Cloud AWS re:Invent Coverage | Swami Sivasubramanian, AWS

Clips
More from Cloud AWS re:Invent Coverage

Swami Sivasubramanian

VP, AI and Data

AWS

play_circle_outline Value of AI agents in automating tasks and increasing productivity

play_circle_outline Use cases of Gen AI in various industries

play_circle_outline Role of data integration and fusion in Gen AI systems

play_circle_outline Optimizing Data Storage Accessibility and Analytics Efficiency Using SageMaker for Personalization Platforms

play_circle_outline Nurturing Resilience: Grounding Gen AI Apps With Data Guardrails and Continuous Mapping Development

Info
Transcript

Swami Sivasubramanian, AWS

Swami Sivasubramanian

VP, AI and Data AWS

At the AWS headquarters in Seattle, Swami discusses advancements in Generative AI, using deep learning neural networks combined with specialized accelerator compute. He highlights the transition to real use cases and deployments in 2024, with examples of Gen AI agents automating processes like software upgrades for time and cost savings. Data is crucial for enabling these agents and boosting productivity. Gen AI is being utilized in industries like healthcare and transportation planning to automate processes and improve efficiency, revolutionizing data analys... Read more

explore Keep Exploring

What business value did the Gen AI agents bring in terms of software upgrades? add

What examples can be given of companies leveraging artificial intelligence in the healthcare and automotive industries? add

What are some examples of how Gen AI technology can be used to analyze data from relational databases and generate insights for businesses? add

What are some key considerations when storing, analyzing, and making use of data in the cloud? add

What innovations are being made in the area of data-driven applications and contextual grounding, specifically in terms of building resilience and security for Gen AI? add

bolt Powered by CUBE AI

Swami Sivasubramanian, AWS

search

>> Hello and welcome to theCUBE here in Seattle at the headquarters for Amazon Web Services. This is the re:Invent building. We're right above the spheres outside. I'm here with Swami, preview of re:Invent. Swami, great to see you here in your home turf. It's an away game for me. Great to be here in the studio at AWS. Good to see you.

Swami Sivasubramanian

>> Hey, great to be here. Thanks for coming over.

>> Yeah, you're looking like you're got spring in your step. You're smiling. Re:Invent's coming up. You're excited for your keynote every year. You got a great keynote always. Always a great watch.

Swami Sivasubramanian

>> Hey, thank you. Thank you. Lots expected to come, and I'm actually literally just walking from a re:Invent keynote table read right now, so very, very excited.

>> I can't wait. I know you won't spill the beans. Now, I won't even ask questions because there's under heavy embargo, but one thing I do want to talk about is the AI wave. Obviously, you've been on theCUBE since theCUBE started and you've done all the work around databases. You were part of that machine learning team that built out the core pre-Gen AI, so you saw it before. In fact, we've talked many times around the role of data and how AI is good work. And then now, Generative AI, which is a kind of new category, it's generating, so it's not like a static thing. There's a lot more action around it. And of course, the hottest thing on the planet is agentic systems or agents. And so, this is where the data becomes super valuable, and this is your wheelhouse. I know you've got a great vision on it. So, what is your vision for the state-of-the-art of as the infrastructure gets levels up to the performance we're seeing that democratizes it, the data layers are going to be harmonized, all kinds of new semantics, all kinds of stuff, is having at the data layer that will enable this agent. Multiple agents. You got AI-distributed infrastructure, you got routing challenges. There's a lot of technology involved to make it all work. What's your vision?

Swami Sivasubramanian

>> Yeah, no, actually, if you look at what's happening in the Generative AI, we are having this moment because a lot of things came together to build up to this moment. If you see, it's like deep learning neural net papers happened like 30 years ago, but then, it required deep amount of computational capability and then it needed huge amount of elasticity. But then, transformers came along and that architecture when that unlocked the ability to learn in an unsupervised manner, but then, cloud and the specialized accelerator compute with things like GPUs and Trinium and whatnot, and the ability to have seamlessly infinite storage suddenly unlocked this whole new wave of large language models. That now can learn from huge amount of data that led to the moment where we are in the industry. But if you look at what's happening in right now, 2023, I would probably call it as the year, where people were doing a lot of proof of concepts, 2024 is where you are now seeing, people are suddenly saying, "Okay, now I got to figure out which ones are actual real ROI and which ones are, how do I actually convert it into real-use cases that benefit to my customer and how do I deploy these Gen AI systems and agents that actually saves money or increases revenue?"

>> Yeah.

