Name: Rahul Pathak, AWS
Uploaded: 2024-10-16T18:36:00.000Z
Duration: 19 min 38 s

Clips
More from AWS Generative AI Analyst Summit

Rahul Pathak

VP of Generative AI & AI/ML GTM

AWS

play_circle_outline Revolutionizing Customer Interactions: GenAI's Role in Boosting Productivity and Saving Major Corporations Millions

play_circle_outline AI's role in augmenting existing business processes and driving value

play_circle_outline Customers investing more in GenAI and seeing tangible results

play_circle_outline Importance of data quality, security, and governance for successful AI deployment

Info
Transcript

Rahul Pathak, AWS

Rahul Pathak

VP of Generative AI & AI/ML GTM AWS

Rahul Pathak, the vice president of generative AI and AIML go-to-market at AWS, believes that GenAI will revolutionize customer interactions. Companies like Pfizer, Amazon, and banks are already benefiting from GenAI. AWS stands out in GenAI spending, although industry spending overall is growing slowly. Customers are seeing productivity gains, cost savings, and improved satisfaction. AWS is focused on making AI models accessible through platforms like Bedrock and Q. Data quality and governance are challenges in moving from POC to production with AI, but AWS ... Read more

explore Keep Exploring

What are some examples of businesses utilizing GenAI in production to drive business outcomes? add

What are some examples of the productivity gains and cost savings seen with the use of AI in various industries? add

What are customers investing in and how are they navigating priorities related to AI technology adoption and cloud services? add

What factors contribute to the ability of customers in regulated segments like financial services or healthcare to move quickly in GenAI implementation? add

bolt Powered by CUBE AI

Rahul Pathak, AWS

>> Welcome back to our special CUBE presentation. We're here at AWS. We're on the ground in Seattle. My name is Dave Vellante, I'm here with Rob Strechay, my co-host, and Rahul Pathak is here. He's the vice president of generative AI and AIML go-to-market at AWS. Good to see you again.

>> It's great to be here, David. Good to see you too.

>> New title, Rahul. Cool. Congratulations.

>> Thank you.

>> How long have you been in this role?

>> I've been in this role about four months and it's fantastic. It's awesome to get to work with our customers and be at the forefront of them figuring out how to put GenAI to work in their businesses.

>> Well, if you can get GenAI in your title, that's good. So where are we at? You guys have been going crazy. Bedrock is rocking, no pun intended. It's changing everything. I think I found one person in the last two and a half years who said, "Ah, I'm not quite sure about this AI thing." Well, if he's right, we're in big trouble, but we're betting the farm on GenAI.

>> Yeah, we're definitely of the belief that GenAI is going to change just about every interaction that our customers have with us across all of our businesses and we're investing with that in mind.

>> How are you seeing that? Because I think, again, big part of AWS is working backwards from the customer, and working backwards docs, sat there and done those myself, so very much, what are some of the use cases and how are some of these customers really seeing productivity gains or what are they doing to get there?

>> Yeah, so it's really very diverse at the moment, and I think what we're seeing this year is really GenAI getting put to work in production, driving business outcomes, whether that's faster time to market, better productivity. So we see customers like Pfizer using Bedrock training it on all of their clinical trial data and they're saving researchers thousands of hours a year in finding relevant information. We've got Q Developer, we've used that within Amazon to upgrade all our Java packages that saved us over 4,000 developer years and a couple of hundred million in CapEx. And we've got customers in the banking space, they're using it to generate customer proposals. So this is directly translating into real customer value and production.

>> So broadly speaking, we look at our data generally speaking, I would say that a big chunk of customers, about 40-plus-percent say that they're stealing from other budgets to fund their GenAI. Okay, maybe that's happening in AWS, but the overall macro is probably growing it spending maybe three and a half percent. Obviously AWS is growing much, much faster than that. So I feel like AWS is a very large outlier, like you're at really at the tip of the spear, you and perhaps other hyperscalers. But what are you seeing in terms of tangible results? Because that's the big narrative now is we're spending know we're putting some time and effort in when I feel like we're not getting the big ROI, we're hitting singles, and that's okay. What are you seeing in your customer base?

