WEBVTT 00:00.000 --> 00:03.080 The following is a conversation with Rajat Manga. 00:03.080 --> 00:04.960 He's an engineering director at Google, 00:04.960 --> 00:06.960 leading the TensorFlow team. 00:06.960 --> 00:09.160 TensorFlow is an open source library 00:09.160 --> 00:11.520 at the center of much of the work going on in the world 00:11.520 --> 00:14.040 in deep learning, both the cutting edge research 00:14.040 --> 00:17.720 and the large scale application of learning based approaches. 00:17.720 --> 00:19.480 But it's quickly becoming much more 00:19.480 --> 00:20.960 than a software library. 00:20.960 --> 00:23.760 It's now an ecosystem of tools for the deployment 00:23.760 --> 00:25.720 of machine learning in the cloud, on the phone, 00:25.720 --> 00:29.840 in the browser, on both generic and specialized hardware. 00:29.840 --> 00:31.920 TPU, GPU, and so on. 00:31.920 --> 00:34.200 Plus, there's a big emphasis on growing 00:34.200 --> 00:36.600 a passionate community of developers. 00:36.600 --> 00:39.760 Rajat, Jeff Dean, and a large team of engineers at Google 00:39.760 --> 00:42.720 Brain are working to define the future of machine learning 00:42.720 --> 00:46.200 with TensorFlow 2.0, which is now in alpha. 00:46.200 --> 00:49.120 I think the decision to open source TensorFlow 00:49.120 --> 00:51.720 is a definitive moment in the tech industry. 00:51.720 --> 00:54.360 It showed that open innovation can be successful 00:54.360 --> 00:56.840 and inspire many companies to open source their code, 00:56.840 --> 00:59.640 to publish, and in general engage in the open exchange 00:59.640 --> 01:01.160 of ideas. 01:01.160 --> 01:03.880 This conversation is part of the artificial intelligence 01:03.880 --> 01:05.000 podcast. 01:05.000 --> 01:07.760 If you enjoy it, subscribe on YouTube, iTunes, 01:07.760 --> 01:09.600 or simply connect with me on Twitter 01:09.600 --> 01:12.640 at Lex Friedman, spelled FRID. 01:12.640 --> 01:17.880 And now, here's my conversation with Rajat Manga. 01:17.880 --> 01:22.440 You were involved with Google Brain since its start in 2011 01:22.440 --> 01:24.800 with Jeff Dean. 01:24.800 --> 01:29.160 It started with disbelief, the proprietary machine learning 01:29.160 --> 01:32.760 library, and turned into TensorFlow 2014, 01:32.760 --> 01:35.760 the open source library. 01:35.760 --> 01:39.040 So what were the early days of Google Brain like? 01:39.040 --> 01:41.760 What were the goals, the missions? 01:41.760 --> 01:45.080 How do you even proceed forward once there's 01:45.080 --> 01:47.680 so much possibilities before you? 01:47.680 --> 01:50.520 It was interesting back then when I started, 01:50.520 --> 01:55.320 or when you were even just talking about it. 01:55.320 --> 01:58.800 The idea of deep learning was interesting 01:58.800 --> 02:00.400 and intriguing in some ways. 02:00.400 --> 02:04.840 It hadn't yet taken off, but it held some promise. 02:04.840 --> 02:08.680 It had shown some very promising and early results. 02:08.680 --> 02:11.360 I think the idea where Andrew and Jeff had started 02:11.360 --> 02:16.160 was what if we can take this, what people are doing in research, 02:16.160 --> 02:21.560 and scale it to what Google has in terms of the compute power, 02:21.560 --> 02:25.240 and also put that kind of data together, what does it mean? 02:25.240 --> 02:28.240 And so far, the results had been if you scale the computer, 02:28.240 --> 02:31.480 scale the data, it does better, and would that work? 02:31.480 --> 02:33.360 And so that was the first year or two. 02:33.360 --> 02:35.080 Can we prove that outright? 02:35.080 --> 02:37.440 And with disbelief, when we started the first year, 02:37.440 --> 02:40.760 we got some early wins, which is always great. 02:40.760 --> 02:41.880 What were the wins like? 02:41.880 --> 02:45.240 What was the wins where there are some problems to this? 02:45.240 --> 02:46.560 This is going to be good. 02:46.560 --> 02:49.640 I think the two early wins were one was speech 02:49.640 --> 02:52.200 that we collaborated very closely with the speech research 02:52.200 --> 02:54.760 team, who was also getting interested in this. 02:54.760 --> 02:57.760 And the other one was on images where 02:57.760 --> 03:03.120 the cat paper, as we call it, that was covered by a lot of folks. 03:03.120 --> 03:07.440 And the birth of Google Brain was around neural networks. 03:07.440 --> 03:09.280 So it was deep learning from the very beginning. 03:09.280 --> 03:10.760 That was the whole mission. 03:10.760 --> 03:18.960 So in terms of scale, what was the dream 03:18.960 --> 03:21.040 of what this could become? 03:21.040 --> 03:24.280 Were there echoes of this open source TensorFlow community 03:24.280 --> 03:26.240 that might be brought in? 03:26.240 --> 03:28.640 Was there a sense of TPUs? 03:28.640 --> 03:31.120 Was there a sense of machine learning 03:31.120 --> 03:33.680 is now going to be at the core of the entire company? 03:33.680 --> 03:36.040 Is going to grow into that direction? 03:36.040 --> 03:38.320 Yeah, I think so that was interesting. 03:38.320 --> 03:41.320 And if I think back to 2012 or 2011, 03:41.320 --> 03:45.240 and first was can we scale it in the year or so, 03:45.240 --> 03:47.520 we had started scaling it to hundreds and thousands 03:47.520 --> 03:48.080 of machines. 03:48.080 --> 03:51.040 In fact, we had some runs even going to 10,000 machines. 03:51.040 --> 03:53.840 And all of those shows great promise. 03:53.840 --> 03:56.760 In terms of machine learning at Google, 03:56.760 --> 03:58.760 the good thing was Google's been doing machine learning 03:58.760 --> 04:00.200 for a long time. 04:00.200 --> 04:02.120 Deep learning was new. 04:02.120 --> 04:05.000 But as we scale this up, we showed that, yes, that was 04:05.000 --> 04:07.840 possible, and it was going to impact lots of things. 04:07.840 --> 04:11.160 Like, we started seeing real products wanting to use this. 04:11.160 --> 04:12.720 Again, speech was the first. 04:12.720 --> 04:15.120 There were image things that photos came out of 04:15.120 --> 04:17.360 in many other products as well. 04:17.360 --> 04:20.120 So that was exciting. 04:20.120 --> 04:23.120 As we went into with that a couple of years, 04:23.120 --> 04:25.760 externally also academia started to, 04:25.760 --> 04:27.760 there was lots of push on, OK, deep learning's 04:27.760 --> 04:30.520 interesting, we should be doing more, and so on. 04:30.520 --> 04:35.560 And so by 2014, we were looking at, OK, this is a big thing. 04:35.560 --> 04:36.680 It's going to grow. 04:36.680 --> 04:39.400 And not just internally, externally as well. 04:39.400 --> 04:42.240 Yes, maybe Google's ahead of where everybody is, 04:42.240 --> 04:43.600 but there's a lot to do. 04:43.600 --> 04:46.640 So a lot of this start to make sense and come together. 04:46.640 --> 04:51.080 So the decision to open source, I was just chatting with Chris 04:51.080 --> 04:53.360 Flattner about this, the decision to go open source 04:53.360 --> 04:57.040 with TensorFlow, I would say for me personally, 04:57.040 --> 05:00.000 seems to be one of the big seminal moments in all 05:00.000 --> 05:01.680 of software engineering ever. 05:01.680 --> 05:04.600 I think that when a large company like Google 05:04.600 --> 05:08.680 decides to take a large project that many lawyers might argue 05:08.680 --> 05:12.840 has a lot of IP, just decide to go open source with it. 05:12.840 --> 05:15.200 And in so doing, lead the entire world in saying, 05:15.200 --> 05:19.280 you know what, open innovation is a pretty powerful thing. 05:19.280 --> 05:22.320 And it's OK to do. 05:22.320 --> 05:26.400 That was, I mean, that's an incredible moment in time. 05:26.400 --> 05:29.280 So do you remember those discussions happening? 05:29.280 --> 05:31.320 Are there open source should be happening? 05:31.320 --> 05:32.600 What was that like? 05:32.600 --> 05:36.840 I would say, I think, so the initial idea came from Jeff, 05:36.840 --> 05:39.320 who was a big proponent of this. 05:39.320 --> 05:42.400 I think it came off of two big things. 05:42.400 --> 05:46.280 One was research wise, we were a research group. 05:46.280 --> 05:50.240 We were putting all our research out there if you wanted to. 05:50.240 --> 05:51.680 We were building on other's research, 05:51.680 --> 05:54.920 and we wanted to push the state of the art forward. 05:54.920 --> 05:56.800 And part of that was to share the research. 05:56.800 --> 05:58.920 That's how I think deep learning and machine learning 05:58.920 --> 06:01.360 has really grown so fast. 06:01.360 --> 06:04.280 So the next step was, OK, now word software 06:04.280 --> 06:05.280 help for that. 06:05.280 --> 06:09.720 And it seemed like they were existing a few libraries 06:09.720 --> 06:12.160 out there, Tiano being one, Torch being another, 06:12.160 --> 06:13.960 and a few others. 06:13.960 --> 06:15.400 But they were all done by academia, 06:15.400 --> 06:19.000 and so the level was significantly different. 06:19.000 --> 06:22.040 The other one was, from a software perspective, 06:22.040 --> 06:27.120 Google had done lots of software that we used internally. 06:27.120 --> 06:29.120 And we published papers. 06:29.120 --> 06:31.680 Often there was an open source project 06:31.680 --> 06:33.600 that came out of that, that somebody else 06:33.600 --> 06:35.440 picked up that paper and implemented, 06:35.440 --> 06:38.280 and they were very successful. 06:38.280 --> 06:40.920 Back then, it was like, OK, there's 06:40.920 --> 06:44.200 Hadoop, which has come off of tech that we've built. 06:44.200 --> 06:46.240 We know that tech we've built is way better 06:46.240 --> 06:47.880 for a number of different reasons. 06:47.880 --> 06:51.680 We've invested a lot of effort in that. 06:51.680 --> 06:54.320 And turns out, we have Google Cloud, 06:54.320 --> 06:57.520 and we are now not really providing our tech, 06:57.520 --> 07:00.520 but we are saying, OK, we have Bigtable, which 07:00.520 --> 07:02.080 is the original thing. 07:02.080 --> 07:05.280 We are going to now provide HBase APIs on top of that, which 07:05.280 --> 07:07.480 isn't as good, but that's what everybody's used to. 07:07.480 --> 07:10.960 So there's like, can we make something that is better 07:10.960 --> 07:12.320 and really just provide? 07:12.320 --> 07:14.320 Helps the community in lots of ways, 07:14.320 --> 07:18.320 but it also helps push the right, a good standard forward. 07:18.320 --> 07:19.960 So how does Cloud fit into that? 07:19.960 --> 07:22.680 There's a TensorFlow open source library. 07:22.680 --> 07:25.800 And how does the fact that you can 07:25.800 --> 07:28.240 use so many of the resources that Google provides 07:28.240 --> 07:31.480 and the Cloud fit into that strategy? 07:31.480 --> 07:34.920 So TensorFlow itself is open, and you can use it anywhere. 07:34.920 --> 07:38.360 And we want to make sure that continues to be the case. 07:38.360 --> 07:42.080 On Google Cloud, we do make sure that there's 07:42.080 --> 07:43.800 lots of integrations with everything else, 07:43.800 --> 07:47.280 and we want to make sure that it works really, really well there. 07:47.280 --> 07:50.080 You're leading the TensorFlow effort. 07:50.080 --> 07:52.360 Can you tell me the history and the timeline of TensorFlow 07:52.360 --> 07:55.880 project in terms of major design decisions, 07:55.880 --> 08:01.240 like the open source decision, but really, what to include 08:01.240 --> 08:01.600 and not? 08:01.600 --> 08:03.600 There's this incredible ecosystem that I'd 08:03.600 --> 08:05.680 like to talk about, there's all these parts. 08:05.680 --> 08:12.120 But if you just some sample moments that 08:12.120 --> 08:15.960 defined what TensorFlow eventually became through its, 08:15.960 --> 08:19.400 I don't know if you were allowed to say history when it's just, 08:19.400 --> 08:21.240 but in deep learning, everything moves so fast 08:21.240 --> 08:23.400 in just a few years, it's already history. 08:23.400 --> 08:24.880 Yes, yes. 08:24.880 --> 08:29.760 So looking back, we were building TensorFlow. 08:29.760 --> 08:34.240 I guess we open sourced it in 2015, November 2015. 08:34.240 --> 08:39.800 We started on it in summer of 2014, I guess. 08:39.800 --> 08:42.960 And somewhere like three to six late 2014, 08:42.960 --> 08:45.320 by then we had decided that, OK, there's 08:45.320 --> 08:47.080 a high likelihood we'll open source it. 08:47.080 --> 08:49.560 So we started thinking about that and making sure 08:49.560 --> 08:51.320 that we're heading down that path. 08:53.960 --> 08:56.280 At that point, by that point, we'd 08:56.280 --> 08:59.280 seen a few lots of different use cases at Google. 