Swami Sivasubramanian

>> So, I'll give an example of my favorite agent that we use within Amazon. So, I use Q, the developer one even now when I write code and whatnot. But the part while everybody gets excited about Gen AI agents in terms of coding and so forth, the actual business value that was disproportionately important when it comes to Q was actually it's agent for doing co-transformation software upgrades. But then Amazon, one of the things we did was actually ask Q, the agent to say, "Take this Java package upgrade from JDK 8 to JDK 17." And what it did is it inspected all the right packages and the dependencies and said, "You know what? To do this, here are the changes you need to make and ship a code review and then do it." And almost like more than up to 90% of these code reviews where it just accepted by our developers, and this year up until now, we saved at least up to 4,500 developer years. People always think it's hours, actually just years. That is a huge amount of savings, let alone more than $250 million worth of CapEx savings. But that is just one example of what an agent can do in a real-world setting. But this is where, again, Q is built on top of Bedrock and all the agentic capabilities we offer. I actually think the world we are about to get into, if we map the business problem to actually your data and actually build with the best-in-class LLMs that are available. Suddenly, the efficiency of productivity is going to be like that TenX, And that is what we are very excited. That's why you already see hundreds of thousands of customers are already leveraging machine learning in a big way on AWS across our generation.

>> Productivity is the killer app here, and that's what we're talking about.

Swami Sivasubramanian

>> Yeah.

>> Productivity. I was talking to a developer and I said, "Hey, what's the impact of some of the agentic stuff." "We get low-hanging fruit we're doing now," use cases that we can clearly see whether it's automating JIRA tickets or doing code transformations as you mentioned. I said, "Okay, what's the value of that?" "No, it's more beer time." It's a direct quote, "More time to drink beer with my friends." And that was like, remember the old days? I don't have to get paged. And then you had that kind of innovation. So, that was the beginning of automation. So now, more beer time means more time to do other things. And the other fallacy that's out there, I want to get your reaction to, is that AI will replace the developer. Well, developers love to code. They don't like to debug or do all the legwork code transformation. That's like a slog. Who wants to do that? But if you know it could be done, again, this is where you start getting into every process. Every single process and every single workflow has an opportunity, whether it's code refresh or my marketing plan or manufacturing. The digital twin concept is moving into what agents will do on their behalf on any problem.

Swami Sivasubramanian

>> Yeah, actually, you're touching on an important one. Just take the software one and then I'll talk about what it means for healthcare and others. What we did with Q is there is a lot of excitement around code generation. As you said, developers really like writing code. They will take the help and Q is the best at it. And if you go see SWE-bench, you're always in the top one or two all the time. But the part actually what we took with approaches to help them also save time doing the drudgery associated with doing security scans or software upgrades or various other things. In the same way, now, one of the big areas we are seeing in terms of our customer adoption is also leveraging Gen AI to do things that are right now, very, very manual and boring. For instance, right now, if you see in the healthcare industry like Pfizer, they came and talked about how they're able to leverage Gen AI to automate huge amount of processes and even realize up to $750,000,000 to $1,000,000,000 savings across their whole value chain of the processes. But it requires essentially mapping bunch of work that they are currently doing and then being able to apply, bring in data and AI together and do this. And this is what they are actually building on top of our Gen AI stack with Bedrock and others. They talked about it even last year, but Toyota is another example. They actually leverage, in this case, Q business and they threw in all their operational run books and various other things, so that developers spend less time actually figuring out what is the best way to actually handle certain scenario. And again, when you have access to data at your fingertips, it just changes the game. Like Praerit Garg, President of Smartsheet, said they deployed Q without a writing single line of code, and it is the built-in NASMI part in Slack, and the 3000 employees can now, in Slack, ask any questions and get response back.

>> I was talking with one of your leaders yesterday here at AWS and she was talking about we go to the end customer. So, you have a customer like say Delta Air Lines, for instance. Their customer is the people who fly and use the service and then work backwards. This is the classic working backwards document. So, you're touching on something I want to get into because at the end of the day, you have customers and customers of the customer and the end-user customer. What you're getting at here and what I love about the cloud, I'll say 1.0, I know you guys don't like that term, but the beginning of the greatness of Amazon was it changed the labor equation. The labor was the developer. I don't have to provision a data center or get a box. I put my credit card down, I start a company called Airbnb or whatever, and next thing you know, it scales and it accelerates, scales and becomes a rocket ship. That's the progression of cloud 1.0. Cloud 2.0 Is a different animal in the sense that you now have a different labor shift. You're getting at the business people. So, inside the end-user customer, like at Delta, they have developers and they use Amazon, right? So, you have developers will always have that relationship to the business. But now, the TenX engineer, which came out of cloud famously quoted, is now the TenX or more business professional. I hear stories of people using the queue and other tools to write queries, SQL queries. They say, "This is what I want to get out of the data. Write me a SQL query," and actually do the SQL query directly. So, this is unprecedented. Who would do that? No one would ever sit down and learn SQL if you are on the business side. So, the labor on the business side is, so I want to get your reaction to that. And two, I want you to talk about the domain expertise value of that individual. That's the new IP is what's in the head of the human. So, the labor shift to productivity to the business. Do you see a TenX step function?

