No Priors 第139期｜与 Snowflake CEO Sridhar Ramaswamy 对谈

Today I'm here with Sridhar Ramaswamy, the CEO of Snowflake, the former founder of Neeva, and the former SVP of Google Ads. We will talk about his first 18 months as CEO, the incredible execution over that time in shifting a company at scale to being AI-first, where the enterprise ROI is, and what happens to cloud service providers and the ads model in the age of AI. Welcome.

Sarah, really excited to be back.

It's a pleasure to talk to you as an old friend and colleague. The last time we spoke, you were on the entrepreneurial journey.

That's right, doing search still.

You're now 18 months into being CEO of Snowflake. It has been a very eventful 18 months. I think the markets reacted in many ways, most recently incredibly well to the execution, but it's been a journey. Tell us a little bit about the journey from taking the mantle from Frank to the first few months and where you are today.

That's right. Snowflake has always been an amazing product company. The original product that Benoit Dageville conceived of 10+ years ago was many years ahead of its time, and it took the world by storm. They had a storied IPO, the biggest software IPO at that time.

I think what happened was the company was a little slow to react to changes from things like machine learning and AI. That was honestly part of the reason why Frank voluntarily pushed for the change, because he felt presciently that we were headed into a time that was a lot more tumultuous from a product perspective, and he wanted someone who was product-first to be in charge of the company.

The last 18 months have really been about embracing that wave of change. If you look back at what's happened in the last 2 years, it's crazy how much change has happened with respect to AI, how it's become commonplace every day in all of our lives, and the speed at which things are still getting driven through.

I think the really amazing thing about Snowflake is that the company embraced this change, transformed itself, and then showed that not only could we do it from a product perspective, which one could have expected, but we have also done significant things to retool our marketing and our go-to-market overall. I think that transformation has been pretty amazing to watch.

Times can be difficult. Last year, there were a lot of doubters, but there were a lot of us who believed both in the value that Snowflake was already creating and in the reason I took this job. I talked to a whole lot of customers before I became CEO, and they all loved Snowflake. That was a big motivation for me to take this job.

I think we have successfully ridden through that and are now at the cutting edge of data and AI for enterprises. It's been an amazing journey to have gone through.

Walk me through some of the orientation and prioritization you did in the first 6 months, and then tell me a little bit more about the long-term vision here.

The first 6 months were a lot of tactical changes, primarily around accountability. Like every company that goes through essentially a rocket-ship phase of growth, growing at 100+% year on year, Snowflake had basically specialized at every layer possible. There was a very long distance between the engineer who did a feature and the customer who made use of that feature, and there were 7 to 10 layers of teams involved.

That works fine when you have perfect product-market fit and you're trying to optimize for every function.

You're just the winning cloud data warehouse.

Yeah. Drive a truck through that. But on the other hand, if you're working in the world of AI, where we can barely tell what's going to come out next month, forget next year, this is the wrong structure to have.

We did a lot of organizing by different areas, making sure that there were accountable people. For example, in product and engineering, that was among the first changes. We organized into different product areas, like AI or the core warehousing and analytics product.

We also wanted a straight line over to our go-to-market team. So we created specialized teams that work closely with product, engineering, and marketing to take these new products to market. That was a lot of the early phase of Snowflake, with an emphasis on faster iteration.

This is something that I've believed in all my life: speed wins. The ability to iterate always trumps carefully laid-out strategies. You shouldn't do dumb things, but realizing any kind of gain requires a lot of iteration.

We made a number of changes on that side, both with respect to how quickly we created products and how quickly we iterated with customers. I would also say we took a little bit of time to find our sweet spot in this AI space. That itself has evolved a lot. We are not a CSP, and we are not a foundation model lab, so what are we?

There was that discovery of ourselves as the AI Data Cloud, as opposed to the Data Cloud. I think it's that kind of clear product insight into what value we add that is setting the stage for the earlier parts of the year and even for what we are about to talk about today.

Now you're announcing Snowflake Intelligence. Tell us about that and how it fits into the broader vision.