08:59.280 --> 09:01.200 So there were things like, OK, yes, 09:01.200 --> 09:04.160 you want to run in at large scale in the data center. 09:04.160 --> 09:07.480 Yes, we need to support different kind of hardware. 09:07.480 --> 09:09.400 We had GPUs at that point. 09:09.400 --> 09:11.880 We had our first GPU at that point 09:11.880 --> 09:15.760 or was about to come out roughly around that time. 09:15.760 --> 09:18.640 So the design included those. 09:18.640 --> 09:21.760 We had started to push on mobile. 09:21.760 --> 09:24.880 So we were running models on mobile. 09:24.880 --> 09:28.080 At that point, people were customizing code. 09:28.080 --> 09:30.280 So we wanted to make sure TensorFlow could support that 09:30.280 --> 09:34.120 as well so that that became part of that overall 09:34.120 --> 09:35.200 design. 09:35.200 --> 09:38.040 When you say mobile, you mean like pretty complicated 09:38.040 --> 09:39.960 algorithms of running on the phone? 09:39.960 --> 09:40.480 That's correct. 09:40.480 --> 09:42.680 So when you have a model that you 09:42.680 --> 09:45.200 deploy on the phone and run it there, right? 09:45.200 --> 09:47.800 So already at that time, there was ideas of running machine 09:47.800 --> 09:48.720 learning on the phone. 09:48.720 --> 09:49.240 That's correct. 09:49.240 --> 09:51.360 We already had a couple of products 09:51.360 --> 09:53.240 that were doing that by then. 09:53.240 --> 09:55.480 And in those cases, we had basically 09:55.480 --> 09:59.280 customized handcrafted code or some internal libraries 09:59.280 --> 10:00.080 that we're using. 10:00.080 --> 10:03.280 So I was actually at Google during this time in a parallel, 10:03.280 --> 10:04.440 I guess, universe. 10:04.440 --> 10:09.240 But we were using Theano and CAFE. 10:09.240 --> 10:11.560 Was there some degree to which you were bouncing, 10:11.560 --> 10:15.440 like trying to see what CAFE was offering people, 10:15.440 --> 10:17.920 trying to see what Theano was offering 10:17.920 --> 10:21.320 that you want to make sure you're delivering on whatever that 10:21.320 --> 10:23.680 is, perhaps the Python part of thing. 10:23.680 --> 10:27.440 Maybe did that influence any design decisions? 10:27.440 --> 10:27.880 Totally. 10:27.880 --> 10:30.840 So when we built this belief, and some of that 10:30.840 --> 10:32.920 was in parallel with some of these libraries 10:32.920 --> 10:36.600 coming up, I mean, Theano itself is older. 10:36.600 --> 10:41.080 But we were building this belief focused on our internal thing 10:41.080 --> 10:42.880 because our systems were very different. 10:42.880 --> 10:44.480 By the time we got to this, we looked 10:44.480 --> 10:47.040 at a number of libraries that were out there. 10:47.040 --> 10:49.240 Theano, there were folks in the group 10:49.240 --> 10:52.080 who had experience with Torch, with Lua. 10:52.080 --> 10:54.720 There were folks here who had seen CAFE. 10:54.720 --> 10:58.800 I mean, actually, Yang Cheng was here as well. 10:58.800 --> 11:02.960 There's what other libraries? 11:02.960 --> 11:04.880 I think we looked at a number of things. 11:04.880 --> 11:06.800 Might even have looked at Jane and her back then. 11:06.800 --> 11:09.320 I'm trying to remember if it was there. 11:09.320 --> 11:12.240 In fact, yeah, we did discuss ideas around, OK, 11:12.240 --> 11:15.280 should we have a graph or not? 11:15.280 --> 11:19.280 And they were supporting all these together 11:19.280 --> 11:21.880 was definitely, you know, there were key decisions 11:21.880 --> 11:22.560 that we wanted. 11:22.560 --> 11:28.680 We had seen limitations in our prior disbelief things. 11:28.680 --> 11:31.320 A few of them were just in terms of research 11:31.320 --> 11:32.280 was moving so fast. 11:32.280 --> 11:34.520 We wanted the flexibility. 11:34.520 --> 11:36.280 We want the hardware was changing fast. 11:36.280 --> 11:39.160 We expected to change that so that those probably were two 11:39.160 --> 11:41.400 things. 11:41.400 --> 11:43.320 And yeah, I think the flexibility in terms 11:43.320 --> 11:45.280 of being able to express all kinds of crazy things 11:45.280 --> 11:46.840 was definitely a big one then. 11:46.840 --> 11:48.920 So what the graph decisions, though, 11:48.920 --> 11:53.720 with moving towards TensorFlow 2.0, there's more, 11:53.720 --> 11:56.680 by default, there'll be eager execution. 11:56.680 --> 11:59.160 So sort of hiding the graph a little bit 11:59.160 --> 12:02.120 because it's less intuitive in terms of the way 12:02.120 --> 12:03.520 people develop and so on. 12:03.520 --> 12:06.720 What was that discussion like with in terms of using graphs? 12:06.720 --> 12:09.320 It seemed it's kind of the theano way. 12:09.320 --> 12:11.600 Did it seem the obvious choice? 12:11.600 --> 12:15.720 So I think where it came from was our disbelief, 12:15.720 --> 12:18.560 had a graph like thing as well. 12:18.560 --> 12:19.720 It wasn't a general graph. 12:19.720 --> 12:23.160 It was more like a straight line thing. 12:23.160 --> 12:25.000 More like what you might think of Cafe, 12:25.000 --> 12:28.840 I guess, in that sense. 12:28.840 --> 12:31.080 And we always cared about the production stuff. 12:31.080 --> 12:33.480 Even with disbelief, we were deploying a whole bunch of stuff 12:33.480 --> 12:34.440 in production. 12:34.440 --> 12:37.960 So graph did come from that when we thought of, OK, 12:37.960 --> 12:40.800 should we do that in Python and we experimented with some ideas 12:40.800 --> 12:44.680 where it looked a lot simpler to use, 12:44.680 --> 12:47.880 but not having a graph meant, OK, how do you deploy now? 12:47.880 --> 12:51.080 So that was probably what tilted the balance for us. 12:51.080 --> 12:52.880 And eventually, we ended up with the graph. 12:52.880 --> 12:55.320 And I guess the question there is, did you? 12:55.320 --> 12:59.800 I mean, production seems to be the really good thing to focus on. 12:59.800 --> 13:02.400 But did you even anticipate the other side of it 13:02.400 --> 13:04.560 where there could be, what is it? 13:04.560 --> 13:05.240 What are the numbers? 13:05.240 --> 13:08.920 Something crazy, 41 million downloads? 13:08.920 --> 13:09.400 Yep. 13:12.680 --> 13:16.240 I mean, was that even like a possibility in your mind 13:16.240 --> 13:19.120 that it would be as popular as it became? 13:19.120 --> 13:24.880 So I think we did see a need for this a lot 13:24.880 --> 13:30.000 from the research perspective and early days of deep learning 13:30.000 --> 13:32.280 in some ways. 13:32.280 --> 13:33.040 41 million? 13:33.040 --> 13:37.640 No, I don't think I imagine this number then. 13:37.640 --> 13:42.760 It seemed like there's a potential future where lots more people 13:42.760 --> 13:43.760 would be doing this. 13:43.760 --> 13:45.640 And how do we enable that? 13:45.640 --> 13:49.560 I would say this kind of growth, I probably 13:49.560 --> 13:53.680 started seeing somewhat after the open sourcing where it was 13:53.680 --> 13:56.240 like, OK, deep learning is actually 13:56.240 --> 13:59.200 growing way faster for a lot of different reasons. 13:59.200 --> 14:02.720 And we are in just the right place to push on that 14:02.720 --> 14:06.040 and leverage that and deliver on lots of things 14:06.040 --> 14:07.440 that people want. 14:07.440 --> 14:09.760 So what changed once the open source? 14:09.760 --> 14:13.320 Like how this incredible amount of attention 14:13.320 --> 14:16.120 from a global population of developers, 14:16.120 --> 14:18.200 how did the projects start changing? 14:18.200 --> 14:21.880 I don't even actually remember it during those times. 14:21.880 --> 14:24.560 I know looking now, there's really good documentation. 14:24.560 --> 14:26.560 There's an ecosystem of tools. 14:26.560 --> 14:31.080 There's a YouTube channel now. 14:31.080 --> 14:33.760 It's very community driven. 14:33.760 --> 14:38.800 Back then, I guess 0.1 version. 14:38.800 --> 14:39.760 Is that the version? 14:39.760 --> 14:42.680 I think we called it 0.6 or 5, something like that. 14:42.680 --> 14:43.720 Something like that. 14:43.720 --> 14:47.200 What changed leading into 1.0? 14:47.200 --> 14:48.480 It's interesting. 14:48.480 --> 14:51.640 I think we've gone through a few things there. 14:51.640 --> 14:53.680 When we started out, when we first came out, 14:53.680 --> 14:56.080 people loved the documentation we have. 14:56.080 --> 14:58.800 Because it was just a huge step up from everything else. 14:58.800 --> 15:01.920 Because all of those were academic projects, people 15:01.920 --> 15:04.560 don't think about documentation. 15:04.560 --> 15:08.040 I think what that changed was instead of deep learning 15:08.040 --> 15:12.560 being a research thing, some people who were just developers 15:12.560 --> 15:15.080 could now suddenly take this out and do 15:15.080 --> 15:16.920 some interesting things with it. 15:16.920 --> 15:20.720 Who had no clue what machine learning was before then. 15:20.720 --> 15:22.520 And that, I think, really changed 15:22.520 --> 15:27.880 how things started to scale up in some ways and pushed on it. 15:27.880 --> 15:30.400 Over the next few months, as we looked at, 15:30.400 --> 15:31.960 how do we stabilize things? 15:31.960 --> 15:33.840 As we look at not just researchers, 15:33.840 --> 15:34.880 now we want stability. 15:34.880 --> 15:36.480 People want to deploy things. 15:36.480 --> 15:38.960 That's how we started planning for 1.0. 15:38.960 --> 15:42.240 And there are certain needs for that perspective. 15:42.240 --> 15:45.320 And so, again, documentation comes up, 15:45.320 --> 15:49.480 designs, more kinds of things to put that together. 15:49.480 --> 15:53.120 And so that was exciting to get that to a stage where 15:53.120 --> 15:56.400 more and more enterprises wanted to buy in and really 15:56.400 --> 15:58.720 get behind that. 15:58.720 --> 16:02.640 And I think post 1.0 and with the next few releases, 16:02.640 --> 16:05.240 their enterprise adoption also started to take off. 16:05.240 --> 16:08.000 I would say between the initial release and 1.0, 16:08.000 --> 16:11.000 it was, OK, researchers, of course. 16:11.000 --> 16:13.720 Then a lot of hobbies and early interest, 16:13.720 --> 16:15.920 people excited about this who started to get on board. 16:15.920 --> 16:19.000 And then over the 1.x thing, lots of enterprises. 16:19.000 --> 16:23.760 I imagine anything that's below 1.0 16:23.760 --> 16:27.160 gets pressured to be enterprise problem or something 16:27.160 --> 16:28.000 that's stable. 16:28.000 --> 16:28.800 Exactly. 16:28.800 --> 16:33.360 And do you have a sense now that TensorFlow is stable? 16:33.360 --> 16:35.520 It feels like deep learning, in general, 16:35.520 --> 16:37.800 is extremely dynamic field. 16:37.800 --> 16:39.680 So much is changing. 16:39.680 --> 16:43.400 Do you have a, and TensorFlow has been growing incredibly. 16:43.400 --> 16:46.720 Do you have a sense of stability at the helm of this? 16:46.720 --> 16:48.360 I mean, I know you're in the midst of it. 16:48.360 --> 16:50.360 Yeah. 16:50.360 --> 16:54.000 I think in the midst of it, it's often easy to forget what 16:54.000 --> 16:58.160 an enterprise wants and what some of the people on that side 16:58.160 --> 16:58.760 want. 16:58.760 --> 17:00.360 There are still people running models 17:00.360 --> 17:02.640 that are three years old, four years old. 17:02.640 --> 17:06.000 So inception is still used by tons of people. 17:06.000 --> 17:08.880 Even less than 50 is what, a couple of years old now or more. 17:08.880 --> 17:12.200 But there are tons of people who use that, and they're fine. 17:12.200 --> 17:16.200 They don't need the last couple of bits of performance or quality. 17:16.200 --> 17:19.600 They want some stability in things that just work. 17:19.600 --> 17:22.720 And so there is value in providing that with that kind 17:22.720 --> 17:25.160 of stability and making it really simpler, 17:25.160 --> 17:27.800 because that allows a lot more people to access it. 17:27.800 --> 17:31.640 And then there's the research crowd, which wants, OK, 17:31.640 --> 17:33.680 they want to do these crazy things exactly like you're 17:33.680 --> 17:37.000 saying, not just deep learning in the straight up models 17:37.000 --> 17:38.400 that used to be there. 