Swami Sivasubramanian

>> Actually, that's a great one to dive into. If you see, you rightly framed it that cloud has changed the equation and made developers like TenX productive that they don't have to do the undifferentiated heavy lifting. Now, if you see what is happening with the Gen AI, well, the business users are now actually can be like TenX productive. I actually had one of the companies which does transportation planning. They built on top of Bedrock, a Gen AI assistant that can tell why is this exit connecting to iFi crowded at this time? And their transportation planner can ask that question and it'll say, "It is because there is an event happening," like a construction happening based on, and it answers it based on the real time streaming data flows into the cloud and it maps it and generates. And then you can ask a subsequent question saying, "How do I actually make this more efficient?" And it says based on actually the past projects that we have seen here is the way to do this better and schedule it at this time. So, this kind of workflow usually used to take weeks to do this. Now, it can be done in a matter of minutes. And to me, these are just the beginning of how the world is about to change. That's what makes this so exciting.

>> I think I'm more excited about this history in all history, this inflection point because it takes all the other ways, almost puts them all together. And you bring up a good point. The role of data is super valuable. That example of the construction and the freeways could be applied to anything. Supply chain, my marketing program, audience consumption, let's say theCUBE videos, everything has data. And what you're getting at here and the focus of this agentic system is for the first time in history, the ability to put data sets together faster.

Swami Sivasubramanian

>> That's right.

>> And harmonize and or integrate is here. And it reminds me of the old days of when APIs came out. Oh, yeah. People think, "Oh, well, APIs are great. I just connect two systems. I got to rest a REST API and then now, I'm connected." And Gentic is like, "Well, I can just connect to that. Now, I have instant value." So, it used to be hard and you've done this work. So, tick-tock through the role of data because it's not just about the LLM, it's the small language model, secure language, it's the sovereign language model. So, you have data now is, I won't say fragmented, but just categorically everywhere and different. So, the integration and the fusion of the data.

Swami Sivasubramanian

>> I think-

>> Talk about your vision on this because I think this is a key piece.

Swami Sivasubramanian

>> So, one of the key things that for us to realize the power of Gen AI is that we need to bring in, so far all the LLM excitement has been all things around unstructured data, purely text and maybe images and so forth. But one of the big things we actually see, especially for enterprises and next generation applications to do is to reason about data sitting in relational databases and data viruses or actually other kinds of data. So, actually, this is where Q and QuickSight is a good example that now you can actually build a PowerPoint presentation like data stories by just asking a question, saying like, "Hey, show me the value of how good is my free trial program." And it actually generates an entire slide deck based on it that used to take a week. Now, a different kind of data and favorite customer of mine, BrainBox AI, they actually had an building efficiency assistant called ARIA, I think. And using data system, they're able to query the interior building schematics and energy consumption and then us saying, "Why is there increased consumption in this floor and what is the reason behind it?" You can dive deep, and then it says, "It's because this has not been serviced this much, so here is the way to fix it." And now, you can actually run this in a much more efficient way using these kind of technologies. So, the ability to get insights on what used to take from months to minutes is going to be a game changer.

>> Scope that alternative that the old way and new way because that's the way I like to think about this wavering. In the old way, what would it take? Just kind of lay out the concept of what the steps would be to merge data together. I go to the data warehouse, I export the data, I build the connector. Take us through the old way and then the new way.

Swami Sivasubramanian

>> Yeah, actually, so let's talk through, actually, I built RDS or Relational Database Service, heard of it along with Dynamo. So, one of the questions-

>> That's why I asked the question.

Swami Sivasubramanian

>> So, if I had to think through what is the best way to determine the optimal free trial program for Dynamo, the way to actually do this is first, figure out what is the actual typical usage pattern on how long does a customer, once they sign up to, when they actually run production use, you actually track this event history, store it in a relational data warehouse, and then work with your data engineer to export this data first into the data warehouse and then run a bunch of queries and then say you make a list on what is a percentage of customers who actually use it on day 1 to day 2 to day 30 to day 60. And then, you actually go through a one-month analysis before you recommend.

>> And you get the schema set up first.

Swami Sivasubramanian

>> With the schemas-

>> Man, that's work.

Swami Sivasubramanian

>> Yeah. So, all this. So, you need almost like a team of data engineers to BI engineers to product managers before you get to business station. Now, foster it to the world that we are in with Q in Quicksight.

>> Oh, that would take how many days, weeks?

Swami Sivasubramanian

>> It's probably like I worked on this exact problem and it took us close to three to four weeks to do this after schema and data warehouse.

