WEBVTT 00:00.000 --> 00:03.120 The following is a conversation with Jeremy Howard. 00:03.120 --> 00:07.080 He's the founder of Fast AI, a research institute dedicated 00:07.080 --> 00:09.760 to making deep learning more accessible. 00:09.760 --> 00:12.560 He's also a distinguished research scientist 00:12.560 --> 00:14.600 at the University of San Francisco, 00:14.600 --> 00:17.600 a former president of Kegel, as well as a top breaking 00:17.600 --> 00:18.800 competitor there. 00:18.800 --> 00:21.680 And in general, he's a successful entrepreneur, 00:21.680 --> 00:25.240 educator, researcher, and an inspiring personality 00:25.240 --> 00:27.000 in the AI community. 00:27.000 --> 00:28.680 When someone asked me, how do I get 00:28.680 --> 00:30.240 started with deep learning? 00:30.240 --> 00:33.360 Fast AI is one of the top places I point them to. 00:33.360 --> 00:34.120 It's free. 00:34.120 --> 00:35.520 It's easy to get started. 00:35.520 --> 00:37.600 It's insightful and accessible. 00:37.600 --> 00:40.960 And if I may say so, it has very little BS. 00:40.960 --> 00:44.160 It can sometimes dilute the value of educational content 00:44.160 --> 00:46.720 on popular topics like deep learning. 00:46.720 --> 00:49.440 Fast AI has a focus on practical application 00:49.440 --> 00:51.600 of deep learning and hands on exploration 00:51.600 --> 00:53.880 of the cutting edge that is incredibly 00:53.880 --> 00:57.960 both accessible to beginners and useful to experts. 00:57.960 --> 01:01.360 This is the Artificial Intelligence Podcast. 01:01.360 --> 01:03.760 If you enjoy it, subscribe on YouTube, 01:03.760 --> 01:06.920 give it five stars on iTunes, support it on Patreon, 01:06.920 --> 01:09.040 or simply connect with me on Twitter. 01:09.040 --> 01:13.280 Alex Friedman, spelled F R I D M A N. 01:13.280 --> 01:18.560 And now, here's my conversation with Jeremy Howard. 01:18.560 --> 01:21.680 What's the first program you ever written? 01:21.680 --> 01:24.800 First program I wrote that I remember 01:24.800 --> 01:29.200 would be at high school. 01:29.200 --> 01:31.240 I did an assignment where I decided 01:31.240 --> 01:36.240 to try to find out if there were some better musical scales 01:36.240 --> 01:40.640 than the normal 12 tone, 12 interval scale. 01:40.640 --> 01:43.680 So I wrote a program on my Commodore 64 in BASIC 01:43.680 --> 01:46.080 that searched through other scale sizes 01:46.080 --> 01:48.440 to see if it could find one where there 01:48.440 --> 01:51.880 were more accurate harmonies. 01:51.880 --> 01:53.040 Like mid tone? 01:53.040 --> 01:56.520 Like you want an actual exactly 3 to 2 ratio, 01:56.520 --> 01:59.400 where else with a 12 interval scale, 01:59.400 --> 02:01.480 it's not exactly 3 to 2, for example. 02:01.480 --> 02:05.080 So that's well tempered, as they say. 02:05.080 --> 02:07.680 And BASIC on a Commodore 64. 02:07.680 --> 02:09.440 Where was the interest in music from? 02:09.440 --> 02:10.480 Or is it just technical? 02:10.480 --> 02:14.640 I did music all my life, so I played saxophone and clarinet 02:14.640 --> 02:18.120 and piano and guitar and drums and whatever. 02:18.120 --> 02:22.200 How does that thread go through your life? 02:22.200 --> 02:24.160 Where's music today? 02:24.160 --> 02:28.320 It's not where I wish it was. 02:28.320 --> 02:30.200 For various reasons, couldn't really keep it going, 02:30.200 --> 02:32.560 particularly because I had a lot of problems with RSI, 02:32.560 --> 02:33.480 with my fingers. 02:33.480 --> 02:37.360 And so I had to cut back anything that used hands 02:37.360 --> 02:39.360 and fingers. 02:39.360 --> 02:43.920 I hope one day I'll be able to get back to it health wise. 02:43.920 --> 02:46.240 So there's a love for music underlying it all. 02:46.240 --> 02:47.840 Sure, yeah. 02:47.840 --> 02:49.480 What's your favorite instrument? 02:49.480 --> 02:50.360 Saxophone. 02:50.360 --> 02:51.000 Sax. 02:51.000 --> 02:52.840 Baritone saxophone. 02:52.840 --> 02:57.440 Well, probably bass saxophone, but they're awkward. 02:57.440 --> 03:00.120 Well, I always love it when music is 03:00.120 --> 03:01.760 coupled with programming. 03:01.760 --> 03:03.800 There's something about a brain that 03:03.800 --> 03:07.520 utilizes those that emerges with creative ideas. 03:07.520 --> 03:11.200 So you've used and studied quite a few programming languages. 03:11.200 --> 03:15.120 Can you give an overview of what you've used? 03:15.120 --> 03:17.920 What are the pros and cons of each? 03:17.920 --> 03:21.960 Well, my favorite programming environment almost certainly 03:21.960 --> 03:26.520 was Microsoft Access back in the earliest days. 03:26.520 --> 03:29.080 So that was a special basic for applications, which 03:29.080 --> 03:30.720 is not a good programming language, 03:30.720 --> 03:33.080 but the programming environment is fantastic. 03:33.080 --> 03:40.120 It's like the ability to create user interfaces and tied data 03:40.120 --> 03:43.720 and actions to them and create reports and all that. 03:43.720 --> 03:46.800 As I've never seen anything as good. 03:46.800 --> 03:48.920 So things nowadays like Airtable, which 03:48.920 --> 03:56.200 are like small subsets of that, which people love for good reason. 03:56.200 --> 04:01.160 But unfortunately, nobody's ever achieved anything like that. 04:01.160 --> 04:03.320 What is that, if you could pause on that for a second? 04:03.320 --> 04:03.840 Oh, Access. 04:03.840 --> 04:04.340 Access. 04:04.340 --> 04:06.320 Is it a fundamental database? 04:06.320 --> 04:09.600 It was a database program that Microsoft produced, 04:09.600 --> 04:13.440 part of Office, and it kind of withered. 04:13.440 --> 04:16.320 But basically, it lets you in a totally graphical way 04:16.320 --> 04:18.480 create tables and relationships and queries 04:18.480 --> 04:24.720 and tie them to forms and set up event handlers and calculations. 04:24.720 --> 04:28.680 And it was a very complete, powerful system designed 04:28.680 --> 04:35.000 for not massive scalable things, but for useful little applications 04:35.000 --> 04:36.400 that I loved. 04:36.400 --> 04:40.240 So what's the connection between Excel and Access? 04:40.240 --> 04:42.160 So very close. 04:42.160 --> 04:47.680 So Access was the relational database equivalent, 04:47.680 --> 04:48.360 if you like. 04:48.360 --> 04:51.080 So people still do a lot of that stuff 04:51.080 --> 04:54.120 that should be in Access in Excel because they know it. 04:54.120 --> 04:56.680 Excel's great as well. 04:56.680 --> 05:01.760 But it's just not as rich a programming model as VBA 05:01.760 --> 05:04.680 combined with a relational database. 05:04.680 --> 05:07.320 And so I've always loved relational databases. 05:07.320 --> 05:11.080 But today, programming on top of relational databases 05:11.080 --> 05:13.840 is just a lot more of a headache. 05:13.840 --> 05:16.680 You generally either need to kind of, 05:16.680 --> 05:19.040 you need something that connects, that runs some kind 05:19.040 --> 05:21.560 of database server, unless you use SQLite, which 05:21.560 --> 05:25.000 has its own issues. 05:25.000 --> 05:26.320 Then you kind of often, if you want 05:26.320 --> 05:27.760 to get a nice programming model, you 05:27.760 --> 05:30.440 need to create an ORM on top. 05:30.440 --> 05:34.360 And then, I don't know, there's all these pieces tied together. 05:34.360 --> 05:37.000 And it's just a lot more awkward than it should be. 05:37.000 --> 05:39.200 There are people that are trying to make it easier, 05:39.200 --> 05:44.480 so in particular, I think of Fsharp, Don Syme, who him 05:44.480 --> 05:49.320 and his team have done a great job of making something 05:49.320 --> 05:51.640 like a database appear in the type system, 05:51.640 --> 05:54.960 so you actually get tab completion for fields and tables 05:54.960 --> 05:57.840 and stuff like that. 05:57.840 --> 05:59.280 Anyway, so that was kind of, anyway, 05:59.280 --> 06:01.880 so that whole VBA Office thing, I guess, 06:01.880 --> 06:04.560 was a starting point, which is your miss. 06:04.560 --> 06:07.800 And I got into Standard Visual Basic, which 06:07.800 --> 06:09.840 that's interesting, just to pause on that for a second. 06:09.840 --> 06:12.600 And it's interesting that you're connecting programming 06:12.600 --> 06:18.200 languages to the ease of management of data. 06:18.200 --> 06:20.600 So in your use of programming languages, 06:20.600 --> 06:24.880 you always had a love and a connection with data. 06:24.880 --> 06:28.640 I've always been interested in doing useful things for myself 06:28.640 --> 06:31.880 and for others, which generally means getting some data 06:31.880 --> 06:34.600 and doing something with it and putting it out there again. 06:34.600 --> 06:38.400 So that's been my interest throughout. 06:38.400 --> 06:41.560 So I also did a lot of stuff with Apple script 06:41.560 --> 06:43.880 back in the early days. 06:43.880 --> 06:47.960 So it's kind of nice being able to get the computer 06:47.960 --> 06:52.960 and computers to talk to each other and to do things for you. 06:52.960 --> 06:56.600 And then I think that one night, the programming language 06:56.600 --> 06:59.960 I most loved then would have been Delphi, which 06:59.960 --> 07:05.960 was Object Pascal created by Anders Halsberg, who previously 07:05.960 --> 07:08.840 did Turbo Pascal and then went on to create.net 07:08.840 --> 07:11.080 and then went on to create TypeScript. 07:11.080 --> 07:16.720 Delphi was amazing because it was like a compiled, fast language 07:16.720 --> 07:20.200 that was as easy to use as Visual Basic. 07:20.200 --> 07:27.480 Delphi, what is it similar to in more modern languages? 07:27.480 --> 07:28.840 Visual Basic. 07:28.840 --> 07:29.680 Visual Basic. 07:29.680 --> 07:32.320 Yeah, that a compiled, fast version. 07:32.320 --> 07:37.080 So I'm not sure there's anything quite like it anymore. 07:37.080 --> 07:42.520 If you took C Sharp or Java and got rid of the virtual machine 07:42.520 --> 07:45.040 and replaced it with something, you could compile a small type 07:45.040 --> 07:46.520 binary. 07:46.520 --> 07:51.680 I feel like it's where Swift could get to with the new Swift 07:51.680 --> 07:56.640 UI and the cross platform development going on. 07:56.640 --> 08:01.600 That's one of my dreams is that we'll hopefully get back 08:01.600 --> 08:02.840 to where Delphi was. 08:02.840 --> 08:08.520 There is actually a free Pascal project nowadays 08:08.520 --> 08:10.320 called Lazarus, which is also attempting 08:10.320 --> 08:13.960 to recreate Delphi. 08:13.960 --> 08:16.080 They're making good progress. 08:16.080 --> 08:21.000 So OK, Delphi, that's one of your favorite programming languages? 08:21.000 --> 08:22.360 Well, it's programming environments. 08:22.360 --> 08:26.280 Again, say Pascal's not a nice language. 08:26.280 --> 08:27.880 If you wanted to know specifically 08:27.880 --> 08:30.360 about what languages I like, I would definitely 08:30.360 --> 08:35.480 pick Jay as being an amazingly wonderful language. 08:35.480 --> 08:37.000 What's Jay? 08:37.000 --> 08:39.600 Jay, are you aware of APL? 08:39.600 --> 08:43.520 I am not, except from doing a little research on the work 08:43.520 --> 08:44.080 you've done. 08:44.080 --> 08:47.280 OK, so not at all surprising you're not 08:47.280 --> 08:49.040 familiar with it because it's not well known, 08:49.040 --> 08:55.480 but it's actually one of the main families of programming 08:55.480 --> 08:57.920 languages going back to the late 50s, early 60s. 08:57.920 --> 09:01.720 So there was a couple of major directions. 09:01.720 --> 09:04.440 One was the kind of lambda, calculus, 09:04.440 --> 09:08.640 Alonzo church direction, which I guess kind of Lisbon scheme 09:08.640 --> 09:12.040 and whatever, which has a history going back 09:12.040 --> 09:13.440 to the early days of computing. 09:13.440 --> 09:17.360 The second was the kind of imperative slash 09:17.360 --> 09:23.240 OO, algo, similar going on to C, C++, so forth. 09:23.240 --> 09:26.960 There was a third, which are called array oriented languages, 09:26.960 --> 09:31.720 which started with a paper by a guy called Ken Iverson, which 09:31.720 --> 09:37.480 was actually a math theory paper, not a programming paper. 09:37.480 --> 09:41.520 It was called Notation as a Tool for Thought. 09:41.520 --> 09:45.320 And it was the development of a new type of math notation. 09:45.320 --> 09:48.560 And the idea is that this math notation was much more 09:48.560 --> 09:54.480 flexible, expressive, and also well defined than traditional 09:54.480 --> 09:56.440 math notation, which is none of those things. 09:56.440 --> 09:59.160 Math notation is awful. 09:59.160 --> 10:02.840 And so he actually turned that into a programming language. 10:02.840 --> 10:06.720 Because this was the late 50s, all the names were available. 10:06.720 --> 10:10.520 So he called his programming language, or APL. 10:10.520 --> 10:11.160 APL, what? 10:11.160 --> 10:15.360 So APL is a implementation of notation 10:15.360 --> 10:18.280 as a tool for thought, by which he means math notation. 10:18.280 --> 10:22.880 And Ken and his son went on to do many things, 10:22.880 --> 10:26.720 but eventually they actually produced a new language that 10:26.720 --> 10:28.440 was built on top of all the learnings of APL. 10:28.440 --> 10:32.800 And that was called J. And J is the most 10:32.800 --> 10:41.040 expressive, composable, beautifully designed language 10:41.040 --> 10:42.400 I've ever seen. 10:42.400 --> 10:44.520 Does it have object oriented components? 10:44.520 --> 10:45.520 Does it have that kind of thing? 10:45.520 --> 10:46.240 Not really. 10:46.240 --> 10:47.720 It's an array oriented language. 10:47.720 --> 10:51.400 It's the third path. 10:51.400 --> 10:52.760 Are you saying array? 10:52.760 --> 10:53.720 Array oriented. 10:53.720 --> 10:54.200 Yeah. 10:54.200 --> 10:55.480 It needs to be array oriented. 10:55.480 --> 10:57.480 So array oriented means that you generally 10:57.480 --> 10:59.520 don't use any loops. 10:59.520 --> 11:02.240 But the whole thing is done with kind 11:02.240 --> 11:06.360 of an extreme version of broadcasting, 11:06.360 --> 11:09.880 if you're familiar with that NumPy slash Python concept. 11:09.880 --> 11:14.240 So you do a lot with one line of code. 11:14.240 --> 11:17.520 It looks a lot like math. 11:17.520 --> 11:20.280 Notation is basically highly compact. 11:20.280 --> 11:22.800 And the idea is that you can kind of, 11:22.800 --> 11:24.760 because you can do so much with one line of code, 11:24.760 --> 11:27.720 a single screen of code is very unlikely to, 11:27.720 --> 11:31.080 you very rarely need more than that to express your program. 11:31.080 --> 11:33.240 And so you can kind of keep it all in your head. 11:33.240 --> 11:36.000 And you can kind of clearly communicate it. 11:36.000 --> 11:41.560 It's interesting that APL created two main branches, K and J. 11:41.560 --> 11:47.920 J is this kind of like open source niche community of crazy 11:47.920 --> 11:49.360 enthusiasts like me. 11:49.360 --> 11:52.120 And then the other path, K, was fascinating. 11:52.120 --> 11:56.600 It's an astonishingly expensive programming language, 11:56.600 --> 12:01.920 which many of the world's most ludicrously rich hedge funds 12:01.920 --> 12:02.840 use. 12:02.840 --> 12:06.640 So the entire K machine is so small, 12:06.640 --> 12:09.320 it sits inside level three cache on your CPU. 12:09.320 --> 12:14.040 And it easily wins every benchmark I've ever seen 12:14.040 --> 12:16.440 in terms of data processing speed. 12:16.440 --> 12:17.840 But you don't come across it very much, 12:17.840 --> 12:22.640 because it's like $100,000 per CPU to run it. 12:22.640 --> 12:26.240 But it's like this path of programming languages 12:26.240 --> 12:29.760 is just so much, I don't know, so much more powerful 12:29.760 --> 12:33.840 in every way than the ones that almost anybody uses every day. 12:33.840 --> 12:37.400 So it's all about computation. 12:37.400 --> 12:38.360 It's really focusing on it. 12:38.360 --> 12:40.640 Pretty heavily focused on computation. 12:40.640 --> 12:44.320 I mean, so much of programming is data processing 12:44.320 --> 12:45.640 by definition. 12:45.640 --> 12:49.000 And so there's a lot of things you can do with it. 12:49.000 --> 12:51.320 But yeah, there's not much work being 12:51.320 --> 12:57.080 done on making user interface toolkills or whatever. 12:57.080 --> 12:59.400 I mean, there's some, but they're not great. 12:59.400 --> 13:03.160 At the same time, you've done a lot of stuff with Perl and Python. 13:03.160 --> 13:08.320 So what does that fit into the picture of J and K and APL 13:08.320 --> 13:08.880 and Python? 13:08.880 --> 13:12.400 Well, it's just much more pragmatic. 13:12.400 --> 13:13.960 In the end, you kind of have to end up 13:13.960 --> 13:17.960 where the libraries are. 13:17.960 --> 13:21.320 Because to me, my focus is on productivity. 13:21.320 --> 13:23.800 I just want to get stuff done and solve problems. 13:23.800 --> 13:27.360 So Perl was great. 13:27.360 --> 13:29.760 I created an email company called Fastmail. 13:29.760 --> 13:35.200 And Perl was great, because back in the late 90s, early 2000s, 13:35.200 --> 13:38.160 it just had a lot of stuff it could do. 13:38.160 --> 13:41.840 I still had to write my own monitoring system 13:41.840 --> 13:43.840 and my own web framework and my own whatever, 13:43.840 --> 13:45.760 because none of that stuff existed. 13:45.760 --> 13:50.280 But it was a super flexible language to do that in. 13:50.280 --> 13:52.720 And you used Perl for Fastmail. 13:52.720 --> 13:54.520 You used it as a back end. 13:54.520 --> 13:55.800 So everything was written in Perl? 13:55.800 --> 13:56.520 Yeah. 13:56.520 --> 13:58.720 Yeah, everything was Perl. 13:58.720 --> 14:04.480 Why do you think Perl hasn't succeeded or hasn't dominated 14:04.480 --> 14:07.120 the market where Python really takes over a lot of the 14:07.120 --> 14:08.200 tests? 14:08.200 --> 14:09.640 Well, I mean, Perl did dominate. 14:09.640 --> 14:13.080 It was everything, everywhere. 