Snowflake CEO Sridhar Ramaswamy on the Future of AI and Data

Thanks for hosting this. It's a cool room. We're in the, what do you call this, the gaming room? This is the game room.

Let's see, joined the day after Memorial Day last year, so like 17, 18 months.

Yeah, when are we doing this? And I also had a few other opportunities that I had to go say no to. And you know, opportunities don't come all that often, and so saying no to some concrete things while saying this will likely work out was super awkward.

Six months, that was my deal with Frank, that I would stay on till the beginning of December and then go figure out what I was going to do with my life.

I started looking, I would say I started talking to a few folks in October of last year. These things take time. Senior roles don't come by that often.

I was head of AI during those six months. And I joke to people that I was like a cabinet position without an actual portfolio, kind of an IC, right? No direct reports. But roughly the mission that we set for ourselves was to make sure that Snowflake was kind of an A-player when it came to data and AI. It was really to craft that product vision along with Vivek, along with others that came like Baris and Asim, about how do we bring AI front and center to Snowflake in a way that stayed true to the product principles of Snowflake: one product, everything is tightly integrated, stuff just works. So that was the bar that we set for ourselves. And so it was a lot of product design, and then there was also a lot of talking to customers about AI, what was coming. So that was roughly what I did.

First of all, I think of AI as a horizontal capability with broad application. It's a little bit similar in that sense to, let's say, the mobile change, where every company that was involved in software, especially if you're a consumer company, knew that you better figure out what mobile meant to your business. A different way in which I crystallize how it is a horizontal is by saying that, for pretty much the first time in computing's history, there is now a seamless, fluid way to go between different kinds of data, whether it is unstructured text to extracting structured information from it, or going from an image to the structure of images, same with video. Things that honestly you had to do software engineering projects for, or you had like specialty areas in computer science that just went and did one thing, all of those capabilities are now exposed to you via AI models. So to make it super concrete, let's say we have a health provider as a Snowflake customer, and let's say they have a bunch of clinician notes from years and years of doctors talking to patients. If you wanted to ask a question like, oh, make me a list of all patients that had, I don't know, cough symptoms, that would be like a custom job. You had to go find some software engineers, they would figure out what model to use, they would tune it, and then maybe three, six months later you got like this other table that said, okay, this is the list of patients that had cough symptoms. Now a 17-year-old kid can write a little prompt and then run a SQL query on Snowflake that'll be like, okay, here's your list of patients that had cough symptoms. And so first we said we want to bring the power of these language models and make it available to anyone that touches Snowflake. And that is Cortex AI. It's the model garden, but it's integrated into our Python environment, into our SQL environment, so that you and I can just write SQL and underneath this AI, you don't need to know about where and how the models are hosted, how fast, how slow, we take care of all of that. It's that baseline of capability. And then beyond that, we think in terms of what are key primitives that use language models, that use AI models, but provide additional functionality. So for example, we implemented search, and this was tech that we had developed at Neeva, but it's now part of Snowflake, where you can create an index on say a bunch of documents. We take care of things like chunking it, generating embeddings, being able to do fast retrieval, so on and so forth. But that's a key primitive that makes chatbots possible. Similarly, we have another primitive for how do you go from spoken text or written text to getting back structured information. And so our vision very much is that there is going to be this base layer, there are important primitives, and people can mix and match them to do all kinds of interesting things. They can build applications, they can do data transformations. And those are things that are still evolving. Absolutely, you can do agents, but part of what we're trying to figure out is what is the boundary of where an agent is still reliable. We place a lot of emphasis, for example, with a chatbot, if you want a business user to use it, you want to be able to tell them, you know, you can engineer this like it'll be right 99% of the time, and if you make a change you can verify that it's still right 99% of the time. Same goes for the analyst. But we think that by building very, very reliable primitives, we unlock the power of AI for business application.

You know, when the Neeva acquisition happened, we first of all made sure that the team, and us, me and Vivek, my co-founder, were very clear about sort of our mission. We shut down Neeva the consumer product, and we told our team, our goal is to make Snowflake successful. Okay? That part is important, where everybody knows that you don't have an agenda, you're here to make things better. And then the founders, Benoit, Theory, Christian, our product lead, were very cognizant of the power of AI. I mean, at the end of the day, you don't have to be a rocket scientist to internalize what I just said about AI being like a magic data transformation tool. They get it, and they're like, we need to make sure that it's available. But much of the work that we did, to be honest with you, was additive to what the rest of the Snowflake team was doing. It was not like that or this. We very much said, this Cortex AI makes Snowflake better, these operators make it easier for people to use the data that is in Snowflake. And in fact, the thing that I tell our customers is, listen, Snowflake does not compete with OpenAI when it comes to foundation models. Let's face it, we're simply nowhere as capable as them. But on the other hand, if you want to use the power of AI to create a chatbot, to do data transformations, to get access to structured data, we make it dirt simple for you. And this message of take Snowflake with all of its glory and power and enhance it with AI, it's a positive cultural message. It's not us coming and saying, you people don't know what you're doing, let us show you how. That's the way in which a startup fails. I've had plenty of founders that have come into my teams with that kind of attitude at Google. And so in many ways, all of the dumb stuff that we did at Google are like good lessons for us in terms of how do you approach change culturally. Organizations adapt, and the Snowflake team very much sees this as a say yes and. And that's a message that goes well.