First of all, when it came to AI, as I said, we had to look hard at ourselves. Early last year, we actually went down the path of creating foundation models. We created a credible model, but we also quickly realized that our ability to compete with the likes of OpenAI or Anthropic was going to be really hard. We simply did not have the capital to be able to invest meaningfully in things like that.

So we pivoted away from that into more of a question of how AI can massively accelerate what can be done with data that is in Snowflake. Over time, that can become a reason to bring more data into Snowflake, which is the phase that we are in now.

A lot of our AI product strategy was actually quite humble. It didn't say, "We are going to rethink everything." It said that an enormous number of customers—something like half of the qualifying Fortune 2000 companies on the planet—are Snowflake customers. They have their most valuable data on Snowflake. What does AI mean for that?

We systematically invested in the components, whether it was search or text-to-SQL, in ways that added value to the things that people were already doing with Snowflake. Snowflake Intelligence is an agentic platform, but it's actually an opinionated agentic platform.

A lot of agentic platforms, for example from the CSPs, will basically say, "You can bring in data from anywhere. You can imagine any kind of workflow that you want, and the one agent will rule them all," which is nice in theory. But in practice, when you have an infinity of things that you can do, it's also hard to figure out what you should actually do.

Snowflake Intelligence is very focused on how you create value from data, whether it's structured or unstructured, a whole lot faster. The kinds of use cases that got us really excited, honestly internal ones, were things like: If we were to take all of the different dashboards that we used in sales and put them into one single interface, what could that be?

We had done 2 or 3 versions of this, but eventually that culminated in this internal product. We call it Raven, but it's basically the sales data assistant. Then we started working with early customers, whether it was Cisco, Fanatics, or the USA Bobsled team, to figure out what all this means for them.

The theme, again, is to get away from the inflexibility of things like dashboards. A dashboard is a 2D view of a complex surface.

It just has no easy answers to the many questions that any reasonable person—you or I—is going to have off of that.

So we wanted to create something that freed people from the 2D style of thinking and was much more flexible in what it gave people access to, but also knew its place. This is not a general-purpose agentic platform to do it all. This is an agentic platform that lets people realize value from data faster and is a great foundation for people to get value from data really quickly in a meaningful way.

I think having this sort of opinionated framework for AI has been super helpful for us.

How does a user consume Snowflake Intelligence? Is it like I ask a question, I get pushed an answer, it builds dashboards for me? How should I imagine that experience?

We should show a demo of Snowflake Intelligence to you. But yes, it's an interactive interface. You can ask questions. There is a set of canned questions to make sure that you don't have a block when it comes to being able to ask questions.

You can ask it, "Hey, what data sets do you have access to? What kind of questions can you answer?" It will do a perfectly reasonable job of that. Our aspiration was for this product to be used by every single employee in the company.

So it's not for people who can write SQL.

It's not for people who can write SQL. We wanted it to be enough of a daily-use product for every single person.

There is not a single customer meeting that I'm going to have without quickly checking up on what's the latest with this customer. Raven, the sales data assistant that I talked about, absolutely has things like: What's our relationship with the customer? What kind of contract have they signed? What is their consumption looking like?

It also has things like: What are the most recent conversations that we have had with them? What came out of these? Are there any outstanding ticketing issues? And so it is a lot of that.

Like many good products, there is breadth. There is value driven to many, many people within a company. But on the other hand, we don't pretend it's a BI dashboard.

There are more things that you can do with Tableau or Sigma than you can do with Snowflake Intelligence. But that's not the goal, because this product also lets you do a bunch of things that you could not easily do in a dashboard and is really meant for any business user.

We place a lot of trust—and a lot of emphasis—in all our AI products on trust. I tell people we need to think of AI the same way we think about software engineering: there's a right and there's a wrong.

Okay.

It cannot be this mode of YOLO AI, where you can get some good answers and some terrible answers and it's your problem. We very much emphasize that you need an eval for every single new thing that you're going to launch. If you want to change the underlying model, you need to be able to quickly verify that you didn't blow up the things that you were already doing.