14:13.080 --> 14:19.920 But then the guy that ran Perl, Larry Wall, 14:19.920 --> 14:22.280 just didn't put the time in anymore. 14:22.280 --> 14:29.680 And no project can be successful if there isn't. 14:29.680 --> 14:32.640 Particularly one that started with a strong leader that 14:32.640 --> 14:35.040 loses that strong leadership. 14:35.040 --> 14:38.040 So then Python has kind of replaced it. 14:38.040 --> 14:45.040 Python is a lot less elegant language in nearly every way. 14:45.040 --> 14:48.880 But it has the data science libraries. 14:48.880 --> 14:51.240 And a lot of them are pretty great. 14:51.240 --> 14:58.280 So I kind of use it because it's the best we have. 14:58.280 --> 15:01.800 But it's definitely not good enough. 15:01.800 --> 15:04.040 What do you think the future of programming looks like? 15:04.040 --> 15:06.880 What do you hope the future of programming looks like if we 15:06.880 --> 15:10.200 zoom in on the computational fields on data science 15:10.200 --> 15:11.800 and machine learning? 15:11.800 --> 15:19.440 I hope Swift is successful because the goal of Swift, 15:19.440 --> 15:21.000 the way Chris Latna describes it, 15:21.000 --> 15:22.640 is to be infinitely hackable. 15:22.640 --> 15:23.480 And that's what I want. 15:23.480 --> 15:26.920 I want something where me and the people I do research with 15:26.920 --> 15:30.360 and my students can look at and change everything 15:30.360 --> 15:32.000 from top to bottom. 15:32.000 --> 15:36.240 There's nothing mysterious and magical and inaccessible. 15:36.240 --> 15:38.600 Unfortunately, with Python, it's the opposite of that 15:38.600 --> 15:42.640 because Python is so slow, it's extremely unhackable. 15:42.640 --> 15:44.840 You get to a point where it's like, OK, from here on down 15:44.840 --> 15:47.320 at C. So your debugger doesn't work in the same way. 15:47.320 --> 15:48.920 Your profiler doesn't work in the same way. 15:48.920 --> 15:50.880 Your build system doesn't work in the same way. 15:50.880 --> 15:53.760 It's really not very hackable at all. 15:53.760 --> 15:55.600 What's the part you like to be hackable? 15:55.600 --> 16:00.120 Is it for the objective of optimizing training 16:00.120 --> 16:02.600 of neural networks, inference of neural networks? 16:02.600 --> 16:04.360 Is it performance of the system? 16:04.360 --> 16:08.440 Or is there some nonperformance related, just creative idea? 16:08.440 --> 16:09.080 It's everything. 16:09.080 --> 16:15.480 I mean, in the end, I want to be productive as a practitioner. 16:15.480 --> 16:18.440 So at the moment, our understanding of deep learning 16:18.440 --> 16:20.080 is incredibly primitive. 16:20.080 --> 16:21.520 There's very little we understand. 16:21.520 --> 16:24.200 Most things don't work very well, even though it works better 16:24.200 --> 16:26.200 than anything else out there. 16:26.200 --> 16:28.760 There's so many opportunities to make it better. 16:28.760 --> 16:34.360 So you look at any domain area like speech recognition 16:34.360 --> 16:37.720 with deep learning or natural language processing 16:37.720 --> 16:39.440 classification with deep learning or whatever. 16:39.440 --> 16:41.960 Every time I look at an area with deep learning, 16:41.960 --> 16:44.480 I always see like, oh, it's terrible. 16:44.480 --> 16:47.560 There's lots and lots of obviously stupid ways 16:47.560 --> 16:50.000 to do things that need to be fixed. 16:50.000 --> 16:53.320 So then I want to be able to jump in there and quickly 16:53.320 --> 16:54.880 experiment and make them better. 16:54.880 --> 16:59.320 Do you think the programming language has a role in that? 16:59.320 --> 17:00.280 Huge role, yeah. 17:00.280 --> 17:07.080 So currently, Python has a big gap in terms of our ability 17:07.080 --> 17:11.880 to innovate particularly around recurrent neural networks 17:11.880 --> 17:16.840 and natural language processing because it's so slow. 17:16.840 --> 17:20.200 The actual loop where we actually loop through words, 17:20.200 --> 17:23.760 we have to do that whole thing in CUDA C. 17:23.760 --> 17:27.600 So we actually can't innovate with the kernel, the heart, 17:27.600 --> 17:31.560 of that most important algorithm. 17:31.560 --> 17:33.680 And it's just a huge problem. 17:33.680 --> 17:36.600 And this happens all over the place. 17:36.600 --> 17:40.080 So we hit research limitations. 17:40.080 --> 17:42.840 Another example, convolutional neural networks, which 17:42.840 --> 17:46.800 are actually the most popular architecture for lots of things, 17:46.800 --> 17:48.920 maybe most things in deep learning. 17:48.920 --> 17:50.360 We almost certainly should be using 17:50.360 --> 17:54.600 sparse convolutional neural networks, but only like two 17:54.600 --> 17:56.800 people are because to do it, you have 17:56.800 --> 17:59.920 to rewrite all of that CUDA C level stuff. 17:59.920 --> 18:04.520 And yeah, just research, just in practitioners, don't. 18:04.520 --> 18:09.240 So there's just big gaps in what people actually research on, 18:09.240 --> 18:11.640 what people actually implement because of the programming 18:11.640 --> 18:13.240 language problem. 18:13.240 --> 18:17.560 So you think it's just too difficult 18:17.560 --> 18:23.480 to write in CUDA C that a higher level programming language 18:23.480 --> 18:30.520 like Swift should enable the easier, 18:30.520 --> 18:33.160 fooling around, create stuff with RNNs, 18:33.160 --> 18:34.920 or sparse convolutional neural networks? 18:34.920 --> 18:35.920 Kind of. 18:35.920 --> 18:38.520 Who is at fault? 18:38.520 --> 18:42.320 Who is at charge of making it easy for a researcher to play around? 18:42.320 --> 18:43.520 I mean, no one's at fault. 18:43.520 --> 18:45.120 Just nobody's got a round to it yet. 18:45.120 --> 18:47.080 Or it's just it's hard. 18:47.080 --> 18:51.800 And I mean, part of the fault is that we ignored that whole APL 18:51.800 --> 18:55.640 kind of direction, or nearly everybody did for 60 years, 18:55.640 --> 18:57.720 50 years. 18:57.720 --> 18:59.920 But recently, people have been starting 18:59.920 --> 19:04.840 to reinvent pieces of that and kind of create some interesting 19:04.840 --> 19:07.400 new directions in the compiler technology. 19:07.400 --> 19:11.760 So the place where that's particularly happening right now 19:11.760 --> 19:14.920 is something called MLIR, which is something that, again, 19:14.920 --> 19:18.000 Chris Lattener, the Swift guy, is leading. 19:18.000 --> 19:20.080 And because it's actually not going 19:20.080 --> 19:22.160 to be Swift on its own that solves this problem. 19:22.160 --> 19:24.880 Because the problem is that currently writing 19:24.880 --> 19:32.360 a acceptably fast GPU program is too complicated, 19:32.360 --> 19:33.680 regardless of what language you use. 19:36.480 --> 19:38.680 And that's just because if you have to deal with the fact 19:38.680 --> 19:43.160 that I've got 10,000 threads and I have to synchronize between them 19:43.160 --> 19:45.360 all, and I have to put my thing into grid blocks 19:45.360 --> 19:47.040 and think about warps and all this stuff, 19:47.040 --> 19:50.720 it's just so much boilerplate that to do that well, 19:50.720 --> 19:52.240 you have to be a specialist at that. 19:52.240 --> 19:58.200 And it's going to be a year's work to optimize that algorithm 19:58.200 --> 19:59.720 in that way. 19:59.720 --> 20:04.640 But with things like TensorFlow Comprehensions, and Tile, 20:04.640 --> 20:08.880 and MLIR, and TVM, there's all these various projects which 20:08.880 --> 20:11.840 are all about saying, let's let people 20:11.840 --> 20:16.080 create domain specific languages for tensor 20:16.080 --> 20:16.880 computations. 20:16.880 --> 20:19.120 These are the kinds of things we do generally 20:19.120 --> 20:21.640 on the GPU for deep learning, and then 20:21.640 --> 20:28.280 have a compiler which can optimize that tensor computation. 20:28.280 --> 20:31.440 A lot of this work is actually sitting on top of a project 20:31.440 --> 20:36.040 called Halide, which is a mind blowing project 20:36.040 --> 20:38.880 where they came up with such a domain specific language. 20:38.880 --> 20:41.240 In fact, two, one domain specific language for expressing, 20:41.240 --> 20:43.840 this is what my tensor computation is. 20:43.840 --> 20:46.320 And another domain specific language for expressing, 20:46.320 --> 20:50.320 this is the way I want you to structure 20:50.320 --> 20:53.040 the compilation of that, and do it block by block 20:53.040 --> 20:54.960 and do these bits in parallel. 20:54.960 --> 20:57.760 And they were able to show how you can compress 20:57.760 --> 21:02.880 the amount of code by 10x compared to optimized GPU 21:02.880 --> 21:05.600 code and get the same performance. 21:05.600 --> 21:08.480 So these are the things that are sitting on top 21:08.480 --> 21:12.240 of that kind of research, and MLIR 21:12.240 --> 21:15.160 is pulling a lot of those best practices together. 21:15.160 --> 21:17.160 And now we're starting to see work done 21:17.160 --> 21:21.400 on making all of that directly accessible through Swift 21:21.400 --> 21:25.040 so that I could use Swift to write those domain specific 21:25.040 --> 21:25.880 languages. 21:25.880 --> 21:29.520 And hopefully we'll get then Swift CUDA kernels 21:29.520 --> 21:31.720 written in a very expressive and concise way that 21:31.720 --> 21:36.280 looks a bit like J in APL, and then Swift layers on top 21:36.280 --> 21:38.360 of that, and then a Swift UI on top of that, 21:38.360 --> 21:42.600 and it'll be so nice if we can get to that point. 21:42.600 --> 21:48.560 Now does it all eventually boil down to CUDA and NVIDIA GPUs? 21:48.560 --> 21:50.120 Unfortunately at the moment it does, 21:50.120 --> 21:52.600 but one of the nice things about MLIR, 21:52.600 --> 21:56.120 if AMD ever gets their act together, which they probably 21:56.120 --> 21:59.040 want, is that they or others could 21:59.040 --> 22:05.000 write MLIR backends for other GPUs 22:05.000 --> 22:10.320 or rather tensor computation devices, of which today 22:10.320 --> 22:15.520 there are increasing number like Graphcore or Vertex AI 22:15.520 --> 22:18.840 or whatever. 22:18.840 --> 22:22.600 So yeah, being able to target lots of backends 22:22.600 --> 22:23.960 would be another benefit of this, 22:23.960 --> 22:26.680 and the market really needs competition, 22:26.680 --> 22:28.680 because at the moment NVIDIA is massively 22:28.680 --> 22:33.640 overcharging for their kind of enterprise class cards, 22:33.640 --> 22:36.720 because there is no serious competition, 22:36.720 --> 22:39.280 because nobody else is doing the software properly. 22:39.280 --> 22:41.400 In the cloud there is some competition, right? 22:41.400 --> 22:45.080 But not really, other than TPUs perhaps, 22:45.080 --> 22:49.040 but TPUs are almost unprogrammable at the moment. 22:49.040 --> 22:51.080 TPUs have the same problem that you can't. 22:51.080 --> 22:51.760 It's even worse. 22:51.760 --> 22:54.800 So TPUs, Google actually made an explicit decision 22:54.800 --> 22:57.200 to make them almost entirely unprogrammable, 22:57.200 --> 22:59.960 because they felt that there was too much IP in there, 22:59.960 --> 23:02.640 and if they gave people direct access to program them, 23:02.640 --> 23:04.360 people would learn their secrets. 23:04.360 --> 23:09.720 So you can't actually directly program 23:09.720 --> 23:12.120 the memory in a TPU. 23:12.120 --> 23:16.360 You can't even directly create code that runs on 23:16.360 --> 23:19.080 and that you look at on the machine that has the TPU. 23:19.080 --> 23:20.920 It all goes through a virtual machine. 23:20.920 --> 23:23.680 So all you can really do is this kind of cookie cutter 23:23.680 --> 23:27.760 thing of like plug in high level stuff together, 23:27.760 --> 23:31.440 which is just super tedious and annoying 23:31.440 --> 23:33.920 and totally unnecessary. 23:33.920 --> 23:40.960 So tell me if you could, the origin story of fast AI. 23:40.960 --> 23:45.760 What is the motivation, its mission, its dream? 23:45.760 --> 23:50.040 So I guess the founding story is heavily 23:50.040 --> 23:51.840 tied to my previous startup, which 23:51.840 --> 23:53.960 is a company called Inletic, which 23:53.960 --> 23:58.280 was the first company to focus on deep learning for medicine. 23:58.280 --> 24:03.240 And I created that because I saw there was a huge opportunity 24:03.240 --> 24:07.960 to, there's about a 10x shortage of the number of doctors 24:07.960 --> 24:12.120 in the world and the developing world that we need. 24:12.120 --> 24:13.840 I expected it would take about 300 years 24:13.840 --> 24:16.120 to train enough doctors to meet that gap. 24:16.120 --> 24:20.760 But I guessed that maybe if we used 24:20.760 --> 24:23.760 deep learning for some of the analytics, 24:23.760 --> 24:25.760 we could maybe make it so you don't need 24:25.760 --> 24:27.320 as highly trained doctors. 24:27.320 --> 24:28.320 For diagnosis? 24:28.320 --> 24:29.840 For diagnosis and treatment planning. 24:29.840 --> 24:33.440 Where's the biggest benefit just before get the fast AI? 24:33.440 --> 24:37.280 Where's the biggest benefit of AI and medicine that you see 24:37.280 --> 24:39.440 today and in the future? 24:39.440 --> 24:41.960 Not much happening today in terms of stuff that's actually 24:41.960 --> 24:42.440 out there. 24:42.440 --> 24:43.160 It's very early. 24:43.160 --> 24:45.320 But in terms of the opportunity, it's 24:45.320 --> 24:51.080 to take markets like India and China and Indonesia, which 24:51.080 --> 24:58.120 have big populations, Africa, small numbers of doctors, 24:58.120 --> 25:02.440 and provide diagnostic, particularly treatment 25:02.440 --> 25:05.160 planning and triage kind of on device 25:05.160 --> 25:10.360 so that if you do a test for malaria or tuberculosis 25:10.360 --> 25:12.800 or whatever, you immediately get something 25:12.800 --> 25:14.840 that even a health care worker that's 25:14.840 --> 25:20.360 had a month of training can get a very high quality 25:20.360 --> 25:23.480 assessment of whether the patient might be at risk 25:23.480 --> 25:27.480 until OK, we'll send them off to a hospital. 25:27.480 --> 25:31.720 So for example, in Africa, outside of South Africa, 25:31.720 --> 25:34.080 there's only five pediatric radiologists 25:34.080 --> 25:35.320 for the entire continent. 25:35.320 --> 25:37.200 So most countries don't have any. 25:37.200 --> 25:39.240 So if your kid is sick and they need something 25:39.240 --> 25:41.200 diagnosed through medical imaging, 25:41.200 --> 25:44.040 the person, even if you're able to get medical imaging done, 25:44.040 --> 25:48.920 the person that looks at it will be a nurse at best. 25:48.920 --> 25:52.480 But actually, in India, for example, and China, 25:52.480 --> 25:54.760 almost no x rays are read by anybody, 25:54.760 --> 25:59.400 by any trained professional, because they don't have enough. 25:59.400 --> 26:02.880 So if instead we had an algorithm that 26:02.880 --> 26:10.080 could take the most likely high risk 5% and say triage, 26:10.080 --> 26:13.280 basically say, OK, somebody needs to look at this, 26:13.280 --> 26:16.240 it would massively change the kind of way 26:16.240 --> 26:20.640 that what's possible with medicine in the developing world. 26:20.640 --> 26:23.680 And remember, increasingly, they have money. 26:23.680 --> 26:24.800 They're the developing world. 26:24.800 --> 26:26.160 They're not the poor world, the developing world. 26:26.160 --> 26:26.920 So they have the money. 26:26.920 --> 26:28.480 So they're building the hospitals. 26:28.480 --> 26:31.960 They're getting the diagnostic equipment. 26:31.960 --> 26:34.880 But there's no way for a very long time 26:34.880 --> 26:38.480 will they be able to have the expertise. 26:38.480 --> 26:39.760 Shortage of expertise. 26:39.760 --> 26:42.720 OK, and that's where the deep learning systems 26:42.720 --> 26:46.040 can step in and magnify the expertise they do have. 26:46.040 --> 26:47.840 Exactly. 26:47.840 --> 26:54.160 So you do see, just to linger a little bit longer, 26:54.160 --> 26:58.520 the interaction, do you still see the human experts still 26:58.520 --> 26:59.840 at the core of the system? 26:59.840 --> 27:00.480 Yeah, absolutely. 27:00.480 --> 27:01.720 Is there something in medicine that 27:01.720 --> 27:03.760 could be automated almost completely? 27:03.760 --> 27:06.360 I don't see the point of even thinking about that, 27:06.360 --> 27:08.480 because we have such a shortage of people. 27:08.480 --> 27:12.160 Why would we want to find a way not to use them? 27:12.160 --> 27:13.840 Like, we have people. 27:13.840 --> 27:17.200 So the idea of, even from an economic point of view, 27:17.200 --> 27:19.800 if you can make them 10x more productive, 27:19.800 --> 27:21.600 getting rid of the person doesn't 27:21.600 --> 27:23.880 impact your unit economics at all. 27:23.880 --> 27:26.680 And it totally involves the fact that there are things 27:26.680 --> 27:28.760 people do better than machines. 27:28.760 --> 27:33.120 So it's just, to me, that's not a useful way 27:33.120 --> 27:34.120 of framing the problem. 27:34.120 --> 27:36.440 I guess, just to clarify, I guess I 27:36.440 --> 27:40.560 meant there may be some problems where you can avoid even 27:40.560 --> 27:42.160 going to the expert ever. 27:42.160 --> 27:46.160 Sort of maybe preventative care or some basic stuff, 27:46.160 --> 27:47.800 the low hanging fruit, allowing the expert 27:47.800 --> 27:51.320 to focus on the things that are really that. 27:51.320 --> 27:52.960 Well, that's what the triage would do, right? 27:52.960 --> 28:00.760 So the triage would say, OK, 99% sure there's nothing here. 28:00.760 --> 28:04.040 So that can be done on device. 28:04.040 --> 28:05.920 And they can just say, OK, go home. 28:05.920 --> 28:10.520 So the experts are being used to look at the stuff which 28:10.520 --> 28:12.240 has some chance it's worth looking at, 28:12.240 --> 28:15.720 which most things is not. 28:15.720 --> 28:16.280 It's fine. 28:16.280 --> 28:19.840 Why do you think we haven't quite made progress on that yet 28:19.840 --> 28:27.480 in terms of the scale of how much AI is applied in the method? 