>> So we're seeing customers get real concrete value, and it comes in a number of buckets. I think the fastest ones to realize are the productivity gains, the ability to do more with less, and I think AI's really been a force multiplier and I think customers' appetites haven't gone down. So what they look at is, "With the teams that we have, how can we really do more and get people more productive and moving more quickly?" And that's coming across both for developers as well as for their knowledge and information workers and their other personnel. And then we also see costs come out of the business in things like support, customer service. You can get faster, more personalized responses that increase customer satisfaction and lower.

>> So thinking about some of the best examples, if you could in your mind? What we're waiting for broadly is for AI GenAI specifically to be self-funding, throwing off enough cash enough ROI, if you will, so that the CFO says, "Okay, keep spending," because making money like Jensen, buy more, save more kind of thing. I love that, Jensen's Law is going to replace Moore's Law, but when you think about your best customers, the real ones that are driving value, is it getting to the point where it is self-funding and it's loosening the purse strings inside these organizations?

>> Absolutely. Our customers are seeing real benefit and they're actually looking to find ways to invest more. And I think as you pointed out, a lot of the growth that customers are investing in is in really getting themselves enabled to take advantage of AI. And that's both spending directly on AI, but also on modernizing their data state in order to be able to plug that into the AI technology they're adopting.

>> So the interesting thing is observing many, many waves in the history of tech, when the new stuff comes in, it's oftentimes not big enough to offset the decline in the old, but this was really pre-cloud. Now cloud is different. It's a lot easier to just bring on new services and use your cloud credits elsewhere. Much more flexible. How do you see that dynamic in terms of, a lot of people are fearful that it's going to change their SaaS spending and the like, and there was some experiments going on. You may have seen the Klarna, I think that was maybe a little bit overstated, but instructive, and clearly AWS continues to grow. It's much, much faster than the overall industry. So how are you seeing customers navigate those priorities?

>> There's a couple of dynamics going on. I think one is AI is additive. So what we're finding is customers are using it to augment their existing businesses and to find ways to do new things. The second is, in order to take advantage of AI, you need access to new compute, you've got to get your data into the cloud. And so there's that movement from on-prem where most of it spend still currently lives into the cloud. And I think that's really contributing to the overall growth that we're seeing. And there's a great story that's used to tell, "But when we invented vacuum cleaners, we didn't spend less time cleaning, we got bigger houses." And so I think it's really about the ability to do more with better technology and our appetites are to continue to grow, to continue to innovate, and AI is really enabling that.

>> So one of the things that we also see as a key to this and really the adoption is really the openness of it or the democratization of it I guess you could say. And I think you know, AWS was early on with saying, "Hey, we're going to democratize it." We're not, "Yes, we have our models, but bring use any of these other models." Are you seeing that customers are really embracing that democratization aspect of it and things in Bedrock and Q and things of that nature?

>> A hundred percent. So we've been believers from the start that there wasn't going to be one model to rule them all. And yes, we're investing in our own. We also want to make the best of what's out there available to our customers. And so Bedrock is all about democratizing access to models, but also providing integration with data through things like knowledge bases, support for agents, support for guardrails. So you can build apps that are stable but yet flexible. And as you see with the models as an arms race every three weeks as a new model on top. And so it's really important for customers to be able to take advantage of the best of what's out there and to maintain that optionality. So that's been the core of Bedrock. And then Q is really about enabling customers to build GenAI into their experiences for developers or their broader employee bases or customers without having to do a lot of custom development. And it's really about customization and embedding. That's been a big win as well.

>> Because under Q, they're not picking the model, it's more, "Hey, we're going to bring the best model for that." So almost like having an SLM for them, that's built on an LLM, that type of thing.

>> That's the core idea. So it's really about Q providing the best overall generative experience to the customer. So with Q Developer, it's how do you enhance the full software development life cycle with generative AI. Q Business, it's how do you interact with your data with generative AI and underneath we're picking the best combination of models that are needed to answer those questions with the right guardrails to keep the answers focused to the job at hand.

>> We're big believers in domain specific models. I wonder what you're seeing there. it's great. These LLMs are unbelievable. It's like magic and they're trained in the internet, but so much data is not on the internet. My proprietary data is not out there. At least I hope it's not well.

>> In fact, I saw a stat that only 1% of corporate data is actually in LLM, so it might be by mistake, right?