We want it to be the trustworthy product for every employee, which is actually a new thing for us, by the way, because Snowflake, for pretty much all of its history, has always been used by the data team to slap a dashboard on top, which then gets exposed to end users. This is a very different motion.

This is why we are working on things like identity-provider integration, so that you don't have to set up Snowflake accounts for each of the many users you're going to have in your company. We are also mindful of the fact that there is subscription fatigue, and so Snowflake Intelligence is very much a consumption product. People pay for what they consume, and we are experimenting with a bunch of things in terms of how we drive broad and deep adoption without having people worry about runaway costs and things like that.

The way you describe Raven, or the sales assistant agent use case, it sounds like an application—or like a lot of applications that I get pitched. How do you draw the line between a data and agent system and an app today?

Back to my point about execution, I tend to be completely emotionless about where the strongest current is. On the other hand, it's absurd if we think we are SAP or Salesforce. We are not. Somebody managing a $100 billion supply-chain ecosystem with a complicated software provider isn't saying, “Hey, I'm going to use SI and I don't need that.” That's really not the goal.

But on the other hand, I think the line between what an agentic system like this is going to be and what pure software is going to be will absolutely be blurry. I can imagine a lot of easy use cases. My sales team has to go update Salesforce quite often because we force them to update it whenever there's a use-case transition and stuff like that. Can that be done with APIs? Absolutely.

Should you be able to file a vacation request on top of Workday using our HR agent? I would say that's a reasonable thing. We very much take this approach: be opportunistic, but again, operate from a position of value and strength, and not just on naked ambition, because I think that doesn't work out.

If, on the other hand, you focus on what value creation means and what these users really want, I think that tends to be much more durable.

So you're describing a bunch of changes for the organization that you executed on very rapidly, right?

It always feels entirely too slow.

Yes. I have always experienced you to be quite impatient, but for scale, it seems pretty fast. What is one tactical thing you are doing from a leadership perspective in terms of moving faster or communicating a new direction internally and getting people on board? This is a broader and different vision for Snowflake than before.

Change is hard. You have to acknowledge that, and driving behavioral changes from lots of people is incredibly difficult. We were measured about how we rolled out changes.

For example, among the first changes were leadership and alignment changes and clearer accountability. That happened within a few quarters because you're not dealing with as many people. You organize the teams under them, but I would say that change also included what we then called the war room, or the pod model, where product and engineering and our go-to-market functions all work together. That was again an early change, and it was done with small groups of people without necessarily disrupting lots of people.

I would say other things—for example, rolling out coding agents to our engineers—that was a project.

Not everybody wants to do it. Some people are skeptical; some people are not.

I'm a big fan of combining bottom-up with top-down approaches. The example with coding agents is that Benoit, our wonderful founder, who fell in love with coding agents, did more to drive coding-agent adoption with the engineers than any number of words from me.

You sort of have to find the right people, find the champions. My take is that every large organization has these forward-thinking, curious, “I'm going to work over the weekends to figure out how to do something” kind of people.

You need to find them, encourage them, elevate them, and use that to drive change. Top-down change can be helpful, but it really needs to come from a bottom-up perspective.

We've rolled out coding agents to all of our solution engineers, and they are excited because that just dramatically lowered the amount of time it takes to create a demo. Usually, our demos used to be canned, and they were not always customizable to a particular customer.

But we can now be like, “Okay, we know the kind of data we think Elad and Sarah are going to have as part of their podcast. Let's create a demo with synthetic datasets just for that.” I think that's the kind of ability that we have gotten.

When you and I first met and got to work together, you were an investor, then you were an entrepreneur. Did either one of those roles change the way you are a leader at scale or a CEO?

I think these things are accretive. They add on to things in ways that you don't always appreciate or like then or ever. It is what it is.

I always complain to my family about the 10 years that I spent doing research and getting a PhD. I was like, “That was a waste of time,” but not really. For example, doing a PhD teaches you to focus on ideas and teaches you to focus on how to convey them crisply. You often spend enormous amounts of time writing 4-line abstracts, but it actually turns out that's incredibly powerful: being able to convey ideas in an easy way.

