This debate of being a rapper or not. Yes, if you're a rapper then you haven't really built much by definition. All you did was take a model and bring it to the customer. But most problems that you try to solve require you to do a lot more work. These AI models are powerful, but as you said before, they are extremely hard to actually use. First of all, they're not like any software that you ever used before. They are completely non-deterministic. You ask the same question five times, you're going to get five different responses. They may all be on the same target, but they are different. So you have to do a lot of work to actually make this model technology useful for whatever application that you have. In fact, when you think about training your own models, I would actually say that for most startups, probably the better thing is to spend all of that energy instead on preparing, for example, the right input that you can actually give to these models so that they can do a good job, and using them effectively in your overall application architecture. My guest today is Arvind Jain, the founder and CEO of Glean. Before Glean, Arvind spent over a decade building Google search infrastructure. He then co-founded Rubrik, which recently passed a billion in run rate revenues. With Glean, Arvind is tackling the long-standing challenge of enterprise search. Yet his vision goes beyond this. He believes every employee should have their own team of AI agents to help them work smarter and achieve more. In our conversation, Arvind shares his journey as a technical founder and offers his unique perspective on what it takes to build a successful startup today. We also discuss where AI is heading and where he sees the biggest opportunities for founders. Here's Arvind. Arvind, welcome to the BW CEO podcast. It's great to have you join us today.
Why don't we just start by talking a little bit about your personal journey? You were at Google in the very early days. Tell us a little bit about that. What did it teach you? What are the lessons you took away from Google?
Yeah, well Google is an amazing company. First, when I joined the company, I was a bit confused. Did I actually join a corporate company or is it a research lab, is it a playground? What is this place? Because I joined and there's nobody who really asked me to do anything. You join as an engineer, but there is no job that somebody has in mind for you. You just go and figure out what technologies you want to build, what innovation you would like to do, and just go and work on it. I think I saw my manager maybe a month after I joined. Before that, I went in and asked people, 'What am I supposed to do here?' And the advice I got from folks who were in the room was, 'Yeah, just go meet people, figure out if there's any problem that you'd like to work on, and then work on it.' So the first problem I found was that we had some problem in shipping our search index to our worldwide set of data centers. For some reason, nobody thought it was a big deal. So I went and actually fixed that system to make sure that a fresh web index can be pushed to all of our servers. And that's how my journey started in terms of working on search. And that's been the problem that I've been working on for many, many years.
Very cool. And then in 2014, you decided to do a startup. You left Google to start Rubrik. What drew you to start a company and become a founder?
Well, I always had a desire to be a founder, to be an entrepreneur. In my early stages of my career, I wanted to do it because it was sexy, that was a cool thing to do, and I wanted to be cool. But actually, I didn't do it. So that was a desire, it was always on my mind. I was always thinking about ideas. But I kept working and I didn't actually start a company. Though I was part of these very early stage startups before Google. I was part of the founding team of Riverbed. It was three of us initially, four of us initially building that company. So I had some good experiences, I enjoyed those. But then after I joined Google, I really enjoyed that whole experience of getting to work on problems that fundamentally change how we live. Things like Google Search, Google Maps, these are technologies that have fundamentally redefined our lives. So getting that opportunity to work on these problems that everybody in the world benefits from was just so amazing. I just loved that and I kept doing it. And that sort of desire to actually build a startup went away from me. Then in 2014, I was at Google for over a decade at that point, and a friend of mine, B. Lee, he told me, 'Let's go and build a startup together.' He wanted to build it in the enterprise data protection space, which had probably nothing to do with me and what I was doing for the last 15 years building consumer products. I asked him, 'Why? I'm not the right person to build this.' But I think that whole desire that I always had, 'Can I go and build something without this massive support system? Can we actually create value?' It was a thing that I was excited about, proving to myself and learning some new skills. As he and I discussed this more, ultimately I got excited about that opportunity and then we started the company together. I guess you always have the desire to build something big and huge, but I think at least for me, my personality is one of an engineer where there's a lot of doubt in the process. So I never imagined how big Rubrik would become. It just recently crossed a billion dollars in ARR.
Which is incredible, it's truly remarkable, and congratulations. Very few companies end up becoming large, enduring public companies. You have one already under your belt and you're on your way to your second. So let's talk about the second one. Talk to us a little bit about Glean. How would you describe what Glean does?