17:38.400 --> 17:41.920 They want RNNs, and even RNNs are maybe old. 17:41.920 --> 17:45.520 They are transformers now, and now it 17:45.520 --> 17:48.720 needs to combine with RL and GANs and so on. 17:48.720 --> 17:52.160 So there's definitely that area, the boundary that's 17:52.160 --> 17:55.120 shifting and pushing the state of the art. 17:55.120 --> 17:57.120 But I think there's more and more of the past 17:57.120 --> 17:59.680 that's much more stable. 17:59.680 --> 18:02.680 And even stuff that was two, three years old 18:02.680 --> 18:04.920 is very, very usable by lots of people. 18:04.920 --> 18:07.440 So that part makes it a lot easier. 18:07.440 --> 18:09.800 So I imagine maybe you can correct me if I'm wrong. 18:09.800 --> 18:12.440 One of the biggest use cases is essentially 18:12.440 --> 18:15.160 taking something like ResNet 50 and doing 18:15.160 --> 18:18.520 some kind of transfer learning on a very particular problem 18:18.520 --> 18:19.600 that you have. 18:19.600 --> 18:24.480 It's basically probably what majority of the world does. 18:24.480 --> 18:27.040 And you want to make that as easy as possible. 18:27.040 --> 18:30.400 So I would say, for the hobbyist perspective, 18:30.400 --> 18:32.800 that's the most common case. 18:32.800 --> 18:34.800 In fact, the apps on phones and stuff 18:34.800 --> 18:37.680 that you'll see, the early ones, that's the most common case. 18:37.680 --> 18:40.320 I would say there are a couple of reasons for that. 18:40.320 --> 18:44.400 One is that everybody talks about that. 18:44.400 --> 18:46.120 It looks great on slides. 18:46.120 --> 18:48.120 That's a great presentation. 18:48.120 --> 18:50.040 Exactly. 18:50.040 --> 18:53.120 What enterprises want is that is part of it, 18:53.120 --> 18:54.480 but that's not the big thing. 18:54.480 --> 18:56.760 Enterprises really have data that they 18:56.760 --> 18:58.040 want to make predictions on. 18:58.040 --> 19:01.160 This is often what they used to do with the people who 19:01.160 --> 19:03.600 were doing ML was just regression models, 19:03.600 --> 19:06.440 linear regression, logistic regression, linear models, 19:06.440 --> 19:09.800 or maybe gradient booster trees and so on. 19:09.800 --> 19:11.760 Some of them still benefit from deep learning, 19:11.760 --> 19:14.440 but they weren't that that's the bread and butter, 19:14.440 --> 19:16.280 like the structured data and so on. 19:16.280 --> 19:18.200 So depending on the audience you look at, 19:18.200 --> 19:19.520 they're a little bit different. 19:19.520 --> 19:23.320 And they just have, I mean, the best of enterprise 19:23.320 --> 19:26.480 probably just has a very large data set 19:26.480 --> 19:28.640 where deep learning can probably shine. 19:28.640 --> 19:29.360 That's correct. 19:29.360 --> 19:30.320 That's right. 19:30.320 --> 19:32.240 And then I think the other pieces 19:32.240 --> 19:34.560 that they wanted, again, to point out 19:34.560 --> 19:36.400 that the developer summit we put together 19:36.400 --> 19:38.200 is that the whole TensorFlow Extended 19:38.200 --> 19:40.600 piece, which is the entire pipeline, 19:40.600 --> 19:43.560 they care about stability across doing their entire thing. 19:43.560 --> 19:46.200 They want simplicity across the entire thing. 19:46.200 --> 19:47.680 I don't need to just train a model. 19:47.680 --> 19:51.280 I need to do that every day again, over and over again. 19:51.280 --> 19:54.720 I wonder to which degree you have a role in, I don't know. 19:54.720 --> 19:57.040 So I teach a course on deep learning. 19:57.040 --> 20:01.320 I have people like lawyers come up to me and say, 20:01.320 --> 20:04.200 when is machine learning going to enter legal, 20:04.200 --> 20:05.560 the legal realm? 20:05.560 --> 20:11.720 The same thing in all kinds of disciplines, immigration, 20:11.720 --> 20:13.800 insurance. 20:13.800 --> 20:17.400 Often when I see what it boils down to is these companies 20:17.400 --> 20:19.760 are often a little bit old school in the way 20:19.760 --> 20:20.840 they organize the data. 20:20.840 --> 20:22.800 So the data is just not ready yet. 20:22.800 --> 20:24.040 It's not digitized. 20:24.040 --> 20:28.160 Do you also find yourself being in the role of an evangelist 20:28.160 --> 20:33.040 for let's organize your data, folks, 20:33.040 --> 20:35.440 and then you'll get the big benefit of TensorFlow? 20:35.440 --> 20:38.000 Do you have those conversations? 20:38.000 --> 20:45.160 Yeah, I get all kinds of questions there from, OK, 20:45.160 --> 20:49.000 what do I need to make this work, right? 20:49.000 --> 20:50.800 Do we really need deep learning? 20:50.800 --> 20:52.240 I mean, there are all these things. 20:52.240 --> 20:54.000 I already used this linear model. 20:54.000 --> 20:55.160 Why would this help? 20:55.160 --> 20:57.160 I don't have enough data, let's say. 20:57.160 --> 20:59.960 Or I want to use machine learning, 20:59.960 --> 21:01.760 but I have no clue where to start. 21:01.760 --> 21:04.920 So it's a great start to all the way to the experts 21:04.920 --> 21:08.520 who wise were very specific things, so it's interesting. 21:08.520 --> 21:09.600 Is there a good answer? 21:09.600 --> 21:12.480 It boils down to oftentimes digitizing data. 21:12.480 --> 21:15.240 So whatever you want automated, whatever data 21:15.240 --> 21:17.480 you want to make prediction based on, 21:17.480 --> 21:21.240 you have to make sure that it's in an organized form. 21:21.240 --> 21:23.920 Like with an intensive flow ecosystem, 21:23.920 --> 21:26.080 there's now you're providing more and more data 21:26.080 --> 21:28.960 sets and more and more pretrained models. 21:28.960 --> 21:32.400 Are you finding yourself also the organizer of data sets? 21:32.400 --> 21:34.480 Yes, I think with TensorFlow data sets 21:34.480 --> 21:38.360 that we just released, that's definitely come up where people 21:38.360 --> 21:39.200 want these data sets. 21:39.200 --> 21:41.560 Can we organize them and can we make that easier? 21:41.560 --> 21:45.320 So that's definitely one important thing. 21:45.320 --> 21:47.680 The other related thing I would say is I often tell people, 21:47.680 --> 21:50.960 you know what, don't think of the most fanciest thing 21:50.960 --> 21:53.320 that the newest model that you see. 21:53.320 --> 21:55.480 Make something very basic work, and then 21:55.480 --> 21:56.360 you can improve it. 21:56.360 --> 21:58.840 There's just lots of things you can do with it. 21:58.840 --> 22:00.080 Yeah, start with the basics. 22:00.080 --> 22:00.580 Sure. 22:00.580 --> 22:03.760 One of the big things that makes TensorFlow even more 22:03.760 --> 22:06.440 accessible was the appearance, whenever 22:06.440 --> 22:12.400 that happened, of Keras, the Keras standard outside of TensorFlow. 22:12.400 --> 22:18.200 I think it was Keras on top of Tiano at first only, 22:18.200 --> 22:22.480 and then Keras became on top of TensorFlow. 22:22.480 --> 22:29.840 Do you know when Keras chose to also add TensorFlow as a back end, 22:29.840 --> 22:33.960 who was it just the community that drove that initially? 22:33.960 --> 22:37.000 Do you know if there was discussions, conversations? 22:37.000 --> 22:40.920 Yeah, so Franco started the Keras project 22:40.920 --> 22:44.560 before he was at Google, and the first thing was Tiano. 22:44.560 --> 22:47.120 I don't remember if that was after TensorFlow 22:47.120 --> 22:49.640 was created or way before. 22:49.640 --> 22:52.000 And then at some point, when TensorFlow 22:52.000 --> 22:54.160 started becoming popular, there were enough similarities 22:54.160 --> 22:56.320 that he decided to create this interface 22:56.320 --> 22:59.200 and put TensorFlow as a back end. 22:59.200 --> 23:03.320 I believe that might still have been before he joined Google. 23:03.320 --> 23:06.720 So we weren't really talking about that. 23:06.720 --> 23:09.720 He decided on his own and thought that was interesting 23:09.720 --> 23:12.760 and relevant to the community. 23:12.760 --> 23:17.080 In fact, I didn't find out about him being at Google 23:17.080 --> 23:19.680 until a few months after he was here. 23:19.680 --> 23:21.840 He was working on some research ideas. 23:21.840 --> 23:24.480 And doing Keras and his nights and weekends project and stuff. 23:24.480 --> 23:25.280 I wish this thing. 23:25.280 --> 23:28.480 So he wasn't part of the TensorFlow. 23:28.480 --> 23:29.680 He didn't join initially. 23:29.680 --> 23:32.240 He joined research, and he was doing some amazing research. 23:32.240 --> 23:35.440 He has some papers on that and research. 23:35.440 --> 23:38.400 He's a great researcher as well. 23:38.400 --> 23:42.400 And at some point, we realized, oh, he's doing this good stuff. 23:42.400 --> 23:45.480 People seem to like the API, and he's right here. 23:45.480 --> 23:48.280 So we talked to him, and he said, OK, 23:48.280 --> 23:50.600 why don't I come over to your team 23:50.600 --> 23:52.800 and work with you for a quarter? 23:52.800 --> 23:55.440 And let's make that integration happen. 23:55.440 --> 23:57.200 And we talked to his manager, and he said, sure, 23:57.200 --> 23:59.720 what, quarter's fine. 23:59.720 --> 24:03.320 And that quarter's been something like two years now. 24:03.320 --> 24:05.040 So he's fully on this. 24:05.040 --> 24:12.000 So Keras got integrated into TensorFlow in a deep way. 24:12.000 --> 24:15.920 And now with TensorFlow 2.0, Keras 24:15.920 --> 24:19.400 is kind of the recommended way for a beginner 24:19.400 --> 24:21.960 to interact with TensorFlow, which 24:21.960 --> 24:24.640 makes that initial sort of transfer learning 24:24.640 --> 24:28.040 or the basic use cases, even for an enterprise, 24:28.040 --> 24:29.320 super simple, right? 24:29.320 --> 24:29.920 That's correct. 24:29.920 --> 24:30.440 That's right. 24:30.440 --> 24:32.040 So what was that decision like? 24:32.040 --> 24:38.640 That seems like it's kind of a bold decision as well. 24:38.640 --> 24:41.200 We did spend a lot of time thinking about that one. 24:41.200 --> 24:46.000 We had a bunch of APIs some bit by us. 24:46.000 --> 24:48.760 There was a parallel layers API that we were building 24:48.760 --> 24:51.560 and when we decided to do Keras in parallel, 24:51.560 --> 24:54.400 so they were like, OK, two things that we are looking at. 24:54.400 --> 24:55.960 And the first thing we was trying to do 24:55.960 --> 25:00.080 is just have them look similar, be as integrated as possible, 25:00.080 --> 25:02.200 share all of that stuff. 25:02.200 --> 25:05.800 There were also three other APIs that others had built over time 25:05.800 --> 25:09.000 because we didn't have a standard one. 25:09.000 --> 25:12.080 But one of the messages that we kept hearing from the community, 25:12.080 --> 25:13.200 OK, which one do we use? 25:13.200 --> 25:15.560 And they kept seeing, OK, here's a model in this one, 25:15.560 --> 25:18.840 and here's a model in this one, which should I pick? 25:18.840 --> 25:22.680 So that's sort of like, OK, we had to address that 25:22.680 --> 25:24.000 straight on with 2.0. 25:24.000 --> 25:26.320 The whole idea was we need to simplify. 25:26.320 --> 25:28.600 We had to pick one. 25:28.600 --> 25:34.600 Based on where we were, we were like, OK, let's see what 25:34.600 --> 25:35.640 are the people like. 25:35.640 --> 25:39.280 And Keras was clearly one that lots of people loved. 25:39.280 --> 25:41.600 There were lots of great things about it. 25:41.600 --> 25:43.880 So we settled on that. 25:43.880 --> 25:44.680 Organically. 25:44.680 --> 25:46.560 That's kind of the best way to do it. 25:46.560 --> 25:47.160 It was great. 25:47.160 --> 25:48.720 But it was surprising, nevertheless, 25:48.720 --> 25:51.120 to sort of bring in and outside. 25:51.120 --> 25:54.440 I mean, there was a feeling like Keras might be almost 25:54.440 --> 25:58.000 like a competitor in a certain kind of a two tensor flow. 25:58.000 --> 26:01.320 And in a sense, it became an empowering element 26:01.320 --> 26:02.200 of tensor flow. 26:02.200 --> 26:03.280 That's right. 26:03.280 --> 26:07.200 Yeah, it's interesting how you can put two things together 26:07.200 --> 26:08.280 which can align right. 26:08.280 --> 26:11.760 And in this case, I think Francois, the team, 26:11.760 --> 26:15.480 and a bunch of us have chatted and I think we all 26:15.480 --> 26:17.320 want to see the same kind of things. 26:17.320 --> 26:20.360 We all care about making it easier for the huge set 26:20.360 --> 26:21.440 of developers out there. 26:21.440 --> 26:23.440 And that makes a difference. 