Neeva is among the hardest and most heartbreaking experiences of my life. It is what it is. Sometimes you are too early. But on the other hand, I probably learned more about hustling, took success far less for granted, and learned more about sales or marketing or any of these other things that you kind of take for granted if you're at Google.

At Google, whatever you did—my first launch at Google, which was entirely my work for 3 months, one person, okay—

Was covered by The New York Times.

Okay. Yeah.

So you just have immediate scale with anything.

Yeah. Distribution, your ideas, your products.

Doing a startup really makes you realize that that's actually special. I think I bring quite a lot of that when it comes to what it takes to hustle and what it takes to win.

Honestly, I think both the Google and the Neeva experiences make me somebody that's just a lot more grateful for my job. We talked earlier about how you have to deal with a bunch of stuff that you don't really want to deal with when it comes to doing something big that you like. I'm a lot more gracious about that because it's just such a privilege to be at a place like Snowflake and to be having the kind of impact that we have.

I remember what you told me after Summit. You said something along the lines of, “Thank you for inviting me to your rock concert.” There was a line that was more than 2 blocks long of people waiting to get into Javits Center in New York for this little conference that we were doing. There's nothing ordinary about it.

You, I think, are perhaps the world's expert on game theory and strategy with the tech elephants, because you have led the elephant, fought the elephant, and now built on top of the elephant, right? I think this is—the analogy broke down at some point—but in terms of the experience of building Snowflake on the cloud service providers, and the analogy for anybody building on foundation models, as you are as well, how do you think about that? What's a framework for creating durable value there?

I think product-market fit continues to be magical. It is the reason that Snowflake exists. Think about it: the 3 hyperscalers would love to just own the data space.

Yes.

Like they own any other space.

Yeah.

But yet there's Snowflake, there's Databricks. That sort of redeeming value is quite unique, and we should all have a lot of humility about what it takes to create that lightning in a bottle.

Having said that, I think the model companies, especially OpenAI, are super interesting because they are in that phase of their growth where they literally think they can do anything.

Yes. Yeah.

I joke to people that these are like empires that have not met their oceans just yet. I think you do have to pay attention to what is likely to be in their immediate path.

For example, I think coding agents are particularly interesting from this perspective because it is very clear that both Anthropic and OpenAI are going to be laser-focused on having the best one—the best one that there is.

So I think thinking about what is the likely trajectory of these companies and whether they really have a right to win, or whether it is something different enough that you don’t really have to worry about. Google, for example, stopped at information. God knows I spent enough time trying to get into physical things like shopping, airline purchases, or hotels. We didn’t really succeed because we didn’t have core competence, really, in some fundamental way beyond the world of information.

I think it’s going to be fascinating to discover what that kind of boundary is for an OpenAI or an Anthropic, but I think there are lots of areas that can be reasonably guessed at with respect to where they’re going to go. I think that’s the one that’s tough. Early patterns show that if you are, for example, a set of prompts on top of one of these models, that’s a problematic space to be in. You need to add value.

I also think a lot about what differentiates us from these model companies, which is an area that they’re likely to be interested in. How do we make sure that we have distance with respect to what we add? A lot of the urgency and change in the products we create in collaboration with these folks all take us toward the data platform as a durable category.

But this is also a time where literally no software company can feel secure about its position in the sun. I actually think that perhaps is just as important as anything else that I just said.

Do you have that orientation of, “We need to continue to earn it”?

We need to continue to earn it. If there is anything that all of us have learned, say, from the CSPs, it is that they have infinite budgets and infinite patience. Unless you innovate and stay ahead—not just be ahead, but stay ahead—you will be Intel.

I think that’s another really useful lesson to remember as a company like Snowflake navigates the current realm. This resonates hugely with me, both on the dimension of thinking about one of the founder-CEOs of one of my favorite companies that’s now a public company, which I thought was unassailable. AI—I hate to be the person to be like, “Well, this changes everything,” but they feel threatened for the first time in many years.

I think that’s a pretty common experience right now as a software CEO. Especially when the technical environment is so fluid, defensibility is built, not strategized.