>> It probably is. So you can imagine the opportunities for companies to build on that proprietary data. What are you seeing in that regard? Maybe LLMs becoming S SLMs and maybe even small or maybe even not so small action models. You've mentioned agents before. What are you seeing there? I know it's early days, but...

>> Definitely a dynamic space. I think you're absolutely right. Data is a key differentiator. So if you follow our strategy of trying to democratize model access, if everyone has access to the same models, how are you going to stand out? You're going to stand out with the things that you uniquely know. And so that data that you have about your business, about your customers, marrying that with the LLM in a secure way is the way to get differentiated business value. And so that's on the data side, and I think you can think of it in a number of ways. There's the rag aspect of it with access to your databases, there's prompt customization and then there's the fine-tuning and distillation of models into smaller models that are more use case-specific. And all of these have a role to play.

>> So you went there, let's go there. Data powers this all, right? And I think that a lot of what we see in the ETR who's a partner of ours, their data around this indicates that the two biggest problems from getting from POC to production still, and it has been for the entire length of the survey for over a year and a half now, I believe, is really been... Data security and data governance are the two biggest things. What are the other data issues and what really are you seeing and how are you helping people get past that? Because that's the important thing to get from POC to production. People have to have that trust.

>> You're totally right about the trust and security has always been job zero to AWS. And then when you think about data, it's security, it's privacy, but it's also good governance. There's a quote that I love which says, "Discipline equals freedom." And that really applies to data governance. When you've got the right guardrails in place, you can actually set the organization free to innovate. And so we're really focused with our customers on thinking through what data do you have, who has access to it, how does that go to the model for the right response? And the other really important pieces that we've really focused on making sure that any customization our customers do is private. So if they fine-tune, it's based on a private copy of the model that's only for them never leaks into the main model, they have full control over the inputs and the outputs. And so once you can establish that secure foundation, then customers can really let their teams loose to find new ways to deliver value for their customers.

>> So I think AWS has established itself as trusted in terms of not leaking data to the LLMs. So let's assume you've got that problem nailed, as well as anybody else in the industry. There's still a data quality issue, which is in some respects not in your control. The customer's got data that's locked in different applications and there's data, there's metadata, there's business logic, and it's not consistent, it's not harmonized. How do you see organizations approaching that in the AI era? In other words, being able to tap data sources that are... Sometimes people talk about the semantic layer, but just semantically consistent or harmonized. So that the example I always use is revenue. What's revenue? "Well, it could be revenue, it could be bookings, it could be AR, could be NRR, could be calendar year, it could be fiscal year." And you spend all your time in the meeting arguing about that, and then the meeting's over. What are you seeing in terms of whether it's technology or process to harmonize that data layer so that you can serve it up to agents in this future AI?

>> So garbage in, garbage out definitely applies. And what we're seeing is the customers that have done the most work to establish a solid data foundation are the ones that are actually able to move fastest in GenAI. And interestingly with this wave, a lot of our customers in regulated segments like financial services or healthcare are actually able to move fast because they've done a lot of the work on governance and entitlements and data quality and data definition. We're also seeing though that you can apply AI to start to make sense of this wide variety of disparate data sources. And so data integration is actually well-suited to being tackled by AI. You can use AI to write pipelines, AI to semantically match different things with higher confidence. So you're totally right in terms of establishing a good foundation, but I think we have better tools to do that than we ever have had. But it is not just a tooling thing, it's also a people process policy type of conversation.

>> I almost feel like the technology's getting better. And also, to your point, people are managing the deployments with low risk stuff they're not starting with... And the example you like to use is autonomous driving. They don't start there. Let's start with something where you can absorb some degree of unreliability in the data and the hallucinations and it's lower risk and then work up from there. Is that a viable sort of mental model as to how this will progress?

