vtt/episode_023_small.vtt

WEBVTT

00:00.000 --> 00:04.720
 The following is a conversation with Gavin Miller, he's the head of Adobe Research.

00:04.720 --> 00:08.960
 Adobe has empowered artists, designers, and creative minds from all professions,

00:08.960 --> 00:14.320
 working in the digital medium for over 30 years with software such as Photoshop, Illustrator,

00:14.320 --> 00:20.560
 Premiere, After Effects, InDesign, Audition, Software that work with images, video, and audio.

00:21.200 --> 00:25.920
 Adobe Research is working to define the future evolution of these products in a way

00:25.920 --> 00:31.360
 that makes the life of creatives easier, automates the tedious tasks, and gives more and more time

00:31.360 --> 00:36.880
 to operate in the idea space instead of pixel space. This is where the cutting edge, deep

00:36.880 --> 00:41.360
 learning methods of the past decade can really shine more than perhaps any other application.

00:42.240 --> 00:47.840
 Gavin is the embodiment of combining tech and creativity. Outside of Adobe Research,

00:47.840 --> 00:53.600
 he writes poetry and builds robots, both things that are near and dear to my heart as well.

00:53.600 --> 00:59.200
 This conversation is part of the Artificial Intelligence Podcast. If you enjoy it, subscribe

00:59.200 --> 01:05.360
 on YouTube, iTunes, or simply connect with me on Twitter at Lex Friedman's spelled F R I D.

01:06.000 --> 01:09.600
 And now here's my conversation with Gavin Miller.

01:11.120 --> 01:15.920
 You're head of Adobe Research, leading a lot of innovative efforts and applications of AI,

01:15.920 --> 01:23.200
 creating images, video, audio, language, but you're also yourself an artist, a poet,

01:23.200 --> 01:28.640
 a writer, and even a roboticist. So while I promise to everyone listening,

01:28.640 --> 01:32.880
 that I will not spend the entire time we have together reading your poetry, which I love.

01:33.440 --> 01:39.200
 I have to sprinkle it in at least a little bit. So some of them are pretty deep and profound,

01:39.200 --> 01:43.520
 and some are light and silly. Let's start with a few lines from the silly variety.

01:43.520 --> 01:56.800
 You write in a beautiful parody, both Edith Piaf's and my web at Frank Sinatra.

01:56.800 --> 02:06.400
 So it opens with, and now dessert is near. It's time to pay the final total. I've tried to slim

02:06.400 --> 02:14.800
 all year, but my diets have been anecdotal. So where does that love for poetry come from

02:14.800 --> 02:20.880
 for you? And if we dissect your mind, how does it all fit together in the bigger puzzle of Dr.

02:20.880 --> 02:27.440
 Gavin Miller? Well, interesting you chose that one. That was a poem I wrote when I'd been to

02:27.440 --> 02:32.400
 my doctor and he said, you really need to lose some weight and go on a diet. And whilst the

02:32.400 --> 02:37.200
 rational part of my brain wanted to do that, the irrational part of my brain was protesting and

02:37.200 --> 02:42.400
 sort of embraced the opposite idea. I regret nothing, hence. Yes, exactly. Taken to an extreme,

02:42.400 --> 02:49.600
 I thought it would be funny. Obviously, it's a serious topic for some people. But I think,

02:49.600 --> 02:53.920
 for me, I've always been interested in writing since I was in high school, as well as doing

02:53.920 --> 02:58.960
 technology and invention. And sometimes the parallel strands in your life that carry on,

02:58.960 --> 03:05.120
 and one is more about your private life and one's more about your technological career.

03:05.680 --> 03:10.640
 And then at sort of happy moments along the way, sometimes the two things touch, one idea informs

03:10.640 --> 03:17.040
 the other. And we can talk about that as we go. Do you think you're writing the art, the poetry

03:17.040 --> 03:23.440
 contribute indirectly or directly to your research, to your work in Adobe? Well, sometimes it does if

03:23.440 --> 03:30.000
 I say, imagine a future in a science fiction kind of way. And then once it exists on paper,

03:30.000 --> 03:37.520
 I think, well, why shouldn't I just build that? There was an example where when realistic voice

03:37.520 --> 03:42.800
 synthesis first started in the 90s at Apple, where I worked in research. I was done by a friend of mine.

03:44.000 --> 03:48.640
 I sort of sat down and started writing a poem which each line I would enter into the voice

03:48.640 --> 03:54.240
 synthesizer and see how it sounded and sort of wrote it for that voice. And at the time,

03:55.040 --> 04:00.160
 the agents weren't very sophisticated. So they'd sort of add random intonation. And I kind of made

04:00.160 --> 04:06.560
 up the poem to sort of match the tone of the voice. And it sounded slightly sad and depressed. So I

04:06.560 --> 04:12.720
 pretended it was a poem written by an intelligent agent, sort of telling the user to go home and

04:12.720 --> 04:16.560
 leave them alone. But at the same time, they were lonely and wanted to have company and learn from

04:16.560 --> 04:21.760
 what the user was saying. And at the time, it was way beyond anything that AI could possibly do.

04:21.760 --> 04:27.520
 But, you know, since then, it's becoming more within the bounds of possibility.