That’s correct. It’s built every single day. You have to keep moving.

One of the questions I would have on the data cloud is: even if you don’t have the CSPs’ budget, you do have the ability to make multiyear plans, and you joined Snowflake because you saw a vision for it to be much more than the data cloud.

As a technologist, when you look out 3 to 5 years, how do you expect people to think of Snowflake, both in the ecosystem and in how customers use it?

Most of our core strength comes in that data platform layer. I sometimes internally talk about being there for our customers from inception to insight—from when data is first conceived to when somebody gets an insight that feeds back into that system.

In fact, the pitch that I make to CEOs or CIOs that I meet for the first time is really that the great companies of this century—a company like Google or Meta—were more data-first companies than purely product-first companies, compared to pretty much any others before. You built cars and then did instrumentation to make sure that they didn’t crash or to track their maintenance records.

Even products—think about it. If you built something like Adobe Photoshop in the ’90s, you did a bunch of research, built the product, sent CDs over to various people, and then waited for some feedback to come back. All data was always a slow afterthought.

Search ads, for example, were magical because the behavior—what people did while interacting with these ads—went back into influencing what happened to that system, pretty much within a few minutes.

Mhm.

My data teams were as large as the product teams. What I tell our customers is that we want to be that companion for all different kinds of data. We want them to have the expertise that the Googles and Metas of the world have.

We see AI as a massive accelerant for things like that because, all of a sudden, CEOs now realize, “Wait, this is not just about data modernization. This is not just about me being able to run more queries or perhaps code up a machine-learning algorithm. This could influence how my business operations work. This could influence what efficiency means for me as a category.”

I think that’s the tailwind that we have from AI, because the value of data just got vastly elevated. That’s our aspiration.

With respect to the CSPs, my take is that a company like Snowflake, which comes data-first as opposed to a set of services-first offerings, with an emphasis on simplicity and integration, can create as large a database as you like on Snowflake, and it’ll be completely shareable within the company. It’ll be completely shareable with your partners.

When we talk about AI in Snowflake, it’s not an afterthought. It will work with all of the governance that you have put in place before. That kind of integrated approach, we think, will have durability because, over time, the idea of buying raw compute and storage and writing code in order to solve a problem—it never gets easier. We are a higher level of abstraction.

Pre-AI, that was my main thesis for why I wanted to be part of Snowflake. I said a data platform that especially spans CSPs has the right—you have to earn it—to be as large as a CSP itself. That was, roughly, why I joined the company and our medium-term vision for what we want to be. As I said, AI is a massive accelerant on how you get value from data faster or how you get better at acting on data quicker.

When you think about the overall data landscape today, there’s the data that’s traditionally been in Snowflake, and then the investments you’ve made and new partnerships you’re announcing. Can you explain why SAP and some of the other partners you’re working with now?

Yeah, this is a good question. For a while, Snowflake had a Snowflake-centric view of the world. Plenty of people brought in data from SAP, or from Workday, or from Salesforce, but what is increasingly happening is that all of these companies realize that this is incredibly valuable data. It’s not quite their data—it’s customer data—but they understand that it is valuable.

They also understand that this line between software and services, and software and data, is blurry in a pretty meaningful way. One quality that I learned from Google, working in areas like payments, which is all about partnerships, was that partnership mentality: How do you pick a set of folks and figure out how you can create value together?

Among the earliest places where this went to work was our relationship with Microsoft, which was okay but not that great because they had a relationship with Databricks, and they were always conflicted about whether Fabric was the answer or Snowflake was the answer. Of course, you know, Satya is the master of how to create winning partnerships.

We took a lesson from how to get out of them and how to adjust them as you need to. We’ve been working on a partnership with Microsoft for the past couple of years. This is both product integration with things like Fabric and figuring out how the companies work together.

I think we’re in a much better place now compared to 18-odd months ago. There’s an understanding that we will compete with some customers, and that’s fine, and we will collaborate on a whole set of other customers where, let’s say, Azure plus Snowflake is a strictly positive combination. That’s the same attitude we have with AWS, and we’re working on a similar sort of arrangement with GCP.