28:27.480 --> 28:28.400 There's a lot of reasons. 28:28.400 --> 28:29.640 I mean, one is it's pretty new. 28:29.640 --> 28:32.040 I only started in late 2014. 28:32.040 --> 28:35.920 And before that, it's hard to express 28:35.920 --> 28:37.760 to what degree the medical world was not 28:37.760 --> 28:40.720 aware of the opportunities here. 28:40.720 --> 28:45.520 So I went to RSNA, which is the world's largest radiology 28:45.520 --> 28:46.240 conference. 28:46.240 --> 28:50.040 And I told everybody I could, like, 28:50.040 --> 28:51.800 I'm doing this thing with deep learning. 28:51.800 --> 28:53.320 Please come and check it out. 28:53.320 --> 28:56.880 And no one had any idea what I was talking about. 28:56.880 --> 28:59.640 No one had any interest in it. 28:59.640 --> 29:05.040 So we've come from absolute zero, which is hard. 29:05.040 --> 29:09.920 And then the whole regulatory framework, education system, 29:09.920 --> 29:13.400 everything is just set up to think of doctoring 29:13.400 --> 29:14.920 in a very different way. 29:14.920 --> 29:16.400 So today, there is a small number 29:16.400 --> 29:22.040 of people who are deep learning practitioners and doctors 29:22.040 --> 29:22.960 at the same time. 29:22.960 --> 29:25.040 And we're starting to see the first ones come out 29:25.040 --> 29:26.520 of their PhD programs. 29:26.520 --> 29:33.960 So Zach Cahane over in Boston, Cambridge 29:33.960 --> 29:41.040 has a number of students now who are data science experts, 29:41.040 --> 29:46.400 deep learning experts, and actual medical doctors. 29:46.400 --> 29:49.480 Quite a few doctors have completed our fast AI course 29:49.480 --> 29:54.920 now and are publishing papers and creating journal reading 29:54.920 --> 29:58.040 groups in the American Council of Radiology. 29:58.040 --> 30:00.280 And it's just starting to happen. 30:00.280 --> 30:02.840 But it's going to be a long process. 30:02.840 --> 30:04.920 The regulators have to learn how to regulate this. 30:04.920 --> 30:08.720 They have to build guidelines. 30:08.720 --> 30:12.120 And then the lawyers at hospitals 30:12.120 --> 30:15.080 have to develop a new way of understanding 30:15.080 --> 30:18.680 that sometimes it makes sense for data 30:18.680 --> 30:24.880 to be looked at in raw form in large quantities 30:24.880 --> 30:27.000 in order to create world changing results. 30:27.000 --> 30:30.080 Yeah, there's a regulation around data, all that. 30:30.080 --> 30:33.840 It sounds probably the hardest problem, 30:33.840 --> 30:36.760 but it sounds reminiscent of autonomous vehicles as well. 30:36.760 --> 30:38.760 Many of the same regulatory challenges, 30:38.760 --> 30:40.560 many of the same data challenges. 30:40.560 --> 30:42.160 Yeah, I mean, funnily enough, the problem 30:42.160 --> 30:44.880 is less the regulation and more the interpretation 30:44.880 --> 30:48.200 of that regulation by lawyers in hospitals. 30:48.200 --> 30:52.560 So HIPAA was actually designed. 30:52.560 --> 30:56.400 The P in HIPAA does not stand for privacy. 30:56.400 --> 30:57.640 It stands for portability. 30:57.640 --> 31:01.200 It's actually meant to be a way that data can be used. 31:01.200 --> 31:04.400 And it was created with lots of gray areas 31:04.400 --> 31:06.560 because the idea is that would be more practical 31:06.560 --> 31:10.480 and it would help people to use this legislation 31:10.480 --> 31:13.680 to actually share data in a more thoughtful way. 31:13.680 --> 31:15.320 Unfortunately, it's done the opposite 31:15.320 --> 31:18.880 because when a lawyer sees a gray area, they see, oh, 31:18.880 --> 31:22.440 if we don't know we won't get sued, then we can't do it. 31:22.440 --> 31:26.360 So HIPAA is not exactly the problem. 31:26.360 --> 31:30.080 The problem is more that hospital lawyers 31:30.080 --> 31:34.720 are not incented to make bold decisions 31:34.720 --> 31:36.520 about data portability. 31:36.520 --> 31:40.480 Or even to embrace technology that saves lives. 31:40.480 --> 31:42.440 They more want to not get in trouble 31:42.440 --> 31:44.280 for embracing that technology. 31:44.280 --> 31:47.840 Also, it is also saves lives in a very abstract way, 31:47.840 --> 31:49.840 which is like, oh, we've been able to release 31:49.840 --> 31:52.360 these 100,000 anonymous records. 31:52.360 --> 31:55.360 I can't point at the specific person whose life that's saved. 31:55.360 --> 31:57.760 I can say like, oh, we've ended up with this paper 31:57.760 --> 32:02.200 which found this result, which diagnosed 1,000 more people 32:02.200 --> 32:04.200 than we would have otherwise, but it's like, 32:04.200 --> 32:07.360 which ones were helped, it's very abstract. 32:07.360 --> 32:09.400 Yeah, and on the counter side of that, 32:09.400 --> 32:13.080 you may be able to point to a life that was taken 32:13.080 --> 32:14.360 because of something that was... 32:14.360 --> 32:18.240 Yeah, or a person whose privacy was violated. 32:18.240 --> 32:20.360 It's like, oh, this specific person, 32:20.360 --> 32:25.480 you know, there was deidentified. 32:25.480 --> 32:27.360 Just a fascinating topic. 32:27.360 --> 32:28.360 We're jumping around. 32:28.360 --> 32:32.880 We'll get back to fast AI, but on the question of privacy, 32:32.880 --> 32:38.160 data is the fuel for so much innovation in deep learning. 32:38.160 --> 32:39.840 What's your sense on privacy, 32:39.840 --> 32:44.080 whether we're talking about Twitter, Facebook, YouTube, 32:44.080 --> 32:48.720 just the technologies like in the medical field 32:48.720 --> 32:53.440 that rely on people's data in order to create impact? 32:53.440 --> 32:58.840 How do we get that right, respecting people's privacy 32:58.840 --> 33:03.360 and yet creating technology that is learned from data? 33:03.360 --> 33:11.480 One of my areas of focus is on doing more with less data, 33:11.480 --> 33:15.000 which so most vendors, unfortunately, are strongly 33:15.000 --> 33:20.000 centred to find ways to require more data and more computation. 33:20.000 --> 33:24.000 So Google and IBM being the most obvious... 33:24.000 --> 33:26.000 IBM. 33:26.000 --> 33:30.600 Yeah, so Watson, you know, so Google and IBM both strongly push 33:30.600 --> 33:35.400 the idea that they have more data and more computation 33:35.400 --> 33:37.800 and more intelligent people than anybody else, 33:37.800 --> 33:39.840 and so you have to trust them to do things 33:39.840 --> 33:42.600 because nobody else can do it. 33:42.600 --> 33:45.360 And Google's very upfront about this, 33:45.360 --> 33:48.680 like Jeff Dain has gone out there and given talks and said, 33:48.680 --> 33:52.840 our goal is to require 1,000 times more computation, 33:52.840 --> 33:55.120 but less people. 33:55.120 --> 34:00.600 Our goal is to use the people that you have better 34:00.600 --> 34:02.960 and the data you have better and the computation you have better. 34:02.960 --> 34:06.000 So one of the things that we've discovered is, 34:06.000 --> 34:11.080 or at least highlighted, is that you very, very, very often 34:11.080 --> 34:13.360 don't need much data at all. 34:13.360 --> 34:16.160 And so the data you already have in your organization 34:16.160 --> 34:19.240 will be enough to get state of the art results. 34:19.240 --> 34:22.600 So like my starting point would be to kind of say around privacy 34:22.600 --> 34:25.760 is a lot of people are looking for ways 34:25.760 --> 34:28.120 to share data and aggregate data, 34:28.120 --> 34:29.920 but I think often that's unnecessary. 34:29.920 --> 34:32.160 They assume that they need more data than they do 34:32.160 --> 34:35.240 because they're not familiar with the basics of transfer 34:35.240 --> 34:38.440 learning, which is this critical technique 34:38.440 --> 34:42.000 for needing orders of magnitude less data. 34:42.000 --> 34:44.680 Is your sense, one reason you might want to collect data 34:44.680 --> 34:50.440 from everyone is like in the recommender system context, 34:50.440 --> 34:54.520 where your individual, Jeremy Howard's individual data 34:54.520 --> 34:58.600 is the most useful for providing a product that's 34:58.600 --> 34:59.880 impactful for you. 34:59.880 --> 35:02.240 So for giving you advertisements, 35:02.240 --> 35:07.640 for recommending to you movies, for doing medical diagnosis. 35:07.640 --> 35:11.720 Is your sense we can build with a small amount of data, 35:11.720 --> 35:16.040 general models that will have a huge impact for most people, 35:16.040 --> 35:19.120 that we don't need to have data from each individual? 35:19.120 --> 35:20.560 On the whole, I'd say yes. 35:20.560 --> 35:26.400 I mean, there are things like, recommender systems 35:26.400 --> 35:30.960 have this cold start problem, where Jeremy is a new customer. 35:30.960 --> 35:33.280 We haven't seen him before, so we can't recommend him things 35:33.280 --> 35:36.520 based on what else he's bought and liked with us. 35:36.520 --> 35:39.440 And there's various workarounds to that. 35:39.440 --> 35:41.160 A lot of music programs will start out 35:41.160 --> 35:44.920 by saying, which of these artists do you like? 35:44.920 --> 35:46.800 Which of these albums do you like? 35:46.800 --> 35:49.800 Which of these songs do you like? 35:49.800 --> 35:51.040 Netflix used to do that. 35:51.040 --> 35:55.320 Nowadays, people don't like that because they think, oh, 35:55.320 --> 35:57.400 we don't want to bother the user. 35:57.400 --> 36:00.560 So you could work around that by having some kind of data 36:00.560 --> 36:04.240 sharing where you get my marketing record from Axiom 36:04.240 --> 36:06.360 or whatever and try to question that. 36:06.360 --> 36:12.360 To me, the benefit to me and to society 36:12.360 --> 36:16.520 of saving me five minutes on answering some questions 36:16.520 --> 36:23.520 versus the negative externalities of the privacy issue 36:23.520 --> 36:24.800 doesn't add up. 36:24.800 --> 36:26.600 So I think a lot of the time, the places 36:26.600 --> 36:30.520 where people are invading our privacy in order 36:30.520 --> 36:35.360 to provide convenience is really about just trying 36:35.360 --> 36:36.880 to make them more money. 36:36.880 --> 36:40.760 And they move these negative externalities 36:40.760 --> 36:44.360 into places that they don't have to pay for them. 36:44.360 --> 36:48.120 So when you actually see regulations 36:48.120 --> 36:50.560 appear that actually cause the companies that 36:50.560 --> 36:52.360 create these negative externalities to have 36:52.360 --> 36:54.320 to pay for it themselves, they say, well, 36:54.320 --> 36:56.160 we can't do it anymore. 36:56.160 --> 36:58.240 So the cost is actually too high. 36:58.240 --> 37:02.280 But for something like medicine, the hospital 37:02.280 --> 37:06.440 has my medical imaging, my pathology studies, 37:06.440 --> 37:08.920 my medical records. 37:08.920 --> 37:11.920 And also, I own my medical data. 37:11.920 --> 37:16.960 So I help a startup called DocAI. 37:16.960 --> 37:19.760 One of the things DocAI does is that it has an app. 37:19.760 --> 37:26.120 You can connect to Sutter Health and Labcore and Walgreens 37:26.120 --> 37:29.840 and download your medical data to your phone 37:29.840 --> 37:33.560 and then upload it, again, at your discretion 37:33.560 --> 37:36.040 to share it as you wish. 37:36.040 --> 37:38.440 So with that kind of approach, we 37:38.440 --> 37:41.160 can share our medical information 37:41.160 --> 37:44.840 with the people we want to. 37:44.840 --> 37:45.720 Yeah, so control. 37:45.720 --> 37:48.240 I mean, really being able to control who you share it with 37:48.240 --> 37:49.760 and so on. 37:49.760 --> 37:53.080 So that has a beautiful, interesting tangent 37:53.080 --> 37:59.360 to return back to the origin story of FastAI. 37:59.360 --> 38:02.520 Right, so before I started FastAI, 38:02.520 --> 38:07.160 I spent a year researching where are the biggest 38:07.160 --> 38:10.400 opportunities for deep learning. 38:10.400 --> 38:14.080 Because I knew from my time at Kaggle in particular 38:14.080 --> 38:17.960 that deep learning had hit this threshold point where it was 38:17.960 --> 38:20.520 rapidly becoming the state of the art approach in every area 38:20.520 --> 38:21.600 that looked at it. 38:21.600 --> 38:25.400 And I'd been working with neural nets for over 20 years. 38:25.400 --> 38:27.440 I knew that from a theoretical point of view, 38:27.440 --> 38:30.760 once it hit that point, it would do that in just about every 38:30.760 --> 38:31.600 domain. 38:31.600 --> 38:34.480 And so I spent a year researching 38:34.480 --> 38:37.120 what are the domains it's going to have the biggest low hanging 38:37.120 --> 38:39.400 fruit in the shortest time period. 38:39.400 --> 38:43.880 I picked medicine, but there were so many I could have picked. 38:43.880 --> 38:47.640 And so there was a level of frustration for me of like, OK, 38:47.640 --> 38:50.840 I'm really glad we've opened up the medical deep learning 38:50.840 --> 38:53.880 world and today it's huge, as you know. 38:53.880 --> 38:58.280 But we can't do, you know, I can't do everything. 38:58.280 --> 39:00.400 I don't even know like, like in medicine, 39:00.400 --> 39:02.760 it took me a really long time to even get a sense of like, 39:02.760 --> 39:05.080 what kind of problems do medical practitioners solve? 39:05.080 --> 39:06.400 What kind of data do they have? 39:06.400 --> 39:08.520 Who has that data? 39:08.520 --> 39:12.480 So I kind of felt like I need to approach this differently 39:12.480 --> 39:16.200 if I want to maximize the positive impact of deep learning. 39:16.200 --> 39:19.480 Rather than me picking an area and trying 39:19.480 --> 39:21.720 to become good at it and building something, 39:21.720 --> 39:24.480 I should let people who are already domain experts 39:24.480 --> 39:29.240 in those areas and who already have the data do it themselves. 39:29.240 --> 39:35.520 So that was the reason for vast AI is to basically try 39:35.520 --> 39:38.840 and figure out how to get deep learning 39:38.840 --> 39:41.800 into the hands of people who could benefit from it 39:41.800 --> 39:45.400 and help them to do so in as quick and easy and effective 39:45.400 --> 39:47.080 a way as possible. 39:47.080 --> 39:47.560 Got it. 39:47.560 --> 39:50.240 So sort of empower the domain experts. 39:50.240 --> 39:51.320 Yeah. 39:51.320 --> 39:54.200 And like partly it's because like, 39:54.200 --> 39:56.280 unlike most people in this field, 39:56.280 --> 39:59.960 my background is very applied and industrial. 39:59.960 --> 40:02.480 Like my first job was at McKinsey & Company. 40:02.480 --> 40:04.640 I spent 10 years of management consulting. 40:04.640 --> 40:10.240 I spend a lot of time with domain experts. 40:10.240 --> 40:12.800 You know, so I kind of respect them and appreciate them. 40:12.800 --> 40:16.440 And I know that's where the value generation in society is. 40:16.440 --> 40:21.560 And so I also know how most of them can't code. 40:21.560 --> 40:26.320 And most of them don't have the time to invest, you know, 40:26.320 --> 40:29.320 three years in a graduate degree or whatever. 40:29.320 --> 40:33.520 So it's like, how do I upskill those domain experts? 40:33.520 --> 40:36.080 I think that would be a super powerful thing, 40:36.080 --> 40:40.200 you know, the biggest societal impact I could have. 40:40.200 --> 40:41.680 So yeah, that was the thinking. 40:41.680 --> 40:45.680 So so much of fast AI students and researchers 40:45.680 --> 40:50.120 and the things you teach are programmatically minded, 40:50.120 --> 40:51.520 practically minded, 40:51.520 --> 40:55.840 figuring out ways how to solve real problems and fast. 40:55.840 --> 40:57.480 So from your experience, 40:57.480 --> 41:02.040 what's the difference between theory and practice of deep learning? 41:02.040 --> 41:03.680 Hmm. 41:03.680 --> 41:07.520 Well, most of the research in the deep mining world 41:07.520 --> 41:09.840 is a total waste of time. 41:09.840 --> 41:11.040 Right. That's what I was getting at. 41:11.040 --> 41:12.200 Yeah. 41:12.200 --> 41:16.240 It's it's a problem in science in general. 41:16.240 --> 41:19.600 Scientists need to be published, 41:19.600 --> 41:21.480 which means they need to work on things 41:21.480 --> 41:24.040 that their peers are extremely familiar with 41:24.040 --> 41:26.200 and can recognize in advance in that area. 41:26.200 --> 41:30.040 So that means that they all need to work on the same thing. 41:30.040 --> 41:33.040 And so it really ink and the thing they work on 41:33.040 --> 41:35.640 is nothing to encourage them to work on things 41:35.640 --> 41:38.840 that are practically useful. 41:38.840 --> 41:41.120 So you get just a whole lot of research, 41:41.120 --> 41:43.200 which is minor advances in stuff 41:43.200 --> 41:44.600 that's been very highly studied 41:44.600 --> 41:49.280 and has no significant practical impact. 41:49.280 --> 41:50.840 Whereas the things that really make a difference 41:50.840 --> 41:52.760 like I mentioned transfer learning, 41:52.760 --> 41:55.560 like if we can do better at transfer learning, 41:55.560 --> 41:58.160 then it's this like world changing thing 41:58.160 --> 42:02.880 where suddenly like lots more people can do world class work 42:02.880 --> 42:06.760 with less resources and less data and. 42:06.760 --> 42:08.480 But almost nobody works on that. 42:08.480 --> 42:10.760 Or another example, active learning, 42:10.760 --> 42:11.880 which is the study of like, 42:11.880 --> 42:15.880 how do we get more out of the human beings in the loop? 42:15.880 --> 42:17.120 That's my favorite topic. 42:17.120 --> 42:18.520 Yeah. So active learning is great, 42:18.520 --> 42:21.160 but it's almost nobody working on it 42:21.160 --> 42:23.800 because it's just not a trendy thing right now. 42:23.800 --> 42:27.040 You know what somebody started to interrupt? 42:27.040 --> 42:29.720 He was saying that nobody is publishing 42:29.720 --> 42:31.520 on active learning, right? 