26:23.440 --> 26:27.280 So Python has Guido van Rossum, who 26:27.280 --> 26:30.320 until recently held the position of benevolent 26:30.320 --> 26:31.960 dictator for life. 26:31.960 --> 26:36.040 Right, so there's a huge successful open source 26:36.040 --> 26:37.320 project like tensor flow. 26:37.320 --> 26:40.680 Need one person who makes a final decision. 26:40.680 --> 26:45.480 So you did a pretty successful tensor flow Dev Summit 26:45.480 --> 26:47.520 just now, last couple of days. 26:47.520 --> 26:51.080 There's clearly a lot of different new features 26:51.080 --> 26:55.480 being incorporated in amazing ecosystem, so on. 26:55.480 --> 26:57.320 How are those design decisions made? 26:57.320 --> 27:00.960 Is there a BDFL in tensor flow? 27:00.960 --> 27:05.800 And or is it more distributed and organic? 27:05.800 --> 27:09.880 I think it's somewhat different, I would say. 27:09.880 --> 27:16.160 I've always been involved in the key design directions. 27:16.160 --> 27:17.560 But there are lots of things that 27:17.560 --> 27:20.960 are distributed where their number of people, Martin 27:20.960 --> 27:24.760 Wick being one who has really driven a lot of our open source 27:24.760 --> 27:27.360 stuff, a lot of the APIs. 27:27.360 --> 27:29.200 And there are a number of other people 27:29.200 --> 27:32.720 who have been pushed and been responsible 27:32.720 --> 27:35.240 for different parts of it. 27:35.240 --> 27:37.840 We do have regular design reviews. 27:37.840 --> 27:40.680 Over the last year, we've really spent a lot of time opening up 27:40.680 --> 27:44.160 to the community and adding transparency. 27:44.160 --> 27:45.880 We're setting more processes in place, 27:45.880 --> 27:49.600 so RFCs, special interest groups, really 27:49.600 --> 27:53.560 grow that community and scale that. 27:53.560 --> 27:57.680 I think the kind of scale that ecosystem is in, 27:57.680 --> 28:00.240 I don't think we could scale with having me as the lone 28:00.240 --> 28:02.320 point of decision maker. 28:02.320 --> 28:03.440 I got it. 28:03.440 --> 28:05.880 So yeah, the growth of that ecosystem, 28:05.880 --> 28:08.040 maybe you can talk about it a little bit. 28:08.040 --> 28:10.720 First of all, when I started with Andre Karpathi 28:10.720 --> 28:13.640 when he first did ComNet.js, the fact 28:13.640 --> 28:15.360 that you can train in your own network 28:15.360 --> 28:18.480 and the browser in JavaScript was incredible. 28:18.480 --> 28:21.000 So now TensorFlow.js is really making 28:21.000 --> 28:26.920 that a serious, a legit thing, a way 28:26.920 --> 28:29.560 to operate, whether it's in the back end or the front end. 28:29.560 --> 28:32.720 Then there's the TensorFlow Extended, like you mentioned. 28:32.720 --> 28:35.360 There's TensorFlow Lite for mobile. 28:35.360 --> 28:37.480 And all of it, as far as I can tell, 28:37.480 --> 28:39.640 it's really converging towards being 28:39.640 --> 28:43.440 able to save models in the same kind of way. 28:43.440 --> 28:46.680 You can move around, you can train on the desktop, 28:46.680 --> 28:48.800 and then move it to mobile, and so on. 28:48.800 --> 28:49.280 That's right. 28:49.280 --> 28:52.320 So this is that cohesiveness. 28:52.320 --> 28:55.240 So can you maybe give me whatever 28:55.240 --> 28:58.840 I missed, a bigger overview of the mission of the ecosystem 28:58.840 --> 29:02.120 that's trying to be built, and where is it moving forward? 29:02.120 --> 29:02.800 Yeah. 29:02.800 --> 29:05.720 So in short, the way I like to think of this 29:05.720 --> 29:09.760 is our goals to enable machine learning. 29:09.760 --> 29:13.320 And in a couple of ways, one is we 29:13.320 --> 29:16.560 have lots of exciting things going on in ML today. 29:16.560 --> 29:18.160 We started with deep learning, but we now 29:18.160 --> 29:21.400 support a bunch of other algorithms too. 29:21.400 --> 29:23.760 So one is to, on the research side, 29:23.760 --> 29:25.360 keep pushing on the state of the art. 29:25.360 --> 29:27.240 Can we, how do we enable researchers 29:27.240 --> 29:28.960 to build the next amazing thing? 29:28.960 --> 29:31.800 So BERT came out recently. 29:31.800 --> 29:34.000 It's great that people are able to do new kinds of research. 29:34.000 --> 29:35.400 There are lots of amazing research 29:35.400 --> 29:37.600 that happens across the world. 29:37.600 --> 29:38.880 So that's one direction. 29:38.880 --> 29:41.400 The other is, how do you take that 29:41.400 --> 29:45.200 across all the people outside who want to take that research 29:45.200 --> 29:47.400 and do some great things with it and integrate it 29:47.400 --> 29:51.800 to build real products, to have a real impact on people? 29:51.800 --> 29:56.720 And so if that's the other axes in some ways. 29:56.720 --> 29:58.520 And a high level, one way I think about it 29:58.520 --> 30:02.480 is there are a crazy number of computer devices 30:02.480 --> 30:04.240 across the world. 30:04.240 --> 30:08.440 And we often used to think of ML and training and all of this 30:08.440 --> 30:10.800 as, OK, something you do either in the workstation 30:10.800 --> 30:13.600 or the data center or cloud. 30:13.600 --> 30:15.720 But we see things running on the phones. 30:15.720 --> 30:17.640 We see things running on really tiny chips. 30:17.640 --> 30:20.760 And we had some demos at the developer summit. 30:20.760 --> 30:25.160 And so the way I think about this ecosystem 30:25.160 --> 30:30.280 is, how do we help get machine learning on every device that 30:30.280 --> 30:32.520 has a compute capability? 30:32.520 --> 30:33.760 And that continues to grow. 30:33.760 --> 30:37.240 And so in some ways, this ecosystem 30:37.240 --> 30:40.280 has looked at various aspects of that 30:40.280 --> 30:42.440 and grown over time to cover more of those. 30:42.440 --> 30:44.640 And we continue to push the boundaries. 30:44.640 --> 30:48.640 In some areas, we've built more tooling and things 30:48.640 --> 30:50.040 around that to help you. 30:50.040 --> 30:52.800 I mean, the first tool we started was TensorBoard. 30:52.800 --> 30:56.920 You want to learn just the training piece, the effects 30:56.920 --> 30:59.840 for TensorFlow Extended to really do your entire ML 30:59.840 --> 31:04.760 pipelines if you care about all that production stuff, 31:04.760 --> 31:09.520 but then going to the edge, going to different kinds of things. 31:09.520 --> 31:11.800 And it's not just us now. 31:11.800 --> 31:15.120 We are a place where there are lots of libraries being built 31:15.120 --> 31:15.840 on top. 31:15.840 --> 31:18.440 So there are some for research, maybe things 31:18.440 --> 31:21.240 like TensorFlow Agents or TensorFlow Probability that 31:21.240 --> 31:23.480 started as research things or for researchers 31:23.480 --> 31:26.160 for focusing on certain kinds of algorithms, 31:26.160 --> 31:30.280 but they're also being deployed or reduced by production folks. 31:30.280 --> 31:34.000 And some have come from within Google, just teams 31:34.000 --> 31:37.040 across Google who wanted to do the build these things. 31:37.040 --> 31:39.680 Others have come from just the community 31:39.680 --> 31:41.840 because there are different pieces 31:41.840 --> 31:44.640 that different parts of the community care about. 31:44.640 --> 31:49.520 And I see our goal as enabling even that. 31:49.520 --> 31:53.240 It's not we cannot and won't build every single thing. 31:53.240 --> 31:54.840 That just doesn't make sense. 31:54.840 --> 31:57.320 But if we can enable others to build the things 31:57.320 --> 32:00.640 that they care about, and there's a broader community that 32:00.640 --> 32:02.880 cares about that, and we can help encourage that, 32:02.880 --> 32:05.240 and that's great. 32:05.240 --> 32:08.600 That really helps the entire ecosystem, not just those. 32:08.600 --> 32:11.280 One of the big things about 2.0 that we're pushing on 32:11.280 --> 32:14.680 is, OK, we have these so many different pieces, right? 32:14.680 --> 32:18.440 How do we help make all of them work well together? 32:18.440 --> 32:21.960 There are a few key pieces there that we're pushing on, 32:21.960 --> 32:23.840 one being the core format in there 32:23.840 --> 32:27.480 and how we share the models themselves through SAVE model 32:27.480 --> 32:30.440 and what TensorFlow Hub and so on. 32:30.440 --> 32:34.000 And a few of the pieces that we really put this together. 32:34.000 --> 32:37.240 I was very skeptical that that's, when TensorFlow.js came out, 32:37.240 --> 32:40.120 it didn't seem or deep learning.js. 32:40.120 --> 32:41.680 Yeah, that was the first. 32:41.680 --> 32:45.040 It seems like technically very difficult project. 32:45.040 --> 32:47.040 As a standalone, it's not as difficult. 32:47.040 --> 32:49.920 But as a thing that integrates into the ecosystem, 32:49.920 --> 32:51.240 it seems very difficult. 32:51.240 --> 32:53.200 So I mean, there's a lot of aspects of this 32:53.200 --> 32:54.200 you're making look easy. 32:54.200 --> 32:58.160 But on the technical side, how many challenges 32:58.160 --> 33:00.560 have to be overcome here? 33:00.560 --> 33:01.520 A lot. 33:01.520 --> 33:03.080 And still have to be overcome. 33:03.080 --> 33:04.840 That's the question here, too. 33:04.840 --> 33:06.160 There are lots of steps to it. 33:06.160 --> 33:08.160 I think we've iterated over the last few years, 33:08.160 --> 33:10.720 so there's a lot we've learned. 33:10.720 --> 33:14.200 I, yeah, and often when things come together well, 33:14.200 --> 33:15.080 things look easy. 33:15.080 --> 33:16.400 And that's exactly the point. 33:16.400 --> 33:18.280 It should be easy for the end user. 33:18.280 --> 33:21.320 But there are lots of things that go behind that. 33:21.320 --> 33:25.320 If I think about still challenges ahead, 33:25.320 --> 33:32.880 there are we have a lot more devices coming on board, 33:32.880 --> 33:35.280 for example, from the hardware perspective. 33:35.280 --> 33:37.640 How do we make it really easy for these vendors 33:37.640 --> 33:42.040 to integrate with something like TensorFlow? 33:42.040 --> 33:43.640 So there's a lot of compiler stuff 33:43.640 --> 33:45.320 that others are working on. 33:45.320 --> 33:48.320 There are things we can do in terms of our APIs 33:48.320 --> 33:50.520 and so on that we can do. 33:50.520 --> 33:55.840 As we, TensorFlow started as a very monolithic system. 33:55.840 --> 33:57.680 And to some extent, it still is. 33:57.680 --> 33:59.400 There are less lots of tools around it, 33:59.400 --> 34:02.960 but the core is still pretty large and monolithic. 34:02.960 --> 34:05.760 One of the key challenges for us to scale that out 34:05.760 --> 34:10.440 is how do we break that apart with clear interfaces? 34:10.440 --> 34:13.720 It's, in some ways, it's software engineering one 34:13.720 --> 34:18.520 one, but for a system that's now four years old, I guess, 34:18.520 --> 34:21.600 or more, and that's still rapidly evolving 34:21.600 --> 34:24.000 and that we're not slowing down with, 34:24.000 --> 34:28.240 it's hard to change and modify and really break apart. 34:28.240 --> 34:29.880 It's sort of like, as people say, right, 34:29.880 --> 34:32.560 it's like changing the engine with a car running 34:32.560 --> 34:33.560 or fixed benefits. 34:33.560 --> 34:35.200 That's exactly what we're trying to do. 34:35.200 --> 34:39.960 So there's a challenge here, because the downside 34:39.960 --> 34:43.840 of so many people being excited about TensorFlow 34:43.840 --> 34:48.600 and becoming to rely on it in many other applications 34:48.600 --> 34:52.200 is that you're kind of responsible. 34:52.200 --> 34:53.520 It's the technical debt. 34:53.520 --> 34:55.640 You're responsible for previous versions 34:55.640 --> 34:57.720 to some degree still working. 34:57.720 --> 34:59.920 So when you're trying to innovate, 34:59.920 --> 35:03.760 I mean, it's probably easier to just start from scratch 35:03.760 --> 35:05.800 every few months. 35:05.800 --> 35:07.160 Absolutely. 35:07.160 --> 35:10.880 So do you feel the pain of that? 35:10.880 --> 35:15.360 2.0 does break some back compatibility, but not too much. 35:15.360 --> 35:18.120 It seems like the conversion is pretty straightforward. 35:18.120 --> 35:20.240 Do you think that's still important, 35:20.240 --> 35:22.880 given how quickly deep learning is changing? 35:22.880 --> 35:26.360 Can you just, the things that you've learned, 35:26.360 --> 35:27.440 can you just start over? 35:27.440 --> 35:30.120 Or is there pressure to not? 35:30.120 --> 35:31.640 It's a tricky balance. 35:31.640 --> 35:36.840 So if it was just a researcher writing a paper who 35:36.840 --> 35:39.400 a year later will not look at that code again, 35:39.400 --> 35:41.560 sure, it doesn't matter. 