04:29.040 --> 04:34.800
 And then at the same time, I had a project at home where I did sort of a smart home. This was

04:34.800 --> 04:40.960
 probably 93, 94. And I had the talking voice who'd remind me when I walked in the door of what

04:40.960 --> 04:45.600
 things I had to do. I had buttons on my washing machine because I was a bachelor and I'd leave

04:45.600 --> 04:49.920
 the clothes in there for three days and they'd go moldy. So as I got up in the morning, I would say,

04:49.920 --> 04:56.640
 don't forget the washing and so on. I made photographic photo albums that used light

04:56.640 --> 05:01.040
 sensors to know which page you were looking at would send that over wireless radio to the agent

05:01.040 --> 05:05.760
 who would then play sounds that matched the image she were looking at in the book. So I was kind of

05:05.760 --> 05:10.480
 in love with this idea of magical realism and whether it was possible to do that with technology.

05:10.480 --> 05:16.080
 So that was a case where the sort of the agent sort of intrigued me from a literary point of

05:16.080 --> 05:22.880
 view and became a personality. I think more recently, I've also written plays and when

05:22.880 --> 05:27.440
 plays you write dialogue and obviously you write a fixed set of dialogue that follows a linear

05:27.440 --> 05:33.360
 narrative. But with modern agents, as you design a personality or a capability for conversation,

05:33.360 --> 05:37.760
 you're sort of thinking of, I kind of have imaginary dialogue in my head. And then I think,

05:37.760 --> 05:43.680
 what would it take not only to have that be real, but for it to really know what it's talking about.

05:44.240 --> 05:49.440
 So it's easy to fall into the uncanny valley with AI where it says something it doesn't really

05:49.440 --> 05:54.560
 understand, but it sounds good to the person. But you rapidly realize that it's kind of just

05:55.520 --> 06:00.000
 stimulus response. It doesn't really have real world knowledge about the thing it's describing.

06:00.640 --> 06:06.320
 And so when you get to that point, it really needs to have multiple ways of talking about

06:06.320 --> 06:10.560
 the same concept. So it sounds as though it really understands it. Now, what really understanding

06:10.560 --> 06:16.160
 means is in the eye of the beholder, right? But if it only has one way of referring to something,

06:16.160 --> 06:21.200
 it feels like it's a canned response. But if it can reason about it, or you can go at it from

06:21.200 --> 06:25.600
 multiple angles and give a similar kind of response that people would, then it starts to

06:26.400 --> 06:30.480
 seem more like there's something there that's sentient.

06:31.040 --> 06:35.600
 You can say the same thing, multiple things from different perspectives. I mean, with the

06:35.600 --> 06:40.000
 automatic image captioning that I've seen the work that you're doing, there's elements of that,

06:40.000 --> 06:46.000
 right? Being able to generate different kinds of... Right. So one in my team, there's a lot of work on

06:46.640 --> 06:52.000
 turning a medium from one form to another, whether it's auto tagging imagery or making up full

06:52.000 --> 06:57.840
 sentences about what's in the image, then changing the sentence, finding another image that matches

06:57.840 --> 07:04.720
 the new sentence or vice versa. And in the modern world of GANs, you sort of give it a description

07:04.720 --> 07:11.360
 and it synthesizes an asset that matches the description. So I've sort of gone on a journey.

07:11.360 --> 07:16.560
 My early days in my career were about 3D computer graphics, the sort of pioneering work sort of

07:16.560 --> 07:22.400
 before movies had special effects done with 3D graphics and sort of rode that revolution. And

07:22.400 --> 07:26.720
 that was very much like the renaissance where people would model light and color and shape

07:26.720 --> 07:32.160
 and everything. And now we're kind of in another wave where it's more impressionistic and it's

07:32.160 --> 07:38.240
 sort of the idea of something can be used to generate an image directly, which is sort of the

07:38.240 --> 07:45.520
 new frontier in computer image generation using AI algorithms. So the creative process is more in

07:45.520 --> 07:49.840
 the space of ideas or becoming more in the space of ideas versus in the raw pixels?

07:50.720 --> 07:55.280
 Well, it's interesting. It depends. I think at Adobe, we really want to span the entire range

07:55.280 --> 08:01.040
 from really, really good, what you might call low level tools by low level, as close to say analog

08:01.040 --> 08:07.040
 workflows as possible. So what we do there is we make up systems that do really realistic oil

08:07.040 --> 08:11.600
 paint and watercolor simulation. So if you want every bristle to behave as it would in the real

08:11.600 --> 08:17.760
 world and leave a beautiful analog trail of water and then flow after you've made the brushstroke,

08:17.760 --> 08:21.600
 you can do that. And that's really important for people who want to create something

08:22.720 --> 08:28.160
 really expressive or really novel because they have complete control. And then a certain other

08:28.160 --> 08:35.600
 task become automated. It frees the artists up to focus on the inspiration and less of the perspiration.

08:35.600 --> 08:43.920
 So thinking about different ideas, obviously, once you finish the design, there's a lot of work to

08:43.920 --> 08:49.840
 say do it for all the different aspect ratio of phones or websites and so on. And that used to

08:49.840 --> 08:55.040
 take up an awful lot of time for artists. It still does for many, what we call content velocity.