42:31.520 --> 42:33.440 But there's people inside companies, 42:33.440 --> 42:36.800 anybody who actually has to solve a problem, 42:36.800 --> 42:39.600 they're going to innovate on active learning. 42:39.600 --> 42:42.080 Yeah. Everybody kind of reinvents active learning 42:42.080 --> 42:43.760 when they actually have to work in practice 42:43.760 --> 42:46.360 because they start labeling things and they think, 42:46.360 --> 42:49.280 gosh, this is taking a long time and it's very expensive. 42:49.280 --> 42:51.200 And then they start thinking, 42:51.200 --> 42:52.640 well, why am I labeling everything? 42:52.640 --> 42:54.840 I'm only, the machine's only making mistakes 42:54.840 --> 42:56.040 on those two classes. 42:56.040 --> 42:56.880 They're the hard ones. 42:56.880 --> 42:58.840 Maybe I'll just start labeling those two classes 42:58.840 --> 43:00.360 and then you start thinking, 43:00.360 --> 43:01.560 well, why did I do that manually? 43:01.560 --> 43:03.000 Why can't I just get the system to tell me 43:03.000 --> 43:04.760 which things are going to be harder steps? 43:04.760 --> 43:06.200 It's an obvious thing to do. 43:06.200 --> 43:11.400 But yeah, it's just like transfer learning. 43:11.400 --> 43:14.120 It's understudied and the academic world 43:14.120 --> 43:17.440 just has no reason to care about practical results. 43:17.440 --> 43:18.360 The funny thing is, like, 43:18.360 --> 43:19.920 I've only really ever written one paper. 43:19.920 --> 43:21.520 I hate writing papers. 43:21.520 --> 43:22.760 And I didn't even write it. 43:22.760 --> 43:25.480 It was my colleague, Sebastian Ruder, who actually wrote it. 43:25.480 --> 43:28.040 I just did the research for it. 43:28.040 --> 43:31.640 But it was basically introducing successful transfer learning 43:31.640 --> 43:34.200 to NLP for the first time. 43:34.200 --> 43:37.000 And the algorithm is called ULMfit. 43:37.000 --> 43:42.320 And I actually wrote it for the course, 43:42.320 --> 43:43.720 for the first day of course. 43:43.720 --> 43:45.360 I wanted to teach people NLP. 43:45.360 --> 43:47.520 And I thought I only want to teach people practical stuff. 43:47.520 --> 43:50.560 And I think the only practical stuff is transfer learning. 43:50.560 --> 43:53.360 And I couldn't find any examples of transfer learning in NLP. 43:53.360 --> 43:54.560 So I just did it. 43:54.560 --> 43:57.320 And I was shocked to find that as soon as I did it, 43:57.320 --> 44:01.080 which, you know, the basic prototype took a couple of days, 44:01.080 --> 44:02.520 smashed the state of the art 44:02.520 --> 44:04.760 on one of the most important data sets in a field 44:04.760 --> 44:06.720 that I knew nothing about. 44:06.720 --> 44:10.400 And I just thought, well, this is ridiculous. 44:10.400 --> 44:13.800 And so I spoke to Sebastian about it. 44:13.800 --> 44:17.680 And he kindly offered to write it up the results. 44:17.680 --> 44:21.360 And so it ended up being published in ACL, 44:21.360 --> 44:25.560 which is the top computational linguistics conference. 44:25.560 --> 44:28.880 So like, people do actually care once you do it. 44:28.880 --> 44:34.160 But I guess it's difficult for maybe junior researchers. 44:34.160 --> 44:37.720 I don't care whether I get citations or papers or whatever. 44:37.720 --> 44:39.640 There's nothing in my life that makes that important, 44:39.640 --> 44:41.240 which is why I've never actually 44:41.240 --> 44:43.040 bothered to write a paper myself. 44:43.040 --> 44:44.400 But for people who do, I guess they 44:44.400 --> 44:50.960 have to pick the kind of safe option, which is like, 44:50.960 --> 44:52.720 yeah, make a slight improvement on something 44:52.720 --> 44:55.160 that everybody's already working on. 44:55.160 --> 44:59.040 Yeah, nobody does anything interesting or succeeds 44:59.040 --> 45:01.240 in life with the safe option. 45:01.240 --> 45:02.960 Well, I mean, the nice thing is nowadays, 45:02.960 --> 45:05.320 everybody is now working on NLP transfer learning. 45:05.320 --> 45:12.240 Because since that time, we've had GPT and GPT2 and BERT. 45:12.240 --> 45:15.400 So yeah, once you show that something's possible, 45:15.400 --> 45:17.680 everybody jumps in, I guess. 45:17.680 --> 45:19.320 I hope to be a part of it. 45:19.320 --> 45:21.600 I hope to see more innovation and active learning 45:21.600 --> 45:22.160 in the same way. 45:22.160 --> 45:24.560 I think transfer learning and active learning 45:24.560 --> 45:27.360 are a fascinating public open work. 45:27.360 --> 45:30.160 I actually helped start a startup called Platform AI, which 45:30.160 --> 45:31.760 is really all about active learning. 45:31.760 --> 45:34.200 And yeah, it's been interesting trying 45:34.200 --> 45:36.920 to kind of see what research is out there 45:36.920 --> 45:37.800 and make the most of it. 45:37.800 --> 45:39.200 And there's basically none. 45:39.200 --> 45:41.040 So we've had to do all our own research. 45:41.040 --> 45:44.240 Once again, and just as you described, 45:44.240 --> 45:47.640 can you tell the story of the Stanford competition, 45:47.640 --> 45:51.520 Dawn Bench, and fast AI's achievement on it? 45:51.520 --> 45:51.960 Sure. 45:51.960 --> 45:55.560 So something which I really enjoy is that I basically 45:55.560 --> 45:59.000 teach two courses a year, the practical deep learning 45:59.000 --> 46:02.120 for coders, which is kind of the introductory course, 46:02.120 --> 46:04.280 and then cutting edge deep learning for coders, which 46:04.280 --> 46:08.080 is the kind of research level course. 46:08.080 --> 46:14.320 And while I teach those courses, I basically 46:14.320 --> 46:18.440 have a big office at the University of San Francisco. 46:18.440 --> 46:19.800 It'd be enough for like 30 people. 46:19.800 --> 46:22.960 And I invite any student who wants to come and hang out 46:22.960 --> 46:25.320 with me while I build the course. 46:25.320 --> 46:26.640 And so generally, it's full. 46:26.640 --> 46:30.880 And so we have 20 or 30 people in a big office 46:30.880 --> 46:33.880 with nothing to do but study deep learning. 46:33.880 --> 46:35.880 So it was during one of these times 46:35.880 --> 46:38.640 that somebody in the group said, oh, there's 46:38.640 --> 46:41.480 a thing called Dawn Bench that looks interesting. 46:41.480 --> 46:42.800 And I say, what the hell is that? 46:42.800 --> 46:44.120 I'm going to set out some competition 46:44.120 --> 46:46.440 to see how quickly you can train a model. 46:46.440 --> 46:50.080 It seems kind of not exactly relevant to what we're doing, 46:50.080 --> 46:51.440 but it sounds like the kind of thing 46:51.440 --> 46:52.480 which you might be interested in. 46:52.480 --> 46:53.960 And I checked it out and I was like, oh, crap. 46:53.960 --> 46:55.840 There's only 10 days till it's over. 46:55.840 --> 46:58.120 It's pretty much too late. 46:58.120 --> 47:01.000 And we're kind of busy trying to teach this course. 47:01.000 --> 47:05.640 But we're like, oh, it would make an interesting case study 47:05.640 --> 47:08.200 for the course like it's all the stuff we're already doing. 47:08.200 --> 47:11.120 Why don't we just put together our current best practices 47:11.120 --> 47:12.480 and ideas. 47:12.480 --> 47:16.880 So me and I guess about four students just decided 47:16.880 --> 47:17.560 to give it a go. 47:17.560 --> 47:19.880 And we focused on this small one called 47:19.880 --> 47:24.640 SciFar 10, which is little 32 by 32 pixel images. 47:24.640 --> 47:26.160 Can you say what Dawn Bench is? 47:26.160 --> 47:29.560 Yeah, so it's a competition to train a model as fast as possible. 47:29.560 --> 47:31.000 It was run by Stanford. 47:31.000 --> 47:32.480 And as cheap as possible, too. 47:32.480 --> 47:34.320 That's also another one for as cheap as possible. 47:34.320 --> 47:38.160 And there's a couple of categories, ImageNet and SciFar 10. 47:38.160 --> 47:42.080 So ImageNet's this big 1.3 million image thing 47:42.080 --> 47:45.400 that took a couple of days to train. 47:45.400 --> 47:51.240 I remember a friend of mine, Pete Warden, who's now at Google. 47:51.240 --> 47:53.760 I remember he told me how he trained ImageNet a few years 47:53.760 --> 47:59.440 ago when he basically had this little granny flat out 47:59.440 --> 48:01.920 the back that he turned into was ImageNet training center. 48:01.920 --> 48:04.240 And after a year of work, he figured out 48:04.240 --> 48:07.040 how to train it in 10 days or something. 48:07.040 --> 48:08.480 It's like that was a big job. 48:08.480 --> 48:10.640 Whereas SciFar 10, at that time, you 48:10.640 --> 48:13.040 could train in a few hours. 48:13.040 --> 48:14.520 It's much smaller and easier. 48:14.520 --> 48:18.120 So we thought we'd try SciFar 10. 48:18.120 --> 48:23.800 And yeah, I've really never done that before. 48:23.800 --> 48:27.920 Like, things like using more than one GPU at a time 48:27.920 --> 48:29.800 was something I tried to avoid. 48:29.800 --> 48:32.160 Because to me, it's very against the whole idea 48:32.160 --> 48:35.080 of accessibility, is she better do things with one GPU? 48:35.080 --> 48:36.480 I mean, have you asked in the past 48:36.480 --> 48:39.680 before, after having accomplished something, 48:39.680 --> 48:42.520 how do I do this faster, much faster? 48:42.520 --> 48:43.240 Oh, always. 48:43.240 --> 48:44.680 But it's always, for me, it's always, 48:44.680 --> 48:47.640 how do I make it much faster on a single GPU 48:47.640 --> 48:50.400 that a normal person could afford in their day to day life? 48:50.400 --> 48:54.760 It's not, how could I do it faster by having a huge data 48:54.760 --> 48:55.280 center? 48:55.280 --> 48:57.240 Because to me, it's all about, like, 48:57.240 --> 48:59.560 as many people should be able to use something as possible 48:59.560 --> 49:04.160 without fussing around with infrastructure. 49:04.160 --> 49:06.080 So anyway, so in this case, it's like, well, 49:06.080 --> 49:10.240 we can use 8GPUs just by renting a AWS machine. 49:10.240 --> 49:11.920 So we thought we'd try that. 49:11.920 --> 49:16.560 And yeah, basically, using the stuff we were already doing, 49:16.560 --> 49:20.360 we were able to get the speed. 49:20.360 --> 49:25.360 Within a few days, we had the speed down to a very small 49:25.360 --> 49:26.040 number of minutes. 49:26.040 --> 49:28.800 I can't remember exactly how many minutes it was, 49:28.800 --> 49:31.440 but it might have been like 10 minutes or something. 49:31.440 --> 49:34.200 And so yeah, we found ourselves at the top of the leaderboard 49:34.200 --> 49:38.720 easily for both time and money, which really shocked me. 49:38.720 --> 49:40.160 Because the other people competing in this 49:40.160 --> 49:41.880 were like Google and Intel and stuff, 49:41.880 --> 49:45.360 where I know a lot more about this stuff than I think we do. 49:45.360 --> 49:46.800 So then we emboldened. 49:46.800 --> 49:50.640 We thought, let's try the ImageNet one too. 49:50.640 --> 49:53.280 I mean, it seemed way out of our league. 49:53.280 --> 49:57.120 But our goal was to get under 12 hours. 49:57.120 --> 49:59.280 And we did, which was really exciting. 49:59.280 --> 50:01.440 And we didn't put anything up on the leaderboard, 50:01.440 --> 50:03.080 but we were down to like 10 hours. 50:03.080 --> 50:10.000 But then Google put in like five hours or something, 50:10.000 --> 50:13.360 and we're just like, oh, we're so screwed. 50:13.360 --> 50:16.880 But we kind of thought, well, keep trying. 50:16.880 --> 50:17.880 If Google can do it in five hours. 50:17.880 --> 50:20.760 I mean, Google did it on five hours on like a TPU pod 50:20.760 --> 50:24.280 or something, like a lot of hardware. 50:24.280 --> 50:26.360 But we kind of like had a bunch of ideas to try. 50:26.360 --> 50:28.920 Like a really simple thing was, why 50:28.920 --> 50:30.480 are we using these big images? 50:30.480 --> 50:36.280 They're like 224, 256 by 256 pixels. 50:36.280 --> 50:37.640 Why don't we try smaller ones? 50:37.640 --> 50:41.360 And just to elaborate, there's a constraint on the accuracy 50:41.360 --> 50:43.080 that your train model is supposed to achieve. 50:43.080 --> 50:45.760 Yeah, you've got to achieve 93%. 50:45.760 --> 50:47.640 I think it was for ImageNet. 50:47.640 --> 50:49.160 Exactly. 50:49.160 --> 50:50.240 Which is very tough. 50:50.240 --> 50:51.240 So you have to repeat that. 50:51.240 --> 50:52.120 Yeah, 93%. 50:52.120 --> 50:54.680 Like they picked a good threshold. 50:54.680 --> 50:58.920 It was a little bit higher than what the most commonly used 50:58.920 --> 51:03.320 ResNet 50 model could achieve at that time. 51:03.320 --> 51:08.080 So yeah, so it's quite a difficult problem to solve. 51:08.080 --> 51:09.920 But yeah, we realized if we actually just 51:09.920 --> 51:16.200 use 64 by 64 images, it trained a pretty good model. 51:16.200 --> 51:17.960 And then we could take that same model 51:17.960 --> 51:19.560 and just give it a couple of epochs 51:19.560 --> 51:21.880 to learn 224 by 224 images. 51:21.880 --> 51:24.440 And it was basically already trained. 51:24.440 --> 51:25.480 It makes a lot of sense. 51:25.480 --> 51:27.200 Like if you teach somebody, like here's 51:27.200 --> 51:30.240 what a dog looks like, and you show them low res versions, 51:30.240 --> 51:33.640 and then you say, here's a really clear picture of a dog. 51:33.640 --> 51:36.000 They already know what a dog looks like. 51:36.000 --> 51:39.920 So that, like, just we jumped to the front, 51:39.920 --> 51:46.400 and we ended up winning parts of that competition. 51:46.400 --> 51:49.680 We actually ended up doing a distributed version 51:49.680 --> 51:51.960 over multiple machines a couple of months later 51:51.960 --> 51:53.560 and ended up at the top of the leaderboard. 51:53.560 --> 51:55.440 We had 18 minutes. 51:55.440 --> 51:56.280 ImageNet. 51:56.280 --> 52:00.560 Yeah, and people have just kept on blasting through again 52:00.560 --> 52:02.320 and again since then. 52:02.320 --> 52:06.760 So what's your view on multi GPU or multiple machine 52:06.760 --> 52:11.960 training in general as a way to speed code up? 52:11.960 --> 52:13.680 I think it's largely a waste of time. 52:13.680 --> 52:15.880 Both multi GPU on a single machine and? 52:15.880 --> 52:17.640 Yeah, particularly multi machines, 52:17.640 --> 52:18.880 because it's just clunky. 52:21.840 --> 52:25.320 Multi GPUs is less clunky than it used to be. 52:25.320 --> 52:28.520 But to me, anything that slows down your iteration speed 52:28.520 --> 52:31.800 is a waste of time. 52:31.800 --> 52:36.960 So you could maybe do your very last perfecting of the model 52:36.960 --> 52:38.960 on multi GPUs if you need to. 52:38.960 --> 52:44.560 But so for example, I think doing stuff on ImageNet 52:44.560 --> 52:46.000 is generally a waste of time. 52:46.000 --> 52:48.240 Why test things on 1.3 million images? 52:48.240 --> 52:51.040 Most of us don't use 1.3 million images. 52:51.040 --> 52:54.360 And we've also done research that shows that doing things 52:54.360 --> 52:56.840 on a smaller subset of images gives you 52:56.840 --> 52:59.280 the same relative answers anyway. 52:59.280 --> 53:02.120 So from a research point of view, why waste that time? 53:02.120 --> 53:06.200 So actually, I released a couple of new data sets recently. 53:06.200 --> 53:08.880 One is called ImageNet. 53:08.880 --> 53:12.920 The French ImageNet, which is a small subset of ImageNet, 53:12.920 --> 53:15.200 which is designed to be easy to classify. 53:15.200 --> 53:17.320 What's how do you spell ImageNet? 53:17.320 --> 53:19.200 It's got an extra T and E at the end, 53:19.200 --> 53:20.520 because it's very French. 53:20.520 --> 53:21.640 Image, OK. 53:21.640 --> 53:24.720 And then another one called ImageWolf, 53:24.720 --> 53:29.840 which is a subset of ImageNet that only contains dog breeds. 53:29.840 --> 53:31.120 But that's a hard one, right? 53:31.120 --> 53:32.000 That's a hard one. 53:32.000 --> 53:34.360 And I've discovered that if you just look at these two 53:34.360 --> 53:39.120 subsets, you can train things on a single GPU in 10 minutes. 53:39.120 --> 53:42.040 And the results you get are directly transferrable 53:42.040 --> 53:44.320 to ImageNet nearly all the time. 53:44.320 --> 53:46.600 And so now I'm starting to see some researchers start 53:46.600 --> 53:48.960 to use these smaller data sets. 53:48.960 --> 53:51.120 I so deeply love the way you think, 53:51.120 --> 53:57.000 because I think you might have written a blog post saying 53:57.000 --> 54:00.200 that going with these big data sets 54:00.200 --> 54:03.920 is encouraging people to not think creatively. 54:03.920 --> 54:04.560 Absolutely. 54:04.560 --> 54:08.320 So year two, it sort of constrains you 54:08.320 --> 54:09.840 to train on large resources. 54:09.840 --> 54:11.280 And because you have these resources, 54:11.280 --> 54:14.040 you think more research will be better. 54:14.040 --> 54:17.760 And then you start to like somehow you kill the creativity. 54:17.760 --> 54:18.000 Yeah. 54:18.000 --> 54:20.760 And even worse than that, Lex, I keep hearing from people 54:20.760 --> 54:23.480 who say, I decided not to get into deep learning 54:23.480 --> 54:26.080 because I don't believe it's accessible to people 54:26.080 --> 54:28.560 outside of Google to do useful work. 54:28.560 --> 54:31.640 So like I see a lot of people make an explicit decision 54:31.640 --> 54:36.000 to not learn this incredibly valuable tool 54:36.000 --> 54:39.840 because they've drunk the Google Kool Aid, which is that only 54:39.840 --> 54:42.440 Google's big enough and smart enough to do it. 54:42.440 --> 54:45.400 And I just find that so disappointing and it's so wrong. 