Yeah, so Glean is actually very simple. Think of it like ChatGPT inside your company. It's a bit more than what ChatGPT does for you. It's an AI assistant. As an employee, you can go to Glean and you can ask any question that you have, and we'll try to answer that question for you. Sometimes we'll use all of the world's knowledge to answer that question for you, that's just like ChatGPT. But if your questions require us to tap into all of your internal company's knowledge, then we will do that as well. So that's the way to think about Glean. It's a Google or ChatGPT that's connected to all of your enterprise information, all of your company's data and knowledge, and it helps people with their day-to-day job. The primary function of course is you come ask questions and we'll answer those for you. But as you see with AI technologies, AI can do a lot more for you than just answering questions. It can do a lot of work for you. So in fact, Glean actually becomes that work AI platform. We can actually help you bring automation and efficiency in many of your business processes.
That's very helpful. Now interestingly, you actually started Glean before ChatGPT in 2019, before the chat interface just became mainstream. So what was your 'why now' thesis in 2019 when you started Glean?
ChatGPT didn't exist, so when people would ask me the same question, 'Hey, what is Glean? What do you want to do?' I would tell them that I'm actually building a Google for people in their work lives. And so that was our ambition. We wanted to make it really easy for people to find information and get answers to questions that they have when they're working. And it's been a frustrating problem for people for a long time. Even when I was at Google, the world's biggest search company, we all spent our engineering careers trying to help people find information on the web, but we never really helped ourselves. It was incredibly hard to find anything inside Google as well when it came to our work. You can think of it as enterprise search. This is a market that has never seen success. There have been a lot of attempts by companies trying to solve this problem, and you never saw anybody really achieve great success, build great products, partly because it was a really, really hard problem to solve. But in 2019, when we started Glean, we saw a few different technology trends that we felt really enabled us to build a first good product in this domain. Over the last 15 years, companies have moved progressively from packaged applications in private data centers to SaaS-based applications. And one of the good things about SaaS is that the whole complexity of software versioning goes away. Also, SaaS products have been quite interoperable. You can pull data out via great APIs. To build a good search product, you need to actually have access to all of your company's data. If you're using 100 SaaS applications, we can build integrations into them, use the open APIs, and there you go, we have all of your company's knowledge to work on. This used to be a problem that in the past would take one year of implementation to get hold of all of your data inside the enterprise. So that was a big trend.
The other big trend, of course, is that search is such a complex problem. Imagine Google, there are 50,000 engineers working over the last 25 years building that search product, which looks and feels the same, a white page with 10 blue links. You probably would wonder what's happening behind the scenes. Why do you need all these engineers? Because behind the scenes, there's a lot of work to be done. Search is like magic. You can come in, you can ask questions, you can just give us one or two words, you can misspell them. Somehow we have to read your mind, figure out what you're looking for, and produce the right answer back to you. So it is a complicated problem and there's a lot of work to be done. And frankly, it's not startup-friendly in some ways, the amount of R&D investment you need. But again, with cloud technology, a lot of core search tech components becoming open source, we were actually able to build on top of this great infrastructure. And the biggest one which I didn't talk about yet is the Transformer. Everybody talks about GenAI and Transformers, and we can see how they can have impact on pretty much every single part of our lives. But in 2018, this was a concept largely restricted to folks on the Google search team. This is where this technology was originally built. And we were already seeing the impact of these models on search, on all the different individual technical components that we have to build inside search. For example, we spent many, many years building a synonym system at Google, lots of technology, lots of machine learning that powered that system. But now you have a Transformer model, and you just put the entire world's data on it, and it actually builds a better model without any tuning or any hard work. So we saw that there was a step function change in what you could do with these language models and understand information and knowledge at a conceptual level that we could never before. So that was a big unlock. We felt that this ability that these models have will allow us to truly understand an enterprise and all the information that lives within that enterprise, and build a search product like one that has never been built before.
Those three things made me believe that this was a good time to build a product like this. But of course, the catalyst was not the technology. The catalyst was me getting frustrated with not being able to find things at Rubrik. In fact, we had a pulse survey that we did in the company where this was one of the biggest problems that people complained about: 'I cannot find anything in this company.' And I'm like, 'I'm now determined to solve this. It's time I fix this problem.'