08:55.040 --> 09:01.360
 And one of the targets of AI is actually to reason about from the first example of what are the

09:01.360 --> 09:06.960
 likely intent for these other formats, maybe if you change the language to German and the words

09:06.960 --> 09:12.160
 are longer, how do you reflow everything so that it looks nicely artistic in that way.

09:12.160 --> 09:17.360
 And so the person can focus on the really creative bit in the middle, which is what is the look and

09:17.360 --> 09:21.200
 style and feel and what's the message and what's the story and the human element.

09:21.200 --> 09:27.920
 So I think creativity is changing. So that's one way in which we're trying to just make it easier

09:27.920 --> 09:32.880
 and faster and cheaper to do so that there can be more of it, more demand, because it's less

09:32.880 --> 09:39.200
 expensive. So everyone wants beautiful artwork for everything from a school website to Hollywood movie.

09:40.800 --> 09:46.480
 On the other side, as some of these things have automatic versions of them, people will

09:46.480 --> 09:52.480
 possibly change role from being the hands on artist and to being either the art director or

09:52.480 --> 09:57.920
 the conceptual artist. And then the computer will be a partner to help create polished examples of

09:57.920 --> 10:02.720
 the idea that they're exploring. Let's talk about Adobe products versus AI and Adobe products.

10:04.000 --> 10:11.200
 Just so you know where I'm coming from, I'm a huge fan of Photoshop for images premiere for video,

10:11.200 --> 10:17.200
 audition for audio. I'll probably use Photoshop to create the thumbnail for this video, premiere

10:17.200 --> 10:24.800
 to edit the video, audition to do the audio. That said, everything I do is really manually. And I

10:24.800 --> 10:30.640
 set up, I use this old school kinesis keyboard and I have auto hotkey that just it's really about

10:30.640 --> 10:37.280
 optimizing the flow of just making sure there's as few clicks as possible. So just being extremely

10:37.280 --> 10:43.920
 efficient. It's something you started to speak to. So before we get into the fun, sort of awesome

10:43.920 --> 10:48.880
 deep learning things, where does AI, if you could speak a little more to it AI or just

10:48.880 --> 10:57.280
 automation in general, do you see in the coming months and years or in general prior in 2018

10:58.560 --> 11:03.840
 fitting into making the life, the low level pixel work flow easier?

11:03.840 --> 11:10.160
 Yeah, that's a great question. So we have a very rich array of algorithms already in Photoshop,

11:10.160 --> 11:17.600
 just classical procedural algorithms as well as ones based on data. In some cases, they end up

11:17.600 --> 11:22.560
 with a large number of sliders and degrees of freedom. So one way in which AI can help is just

11:22.560 --> 11:27.520
 an auto button which comes up with default settings based on the content itself rather than

11:27.520 --> 11:34.080
 default values for the tool. At that point, you then start tweaking. So that's that's a very kind of

11:34.080 --> 11:39.600
 make life easier for people whilst making use of common sense from other example images.

11:39.600 --> 11:40.960
 So like smart defaults.

11:40.960 --> 11:47.760
 Smart defaults, absolutely. Another one is something we've spent a lot of work over the last

11:47.760 --> 11:53.360
 20 years. I've been at Adobe 19 thinking about selection, for instance, where

11:53.360 --> 11:58.720
 you know, with a quick select, you would look at color boundaries and figure out how to sort of

11:58.720 --> 12:02.480
 flood fill into regions that you thought were physically connected in the real world.

12:03.360 --> 12:08.080
 But that algorithm had no visual common sense about what a cat looks like or a dog. It would just do

12:08.080 --> 12:14.080
 it based on rules of thumb, which were applied to graph theory. And it was a big improvement over

12:14.080 --> 12:19.600
 the previous work we had sort of almost click every everything by hand or if it just did similar

12:19.600 --> 12:24.880
 colors, it would do little tiny regions that wouldn't be connected. But in the future,

12:24.880 --> 12:31.120
 using neural nets to actually do a great job with say a single click or even in the case of

12:31.120 --> 12:36.160
 well known categories like people or animals, no click, where you just say select the object and

12:36.160 --> 12:41.440
 it just knows the dominant object as a person in the middle of the photograph. Those kinds of things

12:41.440 --> 12:47.520
 are really valuable if they can be robust enough to give you good quality results.

12:47.520 --> 12:50.720
 Or they can be a great start for like tweaking it.

12:50.720 --> 12:56.080
 So for example, background removal, like one thing I'll, in a thumbnail,

12:56.080 --> 13:00.240
 I'll take a picture of you right now and essentially remove the background behind you.

13:00.240 --> 13:06.480
 And I want to make that as easy as possible. You don't have flowing hair, like rich at the

13:06.480 --> 13:11.040
 moment. Rich sort of. I had it in the past, it may come again in the future, but for now.

13:12.320 --> 13:16.160
 So that sometimes makes it a little more challenging to remove the background.

13:16.160 --> 13:23.680
 How difficult do you think is that problem for AI for basically making the quick selection tool

13:23.680 --> 13:26.960
 smarter and smarter and smarter? Well, we have a lot of research on that already.