35:41.560 --> 35:43.440 There are a lot of production systems 35:43.440 --> 35:45.480 that rely on TensorFlow, both at Google 35:45.480 --> 35:47.240 and across the world. 35:47.240 --> 35:49.760 And people worry about this. 35:49.760 --> 35:53.400 I mean, these systems run for a long time. 35:53.400 --> 35:57.240 So it is important to keep that compatibility and so on. 35:57.240 --> 36:00.960 And yes, it does come with a huge cost. 36:00.960 --> 36:02.920 We have to think about a lot of things 36:02.920 --> 36:06.960 as we do new things and make new changes. 36:06.960 --> 36:09.120 I think it's a trade off, right? 36:09.120 --> 36:12.960 You can, you might slow certain kinds of things down, 36:12.960 --> 36:15.440 but the overall value you're bringing because of that 36:15.440 --> 36:18.440 is much bigger because it's not just 36:18.440 --> 36:20.520 about breaking the person yesterday. 36:20.520 --> 36:24.840 It's also about telling the person tomorrow that, you know what? 36:24.840 --> 36:26.320 This is how we do things. 36:26.320 --> 36:28.520 We're not going to break you when you come on board 36:28.520 --> 36:30.320 because there are lots of new people who are also 36:30.320 --> 36:32.880 going to come on board. 36:32.880 --> 36:34.680 So one way I like to think about this, 36:34.680 --> 36:37.960 and I always push the team to think about it as well, 36:37.960 --> 36:39.640 when you want to do new things, you 36:39.640 --> 36:42.000 want to start with a clean slate, 36:42.000 --> 36:44.880 design with a clean slate in mind, 36:44.880 --> 36:48.160 and then we'll figure out how to make sure all the other things 36:48.160 --> 36:48.640 work. 36:48.640 --> 36:52.160 And yes, we do make compromises occasionally. 36:52.160 --> 36:55.200 But unless you design with the clean slate 36:55.200 --> 36:58.400 and not worry about that, you'll never get to a good place. 36:58.400 --> 36:59.120 That's brilliant. 36:59.120 --> 37:04.080 So even if you are responsible in the idea stage, 37:04.080 --> 37:07.680 when you're thinking of new, just put all that behind you. 37:07.680 --> 37:09.600 OK, that's really well put. 37:09.600 --> 37:12.480 So I have to ask this because a lot of students, developers, 37:12.480 --> 37:16.280 asked me how I feel about PyTorch versus TensorFlow. 37:16.280 --> 37:19.720 So I've recently completely switched my research group 37:19.720 --> 37:20.920 to TensorFlow. 37:20.920 --> 37:23.280 I wish everybody would just use the same thing. 37:23.280 --> 37:26.960 And TensorFlow is as close to that, I believe, as we have. 37:26.960 --> 37:32.000 But do you enjoy competition? 37:32.000 --> 37:35.800 So TensorFlow is leading in many ways, many dimensions 37:35.800 --> 37:39.000 in terms of the ecosystem, in terms of the number of users, 37:39.000 --> 37:41.200 momentum power, production level, so on. 37:41.200 --> 37:46.000 But a lot of researchers are now also using PyTorch. 37:46.000 --> 37:47.520 Do you enjoy that kind of competition, 37:47.520 --> 37:49.440 or do you just ignore it and focus 37:49.440 --> 37:52.320 on making TensorFlow the best that it can be? 37:52.320 --> 37:55.480 So just like research or anything people are doing, 37:55.480 --> 37:58.120 it's great to get different kinds of ideas. 37:58.120 --> 38:01.440 And when we started with TensorFlow, 38:01.440 --> 38:05.480 like I was saying earlier, it was very important for us 38:05.480 --> 38:07.440 to also have production in mind. 38:07.440 --> 38:08.960 We didn't want just research, right? 38:08.960 --> 38:11.280 And that's why we chose certain things. 38:11.280 --> 38:13.480 Now PyTorch came along and said, you know what? 38:13.480 --> 38:14.880 I only care about research. 38:14.880 --> 38:16.320 This is what I'm trying to do. 38:16.320 --> 38:18.400 What's the best thing I can do for this? 38:18.400 --> 38:21.120 And it started iterating and said, OK, 38:21.120 --> 38:22.520 I don't need to worry about graphs. 38:22.520 --> 38:25.200 Let me just run things. 38:25.200 --> 38:27.440 I don't care if it's not as fast as it can be, 38:27.440 --> 38:30.480 but let me just make this part easy. 38:30.480 --> 38:32.560 And there are things you can learn from that, right? 38:32.560 --> 38:36.720 They, again, had the benefit of seeing what had come before, 38:36.720 --> 38:40.520 but also exploring certain different kinds of spaces. 38:40.520 --> 38:43.560 And they had some good things there, 38:43.560 --> 38:46.680 building on, say, things like Jainer and so on before that. 38:46.680 --> 38:49.320 So competition is definitely interesting. 38:49.320 --> 38:51.040 It made us, you know, this is an area 38:51.040 --> 38:53.720 that we had thought about, like I said, very early on. 38:53.720 --> 38:56.600 Over time, we had revisited this a couple of times. 38:56.600 --> 38:59.000 Should we add this again? 38:59.000 --> 39:00.480 At some point, we said, you know what, 39:00.480 --> 39:02.920 here's it seems like this can be done well. 39:02.920 --> 39:04.280 So let's try it again. 39:04.280 --> 39:07.680 And that's how we started pushing on eager execution. 39:07.680 --> 39:09.880 How do we combine those two together, 39:09.880 --> 39:13.080 which has finally come very well together in 2.0, 39:13.080 --> 39:15.720 but it took us a while to get all the things together 39:15.720 --> 39:16.320 and so on. 39:16.320 --> 39:19.320 So let me, I mean, ask, put another way. 39:19.320 --> 39:21.800 I think eager execution is a really powerful thing, 39:21.800 --> 39:22.680 those added. 39:22.680 --> 39:24.320 Do you think he wouldn't have been, 39:25.840 --> 39:28.400 you know, Muhammad Ali versus Frazier, right? 39:28.400 --> 39:31.200 Do you think it wouldn't have been added as quickly 39:31.200 --> 39:33.760 if PyTorch wasn't there? 39:33.760 --> 39:35.440 It might have taken longer. 39:35.440 --> 39:36.280 No longer. 39:36.280 --> 39:38.960 It was, I mean, we had tried some variants of that before. 39:38.960 --> 39:40.920 So I'm sure it would have happened, 39:40.920 --> 39:42.240 but it might have taken longer. 39:42.240 --> 39:44.800 I'm grateful that TensorFlow is part of the way they did. 39:44.800 --> 39:47.760 That's doing some incredible work last couple of years. 39:47.760 --> 39:49.640 What other things that we didn't talk about? 39:49.640 --> 39:51.520 Are you looking forward in 2.0? 39:51.520 --> 39:54.040 That comes to mind. 39:54.040 --> 39:56.520 So we talked about some of the ecosystem stuff, 39:56.520 --> 40:01.440 making it easily accessible to Keras, eager execution. 40:01.440 --> 40:02.880 Is there other things that we missed? 40:02.880 --> 40:07.480 Yeah, so I would say one is just where 2.0 is, 40:07.480 --> 40:10.760 and, you know, with all the things that we've talked about, 40:10.760 --> 40:13.760 I think as we think beyond that, 40:13.760 --> 40:16.640 there are lots of other things that it enables us to do 40:16.640 --> 40:18.760 and that we're excited about. 40:18.760 --> 40:20.720 So what it's setting us up for, 40:20.720 --> 40:22.520 okay, here are these really clean APIs. 40:22.520 --> 40:25.640 We've cleaned up the surface for what the users want. 40:25.640 --> 40:28.320 What it also allows us to do a whole bunch of stuff 40:28.320 --> 40:31.600 behind the scenes once we are ready with 2.0. 40:31.600 --> 40:36.600 So for example, in TensorFlow with graphs 40:36.760 --> 40:37.720 and all the things you could do, 40:37.720 --> 40:40.600 you could always get a lot of good performance 40:40.600 --> 40:43.280 if you spent the time to tune it, right? 40:43.280 --> 40:47.720 And we've clearly shown that, lots of people do that. 40:47.720 --> 40:52.720 With 2.0, with these APIs where we are, 40:53.040 --> 40:55.120 we can give you a lot of performance 40:55.120 --> 40:57.040 just with whatever you do. 40:57.040 --> 41:01.400 You know, because we see these, it's much cleaner. 41:01.400 --> 41:03.720 We know most people are gonna do things this way. 41:03.720 --> 41:05.520 We can really optimize for that 41:05.520 --> 41:09.040 and get a lot of those things out of the box. 41:09.040 --> 41:10.400 And it really allows us, you know, 41:10.400 --> 41:13.880 both for single machine and distributed and so on, 41:13.880 --> 41:17.200 to really explore other spaces behind the scenes 41:17.200 --> 41:19.680 after 2.0 in the future versions as well. 41:19.680 --> 41:23.000 So right now, the team's really excited about that, 41:23.000 --> 41:25.800 that over time, I think we'll see that. 41:25.800 --> 41:27.720 The other piece that I was talking about 41:27.720 --> 41:31.600 in terms of just restructuring the monolithic thing 41:31.600 --> 41:34.320 into more pieces and making it more modular, 41:34.320 --> 41:36.800 I think that's gonna be really important 41:36.800 --> 41:41.800 for a lot of the other people in the ecosystem, 41:41.800 --> 41:44.760 other organizations and so on that wanted to build things. 41:44.760 --> 41:46.360 Can you elaborate a little bit what you mean 41:46.360 --> 41:50.680 by making TensorFlow more ecosystem or modular? 41:50.680 --> 41:55.000 So the way it's organized today is there's one, 41:55.000 --> 41:56.280 there are lots of repositories 41:56.280 --> 41:58.320 in the TensorFlow organization at GitHub, 41:58.320 --> 42:01.080 the core one where we have TensorFlow, 42:01.080 --> 42:04.080 it has the execution engine, 42:04.080 --> 42:08.280 it has, you know, the key backends for CPUs and GPUs, 42:08.280 --> 42:12.560 it has the work to do distributed stuff. 42:12.560 --> 42:14.360 And all of these just work together 42:14.360 --> 42:17.240 in a single library or binary, 42:17.240 --> 42:18.800 there's no way to split them apart easily. 42:18.800 --> 42:19.960 I mean, there are some interfaces, 42:19.960 --> 42:21.600 but they're not very clean. 42:21.600 --> 42:24.800 In a perfect world, you would have clean interfaces where, 42:24.800 --> 42:27.720 okay, I wanna run it on my fancy cluster 42:27.720 --> 42:29.360 with some custom networking, 42:29.360 --> 42:30.960 just implement this and do that. 42:30.960 --> 42:32.640 I mean, we kind of support that, 42:32.640 --> 42:34.560 but it's hard for people today. 42:35.480 --> 42:38.160 I think as we are starting to see more interesting things 42:38.160 --> 42:39.400 in some of these spaces, 42:39.400 --> 42:42.280 having that clean separation will really start to help. 42:42.280 --> 42:47.280 And again, going to the large size of the ecosystem 42:47.360 --> 42:50.120 and the different groups involved there, 42:50.120 --> 42:53.440 enabling people to evolve and push on things 42:53.440 --> 42:56.040 more independently just allows it to scale better. 42:56.040 --> 42:59.080 And by people, you mean individual developers and? 42:59.080 --> 42:59.920 And organizations. 42:59.920 --> 43:00.920 And organizations. 43:00.920 --> 43:01.760 That's right. 43:01.760 --> 43:04.200 So the hope is that everybody sort of major, 43:04.200 --> 43:06.880 I don't know, Pepsi or something uses, 43:06.880 --> 43:11.040 like major corporations go to TensorFlow to this kind of. 43:11.040 --> 43:13.640 Yeah, if you look at enterprise like Pepsi or these, 43:13.640 --> 43:15.520 I mean, a lot of them are already using TensorFlow. 43:15.520 --> 43:18.960 They are not the ones that do the development 43:18.960 --> 43:20.360 or changes in the core. 43:20.360 --> 43:21.920 Some of them do, but a lot of them don't. 43:21.920 --> 43:23.720 I mean, they touch small pieces. 43:23.720 --> 43:26.400 There are lots of these, some of them being, 43:26.400 --> 43:28.200 let's say hardware vendors who are building 43:28.200 --> 43:30.840 their custom hardware and they want their own pieces. 43:30.840 --> 43:34.160 Or some of them being bigger companies, say IBM. 43:34.160 --> 43:37.320 I mean, they're involved in some of our special interest 43:37.320 --> 43:39.960 groups and they see a lot of users 43:39.960 --> 43:42.640 who want certain things and they want to optimize for that. 43:42.640 --> 43:44.480 So folks like that often. 43:44.480 --> 43:46.400 Autonomous vehicle companies, perhaps. 43:46.400 --> 43:48.200 Exactly, yes. 43:48.200 --> 43:50.520 So yeah, like I mentioned, TensorFlow 43:50.520 --> 43:54.120 has been down on it 41 million times, 50,000 commits, 43:54.120 --> 43:58.360 almost 10,000 pull requests, 1,800 contributors. 43:58.360 --> 44:02.160 So I'm not sure if you can explain it, 44:02.160 --> 44:06.840 but what does it take to build a community like that? 44:06.840 --> 44:09.200 In retrospect, what do you think? 