>> Okay. So, yeah. Even longer.

Swami Sivasubramanian

>> This was 15 years ago. So, now, with Q in QuickSight, the ability to actually do this kind of analysis has just changed to a matter of minutes. You can actually, now that you have this kind of the ability to ingest data with Zero-ETL to now being able to now actually have QuickSight query this data through Q, and then build dashboards and then get a graph on, what is the conversion rate for each of this? It is like a game changer. When you think about the speed of decision making. That's why your point on labor shift on across for business user and making that TenX product is going to be a big deal.

>> Yeah. And that's going to have that same step function change that you guys did on the first wave, which was clearly developer labor and then enterprise labor for tech. Now, business, you already have that. That continues to go, it's not like it's going away. So, you've got developer, productivity is still going up. Now, you've got the business side and then now, you have the C-level executives mandating Gen AI. So, you're starting to see a lot of activity around process analysis. So, in the enterprise where it's super hot, and by the way, in the entrepreneurial circles right now, in Silicon Valley and in New York, where we're monitoring as closely is that most of these young entrepreneurs that are under the age of 30 who are going after these go big or go home concepts are all doing enterprise stuff because the end-to-end workload and the problem sets are in the enterprise. So, the data opportunity, they're going at the data or going kernel level system programming. So, it's a system architecture and data. So, the number one opportunity we're hearing is process and what's the data for the process and then what new data could be there, and most customers I talked to say, "We never even thought about this before because it was too hard." It was never even on the table to even think, "Let's merge location data or weather data. What's available now? Give me all the data you got."

Swami Sivasubramanian

>> Yeah, actually that is another thing you'll see more and more as data, cloud, basically made the data storage a lot more accessible. You don't have to actually then pick and choose which data you store, but then just storing them alone is not enough. How do you actually make sense of it and putting together them to do efficient analytics and then being able to do machine learning on top of it. So, it's one of the key areas of focus. And SageMaker has already done a huge amount on this front and being able to bring together all of it. That's why you are already have among the hundreds of thousands of customers using SageMaker, you already see Intuit, for instance, is able to build a personalization platform using it. Or booking.com last year at Reem when they talked about how they were able to build an amazing capability for travel assistant and so forth. But what is changing now is also the kind of data is not just about data sitting in a relational data warehouse and so forth. Multi-modality of data is also now becoming more and more accessible because of these large language models, especially because they can generate embeddings. And the moment you're able to seamlessly combine these, it is now suddenly, you can actually create remarkable meaningful value.

>> Matt Garman said, and we've been saying on theCUBE all along that the killer app is productivity. But when I asked, we asked all the practitioners out there how they view Gen AI. And Matt actually mentioned in my interview with them, Gen AI is just another application. And so, we've all know how applications, you do AppSec review, you do all kinds of stuff, there's security concerns. You know what that work is. So, when you look at applications and what power the old applications and say Gen AI has a new application now, it's different, got all these benefits, the word resilience has not yet been defined. So, I want to ask you, and resilience is discussed a lot in security like ransomware, how do we roll back its recovery basically. LLMs and small language models, as they work together, they got to be accurate and they got to have some sort of SLA. You can't be wrong on a finance app or Q can't be wrong on it's coding. So, quality is critical, but also data. Sometimes, someone might get data in. How do you roll that back? So, what does resilience mean to you? How do you view resilience when you talk about Gen AI? Because people want the resiliency, they want the security, they want to have confidence.

Swami Sivasubramanian

>> Yeah, it's a great question. This is one of those where we are investing serious amount of innovation. I'll just start with first that in this area, first you want to be able to build these data-driven applications and you want them to be contextually grounded with the right data, so that the answer that you generate, you actually have it. Even in New York Summit, we launched like contextual grounding as an example. But I actually think you'll see us doing more and more in the form of Guardrails, where you already see Guardrails first provide the first layer of things where we is able to actually save things like prompt injection attacks and various other things, so that, and it is more than 85% accurate compared to the native methods provided by foundational models. But what is missing, and you'll see us doing more in this area, is actually being able to map the data and the LLM responses and actually continue to provide more and more contextual grounding in this way. And we have some lot of interesting things in the space that you're working through and you saw the beginnings of it in your Summit and you'll see more and more on this front.

>> Well, I'm super excited to have this chat with you and we'll certainly have my agent call your agent for a re:Invent interview. A lot going on. It's a super exciting area and you're leading the charge at AWS on the Gen AI and the whole AI mission. Thanks for taking the time. Appreciate it.

Swami Sivasubramanian

>> Oh, thanks again.

>> Good to see you. Always a pleasure to see you. I'm John Furrier with theCUBE here at the AWS headquarters. This is the building called re:Invent. This is theCUBE coverage. Thanks for watching.