54:45.400 --> 54:49.200 And I think all of the major breakthroughs in AI 54:49.200 --> 54:53.280 in the next 20 years will be doable on a single GPU. 54:53.280 --> 54:57.120 Like I would say, my sense is all the big sort of. 54:57.120 --> 54:58.200 Well, let's put it this way. 54:58.200 --> 55:00.200 None of the big breakthroughs of the last 20 years 55:00.200 --> 55:01.720 have required multiple GPUs. 55:01.720 --> 55:05.920 So like batch norm, value, dropout, 55:05.920 --> 55:08.080 to demonstrate that there's something to them. 55:08.080 --> 55:11.840 Every one of them, none of them has required multiple GPUs. 55:11.840 --> 55:15.800 GANs, the original GANs, didn't require multiple GPUs. 55:15.800 --> 55:18.040 Well, and we've actually recently shown 55:18.040 --> 55:19.680 that you don't even need GANs. 55:19.680 --> 55:23.360 So we've developed GAN level outcomes 55:23.360 --> 55:24.720 without needing GANs. 55:24.720 --> 55:26.880 And we can now do it with, again, 55:26.880 --> 55:29.680 by using transfer learning, we can do it in a couple of hours 55:29.680 --> 55:30.520 on a single GPU. 55:30.520 --> 55:31.600 So you're using a generator model 55:31.600 --> 55:32.960 without the adversarial part? 55:32.960 --> 55:33.440 Yeah. 55:33.440 --> 55:35.880 So we've found loss functions that 55:35.880 --> 55:38.680 work super well without the adversarial part. 55:38.680 --> 55:41.840 And then one of our students, a guy called Jason Antich, 55:41.840 --> 55:44.640 has created a system called Dealtify, 55:44.640 --> 55:47.280 which uses this technique to colorize 55:47.280 --> 55:48.840 old black and white movies. 55:48.840 --> 55:51.480 You can do it on a single GPU, colorize a whole movie 55:51.480 --> 55:52.920 in a couple of hours. 55:52.920 --> 55:56.080 And one of the things that Jason and I did together 55:56.080 --> 56:00.480 was we figured out how to add a little bit of GAN 56:00.480 --> 56:03.000 at the very end, which it turns out for colorization, 56:03.000 --> 56:06.000 makes it just a bit brighter and nicer. 56:06.000 --> 56:07.920 And then Jason did masses of experiments 56:07.920 --> 56:10.000 to figure out exactly how much to do. 56:10.000 --> 56:12.840 But it's still all done on his home machine, 56:12.840 --> 56:15.400 on a single GPU in his lounge room. 56:15.400 --> 56:19.200 And if you think about colorizing Hollywood movies, 56:19.200 --> 56:21.720 that sounds like something a huge studio would have to do. 56:21.720 --> 56:25.280 But he has the world's best results on this. 56:25.280 --> 56:27.040 There's this problem of microphones. 56:27.040 --> 56:28.640 We're just talking to microphones now. 56:28.640 --> 56:29.140 Yeah. 56:29.140 --> 56:32.520 It's such a pain in the ass to have these microphones 56:32.520 --> 56:34.440 to get good quality audio. 56:34.440 --> 56:36.720 And I tried to see if it's possible to plop down 56:36.720 --> 56:39.960 a bunch of cheap sensors and reconstruct higher quality 56:39.960 --> 56:41.840 audio from multiple sources. 56:41.840 --> 56:45.440 Because right now, I haven't seen work from, OK, 56:45.440 --> 56:48.760 we can save inexpensive mics, automatically combining 56:48.760 --> 56:52.280 audio from multiple sources to improve the combined audio. 56:52.280 --> 56:53.200 People haven't done that. 56:53.200 --> 56:55.080 And that feels like a learning problem. 56:55.080 --> 56:56.800 So hopefully somebody can. 56:56.800 --> 56:58.760 Well, I mean, it's evidently doable. 56:58.760 --> 57:01.000 And it should have been done by now. 57:01.000 --> 57:03.640 I felt the same way about computational photography 57:03.640 --> 57:04.480 four years ago. 57:04.480 --> 57:05.240 That's right. 57:05.240 --> 57:08.240 Why are we investing in big lenses when 57:08.240 --> 57:13.160 three cheap lenses plus actually a little bit of intentional 57:13.160 --> 57:16.640 movement, so like take a few frames, 57:16.640 --> 57:19.840 gives you enough information to get excellent subpixel 57:19.840 --> 57:22.440 resolution, which particularly with deep learning, 57:22.440 --> 57:25.840 you would know exactly what you meant to be looking at. 57:25.840 --> 57:28.200 We can totally do the same thing with audio. 57:28.200 --> 57:30.720 I think the madness that it hasn't been done yet. 57:30.720 --> 57:33.320 Has there been progress on photography companies? 57:33.320 --> 57:33.820 Yeah. 57:33.820 --> 57:36.720 Photography is basically a standard now. 57:36.720 --> 57:41.120 So the Google Pixel Nightlight, I 57:41.120 --> 57:43.240 don't know if you've ever tried it, but it's astonishing. 57:43.240 --> 57:45.440 You take a picture and almost pitch black 57:45.440 --> 57:49.120 and you get back a very high quality image. 57:49.120 --> 57:51.440 And it's not because of the lens. 57:51.440 --> 57:55.280 Same stuff with like adding the bokeh to the background 57:55.280 --> 57:55.800 blurring. 57:55.800 --> 57:57.200 It's done computationally. 57:57.200 --> 57:58.520 Just the pics over here. 57:58.520 --> 57:59.020 Yeah. 57:59.020 --> 58:05.000 Basically, everybody now is doing most of the fanciest stuff 58:05.000 --> 58:07.120 on their phones with computational photography 58:07.120 --> 58:10.640 and also increasingly, people are putting more than one lens 58:10.640 --> 58:11.840 on the back of the camera. 58:11.840 --> 58:14.360 So the same will happen for audio, for sure. 58:14.360 --> 58:16.520 And there's applications in the audio side. 58:16.520 --> 58:19.360 If you look at an Alexa type device, 58:19.360 --> 58:21.840 most people I've seen, especially I worked at Google 58:21.840 --> 58:26.000 before, when you look at noise background removal, 58:26.000 --> 58:29.480 you don't think of multiple sources of audio. 58:29.480 --> 58:31.920 You don't play with that as much as I would hope people would. 58:31.920 --> 58:33.640 But I mean, you can still do it even with one. 58:33.640 --> 58:36.120 Like, again, it's not much work's been done in this area. 58:36.120 --> 58:38.440 So we're actually going to be releasing an audio library 58:38.440 --> 58:41.040 soon, which hopefully will encourage development of this 58:41.040 --> 58:43.200 because it's so underused. 58:43.200 --> 58:46.480 The basic approach we used for our super resolution, 58:46.480 --> 58:49.960 in which Jason uses for dealdify of generating 58:49.960 --> 58:51.920 high quality images, the exact same approach 58:51.920 --> 58:53.480 would work for audio. 58:53.480 --> 58:57.160 No one's done it yet, but it would be a couple of months work. 58:57.160 --> 59:01.600 OK, also learning rate in terms of dawn bench. 59:01.600 --> 59:04.280 There's some magic on learning rate that you played around 59:04.280 --> 59:04.760 with. 59:04.760 --> 59:05.800 It's kind of interesting. 59:05.800 --> 59:08.120 Yeah, so this is all work that came from a guy called Leslie 59:08.120 --> 59:09.360 Smith. 59:09.360 --> 59:12.760 Leslie's a researcher who, like us, 59:12.760 --> 59:17.720 cares a lot about just the practicalities of training 59:17.720 --> 59:20.000 neural networks quickly and accurately, 59:20.000 --> 59:22.120 which you would think is what everybody should care about, 59:22.120 --> 59:25.000 but almost nobody does. 59:25.000 --> 59:28.120 And he discovered something very interesting, 59:28.120 --> 59:30.000 which he calls super convergence, which 59:30.000 --> 59:32.360 is there are certain networks that with certain settings 59:32.360 --> 59:34.320 of high parameters could suddenly 59:34.320 --> 59:37.440 be trained 10 times faster by using 59:37.440 --> 59:39.480 a 10 times higher learning rate. 59:39.480 --> 59:44.680 Now, no one published that paper 59:44.680 --> 59:49.520 because it's not an area of active research 59:49.520 --> 59:50.440 in the academic world. 59:50.440 --> 59:52.840 No academics recognize this is important. 59:52.840 --> 59:56.080 And also, deep learning in academia 59:56.080 --> 1:00:00.040 is not considered a experimental science. 1:00:00.040 --> 1:00:02.440 So unlike in physics, where you could say, 1:00:02.440 --> 1:00:05.360 I just saw a subatomic particle do something 1:00:05.360 --> 1:00:07.240 which the theory doesn't explain, 1:00:07.240 --> 1:00:10.440 you could publish that without an explanation. 1:00:10.440 --> 1:00:12.120 And then in the next 60 years, people 1:00:12.120 --> 1:00:14.120 can try to work out how to explain it. 1:00:14.120 --> 1:00:16.200 We don't allow this in the deep learning world. 1:00:16.200 --> 1:00:20.720 So it's literally impossible for Leslie to publish a paper that 1:00:20.720 --> 1:00:23.560 says, I've just seen something amazing happen. 1:00:23.560 --> 1:00:25.680 This thing trained 10 times faster than it should have. 1:00:25.680 --> 1:00:27.080 I don't know why. 1:00:27.080 --> 1:00:28.600 And so the reviewers were like, well, 1:00:28.600 --> 1:00:30.280 you can't publish that because you don't know why. 1:00:30.280 --> 1:00:31.000 So anyway. 1:00:31.000 --> 1:00:32.680 That's important to pause on because there's 1:00:32.680 --> 1:00:36.160 so many discoveries that would need to start like that. 1:00:36.160 --> 1:00:39.280 Every other scientific field I know of works of that way. 1:00:39.280 --> 1:00:42.520 I don't know why ours is uniquely 1:00:42.520 --> 1:00:46.480 disinterested in publishing unexplained 1:00:46.480 --> 1:00:47.680 experimental results. 1:00:47.680 --> 1:00:48.680 But there it is. 1:00:48.680 --> 1:00:51.200 So it wasn't published. 1:00:51.200 --> 1:00:55.080 Having said that, I read a lot more 1:00:55.080 --> 1:00:56.840 unpublished papers and published papers 1:00:56.840 --> 1:01:00.080 because that's where you find the interesting insights. 1:01:00.080 --> 1:01:02.680 So I absolutely read this paper. 1:01:02.680 --> 1:01:08.120 And I was just like, this is astonishingly mind blowing 1:01:08.120 --> 1:01:09.760 and weird and awesome. 1:01:09.760 --> 1:01:12.400 And why isn't everybody only talking about this? 1:01:12.400 --> 1:01:15.520 Because if you can train these things 10 times faster, 1:01:15.520 --> 1:01:18.480 they also generalize better because you're doing less epochs, 1:01:18.480 --> 1:01:20.080 which means you look at the data less, 1:01:20.080 --> 1:01:22.400 you get better accuracy. 1:01:22.400 --> 1:01:24.640 So I've been kind of studying that ever since. 1:01:24.640 --> 1:01:28.520 And eventually Leslie kind of figured out 1:01:28.520 --> 1:01:30.160 a lot of how to get this done. 1:01:30.160 --> 1:01:32.280 And we added minor tweaks. 1:01:32.280 --> 1:01:34.840 And a big part of the trick is starting 1:01:34.840 --> 1:01:37.920 at a very low learning rate, very gradually increasing it. 1:01:37.920 --> 1:01:39.800 So as you're training your model, 1:01:39.800 --> 1:01:42.120 you take very small steps at the start. 1:01:42.120 --> 1:01:44.080 And you gradually make them bigger and bigger 1:01:44.080 --> 1:01:46.440 until eventually you're taking much bigger steps 1:01:46.440 --> 1:01:49.400 than anybody thought was possible. 1:01:49.400 --> 1:01:52.280 There's a few other little tricks to make it work. 1:01:52.280 --> 1:01:55.240 Basically, we can reliably get super convergence. 1:01:55.240 --> 1:01:56.640 And so for the dorm bench thing, 1:01:56.640 --> 1:01:59.320 we were using just much higher learning rates 1:01:59.320 --> 1:02:02.200 than people expected to work. 1:02:02.200 --> 1:02:03.880 What do you think the future of, 1:02:03.880 --> 1:02:05.200 I mean, it makes so much sense for that 1:02:05.200 --> 1:02:08.640 to be a critical hyperparameter learning rate that you vary. 1:02:08.640 --> 1:02:13.480 What do you think the future of learning rate magic looks like? 1:02:13.480 --> 1:02:14.960 Well, there's been a lot of great work 1:02:14.960 --> 1:02:17.400 in the last 12 months in this area. 1:02:17.400 --> 1:02:20.800 And people are increasingly realizing that we just 1:02:20.800 --> 1:02:23.120 have no idea really how optimizers work. 1:02:23.120 --> 1:02:25.840 And the combination of weight decay, 1:02:25.840 --> 1:02:27.480 which is how we regularize optimizers, 1:02:27.480 --> 1:02:30.120 and the learning rate, and then other things 1:02:30.120 --> 1:02:32.760 like the epsilon we use in the atom optimizer, 1:02:32.760 --> 1:02:36.560 they all work together in weird ways. 1:02:36.560 --> 1:02:38.560 And different parts of the model, 1:02:38.560 --> 1:02:40.480 this is another thing we've done a lot of work on, 1:02:40.480 --> 1:02:43.480 is research into how different parts of the model 1:02:43.480 --> 1:02:46.600 should be trained at different rates in different ways. 1:02:46.600 --> 1:02:49.040 So we do something we call discriminative learning rates, 1:02:49.040 --> 1:02:51.040 which is really important, particularly for transfer 1:02:51.040 --> 1:02:53.200 learning. 1:02:53.200 --> 1:02:54.880 So really, I think in the last 12 months, 1:02:54.880 --> 1:02:57.360 a lot of people have realized that all this stuff is important. 1:02:57.360 --> 1:03:00.000 There's been a lot of great work coming out. 1:03:00.000 --> 1:03:02.880 And we're starting to see algorithms 1:03:02.880 --> 1:03:06.880 appear which have very, very few dials, if any, 1:03:06.880 --> 1:03:07.920 that you have to touch. 1:03:07.920 --> 1:03:09.240 So I think what's going to happen 1:03:09.240 --> 1:03:10.840 is the idea of a learning rate, well, 1:03:10.840 --> 1:03:14.360 it almost already has disappeared in the latest research. 1:03:14.360 --> 1:03:18.240 And instead, it's just like, we know enough 1:03:18.240 --> 1:03:22.440 about how to interpret the gradients 1:03:22.440 --> 1:03:23.840 and the change of gradients we see 1:03:23.840 --> 1:03:25.440 to know how to set every parameter of our way. 1:03:25.440 --> 1:03:26.440 There you can automate it. 1:03:26.440 --> 1:03:31.720 So you see the future of deep learning, where really, 1:03:31.720 --> 1:03:34.600 where is the input of a human expert needed? 1:03:34.600 --> 1:03:36.520 Well, hopefully, the input of a human expert 1:03:36.520 --> 1:03:39.680 will be almost entirely unneeded from the deep learning 1:03:39.680 --> 1:03:40.560 point of view. 1:03:40.560 --> 1:03:43.480 So again, Google's approach to this 1:03:43.480 --> 1:03:46.000 is to try and use thousands of times more compute 1:03:46.000 --> 1:03:49.400 to run lots and lots of models at the same time 1:03:49.400 --> 1:03:51.040 and hope that one of them is good. 1:03:51.040 --> 1:03:51.960 A lot of malkana stuff. 1:03:51.960 --> 1:03:56.800 Yeah, a lot of malkana stuff, which I think is insane. 1:03:56.800 --> 1:04:01.720 When you better understand the mechanics of how models learn, 1:04:01.720 --> 1:04:03.800 you don't have to try 1,000 different models 1:04:03.800 --> 1:04:05.680 to find which one happens to work the best. 1:04:05.680 --> 1:04:08.240 You can just jump straight to the best one, which 1:04:08.240 --> 1:04:12.720 means that it's more accessible in terms of compute, cheaper, 1:04:12.720 --> 1:04:14.920 and also with less hyperparameters to set. 1:04:14.920 --> 1:04:16.800 That means you don't need deep learning experts 1:04:16.800 --> 1:04:19.360 to train your deep learning model for you, 1:04:19.360 --> 1:04:22.480 which means that domain experts can do more of the work, which 1:04:22.480 --> 1:04:24.960 means that now you can focus the human time 1:04:24.960 --> 1:04:28.320 on the kind of interpretation, the data gathering, 1:04:28.320 --> 1:04:31.360 identifying model errors, and stuff like that. 1:04:31.360 --> 1:04:32.840 Yeah, the data side. 1:04:32.840 --> 1:04:34.720 How often do you work with data these days 1:04:34.720 --> 1:04:38.680 in terms of the cleaning, Darwin looked 1:04:38.680 --> 1:04:43.120 at different species while traveling about, 1:04:43.120 --> 1:04:45.040 do you look at data? 1:04:45.040 --> 1:04:49.400 Have you, in your roots in Kaggle, just look at data? 1:04:49.400 --> 1:04:51.320 Yeah, I mean, it's a key part of our course. 1:04:51.320 --> 1:04:53.480 It's like before we train a model in the course, 1:04:53.480 --> 1:04:55.160 we see how to look at the data. 1:04:55.160 --> 1:04:57.920 And then the first thing we do after we train our first model, 1:04:57.920 --> 1:05:00.520 which we fine tune an ImageNet model for five minutes. 1:05:00.520 --> 1:05:02.240 And then the thing we immediately do after that 1:05:02.240 --> 1:05:05.760 is we learn how to analyze the results of the model 1:05:05.760 --> 1:05:08.920 by looking at examples of misclassified images, 1:05:08.920 --> 1:05:10.880 and looking at a classification matrix, 1:05:10.880 --> 1:05:15.080 and then doing research on Google 1:05:15.080 --> 1:05:18.000 to learn about the kinds of things that it's misclassifying. 1:05:18.000 --> 1:05:19.520 So to me, one of the really cool things 1:05:19.520 --> 1:05:21.840 about machine learning models in general 1:05:21.840 --> 1:05:24.480 is that when you interpret them, they 1:05:24.480 --> 1:05:27.360 tell you about things like what are the most important features, 1:05:27.360 --> 1:05:29.400 which groups you're misclassifying, 1:05:29.400 --> 1:05:32.440 and they help you become a domain expert more quickly, 1:05:32.440 --> 1:05:34.880 because you can focus your time on the bits 1:05:34.880 --> 1:05:38.680 that the model is telling you is important. 1:05:38.680 --> 1:05:40.760 So it lets you deal with things like data leakage, 1:05:40.760 --> 1:05:43.080 for example, if it says, oh, the main feature I'm looking at 1:05:43.080 --> 1:05:45.240 is customer ID. 1:05:45.240 --> 1:05:47.640 And you're like, oh, customer ID should be predictive. 1:05:47.640 --> 1:05:52.280 And then you can talk to the people that manage customer IDs, 1:05:52.280 --> 1:05:56.840 and they'll tell you, oh, yes, as soon as a customer's application 1:05:56.840 --> 1:05:59.480 is accepted, we add a one on the end of their customer ID 1:05:59.480 --> 1:06:01.200 or something. 1:06:01.200 --> 1:06:04.360 So yeah, looking at data, particularly 1:06:04.360 --> 1:06:06.600 from the lens of which parts of the data the model says 1:06:06.600 --> 1:06:09.400 is important, is super important. 1:06:09.400 --> 1:06:11.480 Yeah, and using kind of using the model 1:06:11.480 --> 1:06:14.240 to almost debug the data to learn more about the data. 1:06:14.240 --> 1:06:16.800 Exactly. 1:06:16.800 --> 1:06:18.600 What are the different cloud options 1:06:18.600 --> 1:06:20.160 for training your networks? 1:06:20.160 --> 1:06:22.000 Last question related to Don Bench. 