44:09.200 --> 44:12.080 What is the critical thing that allowed for this growth 44:12.080 --> 44:14.600 to happen and how does that growth continue? 44:14.600 --> 44:17.920 Yeah, that's an interesting question. 44:17.920 --> 44:20.240 I wish I had all the answers there, I guess, 44:20.240 --> 44:22.520 so you could replicate it. 44:22.520 --> 44:25.520 I think there are a number of things 44:25.520 --> 44:27.880 that need to come together, right? 44:27.880 --> 44:33.720 One, just like any new thing, there's 44:33.720 --> 44:37.960 a sweet spot of timing, what's needed, 44:37.960 --> 44:39.520 does it grow with what's needed. 44:39.520 --> 44:41.960 So in this case, for example, TensorFlow 44:41.960 --> 44:43.640 is not just grown because it has a good tool, 44:43.640 --> 44:46.640 it's also grown with the growth of deep learning itself. 44:46.640 --> 44:49.000 So those factors come into play. 44:49.000 --> 44:53.120 Other than that, though, I think just 44:53.120 --> 44:55.560 hearing, listening to the community, what they're 44:55.560 --> 44:58.400 doing, what they need, being open to, 44:58.400 --> 45:01.080 like in terms of external contributions, 45:01.080 --> 45:04.520 we've spent a lot of time in making sure 45:04.520 --> 45:06.840 we can accept those contributions well, 45:06.840 --> 45:09.400 we can help the contributors in adding those, 45:09.400 --> 45:11.240 putting the right process in place, 45:11.240 --> 45:13.320 getting the right kind of community, 45:13.320 --> 45:16.120 welcoming them, and so on. 45:16.120 --> 45:19.000 Like over the last year, we've really pushed on transparency. 45:19.000 --> 45:22.200 That's important for an open source project. 45:22.200 --> 45:23.760 People want to know where things are going, 45:23.760 --> 45:26.400 and we're like, OK, here's a process for you. 45:26.400 --> 45:29.320 You can do that, here are our seasons, and so on. 45:29.320 --> 45:32.880 So thinking through, there are lots of community aspects 45:32.880 --> 45:36.400 that come into that you can really work on. 45:36.400 --> 45:38.720 As a small project, it's maybe easy to do, 45:38.720 --> 45:42.240 because there's two developers, and you can do those. 45:42.240 --> 45:46.960 As you grow, putting more of these processes in place, 45:46.960 --> 45:49.080 thinking about the documentation, 45:49.080 --> 45:51.400 thinking about what two developers 45:51.400 --> 45:55.080 care about, what kind of tools would they want to use, 45:55.080 --> 45:56.840 all of these come into play, I think. 45:56.840 --> 45:58.400 So one of the big things, I think, 45:58.400 --> 46:02.560 that feeds the TensorFlow fire is people building something 46:02.560 --> 46:07.680 on TensorFlow, and implement a particular architecture 46:07.680 --> 46:09.480 that does something cool and useful, 46:09.480 --> 46:11.080 and they put that on GitHub. 46:11.080 --> 46:15.640 And so it just feeds this growth. 46:15.640 --> 46:19.560 Do you have a sense that with 2.0 and 1.0, 46:19.560 --> 46:21.880 that there may be a little bit of a partitioning like there 46:21.880 --> 46:26.040 is with Python 2 and 3, that there'll be a code base 46:26.040 --> 46:28.320 in the older versions of TensorFlow 46:28.320 --> 46:31.120 that will not be as compatible easily, 46:31.120 --> 46:35.600 or are you pretty confident that this kind of conversion 46:35.600 --> 46:37.960 is pretty natural and easy to do? 46:37.960 --> 46:41.480 So we're definitely working hard to make that very easy to do. 46:41.480 --> 46:44.040 There's lots of tooling that we talked about at the developer 46:44.040 --> 46:46.480 summit this week, and we'll continue 46:46.480 --> 46:48.280 to invest in that tooling. 46:48.280 --> 46:52.560 It's when you think of these significant version changes, 46:52.560 --> 46:55.720 that's always a risk, and we are really pushing hard 46:55.720 --> 46:59.160 to make that transition very, very smooth. 46:59.160 --> 47:03.000 I think, so at some level, people 47:03.000 --> 47:05.520 want to move when they see the value in the new thing. 47:05.520 --> 47:07.640 They don't want to move just because it's a new thing. 47:07.640 --> 47:11.400 And some people do, but most people want a really good thing. 47:11.400 --> 47:13.760 And I think over the next few months, 47:13.760 --> 47:15.400 as people start to see the value, 47:15.400 --> 47:17.640 we'll definitely see that shift happening. 47:17.640 --> 47:20.080 So I'm pretty excited and confident that we 47:20.080 --> 47:22.440 will see people moving. 47:22.440 --> 47:24.680 As you said earlier, this field is also moving rapidly, 47:24.680 --> 47:26.720 so that'll help because we can do more things. 47:26.720 --> 47:28.520 And all the new things will clearly 47:28.520 --> 47:32.280 happen in 2.x, so people will have lots of good reasons to move. 47:32.280 --> 47:36.160 So what do you think TensorFlow 3.0 looks like? 47:36.160 --> 47:40.320 Is there things happening so crazily 47:40.320 --> 47:42.520 that even at the end of this year, 47:42.520 --> 47:45.320 seems impossible to plan for? 47:45.320 --> 47:49.440 Or is it possible to plan for the next five years? 47:49.440 --> 47:50.800 I think it's tricky. 47:50.800 --> 47:55.760 There are some things that we can expect in terms of, OK, 47:55.760 --> 47:59.720 change, yes, change is going to happen. 47:59.720 --> 48:01.680 Are there some things going to stick around 48:01.680 --> 48:03.720 and some things not going to stick around? 48:03.720 --> 48:08.160 I would say the basics of deep learning, 48:08.160 --> 48:12.680 the convolutional models or the basic kind of things, 48:12.680 --> 48:16.280 they'll probably be around in some form still in five years. 48:16.280 --> 48:21.160 Will Aurel and Gans stay very likely based on where they are? 48:21.160 --> 48:22.840 Will we have new things? 48:22.840 --> 48:24.680 Probably, but those are hard to predict. 48:24.680 --> 48:29.080 And some directionally, some things that we can see 48:29.080 --> 48:32.800 is in things that we're starting to do 48:32.800 --> 48:36.560 with some of our projects right now is just 48:36.560 --> 48:39.120 to point out combining eager execution and graphs, 48:39.120 --> 48:42.240 where we're starting to make it more like just your natural 48:42.240 --> 48:43.160 programming language. 48:43.160 --> 48:45.640 You're not trying to program something else. 48:45.640 --> 48:47.240 Similarly, with Swift for TensorFlow, 48:47.240 --> 48:48.280 we're taking that approach. 48:48.280 --> 48:50.040 Can you do something round up? 48:50.040 --> 48:52.080 So some of those ideas seem like, OK, 48:52.080 --> 48:55.000 that's the right direction in five years 48:55.000 --> 48:58.360 we expect to see more in that area. 48:58.360 --> 49:01.760 Other things we don't know is, will hardware accelerators 49:01.760 --> 49:03.200 be the same? 49:03.200 --> 49:09.000 Will we be able to train with four bits instead of 32 bits? 49:09.000 --> 49:11.440 And I think the TPU side of things is exploring. 49:11.440 --> 49:13.960 I mean, TPU is already on version three. 49:13.960 --> 49:17.520 It seems that the evolution of TPU and TensorFlow 49:17.520 --> 49:24.080 are coevolving in terms of both their learning 49:24.080 --> 49:25.720 from each other and from the community 49:25.720 --> 49:29.720 and from the applications where the biggest benefit is achieved. 49:29.720 --> 49:30.560 That's right. 49:30.560 --> 49:33.320 You've been trying with eager with Keras 49:33.320 --> 49:36.480 to make TensorFlow as accessible and easy to use as possible. 49:36.480 --> 49:39.040 What do you think for beginners is the biggest thing 49:39.040 --> 49:40.000 they struggle with? 49:40.000 --> 49:42.080 Have you encountered that? 49:42.080 --> 49:44.280 Or is basically what Keras is solving 49:44.280 --> 49:48.680 is that eager, like we talked about TensorFlow? 49:48.680 --> 49:51.480 For some of them, like you said, the beginners 49:51.480 --> 49:54.840 want to just be able to take some image model. 49:54.840 --> 49:58.040 They don't care if it's inception or rest net or something else 49:58.040 --> 50:00.760 and do some training or transfer learning 50:00.760 --> 50:02.440 on their kind of model. 50:02.440 --> 50:04.400 Being able to make that easy is important. 50:04.400 --> 50:08.560 So in some ways, if you do that by providing them 50:08.560 --> 50:11.360 simple models with, say, in Hub or so on, 50:11.360 --> 50:13.680 they don't care about what's inside that box, 50:13.680 --> 50:15.120 but they want to be able to use it. 50:15.120 --> 50:17.600 So we're pushing on, I think, different levels. 50:17.600 --> 50:20.120 If you look at just a component that you get, which 50:20.120 --> 50:22.800 has the layers already smushed in, 50:22.800 --> 50:25.200 the beginners probably just want that. 50:25.200 --> 50:27.360 Then the next step is, OK, look at building 50:27.360 --> 50:29.000 layers with Keras. 50:29.000 --> 50:30.600 If you go out to research, then they 50:30.600 --> 50:33.120 are probably writing custom layers themselves 50:33.120 --> 50:34.360 or doing their own loops. 50:34.360 --> 50:36.320 So there's a whole spectrum there. 50:36.320 --> 50:38.600 And then providing the preentrain models 50:38.600 --> 50:44.760 seems to really decrease the time from you trying to start. 50:44.760 --> 50:46.800 So you could basically, in a Colab notebook, 50:46.800 --> 50:49.080 achieve what you need. 50:49.080 --> 50:51.280 So I'm basically answering my own question, 50:51.280 --> 50:54.240 because I think what TensorFlow delivered on recently 50:54.240 --> 50:57.000 is trivial for beginners. 50:57.000 --> 51:00.760 So I was just wondering if there was other pain points 51:00.760 --> 51:02.480 you're trying to ease, but I'm not sure there would. 51:02.480 --> 51:04.240 No, those are probably the big ones. 51:04.240 --> 51:07.080 I mean, I see high schoolers doing a whole bunch of things 51:07.080 --> 51:08.840 now, which is pretty amazing. 51:08.840 --> 51:11.360 It's both amazing and terrifying. 51:11.360 --> 51:12.640 Yes. 51:12.640 --> 51:16.920 In a sense that when they grow up, 51:16.920 --> 51:19.280 some incredible ideas will be coming from them. 51:19.280 --> 51:21.800 So there's certainly a technical aspect to your work, 51:21.800 --> 51:24.600 but you also have a management aspect 51:24.600 --> 51:28.000 to your role with TensorFlow, leading the project, 51:28.000 --> 51:31.080 a large number of developers and people. 51:31.080 --> 51:34.680 So what do you look for in a good team? 51:34.680 --> 51:37.400 What do you think Google has been at the forefront 51:37.400 --> 51:40.440 of exploring what it takes to build a good team? 51:40.440 --> 51:45.520 And TensorFlow is one of the most cutting edge technologies 51:45.520 --> 51:46.120 in the world. 51:46.120 --> 51:48.080 So in this context, what do you think 51:48.080 --> 51:50.480 makes for a good team? 51:50.480 --> 51:53.200 It's definitely something I think a fair bit about. 51:53.200 --> 51:59.560 I think in terms of the team being 51:59.560 --> 52:02.120 able to deliver something well, one of the things that's 52:02.120 --> 52:05.800 important is a cohesion across the team. 52:05.800 --> 52:10.400 So being able to execute together and doing things, 52:10.400 --> 52:11.440 it's not an end. 52:11.440 --> 52:14.120 Like at this scale, an individual engineer 52:14.120 --> 52:15.400 can only do so much. 52:15.400 --> 52:18.200 There's a lot more that they can do together, 52:18.200 --> 52:21.640 even though we have some amazing superstars across Google 52:21.640 --> 52:22.600 and in the team. 52:22.600 --> 52:26.200 But there's often the way I see it 52:26.200 --> 52:28.360 is the product of what the team generates 52:28.360 --> 52:34.440 is way larger than the whole individual put together. 52:34.440 --> 52:37.320 And so how do we have all of them work together, 52:37.320 --> 52:40.000 the culture of the team itself? 52:40.000 --> 52:43.000 Hiring good people is important. 52:43.000 --> 52:45.600 But part of that is it's not just that, OK, 52:45.600 --> 52:48.120 we hire a bunch of smart people and throw them together 52:48.120 --> 52:49.720 and let them do things. 52:49.720 --> 52:52.920 It's also people have to care about what they're building. 52:52.920 --> 52:57.320 People have to be motivated for the right kind of things. 52:57.320 --> 53:01.400 That's often an important factor. 53:01.400 --> 53:04.600 And finally, how do you put that together 53:04.600 --> 53:08.840 with a somewhat unified vision of where we want to go? 53:08.840 --> 53:11.200 So are we all looking in the same direction 53:11.200 --> 53:13.520 or just going all over? 53:13.520 --> 53:16.040 And sometimes it's a mix. 53:16.040 --> 53:21.400 Google's a very bottom up organization in some sense. 