>> It's a great mental model and I think there's two things to think about. So one is... I actually am a big fan of AI applied to the mundane aspects of business because I think you can really take a lot of drudgery out and drive a lot of business value. A lot of cost is tied up in that. The second is a lot of key processes already have human in the loop steps. You're not going to submit audited financials without having someone check it, but AI can get you to that review step much faster than you could before. So we're finding that combining these two, the mundane, the enterprise, internal internet chatbot, everybody hates their internet search because you just get back this morass of links. But if you can ask a question that gets you a tailored response to your situation, suddenly employee satisfaction goes up. You don't have to invest as much in people having to answer that Q&A. And so you've got something that is simple and safe and really driving value and creating budget space for customers. And then in the human in the loop, you can get to a draft in seconds instead of in days, and then you can look at that draft and make it pass all your usual downstream checks before you turn it into something of consequence. So there's a ton of value to be had, and I really believe that our customers who are adapting their workflows to take advantage of AI will out-distance their peers very quickly.

>> How do you see, because just building off the human in the loop, and I always think of financial services, and I'm using this to do some analysis on portfolios, but I don't want to go make a recommendation. I don't want the SEC to say, "Well, you're blaming the model for giving the right..." Are you seeing and how are you helping organizations understand? Because the regulatory aspect of it is just changing so rapidly. You have the EU AI Act, you have other things coming out. You have, as you put this into an application, then if you're in France or in the EU, Dora hits and you have to actually... "Now I have to be able to HA my GenAI. How do I do that?" What are some of... There seem to be a new set of problems beyond just regulatory, but regulatory to me seems like one of the fear factors out there for people.

>> It's a bit of a fear factor, but it's a fact of life. I think anytime we have a new technological leap, regulation catches up with it and people learn to adapt. And I think we saw similar things when we had the first wave of machine learning come in. You had ML driven models, you had regulation around keeping traceability, repeatability. We're going to see the same things with LLMs. And what we're really trying to do is to give our customers the tools and the information that they need to have the right conversations with the regulatory regimes in which they operate. So to me, it's something that's just a fact of life in terms of how we do business and we will figure it out together because I think the gains to be had collectively are significant. And I like to say that the most secure computer is one that's not connected to the internet or to electricity. It's also the least useful computer you could possibly have. So we've got such a powerful technology. Yes, there are some risks and yes, we'll figure it out and we'll put it to work.

>> When you think about how LLMs and GenAI are so evident today, everybody's touching in, they see them, that's all we talk about. Do you see that changing? Will they just be subsumed and embedded at some point in time?

>> I think it'll be a bit of both. I think we'll continue to have really interesting, unique, innovative interfaces, but I also think they will get embedded. What we're seeing with agents, as you're getting more autonomous activity, you're seeing things become more ambient. So today most of us are used to a Q&A format. I think we'll get to a place where agents will observe what we're doing and make suggestions. It's just almost inevitable. So I think as things become more mainstream, they'll get more embedded. But there will always be a bleeding edge where we'll have a wow factor of things that we interact with.

>> Well, and there's so many things that we talking about how most of the data or much data is not on the internet. If you think about the processes that we initiate in business, a lot of that... Most of that, actually, can't be hard coded in microservices and so it can't be automated. Do you see eventually agents being able to, you caught my attention, but you said they'll be able to observe human actions and then over time learn and sort replicate them or even initiate a new workflow or present a plan back to humans and say, "Hey, you might want to try this or that," and have some kind of interaction with the agents and an iterative process. Is that a midterm, near term or even midterm vision or is that like 10 years away in your view?

>> I think it's much closer than 10 years away, and we're starting to see it today. The interesting thing with agents is rather than the declarative way we generated workflows before, now we can say, "Here's a goal. Here's a set of things the agent can do. Here's access to a model. Figure out the best way to get from A to B. And, by the way, present the plan that you're going to follow to a human. Let them agree with you that the plan makes sense or let them modify the plan, then let's go execute it." So I do think we will get more goal-oriented behavior with agents and that'll allow us to deal with situations we haven't seen.

>> So it'll understand those top-down goals, it'll understand the metrics by which it's being measured, grow market share, or have higher margin or whatever it is, and then execute a bottoms-up plan.

>> Well, I think we might be a little ways off from the, "Find me my next billion-dollar idea," agent, but I do think that figure out how to maximize this objective. We're probably closer than we think.

>> Well listen, hey, great to see you again. Congratulations on all the good work you're doing and look forward to having you back.

>> It's great to be here. Thanks so much for having me.

>> Yeah, you bet. All right, thank you for watching. This is Dave Vellante with Rob Strechay. We're here on the ground at AWS in Seattle. We'll be right back right after this short break.