1:06:22.000 --> 1:06:24.240 Well, it's part of a lot of the work you do, 1:06:24.240 --> 1:06:27.280 but from a perspective of performance, 1:06:27.280 --> 1:06:29.480 I think you've written this in a blog post. 1:06:29.480 --> 1:06:32.720 There's AWS, there's a TPU from Google. 1:06:32.720 --> 1:06:33.440 What's your sense? 1:06:33.440 --> 1:06:34.520 What the future holds? 1:06:34.520 --> 1:06:37.360 What would you recommend now in terms of training in the cloud? 1:06:37.360 --> 1:06:40.520 So from a hardware point of view, 1:06:40.520 --> 1:06:45.520 Google's TPUs and the best Nvidia GPUs are similar. 1:06:45.520 --> 1:06:47.880 And maybe the TPUs are like 30% faster, 1:06:47.880 --> 1:06:51.160 but they're also much harder to program. 1:06:51.160 --> 1:06:54.720 There isn't a clear leader in terms of hardware right now, 1:06:54.720 --> 1:06:57.840 although much more importantly, the Nvidia's GPUs 1:06:57.840 --> 1:06:59.560 are much more programmable. 1:06:59.560 --> 1:07:01.280 They've got much more written problems. 1:07:01.280 --> 1:07:03.480 That's the clear leader for me and where 1:07:03.480 --> 1:07:08.640 I would spend my time as a researcher and practitioner. 1:07:08.640 --> 1:07:12.280 But then in terms of the platform, 1:07:12.280 --> 1:07:15.680 I mean, we're super lucky now with stuff like Google, 1:07:15.680 --> 1:07:21.520 GCP, Google Cloud, and AWS that you can access a GPU 1:07:21.520 --> 1:07:25.440 pretty quickly and easily. 1:07:25.440 --> 1:07:28.280 But I mean, for AWS, it's still too hard. 1:07:28.280 --> 1:07:33.760 You have to find an AMI and get the instance running 1:07:33.760 --> 1:07:37.080 and then install the software you want and blah, blah, blah. 1:07:37.080 --> 1:07:40.400 GCP is currently the best way to get 1:07:40.400 --> 1:07:42.320 started on a full server environment 1:07:42.320 --> 1:07:46.120 because they have a fantastic fast AI in PyTorch, 1:07:46.120 --> 1:07:51.120 ready to go instance, which has all the courses preinstalled. 1:07:51.120 --> 1:07:53.040 It has Jupyter Notebook prerunning. 1:07:53.040 --> 1:07:57.080 Jupyter Notebook is this wonderful interactive computing 1:07:57.080 --> 1:07:59.440 system, which everybody basically 1:07:59.440 --> 1:08:02.920 should be using for any kind of data driven research. 1:08:02.920 --> 1:08:05.880 But then even better than that, there 1:08:05.880 --> 1:08:09.560 are platforms like Salamander, which we own, 1:08:09.560 --> 1:08:13.600 and Paperspace, where literally you click a single button 1:08:13.600 --> 1:08:17.240 and it pops up and you put a notebook straight away 1:08:17.240 --> 1:08:22.240 without any kind of installation or anything. 1:08:22.240 --> 1:08:25.800 And all the course notebooks are all preinstalled. 1:08:25.800 --> 1:08:28.560 So for me, this is one of the things 1:08:28.560 --> 1:08:34.160 we spent a lot of time curating and working on. 1:08:34.160 --> 1:08:35.960 Because when we first started our courses, 1:08:35.960 --> 1:08:39.560 the biggest problem was people dropped out of lesson one 1:08:39.560 --> 1:08:42.680 because they couldn't get an AWS instance running. 1:08:42.680 --> 1:08:44.880 So things are so much better now. 1:08:44.880 --> 1:08:47.760 And we actually have, if you go to course.fast.ai, 1:08:47.760 --> 1:08:49.040 the first thing it says is, here's 1:08:49.040 --> 1:08:50.480 how to get started with your GPU. 1:08:50.480 --> 1:08:52.120 And it's like, you just click on the link 1:08:52.120 --> 1:08:55.120 and you click start and you're going. 1:08:55.120 --> 1:08:56.240 You will go GCP. 1:08:56.240 --> 1:08:58.760 I have to confess, I've never used the Google GCP. 1:08:58.760 --> 1:09:01.600 Yeah, GCP gives you $300 of compute for free, 1:09:01.600 --> 1:09:04.920 which is really nice. 1:09:04.920 --> 1:09:10.960 But as I say, Salamander and Paperspace are even easier still. 1:09:10.960 --> 1:09:15.120 So from the perspective of deep learning frameworks, 1:09:15.120 --> 1:09:18.400 you work with Fast.ai, if you think of it as framework, 1:09:18.400 --> 1:09:22.960 and PyTorch and TensorFlow, what are the strengths 1:09:22.960 --> 1:09:25.840 of each platform in your perspective? 1:09:25.840 --> 1:09:29.240 So in terms of what we've done our research on and taught 1:09:29.240 --> 1:09:34.400 in our course, we started with Theano and Keras. 1:09:34.400 --> 1:09:38.120 And then we switched to TensorFlow and Keras. 1:09:38.120 --> 1:09:40.400 And then we switched to PyTorch. 1:09:40.400 --> 1:09:43.360 And then we switched to PyTorch and Fast.ai. 1:09:43.360 --> 1:09:47.560 And that kind of reflects a growth and development 1:09:47.560 --> 1:09:52.560 of the ecosystem of deep learning libraries. 1:09:52.560 --> 1:09:57.040 Theano and TensorFlow were great, 1:09:57.040 --> 1:10:01.360 but were much harder to teach and to do research and development 1:10:01.360 --> 1:10:04.560 on because they define what's called a computational graph 1:10:04.560 --> 1:10:06.680 up front, a static graph, where you basically 1:10:06.680 --> 1:10:08.360 have to say, here are all the things 1:10:08.360 --> 1:10:12.040 that I'm going to eventually do in my model. 1:10:12.040 --> 1:10:15.080 And then later on, you say, OK, do those things with this data. 1:10:15.080 --> 1:10:17.160 And you can't debug them. 1:10:17.160 --> 1:10:18.560 You can't do them step by step. 1:10:18.560 --> 1:10:20.160 You can't program them interactively 1:10:20.160 --> 1:10:22.280 in a Jupyter notebook and so forth. 1:10:22.280 --> 1:10:24.320 PyTorch was not the first, but PyTorch 1:10:24.320 --> 1:10:27.400 was certainly the strongest entrant to come along 1:10:27.400 --> 1:10:28.720 and say, let's not do it that way. 1:10:28.720 --> 1:10:31.320 Let's just use normal Python. 1:10:31.320 --> 1:10:32.880 And everything you know about in Python 1:10:32.880 --> 1:10:34.000 is just going to work. 1:10:34.000 --> 1:10:37.880 And we'll figure out how to make that run on the GPU 1:10:37.880 --> 1:10:40.800 as and when necessary. 1:10:40.800 --> 1:10:45.120 That turned out to be a huge leap in terms 1:10:45.120 --> 1:10:46.800 of what we could do with our research 1:10:46.800 --> 1:10:49.720 and what we could do with our teaching. 1:10:49.720 --> 1:10:51.160 Because it wasn't limiting. 1:10:51.160 --> 1:10:52.760 Yeah, I mean, it was critical for us 1:10:52.760 --> 1:10:55.960 for something like Dawnbench to be able to rapidly try things. 1:10:55.960 --> 1:10:58.560 It's just so much harder to be a researcher and practitioner 1:10:58.560 --> 1:11:00.520 when you have to do everything upfront 1:11:00.520 --> 1:11:03.400 and you can't inspect it. 1:11:03.400 --> 1:11:07.360 Problem with PyTorch is it's not at all 1:11:07.360 --> 1:11:09.360 accessible to newcomers because you 1:11:09.360 --> 1:11:11.600 have to write your own training loop 1:11:11.600 --> 1:11:15.680 and manage the gradients and all this stuff. 1:11:15.680 --> 1:11:17.920 And it's also not great for researchers 1:11:17.920 --> 1:11:20.680 because you're spending your time dealing with all this boiler 1:11:20.680 --> 1:11:23.920 plate and overhead rather than thinking about your algorithm. 1:11:23.920 --> 1:11:27.760 So we ended up writing this very multi layered API 1:11:27.760 --> 1:11:31.040 that at the top level, you can train a state of the art neural 1:11:31.040 --> 1:11:33.640 network in three lines of code. 1:11:33.640 --> 1:11:35.920 And which talks to an API, which talks to an API, 1:11:35.920 --> 1:11:38.880 which talks to an API, which you can dive into at any level 1:11:38.880 --> 1:11:45.400 and get progressively closer to the machine levels of control. 1:11:45.400 --> 1:11:47.480 And this is the fast AI library. 1:11:47.480 --> 1:11:51.920 That's been critical for us and for our students 1:11:51.920 --> 1:11:54.200 and for lots of people that have won big learning 1:11:54.200 --> 1:11:58.560 competitions with it and written academic papers with it. 1:11:58.560 --> 1:12:00.680 It's made a big difference. 1:12:00.680 --> 1:12:03.960 We're still limited though by Python. 1:12:03.960 --> 1:12:05.920 And particularly this problem with things 1:12:05.920 --> 1:12:10.640 like our current neural nets say where you just can't change 1:12:10.640 --> 1:12:13.320 things unless you accept it going so slowly 1:12:13.320 --> 1:12:15.680 that it's impractical. 1:12:15.680 --> 1:12:18.320 So in the latest incarnation of the course 1:12:18.320 --> 1:12:20.880 and with some of the research we're now starting to do, 1:12:20.880 --> 1:12:24.480 we're starting to do some stuff in Swift. 1:12:24.480 --> 1:12:28.920 I think we're three years away from that being 1:12:28.920 --> 1:12:31.080 super practical, but I'm in no hurry. 1:12:31.080 --> 1:12:35.480 I'm very happy to invest the time to get there. 1:12:35.480 --> 1:12:38.000 But with that, we actually already 1:12:38.000 --> 1:12:41.840 have a nascent version of the fast AI library for vision 1:12:41.840 --> 1:12:44.720 running on Swift and TensorFlow. 1:12:44.720 --> 1:12:48.040 Because Python for TensorFlow is not going to cut it. 1:12:48.040 --> 1:12:49.920 It's just a disaster. 1:12:49.920 --> 1:12:54.440 What they did was they tried to replicate the bits 1:12:54.440 --> 1:12:56.640 that people were saying they like about PyTorch, 1:12:56.640 --> 1:12:59.160 this kind of interactive computation. 1:12:59.160 --> 1:13:02.760 But they didn't actually change their foundational runtime 1:13:02.760 --> 1:13:03.920 components. 1:13:03.920 --> 1:13:06.640 So they kind of added this like syntax, sugar, 1:13:06.640 --> 1:13:08.560 they call TF Eager, TensorFlow Eager, which 1:13:08.560 --> 1:13:10.880 makes it look a lot like PyTorch. 1:13:10.880 --> 1:13:16.400 But it's 10 times slower than PyTorch to actually do a step. 1:13:16.400 --> 1:13:19.080 So because they didn't invest the time 1:13:19.080 --> 1:13:22.080 in retooling the foundations because their code base 1:13:22.080 --> 1:13:23.520 is so horribly complex. 1:13:23.520 --> 1:13:25.280 Yeah, I think it's probably very difficult 1:13:25.280 --> 1:13:26.440 to do that kind of rejoining. 1:13:26.440 --> 1:13:28.680 Yeah, well, particularly the way TensorFlow was written, 1:13:28.680 --> 1:13:31.480 it was written by a lot of people very quickly 1:13:31.480 --> 1:13:33.320 in a very disorganized way. 1:13:33.320 --> 1:13:36.000 So when you actually look in the code, as I do often, 1:13:36.000 --> 1:13:38.840 I'm always just like, oh, god, what were they thinking? 1:13:38.840 --> 1:13:41.480 It's just, it's pretty awful. 1:13:41.480 --> 1:13:47.080 So I'm really extremely negative about the potential future 1:13:47.080 --> 1:13:52.120 for Python TensorFlow that Swift for TensorFlow 1:13:52.120 --> 1:13:53.760 can be a different beast altogether. 1:13:53.760 --> 1:13:57.560 It can be like, it can basically be a layer on top of MLIR 1:13:57.560 --> 1:14:02.640 that takes advantage of all the great compiler stuff 1:14:02.640 --> 1:14:04.760 that Swift builds on with LLVM. 1:14:04.760 --> 1:14:07.040 And yeah, it could be absolutely. 1:14:07.040 --> 1:14:10.320 I think it will be absolutely fantastic. 1:14:10.320 --> 1:14:11.920 Well, you're inspiring me to try. 1:14:11.920 --> 1:14:17.640 Evan truly felt the pain of TensorFlow 2.0 Python. 1:14:17.640 --> 1:14:19.040 It's fine by me. 1:14:19.040 --> 1:14:21.080 But of course. 1:14:21.080 --> 1:14:23.240 Yeah, I mean, it does the job if you're using 1:14:23.240 --> 1:14:27.720 predefined things that somebody's already written. 1:14:27.720 --> 1:14:29.920 But if you actually compare, like I've 1:14:29.920 --> 1:14:33.680 had to do a lot of stuff with TensorFlow recently, 1:14:33.680 --> 1:14:35.480 you actually compare like, I want 1:14:35.480 --> 1:14:37.360 to write something from scratch. 1:14:37.360 --> 1:14:39.040 And you're like, I just keep finding it's like, oh, 1:14:39.040 --> 1:14:41.560 it's running 10 times slower than PyTorch. 1:14:41.560 --> 1:14:43.800 So is the biggest cost. 1:14:43.800 --> 1:14:47.360 Let's throw running time out the window. 1:14:47.360 --> 1:14:49.640 How long it takes you to program? 1:14:49.640 --> 1:14:51.000 That's not too different now. 1:14:51.000 --> 1:14:54.080 Thanks to TensorFlow Eager, that's not too different. 1:14:54.080 --> 1:14:58.640 But because so many things take so long to run, 1:14:58.640 --> 1:15:00.320 you wouldn't run it at 10 times slower. 1:15:00.320 --> 1:15:03.000 Like, you just go like, oh, this is taking too long. 1:15:03.000 --> 1:15:04.240 And also, there's a lot of things 1:15:04.240 --> 1:15:05.840 which are just less programmable, 1:15:05.840 --> 1:15:09.000 like tf.data, which is the way data processing works 1:15:09.000 --> 1:15:11.400 in TensorFlow, is just this big mess. 1:15:11.400 --> 1:15:13.160 It's incredibly inefficient. 1:15:13.160 --> 1:15:14.800 And they kind of had to write it that way 1:15:14.800 --> 1:15:19.160 because of the TPU problems I described earlier. 1:15:19.160 --> 1:15:24.680 So I just feel like they've got this huge technical debt, 1:15:24.680 --> 1:15:27.960 which they're not going to solve without starting from scratch. 1:15:27.960 --> 1:15:29.440 So here's an interesting question then. 1:15:29.440 --> 1:15:34.720 If there's a new student starting today, 1:15:34.720 --> 1:15:37.480 what would you recommend they use? 1:15:37.480 --> 1:15:39.160 Well, I mean, we obviously recommend 1:15:39.160 --> 1:15:42.760 FastAI and PyTorch because we teach new students. 1:15:42.760 --> 1:15:43.960 And that's what we teach with. 1:15:43.960 --> 1:15:46.080 So we would very strongly recommend that 1:15:46.080 --> 1:15:50.280 because it will let you get on top of the concepts much 1:15:50.280 --> 1:15:51.960 more quickly. 1:15:51.960 --> 1:15:53.160 So then you'll become an action. 1:15:53.160 --> 1:15:56.400 And you'll also learn the actual state of the art techniques. 1:15:56.400 --> 1:15:59.240 So you actually get world class results. 1:15:59.240 --> 1:16:03.000 Honestly, it doesn't much matter what library 1:16:03.000 --> 1:16:09.240 you learn because switching from Shaina to MXNet to TensorFlow 1:16:09.240 --> 1:16:12.000 to PyTorch is going to be a couple of days work 1:16:12.000 --> 1:16:15.280 if you long as you understand the foundation as well. 1:16:15.280 --> 1:16:21.600 But you think we'll Swift creep in there as a thing 1:16:21.600 --> 1:16:22.960 that people start using? 1:16:22.960 --> 1:16:26.400 Not for a few years, particularly because Swift 1:16:26.400 --> 1:16:33.440 has no data science community, libraries, schooling. 1:16:33.440 --> 1:16:39.080 And the Swift community has a total lack of appreciation 1:16:39.080 --> 1:16:41.040 and understanding of numeric computing. 1:16:41.040 --> 1:16:43.640 So they keep on making stupid decisions. 1:16:43.640 --> 1:16:47.480 For years, they've just done dumb things around performance 1:16:47.480 --> 1:16:50.280 and prioritization. 1:16:50.280 --> 1:16:56.360 That's clearly changing now because the developer of Chris 1:16:56.360 --> 1:16:59.960 Lattner is working at Google on Swift for TensorFlow. 1:16:59.960 --> 1:17:03.200 So that's a priority. 1:17:03.200 --> 1:17:05.000 It'll be interesting to see what happens with Apple 1:17:05.000 --> 1:17:10.000 because Apple hasn't shown any sign of caring 1:17:10.000 --> 1:17:12.960 about numeric programming in Swift. 1:17:12.960 --> 1:17:16.600 So hopefully they'll get off their arse 1:17:16.600 --> 1:17:18.840 and start appreciating this because currently all 1:17:18.840 --> 1:17:24.240 of their low level libraries are not written in Swift. 1:17:24.240 --> 1:17:27.640 They're not particularly Swifty at all, stuff like Core ML. 1:17:27.640 --> 1:17:30.840 They're really pretty rubbish. 1:17:30.840 --> 1:17:32.760 So yeah, so there's a long way to go. 1:17:32.760 --> 1:17:35.360 But at least one nice thing is that Swift for TensorFlow 1:17:35.360 --> 1:17:40.000 can actually directly use Python code and Python libraries. 1:17:40.000 --> 1:17:44.240 Literally, the entire lesson one notebook of fast AI 1:17:44.240 --> 1:17:47.800 runs in Swift right now in Python mode. 1:17:47.800 --> 1:17:50.800 So that's a nice intermediate thing. 1:17:50.800 --> 1:17:56.800 How long does it take if you look at the two fast AI courses, 1:17:56.800 --> 1:18:00.360 how long does it take to get from 0.0 to completing 1:18:00.360 --> 1:18:02.360 both courses? 1:18:02.360 --> 1:18:04.800 It varies a lot. 1:18:04.800 --> 1:18:12.360 Somewhere between two months and two years, generally. 1:18:12.360 --> 1:18:15.360 So for two months, how many hours a day on average? 1:18:15.360 --> 1:18:20.360 So like somebody who is a very competent coder 1:18:20.360 --> 1:18:27.360 can can do 70 hours per course and pick up. 1:18:27.360 --> 1:18:28.360 70, 70. 1:18:28.360 --> 1:18:29.360 That's it? 1:18:29.360 --> 1:18:30.360 OK. 1:18:30.360 --> 1:18:36.360 But a lot of people I know take a year off to study fast AI 1:18:36.360 --> 1:18:39.360 full time and say at the end of the year, 1:18:39.360 --> 1:18:42.360 they feel pretty competent. 1:18:42.360 --> 1:18:45.360 Because generally, there's a lot of other things you do. 1:18:45.360 --> 1:18:48.360 Generally, they'll be entering Kaggle competitions. 1:18:48.360 --> 1:18:51.360 They might be reading Ian Goodfellow's book. 1:18:51.360 --> 1:18:54.360 They might be doing a bunch of stuff. 1:18:54.360 --> 1:18:57.360 And often, particularly if they are a domain expert, 1:18:57.360 --> 1:19:01.360 their coding skills might be a little on the pedestrian side. 1:19:01.360 --> 1:19:04.360 So part of it's just like doing a lot more writing. 1:19:04.360 --> 1:19:07.360 What do you find is the bottleneck for people usually, 1:19:07.360 --> 1:19:11.360 except getting started and setting stuff up? 1:19:11.360 --> 1:19:13.360 I would say coding. 1:19:13.360 --> 1:19:17.360 The people who are strong coders pick it up the best. 1:19:17.360 --> 1:19:21.360 Although another bottleneck is people who have a lot of 1:19:21.360 --> 1:19:27.360 experience of classic statistics can really struggle 1:19:27.360 --> 1:19:30.360 because the intuition is so the opposite of what they're used to. 1:19:30.360 --> 1:19:33.360 They're very used to trying to reduce the number of parameters 1:19:33.360 --> 1:19:38.360 in their model and looking at individual coefficients 1:19:38.360 --> 1:19:39.360 and stuff like that. 1:19:39.360 --> 1:19:42.360 So I find people who have a lot of coding background 1:19:42.360 --> 1:19:45.360 and know nothing about statistics are generally 1:19:45.360 --> 1:19:48.360 going to be the best stuff. 