13:28.400 --> 13:34.160
 If you want a sort of quick, cheap and cheerful, look, I'm pretending I'm in Hawaii,

13:34.160 --> 13:38.400
 but it's sort of a joke, then you don't need perfect boundaries. And you can do that today

13:38.400 --> 13:44.960
 with a single click for the algorithms we have. We have other algorithms where with a little bit

13:44.960 --> 13:48.560
 more guidance on the boundaries, like you might need to touch it up a little bit.

13:49.920 --> 13:56.480
 We have other algorithms that can pull a nice mat from a crude selection. So we have combinations

13:56.480 --> 14:04.000
 of tools that can do all of that. And at our recent Max conference at AB Max, we demonstrated how

14:04.880 --> 14:09.680
 very quickly just by drawing a simple polygon around the object of interest, we could not

14:09.680 --> 14:17.920
 only do it for a single still, but we could pull at least a selection mask from a moving target,

14:17.920 --> 14:23.360
 like a person dancing in front of a brick wall or something. And so it's going from hours to

14:23.360 --> 14:29.360
 a few seconds for workflows that are really nice. And then you might go in and touch up a little.

14:30.240 --> 14:33.040
 So that's a really interesting question. You mentioned the word robust.

14:33.040 --> 14:40.000
 You know, there's like a journey for an idea, right? And what you presented probably at Max

14:41.360 --> 14:46.320
 has elements of just sort of it inspires the concept, it can work pretty well in a majority

14:46.320 --> 14:51.520
 of cases. But how do you make something that works? Well, in majority of cases, how do you make

14:51.520 --> 14:56.480
 something that works, maybe in all cases, or it becomes a robust tool?

14:56.480 --> 15:01.760
 There are a couple of things. So that really touches on the difference between academic research

15:01.760 --> 15:06.640
 and industrial research. So in academic research, it's really about who's the person to have the

15:06.640 --> 15:12.480
 great new idea that shows promise. And we certainly love to be those people too. But

15:13.120 --> 15:17.840
 we have sort of two forms of publishing. One is academic peer review, which we do a lot of,

15:17.840 --> 15:24.720
 and we have great success there as much as some universities. But then we also have shipping,

15:24.720 --> 15:30.160
 which is a different type of, and then we get customer review, as well as, you know, product

15:30.160 --> 15:37.840
 critics. And that might be a case where it's not about being perfect every single time, but

15:37.840 --> 15:43.280
 perfect enough at the time, plus a mechanism to intervene and recover where you do have mistakes.

15:43.280 --> 15:47.120
 So we have the luxury of very talented customers. We don't want them to be

15:48.720 --> 15:55.200
 overly taxed doing it every time. But if they can go in and just take it from 99 to 100,

15:55.200 --> 16:01.440
 100 with the touch of a mouse or something, then for the professional end, that's something

16:01.440 --> 16:06.320
 that we definitely want to support as well. And for them, it went from having to do that

16:06.320 --> 16:13.360
 tedious task all the time to much less often. So I think that gives us an out. If it had to be

16:13.360 --> 16:18.640
 100% automatic all the time, then that would delay the time at which we could get to market.

16:18.640 --> 16:26.000
 So on that thread, maybe you can untangle something. Again, I'm sort of just speaking to

16:26.000 --> 16:36.000
 my own experience. Maybe that is the most useful idea. So I think Photoshop, as an example or premiere,

16:37.680 --> 16:44.240
 has a lot of amazing features that I haven't touched. And so what's the, in terms of AI,

16:44.240 --> 16:54.080
 helping make my life or the life of creatives easier? How this collaboration between human

16:54.080 --> 16:58.960
 and machine, how do you learn to collaborate better? How do you learn the new algorithms?

17:00.000 --> 17:04.240
 Is it something that where you have to watch tutorials and you have to watch videos and so

17:04.240 --> 17:10.400
 on? Or do you ever think, do you think about the experience itself through exploration being

17:10.400 --> 17:18.320
 the teacher? We absolutely do. So I'm glad that you brought this up. We sort of think about

17:18.320 --> 17:22.080
 two things. One is helping the person in the moment to do the task that they need to do. But

17:22.080 --> 17:26.960
 the other is thinking more holistically about their journey learning a tool. And when it's like,

17:26.960 --> 17:31.120
 think of it as Adobe University, where you use the tool long enough, you become an expert.

17:31.120 --> 17:34.960
 And not necessarily an expert in everything. It's like living in a city. You don't necessarily

17:34.960 --> 17:40.160
 know every street, but you know, the important ones you need to get to. So we have projects in

17:40.160 --> 17:45.360
 research, which actually look at the thousands of hours of tutorials online and try to understand

17:45.360 --> 17:51.040
 what's being taught in them. And then we had one publication at CHI where it was looking at,

17:52.560 --> 17:57.360
 given the last three or four actions you did, what did other people in tutorials do next?

