It's a long journey that we start from the beginning. I guess I graduated in engineering in next mentoring for Might be getting on and then went to work mostly software in Bangalore for a few years there started the company or the process, which is something like 400,000 find intelligence today. Back then, it was just 1000 people still very small, relatively speaking. Um, and it was fun. Gonna learn the large, uh, you know it. After about a year or so, though, I decided I wanted to do more build new things. Really, That is the sort of services companies for the kind of things that we were doing was very different and end up moving to a smaller company. Much more company, very of building a tool back then for helping with the year 2000 changes that sexual. Uh, I spent some time there than onto another start up in Bangalore in 99 for 1998. Uh, and there we were building this project called, basically trying to get more stick notes on the web on the desktop. You have a lot more options today for keeping track of what you want to do. Back then there was nothing. And so we decided. Okay. Can you get this sink across? Trailed machines. All the fact would stop. So I started doing that. Ah, a few months in being guard, connected there, somebody, facts and Microsoft and was looking to start a new company. And so we decided to join hands with them. And, you know, that's how I moved over to the Bay Area. Came to us back in 99. It was a fun ride. The then from, uh, you know, this is the dark, cold room. So things were looking up. Everything was going great. He went through that, you know, lots of loadings. And then two years in, everything comes crashing down. So now there's lots of things to learn from that. Um, yeah, After that, did another start up who are doing like the geolocation. Basically, through this all what I was enjoying and looking floor was Raya's Can I loaned? Derek and I contributed learning offering because these were startups, you know, if after you nor do they weren't quite going any vary in terms of just a business coach, etcetera, there wasn't that much new stuff for me to do were known. So, you know, there was just a ton off learning in these 1st 56 years, I would say starting instrument for sister Nowhere. What? The mainframes and stuff using COBOL JC on what kinds of things to building. You know, desktop applications to ending a chat service of Arzu are, you know, larger systems that cool and so on. There's just a ton of learning through that process now. So I'm that time 2002 I guess I decided, OK, it's maybe time for me to learn from a bigger company and joined. Ended up joining he they This was because one of the few companies that was doing great in that recession in this area and that was a great time because there was going to the changes where it was going like crazy. And we went from Ah, uh, you know, custom where external search engine to really building on one thing. And I had the chance to work with some reason people to build that out and learn a lot from that process. And so, you know, uh, through that all again, one of the key things for me was. Am I learning? Do I continue to learn? Do I consider add value, of course, is well, that's important. I would say early on that learning component was very big. I think that's kind of knew what I still enjoy that if I'm not learning anything today, that that's just hard for me. It's much fun, you know, After a day I left to join some funders, Justus, they were starting out. Bring out the engineering team of this company going contributor, and their experience was there was a technology. These, of course, better left. There's a lot of stress I had, but just in terms of building a new Demark, where does it take what kind of problems you face? Uh, you're going from big company to such a small company that often just changes that happen in terms of, uh, you know, the expectations that you have, what you get in the bigger company was smaller and how deputy through those. So that was exciting. Kid alone. Exactly. Britain. Amusing systems there regarded very mean, uh, no way. We were looking to find copies of content across the web, the grated back. But it was also important to modify starts the first part guard us a whole bunch of customers. Early on, we did well. But then to go beyond back, we needed to be able to help them monetary side. And that was one thing we weren't successful act. And the company ended up being sold toe Biddy Martin with five British marks and later personally for me in 2009. And we've gone to Google, uh, and spent over a decade. They're starting in advertising. Where again? New area. Lots of lots. For me, too long to business. I'd lesser from the infrastructure side. I would say it was exciting. You know, again never some of my strengths and building things and stuff. There went out a good team, etcetera. Not in 2011. I was exploring work next for me inventing to, um you know, Andrew ng him and Jefty and we're starting getting together and talking about Okay, what's next for deep learning community? Scale it up to something like over what could happen if we do that This is before the current, you know, growth phase and so joined them. In the small group of us, there are many people engineers with them Read out the first black form. Sure. The early research start, uh, really changed the game, especially at Goodwood, even across the board. And then over time, serve the cardamom or the ship role, that dangerous pieces as we've been dark tensorflow And, uh, you know, constantly. Dark personally, really enjoyed that. That aspect of just being able to bring that technology? Uh, not just across Coogan to all the developers, all the products that we had there, but also to the world with your kids are sticking. Welcome back in terms of putting all the people together breaking into lots more people in all the different pieces that it. So that was really excited. Uh, you know, as we were getting to do, boy, No. Last year. Personally, I was thinking Will prepare what next with me? And I ended up leaving Google at the end of last year. Nobody working with start up with him back to my journey on foot, pushing myself getting think uncomfortable, really with things and exploring new ideas, but also trying to build something that hasn't potential have changed signal