1:19:48.360 --> 1:19:51.360 So you taught several courses on deep learning 1:19:51.360 --> 1:19:54.360 and as Feynman says, the best way to understand something 1:19:54.360 --> 1:19:55.360 is to teach it. 1:19:55.360 --> 1:19:58.360 What have you learned about deep learning from teaching it? 1:19:58.360 --> 1:20:00.360 A lot. 1:20:00.360 --> 1:20:03.360 It's a key reason for me to teach the courses. 1:20:03.360 --> 1:20:06.360 Obviously, it's going to be necessary to achieve our goal 1:20:06.360 --> 1:20:09.360 of getting domain experts to be familiar with deep learning, 1:20:09.360 --> 1:20:12.360 but it was also necessary for me to achieve my goal 1:20:12.360 --> 1:20:16.360 of being really familiar with deep learning. 1:20:16.360 --> 1:20:24.360 I mean, to see so many domain experts from so many different 1:20:24.360 --> 1:20:28.360 backgrounds, it's definitely, I wouldn't say taught me, 1:20:28.360 --> 1:20:31.360 but convinced me something that I liked to believe was true, 1:20:31.360 --> 1:20:34.360 which was anyone can do it. 1:20:34.360 --> 1:20:37.360 So there's a lot of kind of snobbishness out there about 1:20:37.360 --> 1:20:39.360 only certain people can learn to code, 1:20:39.360 --> 1:20:42.360 only certain people are going to be smart enough to do AI. 1:20:42.360 --> 1:20:44.360 That's definitely bullshit. 1:20:44.360 --> 1:20:48.360 I've seen so many people from so many different backgrounds 1:20:48.360 --> 1:20:52.360 get state of the art results in their domain areas now. 1:20:52.360 --> 1:20:56.360 It's definitely taught me that the key differentiator 1:20:56.360 --> 1:21:00.360 between people that succeed and people that fail is tenacity. 1:21:00.360 --> 1:21:03.360 That seems to be basically the only thing that matters. 1:21:03.360 --> 1:21:07.360 A lot of people give up. 1:21:07.360 --> 1:21:13.360 But if the ones who don't give up pretty much everybody succeeds, 1:21:13.360 --> 1:21:17.360 even if at first I'm just kind of thinking, 1:21:17.360 --> 1:21:20.360 wow, they really aren't quite getting it yet, are they? 1:21:20.360 --> 1:21:24.360 But eventually people get it and they succeed. 1:21:24.360 --> 1:21:27.360 So I think that's been, I think they're both things I liked 1:21:27.360 --> 1:21:29.360 to believe was true, but I don't feel like I really had 1:21:29.360 --> 1:21:31.360 strong evidence for them to be true, 1:21:31.360 --> 1:21:34.360 but now I can see I've seen it again and again. 1:21:34.360 --> 1:21:39.360 So what advice do you have for someone 1:21:39.360 --> 1:21:42.360 who wants to get started in deep learning? 1:21:42.360 --> 1:21:44.360 Train lots of models. 1:21:44.360 --> 1:21:47.360 That's how you learn it. 1:21:47.360 --> 1:21:51.360 So I think, it's not just me. 1:21:51.360 --> 1:21:53.360 I think our course is very good, 1:21:53.360 --> 1:21:55.360 but also lots of people independently have said it's very good. 1:21:55.360 --> 1:21:58.360 It recently won the CogEx Award for AI courses, 1:21:58.360 --> 1:22:00.360 it's being the best in the world. 1:22:00.360 --> 1:22:02.360 I'd say come to our course, course.fast.ai. 1:22:02.360 --> 1:22:05.360 And the thing I keep on harping on in my lessons is 1:22:05.360 --> 1:22:08.360 train models, print out the inputs to the models, 1:22:08.360 --> 1:22:10.360 print out to the outputs to the models, 1:22:10.360 --> 1:22:14.360 like study, you know, change the inputs a bit, 1:22:14.360 --> 1:22:16.360 look at how the outputs vary, 1:22:16.360 --> 1:22:22.360 just run lots of experiments to get an intuitive understanding 1:22:22.360 --> 1:22:24.360 of what's going on. 1:22:24.360 --> 1:22:28.360 To get hooked, do you think, you mentioned training, 1:22:28.360 --> 1:22:32.360 do you think just running the models inference? 1:22:32.360 --> 1:22:35.360 If we talk about getting started. 1:22:35.360 --> 1:22:37.360 No, you've got to fine tune the models. 1:22:37.360 --> 1:22:39.360 So that's the critical thing, 1:22:39.360 --> 1:22:43.360 because at that point, you now have a model that's in your domain area. 1:22:43.360 --> 1:22:46.360 So there's no point running somebody else's model, 1:22:46.360 --> 1:22:48.360 because it's not your model. 1:22:48.360 --> 1:22:50.360 So it only takes five minutes to fine tune a model 1:22:50.360 --> 1:22:52.360 for the data you care about. 1:22:52.360 --> 1:22:54.360 And in lesson two of the course, 1:22:54.360 --> 1:22:56.360 we teach you how to create your own dataset from scratch 1:22:56.360 --> 1:22:58.360 by scripting Google image search. 1:22:58.360 --> 1:23:02.360 And we show you how to actually create a web application running online. 1:23:02.360 --> 1:23:05.360 So I create one in the course that differentiates 1:23:05.360 --> 1:23:08.360 between a teddy bear, a grizzly bear, and a brown bear. 1:23:08.360 --> 1:23:10.360 And it does it with basically 100% accuracy. 1:23:10.360 --> 1:23:13.360 It took me about four minutes to scrape the images 1:23:13.360 --> 1:23:15.360 from Google search in the script. 1:23:15.360 --> 1:23:18.360 There's a little graphical widgets we have in the notebook 1:23:18.360 --> 1:23:21.360 that help you clean up the dataset. 1:23:21.360 --> 1:23:24.360 There's other widgets that help you study the results 1:23:24.360 --> 1:23:26.360 and see where the errors are happening. 1:23:26.360 --> 1:23:29.360 And so now we've got over a thousand replies 1:23:29.360 --> 1:23:32.360 in our Share Your Work Here thread of students saying, 1:23:32.360 --> 1:23:34.360 here's the thing I built. 1:23:34.360 --> 1:23:36.360 And so there's people who, like, 1:23:36.360 --> 1:23:38.360 and a lot of them are state of the art. 1:23:38.360 --> 1:23:40.360 Like somebody said, oh, I tried looking at Dev and Gary characters 1:23:40.360 --> 1:23:42.360 and I couldn't believe it. 1:23:42.360 --> 1:23:44.360 The thing that came out was more accurate 1:23:44.360 --> 1:23:46.360 than the best academic paper after lesson one. 1:23:46.360 --> 1:23:48.360 And then there's others which are just more kind of fun, 1:23:48.360 --> 1:23:53.360 like somebody who's doing Trinidad and Tobago hummingbirds. 1:23:53.360 --> 1:23:55.360 So that's kind of their national bird. 1:23:55.360 --> 1:23:57.360 And Susie's got something that can now classify Trinidad 1:23:57.360 --> 1:23:59.360 and Tobago hummingbirds. 1:23:59.360 --> 1:24:02.360 So yeah, train models, fine tune models with your dataset 1:24:02.360 --> 1:24:05.360 and then study their inputs and outputs. 1:24:05.360 --> 1:24:07.360 How much is Fast AI courses? 1:24:07.360 --> 1:24:09.360 Free. 1:24:09.360 --> 1:24:11.360 Everything we do is free. 1:24:11.360 --> 1:24:13.360 We have no revenue sources of any kind. 1:24:13.360 --> 1:24:15.360 It's just a service to the community. 1:24:15.360 --> 1:24:17.360 You're a saint. 1:24:17.360 --> 1:24:20.360 Okay, once a person understands the basics, 1:24:20.360 --> 1:24:22.360 trains a bunch of models, 1:24:22.360 --> 1:24:25.360 if we look at the scale of years, 1:24:25.360 --> 1:24:27.360 what advice do you have for someone wanting 1:24:27.360 --> 1:24:30.360 to eventually become an expert? 1:24:30.360 --> 1:24:32.360 Train lots of models. 1:24:32.360 --> 1:24:35.360 Specifically, train lots of models in your domain area. 1:24:35.360 --> 1:24:37.360 So an expert, what, right? 1:24:37.360 --> 1:24:40.360 We don't need more expert, like, 1:24:40.360 --> 1:24:45.360 create slightly evolutionary research in areas 1:24:45.360 --> 1:24:47.360 that everybody's studying. 1:24:47.360 --> 1:24:50.360 We need experts at using deep learning 1:24:50.360 --> 1:24:52.360 to diagnose malaria. 1:24:52.360 --> 1:24:55.360 Well, we need experts at using deep learning 1:24:55.360 --> 1:25:00.360 to analyze language to study media bias. 1:25:00.360 --> 1:25:08.360 So we need experts in analyzing fisheries 1:25:08.360 --> 1:25:11.360 to identify problem areas and the ocean. 1:25:11.360 --> 1:25:13.360 That's what we need. 1:25:13.360 --> 1:25:17.360 So become the expert in your passion area. 1:25:17.360 --> 1:25:21.360 And this is a tool which you can use for just about anything, 1:25:21.360 --> 1:25:24.360 and you'll be able to do that thing better than other people, 1:25:24.360 --> 1:25:26.360 particularly by combining it with your passion 1:25:26.360 --> 1:25:27.360 and domain expertise. 1:25:27.360 --> 1:25:28.360 So that's really interesting. 1:25:28.360 --> 1:25:30.360 Even if you do want to innovate on transfer learning 1:25:30.360 --> 1:25:32.360 or active learning, 1:25:32.360 --> 1:25:34.360 your thought is, I mean, 1:25:34.360 --> 1:25:38.360 what I certainly share is you also need to find 1:25:38.360 --> 1:25:41.360 a domain or data set that you actually really care for. 1:25:41.360 --> 1:25:42.360 Right. 1:25:42.360 --> 1:25:45.360 If you're not working on a real problem that you understand, 1:25:45.360 --> 1:25:47.360 how do you know if you're doing it any good? 1:25:47.360 --> 1:25:49.360 How do you know if your results are good? 1:25:49.360 --> 1:25:51.360 How do you know if you're getting bad results? 1:25:51.360 --> 1:25:52.360 Why are you getting bad results? 1:25:52.360 --> 1:25:54.360 Is it a problem with the data? 1:25:54.360 --> 1:25:57.360 How do you know you're doing anything useful? 1:25:57.360 --> 1:26:00.360 Yeah, to me, the only really interesting research is, 1:26:00.360 --> 1:26:03.360 not the only, but the vast majority of interesting research 1:26:03.360 --> 1:26:06.360 is try and solve an actual problem and solve it really well. 1:26:06.360 --> 1:26:10.360 So both understanding sufficient tools on the deep learning side 1:26:10.360 --> 1:26:14.360 and becoming a domain expert in a particular domain 1:26:14.360 --> 1:26:18.360 are really things within reach for anybody. 1:26:18.360 --> 1:26:19.360 Yeah. 1:26:19.360 --> 1:26:23.360 To me, I would compare it to studying self driving cars, 1:26:23.360 --> 1:26:26.360 having never looked at a car or been in a car 1:26:26.360 --> 1:26:29.360 or turned a car on, which is like the way it is 1:26:29.360 --> 1:26:30.360 for a lot of people. 1:26:30.360 --> 1:26:33.360 They'll study some academic data set 1:26:33.360 --> 1:26:36.360 where they literally have no idea about that. 1:26:36.360 --> 1:26:37.360 By the way, I'm not sure how familiar 1:26:37.360 --> 1:26:39.360 you are with autonomous vehicles, 1:26:39.360 --> 1:26:42.360 but that is literally, you describe a large percentage 1:26:42.360 --> 1:26:45.360 of robotics folks working in self driving cars, 1:26:45.360 --> 1:26:48.360 as they actually haven't considered driving. 1:26:48.360 --> 1:26:50.360 They haven't actually looked at what driving looks like. 1:26:50.360 --> 1:26:51.360 They haven't driven. 1:26:51.360 --> 1:26:52.360 And it applies. 1:26:52.360 --> 1:26:54.360 It's a problem because you know when you've actually driven, 1:26:54.360 --> 1:26:57.360 these are the things that happened to me when I was driving. 1:26:57.360 --> 1:26:59.360 There's nothing that beats the real world examples 1:26:59.360 --> 1:27:02.360 or just experiencing them. 1:27:02.360 --> 1:27:04.360 You've created many successful startups. 1:27:04.360 --> 1:27:08.360 What does it take to create a successful startup? 1:27:08.360 --> 1:27:12.360 Same thing as becoming successful deep learning practitioner, 1:27:12.360 --> 1:27:14.360 which is not giving up. 1:27:14.360 --> 1:27:22.360 So you can run out of money or run out of time 1:27:22.360 --> 1:27:24.360 or run out of something, you know, 1:27:24.360 --> 1:27:27.360 but if you keep costs super low 1:27:27.360 --> 1:27:29.360 and try and save up some money beforehand 1:27:29.360 --> 1:27:34.360 so you can afford to have some time, 1:27:34.360 --> 1:27:37.360 then just sticking with it is one important thing. 1:27:37.360 --> 1:27:42.360 Doing something you understand and care about is important. 1:27:42.360 --> 1:27:44.360 By something, I don't mean... 1:27:44.360 --> 1:27:46.360 The biggest problem I see with deep learning people 1:27:46.360 --> 1:27:49.360 is they do a PhD in deep learning 1:27:49.360 --> 1:27:52.360 and then they try and commercialize their PhD. 1:27:52.360 --> 1:27:53.360 It does a waste of time 1:27:53.360 --> 1:27:55.360 because that doesn't solve an actual problem. 1:27:55.360 --> 1:27:57.360 You picked your PhD topic 1:27:57.360 --> 1:28:00.360 because it was an interesting kind of engineering 1:28:00.360 --> 1:28:02.360 or math or research exercise. 1:28:02.360 --> 1:28:06.360 But yeah, if you've actually spent time as a recruiter 1:28:06.360 --> 1:28:10.360 and you know that most of your time was spent sifting through resumes 1:28:10.360 --> 1:28:12.360 and you know that most of the time 1:28:12.360 --> 1:28:14.360 you're just looking for certain kinds of things 1:28:14.360 --> 1:28:19.360 and you can try doing that with a model for a few minutes 1:28:19.360 --> 1:28:21.360 and see whether that's something which a model 1:28:21.360 --> 1:28:23.360 seems to be able to do as well as you could, 1:28:23.360 --> 1:28:27.360 then you're on the right track to creating a startup. 1:28:27.360 --> 1:28:30.360 And then I think just being... 1:28:30.360 --> 1:28:34.360 Just be pragmatic and... 1:28:34.360 --> 1:28:36.360 try and stay away from venture capital money 1:28:36.360 --> 1:28:38.360 as long as possible, preferably forever. 1:28:38.360 --> 1:28:42.360 So yeah, on that point, do you... 1:28:42.360 --> 1:28:43.360 venture capital... 1:28:43.360 --> 1:28:46.360 So were you able to successfully run startups 1:28:46.360 --> 1:28:48.360 with self funded for quite a while? 1:28:48.360 --> 1:28:50.360 Yeah, so my first two were self funded 1:28:50.360 --> 1:28:52.360 and that was the right way to do it. 1:28:52.360 --> 1:28:53.360 Is that scary? 1:28:53.360 --> 1:28:55.360 No. 1:28:55.360 --> 1:28:57.360 VCs startups are much more scary 1:28:57.360 --> 1:29:00.360 because you have these people on your back 1:29:00.360 --> 1:29:01.360 who do this all the time 1:29:01.360 --> 1:29:03.360 and who have done it for years 1:29:03.360 --> 1:29:05.360 telling you grow, grow, grow, grow. 1:29:05.360 --> 1:29:07.360 And they don't care if you fail. 1:29:07.360 --> 1:29:09.360 They only care if you don't grow fast enough. 1:29:09.360 --> 1:29:10.360 So that's scary. 1:29:10.360 --> 1:29:13.360 We're else doing the ones myself 1:29:13.360 --> 1:29:17.360 with partners who were friends. 1:29:17.360 --> 1:29:20.360 It's nice because we just went along 1:29:20.360 --> 1:29:22.360 at a pace that made sense 1:29:22.360 --> 1:29:24.360 and we were able to build it to something 1:29:24.360 --> 1:29:27.360 which was big enough that we never had to work again 1:29:27.360 --> 1:29:29.360 but was not big enough that any VC 1:29:29.360 --> 1:29:31.360 would think it was impressive 1:29:31.360 --> 1:29:35.360 and that was enough for us to be excited. 1:29:35.360 --> 1:29:38.360 So I thought that's a much better way 1:29:38.360 --> 1:29:40.360 to do things for most people. 1:29:40.360 --> 1:29:42.360 And generally speaking now for yourself 1:29:42.360 --> 1:29:44.360 but how do you make money during that process? 1:29:44.360 --> 1:29:47.360 Do you cut into savings? 1:29:47.360 --> 1:29:49.360 So yeah, so I started Fast Mail 1:29:49.360 --> 1:29:51.360 and Optimal Decisions at the same time 1:29:51.360 --> 1:29:54.360 in 1999 with two different friends. 1:29:54.360 --> 1:29:59.360 And for Fast Mail, 1:29:59.360 --> 1:30:03.360 I guess I spent $70 a month on the server. 1:30:03.360 --> 1:30:06.360 And when the server ran out of space 1:30:06.360 --> 1:30:09.360 I put a payments button on the front page 1:30:09.360 --> 1:30:11.360 and said if you want more than 10 meg of space 1:30:11.360 --> 1:30:15.360 you have to pay $10 a year. 1:30:15.360 --> 1:30:18.360 So run low like I keep your cost down. 1:30:18.360 --> 1:30:19.360 Yeah, so I kept my cost down 1:30:19.360 --> 1:30:22.360 and once I needed to spend more money 1:30:22.360 --> 1:30:25.360 I asked people to spend the money for me 1:30:25.360 --> 1:30:29.360 and that was that basically from then on. 1:30:29.360 --> 1:30:34.360 We were making money and I was profitable from then. 1:30:34.360 --> 1:30:37.360 For Optimal Decisions it was a bit harder 1:30:37.360 --> 1:30:40.360 because we were trying to sell something 1:30:40.360 --> 1:30:42.360 that was more like a $1 million sale 1:30:42.360 --> 1:30:46.360 but what we did was we would sell scoping projects 1:30:46.360 --> 1:30:50.360 so kind of like prototypy projects 1:30:50.360 --> 1:30:51.360 but rather than doing it for free 1:30:51.360 --> 1:30:54.360 we would sell them $50,000 to $100,000. 1:30:54.360 --> 1:30:57.360 So again we were covering our costs 1:30:57.360 --> 1:30:58.360 and also making the client feel like 1:30:58.360 --> 1:31:00.360 we were doing something valuable. 1:31:00.360 --> 1:31:06.360 So in both cases we were profitable from six months in. 1:31:06.360 --> 1:31:08.360 Nevertheless it's scary. 1:31:08.360 --> 1:31:10.360 I mean, yeah, sure. 1:31:10.360 --> 1:31:13.360 I mean it's scary before you jump in 1:31:13.360 --> 1:31:18.360 and I guess I was comparing it to the scaredyness of VC. 1:31:18.360 --> 1:31:20.360 I felt like with VC stuff it was more scary. 1:31:20.360 --> 1:31:24.360 Much more in somebody else's hands. 1:31:24.360 --> 1:31:26.360 Will they fund you or not? 1:31:26.360 --> 1:31:28.360 What do they think of what you're doing? 1:31:28.360 --> 1:31:30.360 I also found it very difficult with VC's back startups 1:31:30.360 --> 1:31:33.360 to actually do the thing which I thought was important 1:31:33.360 --> 1:31:35.360 for the company rather than doing the thing 1:31:35.360 --> 1:31:38.360 which I thought would make the VC happy. 