17:57.360 --> 18:01.680
 So if you want some inspiration for what you might do next, or you just want to watch the

18:01.680 --> 18:06.800
 tutorial and see, learn from people who are doing similar workflows to you, you can without having

18:06.800 --> 18:13.360
 to go and search on keywords and everything. So really trying to use the context of your use of

18:13.360 --> 18:18.640
 the app to make intelligent suggestions, either about choices that you might make,

18:20.800 --> 18:26.480
 or in a more assistive way where it could say, if you did this next, we could show you. And that's

18:26.480 --> 18:31.360
 basically the frontier that we're exploring now, which is, if we really deeply understand the

18:31.360 --> 18:38.480
 domain in which designers and creative people work, can we combine that with AI and pattern

18:38.480 --> 18:47.520
 matching of behavior to make intelligent suggestions, either through verbal possibilities or just

18:47.520 --> 18:53.840
 showing the results of if you try this. And that's really the sort of, I was in a meeting today

18:53.840 --> 18:58.880
 thinking about these things. So it's still a grand challenge. We'd all love

18:58.880 --> 19:05.920
 an artist over one shoulder and a teacher over the other, right? And we hope to get there. And

19:05.920 --> 19:10.640
 the right thing to do is to give enough at each stage that it's useful in itself, but it builds

19:10.640 --> 19:19.120
 a foundation for the next level of expectation. Are you aware of this gigantic medium of YouTube

19:19.120 --> 19:26.240
 that's creating just a bunch of creative people, both artists and teachers of different kinds?

19:26.240 --> 19:31.440
 Absolutely. And the more we can understand those media types, both visually and in terms of

19:31.440 --> 19:37.200
 transcripts and words, the more we can bring the wisdom that they embody into the guidance that's

19:37.200 --> 19:42.960
 embedded in the tool. That would be brilliant to remove the barrier from having to yourself type

19:42.960 --> 19:49.600
 in the keyword, searching, so on. Absolutely. And then in the longer term, an interesting

19:49.600 --> 19:54.000
 discussion is, does it ultimately not just assist with learning the interface we have,

19:54.000 --> 19:59.200
 but does it modify the interface to be simpler? Or do you fragment into a variety of tools,

19:59.200 --> 20:05.600
 each of which has a different level of visibility of the functionality? I like to say that if you

20:05.600 --> 20:12.640
 add a feature to a GUI, you have to have yet more visual complexity confronting the new user.

20:12.640 --> 20:17.120
 Whereas if you have an assistant with a new skill, if you know they have it, so you know

20:17.120 --> 20:23.280
 to ask for it, then it's sort of additive without being more intimidating. So we definitely think

20:23.280 --> 20:29.120
 about new users and how to onboard them. Many actually value the idea of being able to master

20:29.120 --> 20:34.720
 that complex interface and keyboard shortcuts, like you were talking about earlier, because

20:35.520 --> 20:39.680
 with great familiarity, it becomes a musical instrument for expressing your visual ideas.

20:40.480 --> 20:45.840
 And other people just want to get something done quickly in the simplest way possible,

20:45.840 --> 20:50.400
 and that's where a more assistive version of the same technology might be useful,

20:50.400 --> 20:54.560
 maybe on a different class of device, which is more in context for capture, say,

20:55.920 --> 21:01.680
 whereas somebody who's in a deep post production workflow maybe want to be on a laptop or a big

21:01.680 --> 21:10.560
 screen desktop and have more knobs and dials to really express the subtlety of what they want to do.

21:12.160 --> 21:16.320
 So there's so many exciting applications of computer vision and machine learning

21:16.320 --> 21:21.920
 that Adobe is working on, like scene stitching, sky replacement, foreground,

21:21.920 --> 21:26.880
 background removal, spatial object based image search, automatic image captioning,

21:26.880 --> 21:31.280
 like we mentioned, project cloak, project deep fill filling in parts of the images,

21:31.920 --> 21:38.640
 project scribbler, style transfer video, style transfer faces and video with Project Puppetron,

21:38.640 --> 21:49.040
 best name ever. Can you talk through a favorite or some of them or examples that pop in mind?

21:49.040 --> 21:54.800
 I'm sure I'll be able to provide links to other ones we don't talk about because there's visual

21:54.800 --> 22:00.640
 elements to all of them that are exciting. Why they're interesting for different reasons might

22:00.640 --> 22:06.640
 be a good way to go. So I think sky replace is interesting because we talked about selection

22:06.640 --> 22:11.440
 being sort of an atomic operation. It's almost like a, if you think of an assembly language,

22:11.440 --> 22:17.840
 it's like a single instruction. Whereas sky replace is a compound action where you automatically

22:17.840 --> 22:22.720
 select the sky, you look for stock content that matches the geometry of the scene.

22:24.000 --> 22:27.600
 You try to have variety in your choices so that you do coverage of different moods.

22:28.160 --> 22:35.600
 It then mats in the sky behind the foreground, but then importantly it uses the foreground

22:35.600 --> 22:40.560
 of the other image that you just searched on to recolor the foreground of the image that

22:40.560 --> 22:47.840
 you're editing. So if you say go from a midday sky to an evening sky, it will actually add

22:47.840 --> 22:53.760
 sort of an orange glow to the foreground objects as well. I was a big fan in college of Magritte

22:53.760 --> 22:59.600
 and he has a number of paintings where it's surrealism because he'll like do a composite,

22:59.600 --> 23:03.440
 but the foreground building will be at night and the sky will be during the day. There's one

23:03.440 --> 23:09.120
 called The Empire of Light which was on my wall in college and we're trying not to do surrealism.