53:21.400 --> 53:24.680 Also research even more so. 53:24.680 --> 53:26.320 And that's how we started. 53:26.320 --> 53:30.840 But as we've become this larger product and ecosystem, 53:30.840 --> 53:35.040 I think it's also important to combine that well with a mix 53:35.040 --> 53:37.920 of, OK, here's the direction we want to go in. 53:37.920 --> 53:39.880 There is exploration we'll do around that. 53:39.880 --> 53:43.320 But let's keep staying in that direction, not just 53:43.320 --> 53:44.360 all over the place. 53:44.360 --> 53:46.880 And is there a way you monitor the health of the team? 53:46.880 --> 53:51.920 Sort of like, is there a way you know you did a good job? 53:51.920 --> 53:53.000 The team is good. 53:53.000 --> 53:56.960 I mean, you're saying nice things, but it's sometimes 53:56.960 --> 54:01.120 difficult to determine how aligned. 54:01.120 --> 54:04.480 Because it's not binary, it's not like there's tensions 54:04.480 --> 54:06.680 and complexities and so on. 54:06.680 --> 54:09.400 And the other element of this is the mesh of superstars. 54:09.400 --> 54:12.880 There's so much, even at Google, such a large percentage 54:12.880 --> 54:16.000 of work is done by individual superstars too. 54:16.000 --> 54:19.920 So there's a, and sometimes those superstars 54:19.920 --> 54:25.120 could be against the dynamic of a team and those tensions. 54:25.120 --> 54:27.320 I mean, I'm sure TensorFlow might be a little bit easier 54:27.320 --> 54:31.720 because the mission of the project is so beautiful. 54:31.720 --> 54:34.760 You're at the cutting edge, so it's exciting. 54:34.760 --> 54:36.640 But have you had struggle with that? 54:36.640 --> 54:38.360 Has there been challenges? 54:38.360 --> 54:39.800 There are always people challenges 54:39.800 --> 54:41.240 in different kinds of ways. 54:41.240 --> 54:44.520 That said, I think we've been what's 54:44.520 --> 54:49.320 good about getting people who care and have 54:49.320 --> 54:51.440 the same kind of culture, and that's Google in general 54:51.440 --> 54:53.480 to a large extent. 54:53.480 --> 54:56.760 But also, like you said, given that the project has had 54:56.760 --> 54:59.160 so many exciting things to do, there's 54:59.160 --> 55:02.080 been room for lots of people to do different kinds of things 55:02.080 --> 55:06.440 and grow, which does make the problem a bit easier, I guess. 55:06.440 --> 55:09.920 And it allows people, depending on what they're doing, 55:09.920 --> 55:13.120 if there's room around them, then that's fine. 55:13.120 --> 55:19.160 But yes, we do care about whether a superstar or not 55:19.160 --> 55:22.560 that they need to work well with the team across Google. 55:22.560 --> 55:23.760 That's interesting to hear. 55:23.760 --> 55:27.960 So it's like superstar or not, the productivity broadly 55:27.960 --> 55:30.520 is about the team. 55:30.520 --> 55:31.520 Yeah. 55:31.520 --> 55:32.960 I mean, they might add a lot of value, 55:32.960 --> 55:35.720 but if they're hurting the team, then that's a problem. 55:35.720 --> 55:38.720 So in hiring engineers, it's so interesting, right? 55:38.720 --> 55:41.840 The high rank process, what do you look for? 55:41.840 --> 55:44.240 How do you determine a good developer 55:44.240 --> 55:47.280 or a good member of a team from just a few minutes 55:47.280 --> 55:50.320 or hours together? 55:50.320 --> 55:51.920 Again, no magic answers, I'm sure. 55:51.920 --> 55:52.760 Yeah. 55:52.760 --> 55:56.240 And Google has a hiring process that we've refined 55:56.240 --> 56:00.880 over the last 20 years, I guess, and that you've probably 56:00.880 --> 56:02.200 heard and seen a lot about. 56:02.200 --> 56:05.280 So we do work with the same hiring process in that. 56:05.280 --> 56:08.280 That's really helped. 56:08.280 --> 56:10.880 For me in particular, I would say, 56:10.880 --> 56:14.200 in addition to the core technical skills, 56:14.200 --> 56:17.560 what does matter is their motivation 56:17.560 --> 56:19.560 in what they want to do. 56:19.560 --> 56:22.960 Because if that doesn't align well with where we want to go, 56:22.960 --> 56:25.320 that's not going to lead to long term success 56:25.320 --> 56:27.640 for either them or the team. 56:27.640 --> 56:30.640 And I think that becomes more important the more senior 56:30.640 --> 56:33.520 the person is, but it's important at every level. 56:33.520 --> 56:34.920 Like even the junior most engineer, 56:34.920 --> 56:37.680 if they're not motivated to do well at what they're trying to do, 56:37.680 --> 56:39.080 however smart they are, it's going 56:39.080 --> 56:40.320 to be hard for them to succeed. 56:40.320 --> 56:44.520 Does the Google hiring process touch on that passion? 56:44.520 --> 56:46.440 So like trying to determine. 56:46.440 --> 56:48.440 Because I think as far as I understand, 56:48.440 --> 56:52.000 maybe you can speak to it that the Google hiring process sort 56:52.000 --> 56:56.360 of helps the initial like determines the skill set there, 56:56.360 --> 56:59.840 is your puzzle solving ability, problem solving ability good. 56:59.840 --> 57:05.000 But I'm not sure, but it seems that the determining 57:05.000 --> 57:07.560 whether the person is like fire inside them 57:07.560 --> 57:09.840 that burns to do anything really doesn't really matter. 57:09.840 --> 57:11.520 It's just some cool stuff. 57:11.520 --> 57:15.320 I'm going to do it that I don't know. 57:15.320 --> 57:17.000 Is that something that ultimately ends up 57:17.000 --> 57:18.840 when they have a conversation with you 57:18.840 --> 57:22.600 or once it gets closer to the team? 57:22.600 --> 57:25.400 So one of the things we do have as part of the process 57:25.400 --> 57:28.600 is just a culture fit, like part of the interview process 57:28.600 --> 57:31.040 itself, in addition to just the technical skills. 57:31.040 --> 57:34.240 And each engineer or whoever the interviewer is, 57:34.240 --> 57:38.800 is supposed to rate the person on the culture and the culture 57:38.800 --> 57:39.960 fit with Google and so on. 57:39.960 --> 57:42.160 So that is definitely part of the process. 57:42.160 --> 57:45.800 Now, there are various kinds of projects 57:45.800 --> 57:46.960 and different kinds of things. 57:46.960 --> 57:50.040 So there might be variants in the kind of culture 57:50.040 --> 57:51.320 you want there and so on. 57:51.320 --> 57:52.720 And yes, that does vary. 57:52.720 --> 57:54.920 So for example, TensorFlow has always 57:54.920 --> 57:56.920 been a fast moving project. 57:56.920 --> 58:00.920 And we want people who are comfortable with that. 58:00.920 --> 58:02.640 But at the same time now, for example, 58:02.640 --> 58:05.200 we are at a place where we are also very full fledged product. 58:05.200 --> 58:08.440 And we want to make sure things that work really, really 58:08.440 --> 58:09.320 work right. 58:09.320 --> 58:11.680 You can't cut corners all the time. 58:11.680 --> 58:14.320 So balancing that out and finding the people 58:14.320 --> 58:17.560 who are the right fit for those is important. 58:17.560 --> 58:19.720 And I think those kind of things do vary a bit 58:19.720 --> 58:23.200 across projects and teams and product areas across Google. 58:23.200 --> 58:25.240 And so you'll see some differences there 58:25.240 --> 58:27.640 in the final checklist. 58:27.640 --> 58:29.600 But a lot of the core culture, it 58:29.600 --> 58:32.200 comes along with just the engineering, excellence, 58:32.200 --> 58:34.720 and so on. 58:34.720 --> 58:39.680 What is the hardest part of your job? 58:39.680 --> 58:41.920 I'll take your pick, I guess. 58:41.920 --> 58:44.440 It's fun, I would say. 58:44.440 --> 58:45.520 Hard, yes. 58:45.520 --> 58:47.240 I mean, lots of things at different times. 58:47.240 --> 58:49.160 I think that does vary. 58:49.160 --> 58:52.640 So let me clarify that difficult things are fun 58:52.640 --> 58:55.720 when you solve them, right? 58:55.720 --> 58:57.480 It's fun in that sense. 58:57.480 --> 59:02.600 I think the key to a successful thing across the board, 59:02.600 --> 59:05.320 and in this case, it's a large ecosystem now, 59:05.320 --> 59:09.800 but even a small product, is striking that fine balance 59:09.800 --> 59:12.000 across different aspects of it. 59:12.000 --> 59:17.000 Sometimes it's how fast you go versus how perfect it is. 59:17.000 --> 59:21.400 Sometimes it's how do you involve this huge community? 59:21.400 --> 59:22.360 Who do you involve? 59:22.360 --> 59:25.440 Or do you decide, OK, now is not a good time to involve them 59:25.440 --> 59:30.160 because it's not the right fit? 59:30.160 --> 59:33.640 Sometimes it's saying no to certain kinds of things. 59:33.640 --> 59:36.880 Those are often the hard decisions. 59:36.880 --> 59:41.000 Some of them you make quickly because you don't have the time. 59:41.000 --> 59:43.200 Some of them you get time to think about them, 59:43.200 --> 59:44.480 but they're always hard. 59:44.480 --> 59:49.200 So both choices are pretty good, those decisions. 59:49.200 --> 59:50.360 What about deadlines? 59:50.360 --> 59:58.200 Is this defined TensorFlow to be driven by deadlines 59:58.200 --> 1:00:00.360 to a degree that a product might? 1:00:00.360 --> 1:00:04.920 Or is there still a balance to where it's less deadline? 1:00:04.920 --> 1:00:08.920 You had the Dev Summit, they came together incredibly. 1:00:08.920 --> 1:00:11.440 Looked like there's a lot of moving pieces and so on. 1:00:11.440 --> 1:00:15.080 So did that deadline make people rise to the occasion, 1:00:15.080 --> 1:00:18.360 releasing TensorFlow 2.0 Alpha? 1:00:18.360 --> 1:00:20.360 I'm sure that was done last minute as well. 1:00:20.360 --> 1:00:25.600 I mean, up to the last point. 1:00:25.600 --> 1:00:28.600 Again, it's one of those things that you 1:00:28.600 --> 1:00:29.960 need to strike the good balance. 1:00:29.960 --> 1:00:32.040 There's some value that deadlines bring 1:00:32.040 --> 1:00:33.920 that does bring a sense of urgency 1:00:33.920 --> 1:00:35.720 to get the right things together. 1:00:35.720 --> 1:00:38.280 Instead of getting the perfect thing out, 1:00:38.280 --> 1:00:41.280 you need something that's good and works well. 1:00:41.280 --> 1:00:43.720 And the team definitely did a great job in putting that 1:00:43.720 --> 1:00:46.560 together, so it was very amazed and excited by everything, 1:00:46.560 --> 1:00:48.680 how that came together. 1:00:48.680 --> 1:00:50.640 That said, across the year, we try not 1:00:50.640 --> 1:00:52.520 to put out official deadlines. 1:00:52.520 --> 1:00:56.960 We focus on key things that are important, 1:00:56.960 --> 1:01:00.600 figure out how much of it's important, 1:01:00.600 --> 1:01:05.760 and we are developing in the open, internally and externally, 1:01:05.760 --> 1:01:07.920 everything's available to everybody. 1:01:07.920 --> 1:01:11.120 So you can pick and look at where things are. 1:01:11.120 --> 1:01:13.160 We do releases at a regular cadence, 1:01:13.160 --> 1:01:16.320 so fine if something doesn't necessarily end up with this 1:01:16.320 --> 1:01:19.600 month, it'll end up in the next release in a month or two. 1:01:19.600 --> 1:01:22.840 And that's OK, but we want to keep moving 1:01:22.840 --> 1:01:26.520 as fast as we can in these different areas. 1:01:26.520 --> 1:01:30.080 Because we can iterate and improve on things, sometimes 1:01:30.080 --> 1:01:32.920 it's OK to put things out that aren't fully ready. 1:01:32.920 --> 1:01:35.640 If you make sure it's clear that, OK, this is experimental, 1:01:35.640 --> 1:01:37.960 but it's out there if you want to try and give feedback. 1:01:37.960 --> 1:01:39.400 That's very, very useful. 1:01:39.400 --> 1:01:43.560 I think that quick cycle and quick iteration is important. 1:01:43.560 --> 1:01:47.200 That's what we often focus on rather than here's 1:01:47.200 --> 1:01:49.200 a deadline where you get everything else. 1:01:49.200 --> 1:01:52.880 It's 2.0, is there pressure to make that stable? 1:01:52.880 --> 1:01:57.760 Or like, for example, WordPress 5.0 just came out, 1:01:57.760 --> 1:02:01.760 and there was no pressure to, it was a lot of build updates 1:02:01.760 --> 1:02:04.960 that delivered way too late. 1:02:04.960 --> 1:02:06.440 And they said, OK, well, we're going 1:02:06.440 --> 1:02:09.680 to release a lot of updates really quickly to improve it. 1:02:09.680 --> 1:02:12.240 Do you see TensorFlow 2.0 in that same kind of way, 1:02:12.240 --> 1:02:15.240 or is there this pressure to once it hits 2.0, 1:02:15.240 --> 1:02:16.760 once you get to the release candidate, 1:02:16.760 --> 1:02:19.440 and then you get to the final, that's 1:02:19.440 --> 1:02:22.480 going to be the stable thing? 