1:31:38.360 --> 1:31:40.360 Now, VC's always tell you not to do the thing 1:31:40.360 --> 1:31:41.360 that makes them happy 1:31:41.360 --> 1:31:43.360 but then if you don't do the thing that makes them happy 1:31:43.360 --> 1:31:45.360 they get sad. 1:31:45.360 --> 1:31:48.360 And do you think optimizing for the whatever they call it 1:31:48.360 --> 1:31:52.360 the exit is a good thing to optimize for? 1:31:52.360 --> 1:31:54.360 I mean it can be but not at the VC level 1:31:54.360 --> 1:31:59.360 because the VC exit needs to be, you know, a thousand X. 1:31:59.360 --> 1:32:02.360 So where else the lifestyle exit 1:32:02.360 --> 1:32:04.360 if you can sell something for $10 million 1:32:04.360 --> 1:32:06.360 then you've made it, right? 1:32:06.360 --> 1:32:08.360 So it depends. 1:32:08.360 --> 1:32:10.360 If you want to build something that's going to, 1:32:10.360 --> 1:32:13.360 you're kind of happy to do forever then fine. 1:32:13.360 --> 1:32:16.360 If you want to build something you want to sell 1:32:16.360 --> 1:32:18.360 then three years time that's fine too. 1:32:18.360 --> 1:32:21.360 I mean they're both perfectly good outcomes. 1:32:21.360 --> 1:32:24.360 So you're learning Swift now? 1:32:24.360 --> 1:32:26.360 In a way, I mean you already. 1:32:26.360 --> 1:32:31.360 And I read that you use at least in some cases 1:32:31.360 --> 1:32:34.360 space repetition as a mechanism for learning new things. 1:32:34.360 --> 1:32:38.360 I use Anki quite a lot myself. 1:32:38.360 --> 1:32:41.360 I actually don't never talk to anybody about it. 1:32:41.360 --> 1:32:44.360 Don't know how many people do it 1:32:44.360 --> 1:32:46.360 and it works incredibly well for me. 1:32:46.360 --> 1:32:48.360 Can you talk to your experience? 1:32:48.360 --> 1:32:52.360 Like how did you, what do you, first of all, okay, 1:32:52.360 --> 1:32:53.360 let's back it up. 1:32:53.360 --> 1:32:55.360 What is space repetition? 1:32:55.360 --> 1:33:00.360 So space repetition is an idea created 1:33:00.360 --> 1:33:03.360 by a psychologist named Ebbinghaus, 1:33:03.360 --> 1:33:06.360 I don't know, must be a couple hundred years ago 1:33:06.360 --> 1:33:08.360 or something 150 years ago. 1:33:08.360 --> 1:33:11.360 He did something which sounds pretty damn tedious. 1:33:11.360 --> 1:33:16.360 He found random sequences of letters on cards 1:33:16.360 --> 1:33:21.360 and tested how well he would remember those random sequences 1:33:21.360 --> 1:33:23.360 a day later, a week later, whatever. 1:33:23.360 --> 1:33:26.360 He discovered that there was this kind of a curve 1:33:26.360 --> 1:33:29.360 where his probability of remembering one of them 1:33:29.360 --> 1:33:31.360 would be dramatically smaller the next day 1:33:31.360 --> 1:33:32.360 and then a little bit smaller the next day 1:33:32.360 --> 1:33:34.360 and a little bit smaller the next day. 1:33:34.360 --> 1:33:37.360 What he discovered is that if he revised those cards 1:33:37.360 --> 1:33:42.360 a day, the probabilities would decrease at a smaller rate 1:33:42.360 --> 1:33:44.360 and then if he revised them again a week later, 1:33:44.360 --> 1:33:46.360 they would decrease at a smaller rate again. 1:33:46.360 --> 1:33:51.360 And so he basically figured out a roughly optimal equation 1:33:51.360 --> 1:33:56.360 for when you should revise something you want to remember. 1:33:56.360 --> 1:34:00.360 So space repetition learning is using this simple algorithm, 1:34:00.360 --> 1:34:03.360 just something like revise something after a day 1:34:03.360 --> 1:34:06.360 and then three days and then a week and then three weeks 1:34:06.360 --> 1:34:07.360 and so forth. 1:34:07.360 --> 1:34:10.360 And so if you use a program like Anki, as you know, 1:34:10.360 --> 1:34:12.360 it will just do that for you. 1:34:12.360 --> 1:34:14.360 And it will say, did you remember this? 1:34:14.360 --> 1:34:18.360 And if you say no, it will reschedule it back to be 1:34:18.360 --> 1:34:22.360 appear again like 10 times faster than it otherwise would have. 1:34:22.360 --> 1:34:27.360 It's a kind of a way of being guaranteed to learn something 1:34:27.360 --> 1:34:30.360 because by definition, if you're not learning it, 1:34:30.360 --> 1:34:33.360 it will be rescheduled to be revised more quickly. 1:34:33.360 --> 1:34:37.360 Unfortunately though, it doesn't let you fool yourself. 1:34:37.360 --> 1:34:42.360 If you're not learning something, you know your revisions 1:34:42.360 --> 1:34:44.360 will just get more and more. 1:34:44.360 --> 1:34:48.360 So you have to find ways to learn things productively 1:34:48.360 --> 1:34:50.360 and effectively treat your brain well. 1:34:50.360 --> 1:34:57.360 So using mnemonics and stories and context and stuff like that. 1:34:57.360 --> 1:34:59.360 So yeah, it's a super great technique. 1:34:59.360 --> 1:35:01.360 It's like learning how to learn is something 1:35:01.360 --> 1:35:05.360 which everybody should learn before they actually learn anything. 1:35:05.360 --> 1:35:07.360 But almost nobody does. 1:35:07.360 --> 1:35:10.360 Yes, so what have you, so it certainly works well 1:35:10.360 --> 1:35:14.360 for learning new languages, for, I mean, for learning, 1:35:14.360 --> 1:35:16.360 like small projects almost. 1:35:16.360 --> 1:35:19.360 But do you, you know, I started using it for, 1:35:19.360 --> 1:35:22.360 I forget who wrote a blog post about this inspired me. 1:35:22.360 --> 1:35:25.360 It might have been you, I'm not sure. 1:35:25.360 --> 1:35:28.360 I started when I read papers. 1:35:28.360 --> 1:35:31.360 I'll, concepts and ideas, I'll put them. 1:35:31.360 --> 1:35:32.360 Was it Michael Nielsen? 1:35:32.360 --> 1:35:33.360 It was Michael Nielsen. 1:35:33.360 --> 1:35:34.360 Yeah, it was Michael Nielsen. 1:35:34.360 --> 1:35:36.360 Michael started doing this recently 1:35:36.360 --> 1:35:39.360 and has been writing about it. 1:35:39.360 --> 1:35:44.360 I, so the kind of today's ebbing house is a guy called Peter Wozniak 1:35:44.360 --> 1:35:47.360 who developed a system called Super Memo. 1:35:47.360 --> 1:35:51.360 And he's been basically trying to become like 1:35:51.360 --> 1:35:55.360 the world's greatest renaissance man over the last few decades. 1:35:55.360 --> 1:36:00.360 He's basically lived his life with space repetition learning 1:36:00.360 --> 1:36:03.360 for everything. 1:36:03.360 --> 1:36:07.360 I, and sort of like Michael's only very recently got into this, 1:36:07.360 --> 1:36:09.360 but he started really getting excited about doing it 1:36:09.360 --> 1:36:10.360 for a lot of different things. 1:36:10.360 --> 1:36:14.360 For me personally, I actually don't use it 1:36:14.360 --> 1:36:16.360 for anything except Chinese. 1:36:16.360 --> 1:36:21.360 And the reason for that is that Chinese is specifically a thing. 1:36:21.360 --> 1:36:26.360 I made a conscious decision that I want to continue to remember 1:36:26.360 --> 1:36:29.360 even if I don't get much of a chance to exercise it 1:36:29.360 --> 1:36:33.360 because like I'm not often in China, so I don't. 1:36:33.360 --> 1:36:37.360 Or else something like programming languages or papers, 1:36:37.360 --> 1:36:39.360 they have a very different approach, 1:36:39.360 --> 1:36:42.360 which is I try not to learn anything from them, 1:36:42.360 --> 1:36:46.360 but instead I try to identify the important concepts 1:36:46.360 --> 1:36:48.360 and like actually ingest them. 1:36:48.360 --> 1:36:53.360 So like really understand that concept deeply 1:36:53.360 --> 1:36:54.360 and study it carefully. 1:36:54.360 --> 1:36:56.360 Well, decide if it really is important. 1:36:56.360 --> 1:37:00.360 If it is like incorporate it into our library, 1:37:00.360 --> 1:37:03.360 you know, incorporate it into how I do things 1:37:03.360 --> 1:37:06.360 or decide it's not worth it. 1:37:06.360 --> 1:37:12.360 So I find I then remember the things that I care about 1:37:12.360 --> 1:37:15.360 because I'm using it all the time. 1:37:15.360 --> 1:37:19.360 So for the last 25 years, 1:37:19.360 --> 1:37:23.360 I've committed to spending at least half of every day 1:37:23.360 --> 1:37:25.360 learning or practicing something new, 1:37:25.360 --> 1:37:28.360 which is all my colleagues have always hated 1:37:28.360 --> 1:37:30.360 because it always looks like I'm not working on 1:37:30.360 --> 1:37:31.360 what I'm meant to be working on, 1:37:31.360 --> 1:37:34.360 but that always means I do everything faster 1:37:34.360 --> 1:37:36.360 because I've been practicing a lot of stuff. 1:37:36.360 --> 1:37:39.360 So I kind of give myself a lot of opportunity 1:37:39.360 --> 1:37:41.360 to practice new things. 1:37:41.360 --> 1:37:47.360 And so I find now I don't often kind of find myself 1:37:47.360 --> 1:37:50.360 wishing I could remember something 1:37:50.360 --> 1:37:51.360 because if it's something that's useful, 1:37:51.360 --> 1:37:53.360 then I've been using it a lot. 1:37:53.360 --> 1:37:55.360 It's easy enough to look it up on Google. 1:37:55.360 --> 1:37:59.360 But speaking Chinese, you can't look it up on Google. 1:37:59.360 --> 1:38:01.360 Do you have advice for people learning new things? 1:38:01.360 --> 1:38:04.360 What have you learned as a process? 1:38:04.360 --> 1:38:07.360 I mean, it all starts just making the hours 1:38:07.360 --> 1:38:08.360 and the day available. 1:38:08.360 --> 1:38:10.360 Yeah, you've got to stick with it, 1:38:10.360 --> 1:38:12.360 which is, again, the number one thing 1:38:12.360 --> 1:38:14.360 that 99% of people don't do. 1:38:14.360 --> 1:38:16.360 So the people I started learning Chinese with, 1:38:16.360 --> 1:38:18.360 none of them were still doing it 12 months later. 1:38:18.360 --> 1:38:20.360 I'm still doing it 10 years later. 1:38:20.360 --> 1:38:22.360 I tried to stay in touch with them, 1:38:22.360 --> 1:38:24.360 but they just, no one did it. 1:38:24.360 --> 1:38:26.360 For something like Chinese, 1:38:26.360 --> 1:38:28.360 like study how human learning works. 1:38:28.360 --> 1:38:31.360 So every one of my Chinese flashcards 1:38:31.360 --> 1:38:33.360 is associated with a story, 1:38:33.360 --> 1:38:36.360 and that story is specifically designed to be memorable. 1:38:36.360 --> 1:38:38.360 And we find things memorable, 1:38:38.360 --> 1:38:41.360 funny or disgusting or sexy 1:38:41.360 --> 1:38:44.360 or related to people that we know or care about. 1:38:44.360 --> 1:38:47.360 So I try to make sure all the stories that are in my head 1:38:47.360 --> 1:38:50.360 have those characteristics. 1:38:50.360 --> 1:38:52.360 Yeah, so you have to, you know, 1:38:52.360 --> 1:38:55.360 you won't remember things well if they don't have some context. 1:38:55.360 --> 1:38:57.360 And yeah, you won't remember them well 1:38:57.360 --> 1:39:00.360 if you don't regularly practice them, 1:39:00.360 --> 1:39:02.360 whether it be just part of your day to day life 1:39:02.360 --> 1:39:05.360 for the Chinese and me flashcards. 1:39:05.360 --> 1:39:09.360 I mean, the other thing is, let yourself fail sometimes. 1:39:09.360 --> 1:39:11.360 So like, I've had various medical problems 1:39:11.360 --> 1:39:13.360 over the last few years, 1:39:13.360 --> 1:39:16.360 and basically my flashcards just stopped 1:39:16.360 --> 1:39:18.360 for about three years. 1:39:18.360 --> 1:39:21.360 And then there've been other times I've stopped 1:39:21.360 --> 1:39:24.360 for a few months, and it's so hard because you get back to it, 1:39:24.360 --> 1:39:27.360 and it's like, you have 18,000 cards due. 1:39:27.360 --> 1:39:30.360 It's like, and so you just have to go, 1:39:30.360 --> 1:39:33.360 all right, well, I can either stop and give up everything 1:39:33.360 --> 1:39:36.360 or just decide to do this every day for the next two years 1:39:36.360 --> 1:39:38.360 until I get back to it. 1:39:38.360 --> 1:39:41.360 The amazing thing has been that even after three years, 1:39:41.360 --> 1:39:45.360 I, you know, the Chinese were still in there. 1:39:45.360 --> 1:39:47.360 Like, it was so much faster to relearn 1:39:47.360 --> 1:39:49.360 than it was to mine the first time. 1:39:49.360 --> 1:39:51.360 Yeah, absolutely. 1:39:51.360 --> 1:39:52.360 It's in there. 1:39:52.360 --> 1:39:55.360 I have the same with guitar, with music and so on. 1:39:55.360 --> 1:39:58.360 It's sad because work sometimes takes away 1:39:58.360 --> 1:40:00.360 and then you won't play for a year. 1:40:00.360 --> 1:40:03.360 But really, if you then just get back to it every day, 1:40:03.360 --> 1:40:05.360 you're right there again. 1:40:05.360 --> 1:40:08.360 What do you think is the next big breakthrough 1:40:08.360 --> 1:40:09.360 in artificial intelligence? 1:40:09.360 --> 1:40:12.360 What are your hopes in deep learning or beyond 1:40:12.360 --> 1:40:14.360 that people should be working on, 1:40:14.360 --> 1:40:16.360 or you hope there'll be breakthroughs? 1:40:16.360 --> 1:40:18.360 I don't think it's possible to predict. 1:40:18.360 --> 1:40:20.360 I think what we already have 1:40:20.360 --> 1:40:23.360 is an incredibly powerful platform 1:40:23.360 --> 1:40:26.360 to solve lots of societally important problems 1:40:26.360 --> 1:40:28.360 that are currently unsolved. 1:40:28.360 --> 1:40:30.360 I just hope that people will, lots of people 1:40:30.360 --> 1:40:33.360 will learn this toolkit and try to use it. 1:40:33.360 --> 1:40:36.360 I don't think we need a lot of new technological breakthroughs 1:40:36.360 --> 1:40:39.360 to do a lot of great work right now. 1:40:39.360 --> 1:40:42.360 And when do you think we're going to create 1:40:42.360 --> 1:40:44.360 a human level intelligence system? 1:40:44.360 --> 1:40:45.360 Do you think? 1:40:45.360 --> 1:40:46.360 I don't know. 1:40:46.360 --> 1:40:47.360 How hard is it? 1:40:47.360 --> 1:40:48.360 How far away are we? 1:40:48.360 --> 1:40:49.360 I don't know. 1:40:49.360 --> 1:40:50.360 I have no way to know. 1:40:50.360 --> 1:40:51.360 I don't know. 1:40:51.360 --> 1:40:53.360 Like, I don't know why people make predictions about this 1:40:53.360 --> 1:40:57.360 because there's no data and nothing to go on. 1:40:57.360 --> 1:40:59.360 And it's just like, 1:40:59.360 --> 1:41:03.360 there's so many societally important problems 1:41:03.360 --> 1:41:04.360 to solve right now, 1:41:04.360 --> 1:41:08.360 I just don't find it a really interesting question 1:41:08.360 --> 1:41:09.360 to even answer. 1:41:09.360 --> 1:41:12.360 So in terms of societally important problems, 1:41:12.360 --> 1:41:15.360 what's the problem that is within reach? 1:41:15.360 --> 1:41:17.360 Well, I mean, for example, 1:41:17.360 --> 1:41:19.360 there are problems that AI creates, right? 1:41:19.360 --> 1:41:21.360 So more specifically, 1:41:22.360 --> 1:41:26.360 labor force displacement is going to be huge 1:41:26.360 --> 1:41:28.360 and people keep making this 1:41:28.360 --> 1:41:31.360 frivolous econometric argument of being like, 1:41:31.360 --> 1:41:33.360 oh, there's been other things that aren't AI 1:41:33.360 --> 1:41:34.360 that have come along before 1:41:34.360 --> 1:41:37.360 and haven't created massive labor force displacement. 1:41:37.360 --> 1:41:39.360 Therefore, AI won't. 1:41:39.360 --> 1:41:41.360 So that's a serious concern for you? 1:41:41.360 --> 1:41:42.360 Oh, yeah. 1:41:42.360 --> 1:41:43.360 Andrew Yang is running on it. 1:41:43.360 --> 1:41:44.360 Yeah. 1:41:44.360 --> 1:41:46.360 It's desperately concerned. 1:41:46.360 --> 1:41:52.360 And you see already that the changing workplace 1:41:52.360 --> 1:41:55.360 has lived to a hollowing out of the middle class. 1:41:55.360 --> 1:41:58.360 You're seeing that students coming out of school today 1:41:58.360 --> 1:42:03.360 have a less rosy financial future ahead of them 1:42:03.360 --> 1:42:04.360 than the parents did, 1:42:04.360 --> 1:42:06.360 which has never happened in recent, 1:42:06.360 --> 1:42:08.360 in the last 300 years. 1:42:08.360 --> 1:42:11.360 We've always had progress before. 1:42:11.360 --> 1:42:16.360 And you see this turning into anxiety and despair 1:42:16.360 --> 1:42:19.360 and even violence. 1:42:19.360 --> 1:42:21.360 So I very much worry about that. 1:42:21.360 --> 1:42:24.360 You've written quite a bit about ethics, too. 1:42:24.360 --> 1:42:27.360 I do think that every data scientist 1:42:27.360 --> 1:42:32.360 working with deep learning needs to recognize 1:42:32.360 --> 1:42:34.360 they have an incredibly high leverage tool 1:42:34.360 --> 1:42:36.360 that they're using that can influence society 1:42:36.360 --> 1:42:37.360 in lots of ways. 1:42:37.360 --> 1:42:38.360 And if they're doing research, 1:42:38.360 --> 1:42:41.360 that research is going to be used by people 1:42:41.360 --> 1:42:42.360 doing this kind of work 1:42:42.360 --> 1:42:44.360 and they have a responsibility 1:42:44.360 --> 1:42:46.360 to consider the consequences 1:42:46.360 --> 1:42:49.360 and to think about things like 1:42:49.360 --> 1:42:53.360 how will humans be in the loop here? 1:42:53.360 --> 1:42:55.360 How do we avoid runaway feedback loops? 1:42:55.360 --> 1:42:58.360 How do we ensure an appeals process for humans 1:42:58.360 --> 1:43:00.360 that are impacted by my algorithm? 1:43:00.360 --> 1:43:04.360 How do I ensure that the constraints of my algorithm 1:43:04.360 --> 1:43:08.360 are adequately explained to the people that end up using them? 1:43:08.360 --> 1:43:11.360 There's all kinds of human issues, 1:43:11.360 --> 1:43:13.360 which only data scientists 1:43:13.360 --> 1:43:17.360 are actually in the right place to educate people about, 1:43:17.360 --> 1:43:21.360 but data scientists tend to think of themselves as 1:43:21.360 --> 1:43:22.360 just engineers 1:43:22.360 --> 1:43:24.360 and that they don't need to be part of that process, 1:43:24.360 --> 1:43:26.360 which is wrong. 1:43:26.360 --> 1:43:29.360 Well, you're in the perfect position to educate them better, 1:43:29.360 --> 1:43:32.360 to read literature, to read history, 1:43:32.360 --> 1:43:35.360 to learn from history. 1:43:35.360 --> 1:43:38.360 Well, Jeremy, thank you so much for everything you do 1:43:38.360 --> 1:43:40.360 for inspiring a huge amount of people, 1:43:40.360 --> 1:43:42.360 getting them into deep learning 1:43:42.360 --> 1:43:44.360 and having the ripple effects, 1:43:44.360 --> 1:43:48.360 the flap of a butterfly's wings that will probably change the world. 1:43:48.360 --> 1:44:17.360 So thank you very much.