23:09.120 --> 23:15.680
 It can be a choice, but we'd rather have it be natural by default rather than it looking fake

23:15.680 --> 23:19.760
 and then you have to do a whole bunch of post production to fix it. So that's a case where

23:19.760 --> 23:25.120
 we're kind of capturing an entire workflow into a single action and doing it in about a second

23:25.120 --> 23:30.720
 rather than a minute or two. And when you do that, you can not just do it once, but you can do it

23:30.720 --> 23:36.960
 for say like 10 different backgrounds and then you're almost back to this inspiration idea of

23:36.960 --> 23:41.840
 I don't know quite what I want, but I'll know it when I see it. And you can just explore the

23:41.840 --> 23:47.200
 design space as close to final production value as possible. And then when you really pick one,

23:47.200 --> 23:51.120
 you might go back and slightly tweak the selection mask just to make it perfect and

23:51.120 --> 23:54.320
 do that kind of polish that professionals like to bring to their work.

23:54.320 --> 24:01.040
 So then there's this idea of, as you mentioned, the sky replacing it to different stock images of

24:01.040 --> 24:07.040
 the sky. In general, you have this idea or it could be on your disk or whatever. But making even

24:07.040 --> 24:12.400
 more intelligent choices about ways to search stock images which is really interesting. It's

24:12.400 --> 24:19.120
 kind of spatial. Absolutely. Right. So that was something we called Concept Canvas. So normally

24:19.120 --> 24:26.240
 when you do say an image search, I assume it's just based on text. You would give the keywords

24:26.240 --> 24:30.080
 of the things you want to be in the image and it would find the nearest one that had those tags.

24:32.720 --> 24:37.200
 For many tasks, you really want to be able to say I want a big person in the middle or in a

24:37.200 --> 24:41.280
 dog to the right and umbrella above the left because you want to leave space for the text or

24:41.280 --> 24:48.640
 whatever. And so Concept Canvas lets you assign spatial regions to the keywords. And then we've

24:48.640 --> 24:54.560
 already preindexed the images to know where the important concepts are in the picture. So we then

24:54.560 --> 25:01.200
 go through that index matching to assets. And even though it's just another form of search,

25:01.200 --> 25:06.480
 because you're doing spatial design or layout, it starts to feel like design. You sort of feel

25:06.480 --> 25:13.120
 oddly responsible for the image that comes back as if you invented it a little bit. So it's a good

25:13.120 --> 25:18.960
 example where giving enough control starts to make people have a sense of ownership over the

25:18.960 --> 25:23.280
 outcome of the event. And then we also have technologies in Photoshop where you physically

25:23.280 --> 25:29.440
 can move the dog in post as well. But for Concept Canvas, it was just a very fast way to sort of

25:29.440 --> 25:38.560
 loop through and be able to lay things out. In terms of being able to remove objects from a scene

25:38.560 --> 25:45.920
 and fill in the background automatically. So that's extremely exciting. And that's

25:45.920 --> 25:51.200
 so neural networks are stepping in there. I just talked this week with Ian Goodfellow.

25:51.200 --> 25:55.360
 So the GANS for doing that is definitely one approach.

25:55.360 --> 25:59.440
 So is that a really difficult problem? Is it as difficult as it looks,

25:59.440 --> 26:06.080
 again, to take it to a robust product level? Well, there are certain classes of image for

26:06.080 --> 26:10.800
 which the traditional algorithms like Content Aware Fill work really well. Like if you have

26:10.800 --> 26:15.200
 a naturalistic texture like a gravel path or something, because it's patch based, it will

26:15.200 --> 26:19.840
 make up a very plausible looking intermediate thing and fill in the hole. And then we use some

26:20.960 --> 26:25.200
 algorithms to sort of smooth out the lighting so you don't see any brightness contrasts in that

26:25.200 --> 26:29.680
 region or you've gradually ramped from dark to light if it straddles the boundary.

26:29.680 --> 26:37.600
 Where it gets complicated is if you have to infer invisible structure behind the person in front.

26:37.600 --> 26:41.920
 And that really requires a common sense knowledge of the world to know what,

26:42.480 --> 26:47.040
 if I see three quarters of a house, do I have a rough sense of what the rest of the house looks

26:47.040 --> 26:51.840
 like? If you just fill it in with patches, it can end up sort of doing things that make sense

26:51.840 --> 26:56.480
 locally. But you look at the global structure and it looks like it's just sort of crumpled or messed

26:56.480 --> 27:02.800
 up. And so what GANs and neural nets bring to the table is this common sense learned from the

27:02.800 --> 27:10.640
 training set. And the challenge right now is that the generative methods that can make up

27:10.640 --> 27:14.960
 missing holes using that kind of technology are still only stable at low resolutions.

27:15.520 --> 27:19.680
 And so you either need to then go from a low resolution to a high resolution using some other

27:19.680 --> 27:24.720
 algorithm or we need to push the state of the art and it's still in research to get to that point.

27:24.720 --> 27:30.720
 Right. Of course, if you show it something, say it's trained on houses and then you show it in

27:30.720 --> 27:37.360
 octopus, it's not going to do a very good job of showing common sense about octopuses. So

27:39.360 --> 27:45.600
 again, you're asking about how you know that it's ready for prime time. You really need a very

27:45.600 --> 27:52.880
 diverse training set of images. And ultimately, that may be a case where you put it out there

27:52.880 --> 28:00.400
 with some guard rails where you might do a detector which looks at the image and

28:00.960 --> 28:05.280
 sort of estimates its own competence of how well a job could this algorithm do.