said was interesting. There were different points in time where, like based on changes, for example, early on, after we launched tensorflow externally, they grew like crazy. There was so much to do that we just needed more people to do things. And we clearly saw that there's a huge thing for that make sense to invest in there. So we grew like crazy. You had a small doing the products in 15 people. We like double the Demon 69. And so at that point I was being stretched in beyond my limits. I didn't really have the time to work with everybody on the team. That is a lot of your new on the team. So they needed a more help. And, uh, so it took me a while to get that under control there, you know, I have to find the right people to take on more leadership pills in the organization. Some hand off a lot of the things that I was trying to do myself, in some cases, get people from the outside, do they hold back? And so over the next six months to a year, I think a lot of that got some help across folks improving and heart doing out of love. And that really hope so. You know, one lesson I would say that is, uh, you know, as you think of broad and especially if you're not, I fear it's good to think about, you know, assuming you're gonna grow. How are you gonna plan that? Things. How can you hand off things that you've been growing doing yourself as a leader to others so that allows you to take on more things as well, but also makes you less of a bottleneck. Allows organization to scale a lot. So that was sort of one side of leadership and say another example is from more, uh, product oriented thing grab when we started out with tensorflow are clear. Goal waas. Okay, The one oven do nearly new kind of research segments to enable all research. But we also want to deploy this these markings in all kinds of places, a force in the data center. But we're starting to see addressing the blowing in the forms in leveraging accelerators in the data center. So we built this, uh, the infrastructure than dark detector to support that. But that said a urine be realized mostly from feedback from users were starting to do customization and trying to build their own. Thanks for the flying things on before we have a modern You wanted Ukraine that they were using tensorflow for training the money but they were exploiting should be just a blood do a custom hand written for deployment and we're not. When I saw this and you heard this from other, more than one place from other places, it was clear that what we had wasn't solving, you know? Yes, we had a long term plan of Hungary evolved from everywhere to get to a much better stage for that where it was going to think about. And so what are the decisions? We made a back time waas. Okay, let's kick off Project. That's gonna focus on exactly the problem that these today and you actually deliver something in the next six months and saw the immediately It's why don't we take maybe a year or two to get to have the core system than you? Convergence sword. And yes, that was our short term approaching some sense. But it did well because it really sold renewed for these folks in the lower orders to really go in that sector. Intensive The light is something that we created back then and it's really the leading solution for mobile today. If he hadn't done that, I think we would still be lacking. That was an interesting, uh, card on. When do you decide to not be died? Do OK, yes, I know I can do mobile doesn't mean that's the best way to do it. Can you really go out there and try out something yourself? And so that was a interesting in learning experience on just how it hoping this go Well, another one, I would say. Which was similar in around the same time. There I had, uh, you know, I was exploring pushing one new areas and they were getting feedback was when people wanted this whole idea that started with these graphs approach the books who understand that. And there was this new framework that had come out back then. Pytorch has been fairly successful in the especially research community and come out And even before that, there were people who were asking us Why do we need grafts? Why can't we do this? And this is primarily from the research community from our site, and we don't some basic experiments, but never really become serious about that, um, park over that around that time, the sort of floored that you know what? Yes, there are things that we can learn from that. They are pieces. Do not everybody was brought onto that. There were some things still wanted the other stuff instead, because there was value in there and it took us a vying to get you released eventually regarding for example, that's for to really put back together last year. But it took us a while to get, and there were some good reasons for that for something that people didn't just have accelerated early on. So being mindful of how fast you work on some of these things, how fast you you think about these are just important, especially in fast changing Cheerios for other areas that grow slowly. It's a very migrants ness. You have more time, but often as an encumbering as a leader, you miss these things, and it's important to think aboutYeah, in the case. Off in the case off tensorflow lore, it was less like a comedy. We were running it like a project that's yes, family down there. Yes, we had a lot of external influence. Of course, we were working very closely with our partners, like Intel in video on the hardware side. In some cases, other big blouse is wiling because of Turkey piece that they wanted, and they would contribute Bacchus well. And all of those was important. So it was important to think about our resolving the needs for the end user. It's anyone important thing to keep in mind. Ear's Who's the Injuries are. Who are you sold it for? And then, if you have agreement on that, whether it's people within the company or outside, then things get easier in some cases, with other companies involved, etcetera, that can get pretty and people having different goals in mind. So, for example, when she talked about commercializing that grew flowers one direction they want to go to, maybe AWS wants to go in a different direction with the NVIDIA wants to go in a different direction, so on until bringing them on board is something that you have to work through. It set from the engineering perspective. It does change and make the process more much more complex, and you will have its foreign tools to manage the development cycles, etcetera, that has externally on open source project, especially if you something I get up. There are certain kinds of tools that people use. Just how do we bring that together? And that's something we just ended up spending a lot of time and make that better. So it goes up. I wouldn't say it ever got chiropractic? Never. This continues to be a work in progress, but we made huge gains over the last few years.

for me. I was according much by the time I left. You know, managing the group so personally I was wasn't really using many of the students. I spent three more diamond documented in emails and then called itself well, there was involved in the designs themselves. I have lots of this guy shut down back. That said, from the team's perspective, uh, you know, Tensorflow itself is primarily that men C plus plus with a pipe in there. So it's a mix of people programming does the ideas. I think for these do their several ideas that different people like to use at Google. Max and I are probably the most popular given book. There are three ideas that they're polite. Some of those functionalities, I guess, uh, and in terms off the the you know, other pieces they invite in that sector again, there's some there. Some things at Google and External World uses similarly, and the others that are different in severe would support both in some sense, get have ended up in the place where you know, like a word was and that, you know primarily would keep everything. But there was a copy of the court internally as well. Two year periods to super PACs. Keep that completely. Six. So it wasn't suffragette. All that wasn't before cold ever. Um, And then, uh, you know, get have you use? There was some people who would use diuretics on an open source, tools from the others who have been abusing more Google centric because they just help put some of the other pieces.