I think the software providers are different. As I said, they understand that the world is changing. With folks like SAP, we are thinking much harder about what that 1 plus 1 equals 3. With SAP, I think it’s going to be absolutely bidirectional data sharing. But can we also collaborate in the area of analytics, AI, and agents, and make it easier for people to create these on top of SAP data?

I think that can also become a leverage point for us to expand out to more companies because, as you know, SAP has an incredible presence throughout the globe. I think it represents a maturing of how we think about partnerships.

We absolutely want to do this with a few other key folks. This is not the kind of thing that you can do with every company, but I think that partnership mentality—create value together—is something that will stand us in good stead and hopefully also be profitable for us with respect to generating more business.

I want to close out with 2 of the most common questions I get that I think you are more prepared to answer. The first is: with every enterprise customer I talk to, one of the first 2 questions is going to be, “Where are the highest-ROI use cases for AI for my business?” You run a large business, and you serve large businesses. What is your stack rank here? How do you think people should address it?

Every company now has a set of technologists. Even if they’re not software companies, they need to deal with technology.

Mhm.

I would say that coding agents are probably the easiest ROI, just in terms of making new projects faster and demystifying technology so that more people can get at it.

As I said earlier, we are large users of coding agents, and we're working on coding agents that are going to be part of Snowflake because we want to make it easier for people to use Snowflake itself. What's good for other people is also good for Snowflake.

I think areas like customer support absolutely fit the pattern: here is a repository of human knowledge, here is an easy backup in case the AI cannot do something, plus its ability for AI models to effortlessly tap into voice and typed questions and generate answers. That's an area where there is clearly a whole set of obvious ROI: faster, easier, more seamless access to data, especially when it's combined with not having to pay for a $50-per-user-per-month license. That's another easy ROI item, and that's part of our motivation for Snowflake Intelligence: democratizing data access.

These are among the areas where it's more or less guaranteed ROI. But the other way to think about this is that obsessing about ROI too quickly is also a bad idea for many companies, because you don't want your first step to be 100 feet. You want to do a lot of little things that prove value. People can get plenty of value from using ChatGPT, even the free ones, and tools like that for many day-to-day things that we do.

I think the more companies demystify what it is to use AI, the more chance they have of figuring out how to get value, because you take the risk. I place a lot of emphasis on how many shots you take on goal.

How many projects can you run very quickly so that you get a feel for what the landscape of change is? The sales data assistant had 3 versions that came before it.

They all stuck. They all added more and more things on top. The first one was just on enablement. The second one was a little bit more about customer information. We also had an app called Customer 360. It was a Streamlit app, a Python app that you could get that kind of information from. All of these then culminated in the sales assistant, which is all of these things combined.

To me, it's that journey of iteration that's often just as important as having that big one thing that I managed to launch. I prefer not to take big bets, whether it is in getting engineering projects done or these kinds of projects. I think iterating and creating value every step of the way is the key.

This is the same advice that I give to our customers. I say, you should not spend a lot of money on AI with Snowflake. You should do it $1,000 at a time, and when you have significant value that you feel good about, then you can wrap it up.

This is one reason I'm very bullish on applied companies that know what their immediate usefulness is, because the landscape of what you could do with AI, if you have customer trust and you understand the workflows, is very large. I think there's just land for the taking for people who have the velocity and also the paranoia to keep expanding into that.

But it is also an argument for why that layer should exist, because you're just reducing the time to value versus somebody building it themselves with a generic framework or just straight APIs and engineering work as well. What a lot of our customers have found, for example, is that creating something like Cortex Analyst—which in some simple way is text-to-SQL—is actually a much harder problem than they think.

There's more trust in Snowflake because we did many, many analyst projects before we ever got into something like Snowflake Intelligence, which is a step-level increase both in capability and in complexity. Sarah Guo

The other question that I get asked a great deal is, what do you think happens to ads on the internet if we have chat interfaces that are much more directed instead of offering you 10 blue links? I have to ask you.