1:02:22.480 --> 1:02:26.680 So it's going to be stable in just like 1.0X 1:02:26.680 --> 1:02:32.080 was where every API that's there is going to remain in work. 1:02:32.080 --> 1:02:34.800 It doesn't mean we can't change things under the covers. 1:02:34.800 --> 1:02:36.720 It doesn't mean we can't add things. 1:02:36.720 --> 1:02:39.200 So there's still a lot more for us to do, 1:02:39.200 --> 1:02:41.080 and we continue to have more releases. 1:02:41.080 --> 1:02:42.920 So in that sense, there's still, I 1:02:42.920 --> 1:02:44.680 don't think we'd be done in like two months 1:02:44.680 --> 1:02:46.160 when we release this. 1:02:46.160 --> 1:02:49.880 I don't know if you can say, but is there, you know, 1:02:49.880 --> 1:02:53.680 there's not external deadlines for TensorFlow 2.0, 1:02:53.680 --> 1:02:58.520 but is there internal deadlines, artificial or otherwise, 1:02:58.520 --> 1:03:00.840 that you're trying to set for yourself, 1:03:00.840 --> 1:03:03.080 or is it whenever it's ready? 1:03:03.080 --> 1:03:05.680 So we want it to be a great product, right? 1:03:05.680 --> 1:03:09.880 And that's a big, important piece for us. 1:03:09.880 --> 1:03:11.160 TensorFlow is already out there. 1:03:11.160 --> 1:03:13.720 We have 41 million downloads for 1.x, 1:03:13.720 --> 1:03:15.880 so it's not like we have to have this. 1:03:15.880 --> 1:03:17.280 Yeah, exactly. 1:03:17.280 --> 1:03:19.320 So it's not like a lot of the features 1:03:19.320 --> 1:03:22.080 that we've really polishing and putting them together 1:03:22.080 --> 1:03:26.240 are there, we don't have to rush that just because. 1:03:26.240 --> 1:03:28.040 So in that sense, we want to get it right 1:03:28.040 --> 1:03:29.920 and really focus on that. 1:03:29.920 --> 1:03:31.520 That said, we have said that we are 1:03:31.520 --> 1:03:33.520 looking to get this out in the next few months, 1:03:33.520 --> 1:03:37.120 in the next quarter, and as far as possible, 1:03:37.120 --> 1:03:40.000 we'll definitely try to make that happen. 1:03:40.000 --> 1:03:44.360 Yeah, my favorite line was, spring is a relative concept. 1:03:44.360 --> 1:03:45.960 I love it. 1:03:45.960 --> 1:03:47.680 Spoken like a true developer. 1:03:47.680 --> 1:03:50.200 So something I'm really interested in, 1:03:50.200 --> 1:03:53.840 and your previous line of work is, before TensorFlow, 1:03:53.840 --> 1:03:57.720 you let a team and Google on search ads. 1:03:57.720 --> 1:04:02.840 I think this is a very interesting topic on every level, 1:04:02.840 --> 1:04:07.200 on a technical level, because if their best ads connect people 1:04:07.200 --> 1:04:10.080 to the things they want and need, 1:04:10.080 --> 1:04:12.280 and that they're worse, they're just these things 1:04:12.280 --> 1:04:15.840 that annoy the heck out of you to the point of ruining 1:04:15.840 --> 1:04:20.240 the entire user experience of whatever you're actually doing. 1:04:20.240 --> 1:04:23.600 So they have a bad rep, I guess. 1:04:23.600 --> 1:04:28.080 And on the other end, so that this connecting users 1:04:28.080 --> 1:04:32.120 to the thing they need to want is a beautiful opportunity 1:04:32.120 --> 1:04:35.360 for machine learning to shine, like huge amounts of data 1:04:35.360 --> 1:04:36.720 that's personalized, and you've got 1:04:36.720 --> 1:04:40.400 to map to the thing they actually won't get annoyed. 1:04:40.400 --> 1:04:43.760 So what have you learned from this Google that's 1:04:43.760 --> 1:04:45.160 leading the world in this aspect? 1:04:45.160 --> 1:04:47.560 What have you learned from that experience? 1:04:47.560 --> 1:04:51.520 And what do you think is the future of ads? 1:04:51.520 --> 1:04:54.040 Take you back to the end of that. 1:04:54.040 --> 1:04:59.720 Yes, it's been a while, but I totally agree with what you said. 1:04:59.720 --> 1:05:03.200 I think the search ads, the way it was always looked at, 1:05:03.200 --> 1:05:05.520 and I believe it still is, is it's 1:05:05.520 --> 1:05:08.240 an extension of what search is trying to do. 1:05:08.240 --> 1:05:10.560 The goal is to make the information 1:05:10.560 --> 1:05:14.680 and make the world's information accessible. 1:05:14.680 --> 1:05:17.120 With ads, it's not just information, 1:05:17.120 --> 1:05:19.120 but it may be products or other things 1:05:19.120 --> 1:05:20.800 that people care about. 1:05:20.800 --> 1:05:23.360 And so it's really important for them 1:05:23.360 --> 1:05:26.480 to align with what the users need. 1:05:26.480 --> 1:05:30.920 And in search ads, there's a minimum quality level 1:05:30.920 --> 1:05:32.320 before that ad would be shown. 1:05:32.320 --> 1:05:34.200 If we don't have an ad that hits that quality bar, 1:05:34.200 --> 1:05:35.960 it will not be shown, even if we have it. 1:05:35.960 --> 1:05:38.080 And OK, maybe we lose some money there. 1:05:38.080 --> 1:05:39.560 That's fine. 1:05:39.560 --> 1:05:41.200 That is really, really important, 1:05:41.200 --> 1:05:43.000 and I think that that is something I really 1:05:43.000 --> 1:05:45.040 liked about being there. 1:05:45.040 --> 1:05:48.120 Advertising is a key part. 1:05:48.120 --> 1:05:51.680 I mean, as a model, it's been around for ages, right? 1:05:51.680 --> 1:05:52.920 It's not a new model. 1:05:52.920 --> 1:05:57.440 It's been adapted to the web and became a core part of search 1:05:57.440 --> 1:06:02.120 and in many other search engines across the world. 1:06:02.120 --> 1:06:05.920 I do hope, like I said, there are aspects of ads 1:06:05.920 --> 1:06:06.680 that are annoying. 1:06:06.680 --> 1:06:09.600 And I go to a website, and if it just 1:06:09.600 --> 1:06:12.160 keeps popping an ad in my face, not to let me read, 1:06:12.160 --> 1:06:13.800 that's going to be annoying clearly. 1:06:13.800 --> 1:06:22.080 So I hope we can strike that balance between showing a good 1:06:22.080 --> 1:06:25.040 ad where it's valuable to the user 1:06:25.040 --> 1:06:30.960 and provides the monetization to the service. 1:06:30.960 --> 1:06:32.000 And this might be search. 1:06:32.000 --> 1:06:33.680 This might be a website. 1:06:33.680 --> 1:06:37.320 All of these, they do need the monetization for them 1:06:37.320 --> 1:06:39.640 to provide that service. 1:06:39.640 --> 1:06:45.720 But if it's done in a good balance between showing 1:06:45.720 --> 1:06:48.040 just some random stuff that's distracting 1:06:48.040 --> 1:06:50.920 versus showing something that's actually valuable. 1:06:50.920 --> 1:06:55.360 So do you see it moving forward as to continue 1:06:55.360 --> 1:07:00.960 being a model that funds businesses like Google? 1:07:00.960 --> 1:07:05.160 That's a significant revenue stream. 1:07:05.160 --> 1:07:08.080 Because that's one of the most exciting things, 1:07:08.080 --> 1:07:09.680 but also limiting things on the internet 1:07:09.680 --> 1:07:12.200 is nobody wants to pay for anything. 1:07:12.200 --> 1:07:15.360 And advertisements, again, coupled at their best 1:07:15.360 --> 1:07:17.360 are actually really useful and not annoying. 1:07:17.360 --> 1:07:22.320 Do you see that continuing and growing and improving? 1:07:22.320 --> 1:07:26.680 Or is there GC sort of more Netflix type models 1:07:26.680 --> 1:07:28.960 where you have to start to pay for content? 1:07:28.960 --> 1:07:31.000 I think it's a mix. 1:07:31.000 --> 1:07:32.840 I think it's going to take a long while for everything 1:07:32.840 --> 1:07:35.320 to be paid on the internet, if at all. 1:07:35.320 --> 1:07:36.160 Probably not. 1:07:36.160 --> 1:07:37.400 I mean, I think there's always going 1:07:37.400 --> 1:07:40.760 to be things that are sort of monetized with things like ads. 1:07:40.760 --> 1:07:42.800 But over the last few years, I would say 1:07:42.800 --> 1:07:44.760 we've definitely seen that transition 1:07:44.760 --> 1:07:48.560 towards more paid services across the web 1:07:48.560 --> 1:07:50.360 and people are willing to pay for them 1:07:50.360 --> 1:07:51.760 because they do see the value. 1:07:51.760 --> 1:07:53.600 I mean, Netflix is a great example. 1:07:53.600 --> 1:07:56.520 I mean, we have YouTube doing things. 1:07:56.520 --> 1:07:59.720 People pay for the apps they buy, more people 1:07:59.720 --> 1:08:03.120 they find are willing to pay for newspaper content, 1:08:03.120 --> 1:08:07.240 for the good news websites across the web. 1:08:07.240 --> 1:08:11.040 That wasn't the case even a few years ago, I would say. 1:08:11.040 --> 1:08:13.280 And I just see that change in myself as well 1:08:13.280 --> 1:08:14.840 and just lots of people around me. 1:08:14.840 --> 1:08:19.240 So definitely hopeful that we'll transition to that mix model 1:08:19.240 --> 1:08:23.400 where maybe you get to try something out for free, 1:08:23.400 --> 1:08:24.120 maybe with ads. 1:08:24.120 --> 1:08:27.080 But then there is a more clear revenue model 1:08:27.080 --> 1:08:30.600 that sort of helps go beyond that. 1:08:30.600 --> 1:08:34.760 So speaking of revenue, how is it 1:08:34.760 --> 1:08:39.400 that a person can use the TPU in a Google Colab for free? 1:08:39.400 --> 1:08:43.920 So what's the, I guess, the question is, 1:08:43.920 --> 1:08:48.880 what's the future of TensorFlow in terms of empowering, 1:08:48.880 --> 1:08:51.880 say, a class of 300 students? 1:08:51.880 --> 1:08:55.920 And I'm asked by MIT, what is going 1:08:55.920 --> 1:08:58.640 to be the future of them being able to do their homework 1:08:58.640 --> 1:09:00.200 in TensorFlow? 1:09:00.200 --> 1:09:02.800 Where are they going to train these networks, right? 1:09:02.800 --> 1:09:07.720 What's that future look like with TPUs, with cloud services, 1:09:07.720 --> 1:09:08.920 and so on? 1:09:08.920 --> 1:09:10.240 I think a number of things there. 1:09:10.240 --> 1:09:12.600 I mean, any TensorFlow open source, 1:09:12.600 --> 1:09:13.640 you can run it wherever. 1:09:13.640 --> 1:09:15.880 You can run it on your desktop, and your desktops 1:09:15.880 --> 1:09:19.480 always keep getting more powerful, so maybe you can do more. 1:09:19.480 --> 1:09:22.040 My phone is like, I don't know how many times more powerful 1:09:22.040 --> 1:09:23.520 than my first desktop. 1:09:23.520 --> 1:09:25.200 You'll probably train it on your phone, though. 1:09:25.200 --> 1:09:26.200 Yeah, that's true. 1:09:26.200 --> 1:09:28.080 Right, so in that sense, the power 1:09:28.080 --> 1:09:31.440 you have in your hand is a lot more. 1:09:31.440 --> 1:09:34.400 Clouds are actually very interesting from, say, 1:09:34.400 --> 1:09:37.840 students or courses perspective, because they 1:09:37.840 --> 1:09:40.040 make it very easy to get started. 1:09:40.040 --> 1:09:42.040 I mean, Colab, the great thing about it 1:09:42.040 --> 1:09:45.120 is go to a website, and it just works. 1:09:45.120 --> 1:09:47.560 No installation needed, nothing to, you know, 1:09:47.560 --> 1:09:49.960 you're just there, and things are working. 1:09:49.960 --> 1:09:52.280 That's really the power of cloud, as well. 1:09:52.280 --> 1:09:55.320 And so I do expect that to grow. 1:09:55.320 --> 1:09:57.920 Again, Colab is a free service. 1:09:57.920 --> 1:10:00.840 It's great to get started, to play with things, 1:10:00.840 --> 1:10:03.080 to explore things. 1:10:03.080 --> 1:10:08.200 That said, with free, you can only get so much, maybe. 1:10:08.200 --> 1:10:11.080 So just like we were talking about free versus paid, 1:10:11.080 --> 1:10:15.280 and there are services you can pay for and get a lot more. 1:10:15.280 --> 1:10:16.000 Great. 1:10:16.000 --> 1:10:18.480 So if I'm a complete beginner interested in machine 1:10:18.480 --> 1:10:21.560 learning and TensorFlow, what should I do? 1:10:21.560 --> 1:10:24.240 Probably start with going to a website and playing there. 1:10:24.240 --> 1:10:26.560 Just go to TensorFlow.org and start clicking on things. 1:10:26.560 --> 1:10:28.440 Yep, check out tutorials and guides. 1:10:28.440 --> 1:10:30.680 There's stuff you can just click there and go to Colab 1:10:30.680 --> 1:10:31.320 and do things. 1:10:31.320 --> 1:10:32.360 No installation needed. 1:10:32.360 --> 1:10:34.040 You can get started right there. 1:10:34.040 --> 1:10:34.840 OK, awesome. 1:10:34.840 --> 1:10:36.720 Roger, thank you so much for talking today. 1:10:36.720 --> 1:10:37.440 Thank you, Lex. 1:10:37.440 --> 1:10:46.680 Have fun this week.