28:05.920 --> 28:10.400
 So eventually, there may be this idea of what we call an ensemble of experts where

28:11.120 --> 28:15.440
 any particular expert is specialized in certain things and then there's sort of a

28:15.440 --> 28:19.360
 either they vote to say how confident they are about what to do. This is sort of more future

28:19.360 --> 28:24.080
 looking or there's some dispatcher which says you're good at houses, you're good at trees.

28:27.120 --> 28:31.520
 So I mean, all this adds up to a lot of work because each of those models will be a whole

28:31.520 --> 28:38.320
 bunch of work. But I think over time, you'd gradually fill out the set and initially focus

28:38.320 --> 28:41.520
 on certain workflows and then sort of branch out as you get more capable.

28:41.520 --> 28:48.640
 So you mentioned workflows and have you considered maybe looking far into the future?

28:50.000 --> 28:57.680
 First of all, using the fact that there is a huge amount of people that use Photoshop,

28:57.680 --> 29:03.520
 for example, they have certain workflows, being able to collect the information by which

29:04.880 --> 29:09.440
 they basically get information about their workflows, about what they need,

29:09.440 --> 29:15.120
 what can the ways to help them, whether it is houses or octopus that people work on more.

29:16.000 --> 29:23.440
 Basically getting a beat on what kind of data is needed to be annotated and collected for people

29:23.440 --> 29:26.320
 to build tools that actually work well for people.

29:26.320 --> 29:31.680
 Absolutely. And this is a big topic and the whole world of AI is what data can you gather and why.

29:33.200 --> 29:39.120
 At one level, the way to think about it is we not only want to train our customers in how to use

29:39.120 --> 29:44.160
 our products, but we want them to teach us what's important and what's useful. At the same time,

29:44.160 --> 29:51.120
 we want to respect their privacy and obviously we wouldn't do things without their explicit permission.

29:52.800 --> 29:57.440
 And I think the modern spirit of the age around this is you have to demonstrate to somebody

29:57.440 --> 30:02.720
 how they're benefiting from sharing their data with the tool. Either it's helping in the short

30:02.720 --> 30:07.120
 term to understand their intent so you can make better recommendations or if they're

30:07.120 --> 30:11.840
 there friendly to your cause or you're tall or they want to help you evolve quickly because

30:11.840 --> 30:16.320
 they depend on you for their livelihood, they may be willing to share some of their

30:17.360 --> 30:23.360
 workflows or choices with the dataset to be then trained.

30:24.720 --> 30:30.560
 There are technologies for looking at learning without necessarily storing all the information

30:30.560 --> 30:36.080
 permanently so that you can learn on the fly but not keep a record of what somebody did.

30:36.080 --> 30:38.720
 So, we're definitely exploring all of those possibilities.

30:38.720 --> 30:45.520
 And I think Adobe exists in a space where Photoshop, if I look at the data I've created

30:45.520 --> 30:51.840
 in OWN, I'm less comfortable sharing data with social networks than I am with Adobe because

30:51.840 --> 31:01.360
 there's just exactly as you said, there's an obvious benefit for sharing the data that I use

31:01.360 --> 31:05.440
 to create in Photoshop because it's helping improve the workflow in the future.

31:05.440 --> 31:06.080
 Right.

31:06.080 --> 31:09.440
 As opposed to it's not clear what the benefit is in social networks.

31:10.000 --> 31:14.000
 It's nice of you to say that. I mean, I think there are some professional workflows where

31:14.000 --> 31:17.360
 people might be very protective of what they're doing such as if I was preparing

31:18.240 --> 31:22.640
 evidence for a legal case, I wouldn't want any of that, you know,

31:22.640 --> 31:25.440
 phoning home to help train the algorithm or anything.

31:26.560 --> 31:30.720
 There may be other cases where people say having a trial version or they're doing some,

31:30.720 --> 31:35.280
 I'm not saying we're doing this today, but there's a future scenario where somebody has a more

31:35.280 --> 31:40.400
 permissive relationship with Adobe where they explicitly say, I'm fine, I'm only doing hobby

31:40.400 --> 31:48.000
 projects or things which are non confidential and in exchange for some benefit tangible or

31:48.000 --> 31:51.200
 otherwise, I'm willing to share very fine grain data.

31:51.840 --> 31:57.920
 So, another possible scenario is to capture relatively crude high level things from more

31:57.920 --> 32:02.160
 people and then more detailed knowledge from people who are willing to participate.

32:02.160 --> 32:07.280
 We do that today with explicit customer studies where, you know, we go and visit somebody and

32:07.280 --> 32:10.640
 ask them to try the tool and we human observe what they're doing.

32:12.000 --> 32:15.760
 In the future, to be able to do that enough to be able to train an algorithm,

32:16.320 --> 32:20.240
 we'd need a more systematic process, but we'd have to do it very consciously because

32:21.200 --> 32:24.560
 one of the things people treasure about Adobe is a sense of trust

32:24.560 --> 32:28.880
 and we don't want to endanger that through overly aggressive data collection.