It's a great question. I think advertising is just an incredibly powerful medium and an incredibly powerful business. I'm actually 7 years out from Google. I'm more proud of the work that we did in the search ads team now than I was when I left Google.

I think there are good ways of doing it, and you know it when you see it. It's very clear. It will reinvent itself in the chat world. Let's just hope it doesn't become more insidious in terms of discoverability and being able to tell what's an ad or not.

It would sure be creepy for you to have a psychiatrist that has a certain affinity for prescribing one medication versus another. Those are the kinds of unfortunate things that we will discover. But the ad model is here to stay. It will just come in different forms. I think as long as it's done well, it's a reasonable model.

As a consumer, you also have to be smart about what's in these things for you. I think more than ever before, there's an increased premium on preserving our agency. I think that is what we all have to do as individuals.

I am really encouraged by how strongly citations and sourcing in models have taken off in different experiences. I certainly think that the set of things that are being presented to consumers has narrowed a great deal, but the fact that people want to go look at primary sources and understand where information has come from, even given all of this reasoning, is a useful indicator.

It's a very positive thing, I think. The nice thing about some of this is that it is not a whole lot of work for you to take a deep research article written by Gemini, paste it into ChatGPT, and ask it to verify all of the links that are there.

We have talked about how we did citations at Neeva and how proud we were when we launched it in early 2023. I think that's an idea whose relevance is still as strong as ever. Products like ChatGPT Deep Research are truly amazing in terms of the value that they can create.

Anyone—you and I—can get an expert paper literally on any topic. We just have to have the brainpower to be able to digest it. I think that's pretty amazing, and it's really fun to see these core technologies embrace things like that, as opposed to just writing, “Here's this article. Take it or leave it.”

I want to ask you one last architectural question, because you have worked for such a long time on information retrieval and search. You also understand LLMs, and you work with a lot of structured and unstructured data today, so you have a very well-rounded point of view.

I think there is a contingent of folks who believe that traditional information retrieval techniques and indexing are less and less relevant as more data is available through the model, even in enterprise use cases or non-consumer use cases. How do you think about that?

It is tempting to trivialize things like search as just information retrieval. The insight that powered Google was PageRank. It was a way of harnessing the power of the entire internet to figure out what was popular and what was not. But PageRank ran out of juice in 6 years, around 2004 or 2005.

While Google never really liked to talk about it, the kinds of things that became more and more relevant were the click behaviors on top of the search results that Google presented. It was that feedback loop that eventually gave it so much value.

When it comes to AI systems, including Snowflake Intelligence, remember I talked about eval loops. That's a fundamental construct that you need to be able to launch some meaningful product, but it will also turn out that that's the construct you need for that product to get better and better over time. Perhaps we will figure out a way to encode that as well into the context that's presented to the model.

But to me, right now, it's similar to asking, “Should LLMs be able to do math?”

Mhm.

You can argue, yes, they should be able to do math. They're so powerful. But as any reasonable person will tell you, a smarter person is going to say no, they should not do math. Instead, I should write the 2 lines of Python that I know how to direct and run in order to solve the math problem.

I think of trust in a very similar way. There are well-known solutions for figuring out what is the most trustworthy when it comes to a question that you want to answer. Why would you not use that and think of that as another tool that whatever AI system or agentic system that you're building is going to use, rather than be in this maximalist mode where the AI can solve everything?

I think all practical people will use the best tools available to them at a given point in time. At least at this point in time, there's enough value from these outside tools, including search, that I don't see the point of trying to dismiss them right now.

I mean, it's a very principled point of view: a maximal intelligence will use reliable tools wherever they are available.

100%, wherever it is available. There is no way to be so smart that you don't use the computer.

You cannot be so smart that you don't use the computer. [laughter]

Exactly. There is no bravery in just hard work if something can be done easily. You get to focus your energy on other things.

I think that will very much be the case. The prevalence of things like search APIs is actually a testament to the fact that all of these models benefit from things like that, because they provide external information that is not easily internalizable into the AI model just yet.

Sridhar, thank you so much for doing this.

Thank you. Thank you, Sarah.