Amazon Connect Agent Workspace now supports audio optimizations for Citrix and Amazon Workspaces virtual desktops.

[A]

Next up, we have a few updates under Database Amazon timestream for InfluxDB now supports storage scaling, so this allows you to scale your allocated storage and change your storage tiers as needed. With storage scaling in few simple steps, you have greater flexibility and control over your time series data processing and analysis. Amazon Aurora PostgreSQL limitless database now supports PostgreSQL 16.6. Amazon Elasticache now supports one click connectivity setup between EC2 and your cache. Amazon Neptune now supports Open Source Graph Rack Toolkit this is a new capability that enhances generative AI applications by providing more comprehensive, relevant and explainable responses using the RAC technique. Combined with Graph data, this toolkit provides an open source framework for automating the construction of a graph from unstructured data and composing question answering strategies that quell this graph when answering user questions.

[B]

I really like this one because I've noticed from the customers I've worked with there's a lot of people that are both new to RAG and they're new to Graph as well. So I think now having a toolkit that can just help you to be able to get faster, especially as you're learning both at the same time, it's just going to help you to be able to build out your application and provide value faster. Now let's talk about developer tools. AWS Code Build announces support for CodeBuild Project ARN and CodeBuild Build ARN as IAM condition keys. These two new condition keys can be used in IAM policies to restrict the ARN of the project or build that originated the request. Starting today, Codebuild will automatically add the new CodeBuild Project ARN and CodeBuild Build ARN condition keys to the request context of all eight AWS API calls made within the build. You can use the condition element in your IAM policy to compare the codebuild project on condition key in the request contact with values that you specify in your policy. Anything that can help me write better IAM I am totally in favor of since I've had many spent many hours trying to bang my head figuring out how to write the right IAM policies to do certain things. Amazon Corretto January 2025 has a couple of updates so Amazon announced quarterly security and critical updates for Corretto long term supported and feature released versions of Open JDK Credo 23.0.2 21 0.6 17.0.1 4 11.0.268 U 442 are now available for download. Wow these folks are clearly busy. Amazon Corretto if case you are also wondering like me what this is. It is a no cost multi platform production ready distribution of open jdk. Now we're going to switch gears and talk about front end web and mobile AWS Amplify now enables developers to use the Amplify data client within AWS Lambda Functions. This new capability allows you to leverage the same type shape data operations you use in your front end applications directly in your Lambda functions, eliminating the need to write raw GraphQL queries.

[A]

We have one quick update under Internet of Things AWS IoT sitewise now supports Null and NAN or not a number data of bad or uncertain data quality from industrial data sources. Sitewise is a managed service that makes it easy to collect, store, organize and analyze data from industrial equipment at scale, and supporting these two different data types, the Null and nan increases its capability to handle a wider range of data, improving its utility for industrial applications.

[B]

All right, management and governance CloudWatch database insights now supports the analysis of historical snapshots of OS processes running on your databases, allowing you to correlate a spike in database load with OSS process metrics. Amazon CloudWatch Synthetics adds IPv6 support Amazon CloudWatch allows alarming on data up to 7 days old Amazon CloudWatch observability add on launches one step onboarding for EKS workloads and now we're announcing the general availability of AWS Managed Notifications, a new feature of AWS User notifications that enhances how customers receive and manage AWS Health notifications. This feature allows you to view and modify default health notifications in the console. And we have one more on AWS Health. AWS Health customers can now be able to use IPv6. AWS is announcing the general availability of AWS Managed Notifications in the AWS console. Mobile Application AWS Resource Groups is adding support for an additional 172 resource types for tag based resource groups. Now onto Media Services AWS Elemental Media Connect now supports a set of diagnostic metrics designed to provide visibility into the core quality of your video and audio streams. The new metrics detect black frames, frozen video and audio silence, allowing you to quickly identify and address potential disruptions.

[A]

All right, next up we have a couple of updates under Migration and modernization. AWS datasync now supports Kerberos authentication for self managed file servers that use the Server Message Block or SMB network protocol. AWS Transfer Family supports custom directory locations to store AS2 files or the applicability statement two files, including the inbound AS2 messages, message disposition notifications and other metadata files.

[B]

Now onto networking and content delivery. AWS announces the general availability of concurrent VPN connections for AWS client VPN all.

[A]

Right, well that networking and content delivery was our last update for today. Gillian, how can our listeners find you?

[B]

They can find me on LinkedIn. Ilianford Awesome.

[A]

And I am Shruti Koparkar on LinkedIn or also on Twitter or X. And that's it for now, everybody. Until next time, keep on building.