People are endlessly inventive, and it's the composability that's also super, super cool. We have run into applications, obviously, you know, the obvious ones, like a company using us because they make 100,000 devices and searching through PDF manuals of 100,000 devices, like you know, anyone that's looked for a washing machine manual knows it's kind of annoying, kind of hard. In some sense, that's an obvious sort of use case. But even internally, people have done stuff like, oh, we can annotate use cases sitting in Salesforce. This is just like, we record every new thing that we want to launch within a customer as a use case, and tagging them with the right product, with the right feature, is error-prone and also sometimes our sales teams don't do it. But what used to be difficult before in manual tagging, as I said, it's now like a single SQL prompt to be able to do this. We have hedge funds that are using it to grind through analyst reports and say, hey, if this company gets mentioned in this report, we'll be able to extract it. There are lots of different ways in which people are mixing and matching what you can do with different kinds of models. You also have great partners like Landing AI, and that is sort of truly magical because they run within our security parameter, but they can do things like analyze images and extract structured information from it. It's that sort of composability that is incredibly cool to see.

I think it was March of this year.

100% the latter. You know, I think that actually knowing how to use language models is, I would say, almost a required computing skill for any meaningful infrastructure-level company. Just like at the turn of the century, dealing with big data, like data that could no longer fit in a machine, you kind of needed to know how to do that. We thought of Arctic very much as, what will it take to train a modestly sized model? We spent a few million dollars on it, nothing like the billions that are being spent now, but I think it was a very good experience for us. And it also taught us a lot about what does inference mean, what does inference optimization mean, and those skills continue to come in handy. We have folks from teams like VM, which pioneered a lot of efficient inferencing work. And so I'm pretty happy that we have expertise up and down the AI stack, so to say.

I would roughly say for very, very common functions, I don't know, like sentiment detection, people know that small models work just fine, and they're sort of happy to go with defaults, it's very inexpensive. But then I would say that when the requirements are a little bit more stringent, people like the name brand models. Interesting. Which is sort of a weird setup. I think it's the reason why, for example, OpenAI continues to get heavy, heavy, heavy usage. Yes, they make the process of working with them easy, but I think people like to go with something that is known and trusted, even if it's just a smidgen beyond what you actually need. But why bother figuring out what you actually need when you can just use the name?

It's surprising, but yeah, there's first-mover advantage sort of even here.

I think there is still a big gap when it comes to, here is an engineered solution. I think the AI industry, if there is such a term, does itself no favors by simply not talking about things like hallucination rates. If you look, no one publishes hallucination rates on their models or on their solutions. It's like, look, we're so cool, you should just use us. But I think once you, you know, I was actually showing one of my board members an app that we have where we can ask questions about our financial data. It's pretty cool because I can say things like, hey, how did James P., who is our GVP of EMEA, how has he been doing over the past seven days? And it'll give you a table, it'll run the SQL query and show you things like week-on-week growth and year-on-year growth and numbers and so on. But before we give it to my CFO, I'll tell the team, you better be sure that you're doing the right thing. I think it is that engineering mentality that is still a little bit of an issue. And so part of what we end up doing is educating our customers on, okay, so you need a release process, you need to actually be sure that something like 95% accuracy is okay for you, or if you use a chatbot, you need to make sure that you provide the citations or the place in the document where you can go verify the answer. Because business use, you know, people take seriously, you can't make mistakes, especially when it comes to critical applications, for example, like the one that queries revenue data on which people can actually make decisions. I think it's stuff like that that continues to be obstacles for broad use. But on the other hand, I think AI has a ton of use cases to solve problems where you will happily accept a certain amount of mistakes. So for example, if somebody sends me articles all the time, right, I find it annoying, it's sort of virtue signaling, they like, here's a great article you should read. Like, you know, my head hurts, why another one? So often times I'll copy the text in the article, paste it into Claude or into ChatGPT and be like, you know, can you give me a summary? If it's off a little bit, I don't really mind, it's not a huge deal, because the time saving that I get from not actually having to crank through the article, there's so much value. Similarly, the example that I gave you of my team, I think they updated 5,000 different use cases just by running a SQL query, figuring out a bunch of product categories and updating the underlying table. I could never have persuaded them to do this manually, and they're like, off by a little bit, it's fine.

It depends on the application. So for example, with something like a finance chatbot, I'd be like 100%, yeah, please be accurate, or begin to think about how errors are evident. Remember, the insidious thing about hallucinations is not that the model is getting 5% of the answers wrong, it's that you don't know which 5% is wrong. And that's like a trust issue. And so I think it's going to depend on the application. But the good news for all of us as technologists is that there are just a whole pile of these applications where AI will do stuff that you and I have zero patience for doing. I love the little, I'm a Gemini subscriber as well, and I love the little button that I get on my Gmail that's a summarize. Because for things that I want to read quickly, I will happily take the few problems that I run into. So I think there are lots of places where the stakes are not that high where AI is already useful. But I think it's also important that we understand that there is a software engineering aspect to how do you create reliable applications.