Exactly, you're spot on. Enterprise search has been a problem going back to the early 2000s, and you probably know this from your days at Google. Google had a search box, they had an enterprise search product. But it was just the time. A lot of things had to come together, as you rightly articulated, for the market to be ready and for the technology to be at a stage of maturity that that problem would be solved. Talk to us a little bit, and you started to allude to it with the Transformer architecture and this interdependence between search and generative AI. In some way, that's really been the unlock. Can you talk about that in a little more detail?
These models and the technology, the original motivation for that was to make Google search better. That was one of the core reasons why all that investment was made. And you can see it because search, if you think about it, is a problem that requires you to understand a user's question and requires you to understand all the knowledge that you're going to be using to actually help them with that question. So you have to understand that knowledge at a semantic, at a conceptual level. So fundamentally, what these language models do is exactly that. You can think of it as a search problem anyway. And so therefore, today the language models are actually playing a very key role in making search experiences better at a fundamental level, but also at a user interaction level because you can actually go one or two or three steps further than what traditional search could do for you before. If you think about Google 20 years back, when you asked a question, it would surface some links to you. It never made a claim that 'here's the answer to this question.' It was good enough to match some keywords, you could figure out what's a high-quality document that seems to have those words. That was the level of search technology at that time. But today, it's more about doing this matching in an abstract high-dimensional space where we are conceptually trying to take a question and match it with information at a semantic level. So the language models have played a big role in making search experiences better. So that's a given.
So now that you have this great search product, can it actually help you do more with AI? That's actually the thing that I find the most interesting. All these language models, whether you think about GPT-4 or Claude or Gemini, all of these models have been trained on the world's public knowledge, and with that training they've become like smart humans. You can think of them as a human that is smart and that has read a lot. But these models don't know anything about private information that you have inside companies. So now if I think about a large enterprise that wants to use this technology to actually improve its business, make its employees more productive, build better products, how do I actually use this technology that knows nothing about my company? That's where search comes in. Take an example: there's an employee, they have a question, they want to get an answer. The search is the glue between all of that enterprise knowledge and the reasoning and generation capabilities that these models have. So that's how the Glean product works. Given any question that a user has, Glean will do the hard work of looking through the hundreds of millions of documents inside your company, find the right ones, the right sort of materials that we think can potentially answer these questions. But then we can actually take all of that and the original question and give it all to a model and make the model work on it. So that's how you combine your knowledge with the reasoning capabilities of these models and actually solve a user problem. In fact, every enterprise application today has this kind of architecture. You can call it a RAG-based AI application, but typically this retrieval, search plays a big role in it. And therefore, search has become a fundamental component of your AI application architecture.
Makes all the sense in the world. I'm going to zoom out from the specifics of Glean a little bit to talk about your experience of building an AI startup and what are the lessons that other entrepreneurs could learn. Then we'll come back to Glean in a bit. But what are some of the unique challenges you've had because you started building Glean before the hype? Now it's cool, everyone's building an AI app, and you are clearly a leader, one of the leaders of the pack. But it wasn't that cool in 2019 and 2018 when you started.
That's right. It's been a very interesting journey. When we started, we already knew that we were solving a problem that enterprise leaders didn't really care about that much because they hadn't seen great products in that area and they were very skeptical. Search as a problem inside the enterprise was also never the top priority for anyone, and it was very hard to actually show ROI, make it a project that makes a cut in terms of priorities. We knew that we had to educate the market, but we had no doubt in our mind that we were solving a useful problem. Whether an engineer or salesperson or marketer, every single person that I talk to had this pain. As a knowledge worker, you are constantly in need of information and you cannot find it. One-third of all employee time is actually just spent looking for information, often doing duplicate work and things like that. So fundamentally, we believed that we were solving an important problem, but we knew that we had to go and create a market, we had to convince, we had to evangelize. So our initial journey was very much like that. We had to do a lot of work finding those leaders who were willing to take a chance on us and allocate budget towards a problem that they had no intention or plans of solving. So that's what our early journey looked like for the first almost four years of the company. We were in this market creation mode. And I think we got good success by that time. Some of the leading technology companies, for example, were using our product and they loved it. They could see how much productivity gains and also more than that, the happiness that we were bringing to that employee base, reducing those frustrations. So we were seeing good success. But then the ChatGPT moment happened, and that fundamentally changed how people perceive of us, what the market looks like, and it has its own kinds of new challenges which I can talk about. But yeah, those were the initial days.
That's super helpful. What are some of the other challenges that from your perspective other AI startups face today? As a founder who's starting an AI company today, what are the two or three things you would warn them about?