32:28.880 --> 32:35.520
 So, we have a Chief Privacy Officer and it's definitely front and center of thinking about AI

32:35.520 --> 32:39.920
 rather than an afterthought. Well, when you start that program, sign me up.

32:39.920 --> 32:41.040
 Okay, happy to.

32:42.800 --> 32:48.640
 Is there other projects that you wanted to mention that I didn't perhaps that pop into mind?

32:48.640 --> 32:51.840
 Well, you covered the number. I think you mentioned Project Puppetron.

32:51.840 --> 32:58.480
 I think that one is interesting because you might think of Adobe as only thinking in 2D

32:59.760 --> 33:04.800
 and that's a good example where we're actually thinking more three dimensionally about how to

33:04.800 --> 33:09.440
 assign features to faces so that we can, you know, if you take, so what Puppetron does,

33:09.440 --> 33:16.320
 it takes either a still or a video of a person talking and then it can take a painting of somebody

33:16.320 --> 33:20.160
 else and then apply the style of the painting to the person who's talking in the video.

33:20.160 --> 33:29.600
 And it's unlike a sort of screen door post filter effect that you sometimes see online.

33:30.320 --> 33:36.080
 It really looks as though it's sort of somehow attached or reflecting the motion of the face.

33:36.080 --> 33:40.320
 And so that's the case where even to do a 2D workflow like stylization,

33:40.320 --> 33:44.160
 you really need to infer more about the 3D structure of the world.

33:44.160 --> 33:48.560
 And I think as 3D computer vision algorithms get better,

33:48.560 --> 33:52.960
 initially they'll focus on particular domains like faces where you have a lot of

33:52.960 --> 33:57.680
 prior knowledge about structure and you can maybe have a parameterized template that you fit to the

33:57.680 --> 34:03.600
 image. But over time, this should be possible for more general content. And it might even be

34:03.600 --> 34:09.360
 invisible to the user that you're doing 3D reconstruction but under the hood, but it might

34:09.360 --> 34:15.200
 then let you do edits much more reliably or correctly than you would otherwise.

34:15.200 --> 34:20.800
 And you know, the face is a very important application, right?

34:20.800 --> 34:22.480
 So making things work.

34:22.480 --> 34:26.640
 And a very sensitive one. If you do something uncanny, it's very disturbing.

34:26.640 --> 34:30.080
 That's right. You have to get it. You have to get it right.

34:30.080 --> 34:39.040
 So in the space of augmented reality and virtual reality, what do you think is the role of AR and

34:39.040 --> 34:45.360
 VR and in the content we consume as people as consumers and the content we create as creators?

34:45.360 --> 34:47.920
 No, that's a great question. Let me think about this a lot too.

34:48.720 --> 34:55.360
 So I think VR and AR serve slightly different purposes. So VR can really transport you to an

34:55.360 --> 35:02.400
 entire immersive world, no matter what your personal situation is. To that extent, it's a bit like

35:02.400 --> 35:06.720
 a really, really widescreen television where it sort of snaps you out of your context and puts you

35:06.720 --> 35:13.360
 in a new one. And I think it's still evolving in terms of the hardware I actually worked on,

35:13.360 --> 35:17.680
 VR in the 90s, trying to solve the latency and sort of nausea problem, which we did,

35:17.680 --> 35:23.040
 but it was very expensive and a bit early. There's a new wave of that now. I think

35:23.600 --> 35:27.600
 and increasingly those devices are becoming all in one rather than something that's tethered to a

35:27.600 --> 35:34.480
 box. I think the market seems to be bifurcating into things for consumers and things for professional

35:34.480 --> 35:40.080
 use cases, like for architects and people designing where your product is a building and you really

35:40.080 --> 35:45.200
 want to experience it better than looking at a scale model or a drawing, I think,

35:45.920 --> 35:50.320
 or even than a video. So I think for that, where you need a sense of scale and spatial

35:50.320 --> 35:59.600
 relationships, it's great. I think AR holds the promise of sort of taking digital assets off the

35:59.600 --> 36:03.600
 screen and putting them in context in the real world on the table in front of you on the wall

36:03.600 --> 36:10.400
 behind you. And that has the corresponding need that the assets need to adapt to the physical

36:10.400 --> 36:15.680
 context in which they're being placed. I mean, it's a bit like having a live theater troupe come

36:15.680 --> 36:20.880
 to your house and put on Hamlet. My mother had a friend who used to do this at Stately Homes in

36:20.880 --> 36:26.640
 England for the National Trust. And they would adapt the scenes and even they'd walk the audience

36:26.640 --> 36:32.880
 through the rooms to see the action based on the country house they found themselves in for two

36:32.880 --> 36:38.720
 days. And I think AR will have the same issue that if you have a tiny table in a big living room

36:38.720 --> 36:44.960
 or something, it'll try to figure out what can you change and what's fixed. And there's a little

36:44.960 --> 36:52.560
 bit of a tension between fidelity, where if you captured Senior Eye of doing a fantastic ballet,

36:52.560 --> 36:56.640
 you'd want it to be sort of exactly reproduced. And maybe all you could do is scale it down.

36:56.640 --> 37:03.760
 Whereas somebody telling you a story might be walking around the room doing some gestures

37:03.760 --> 37:07.040
 and that could adapt to the room in which they were telling the story.

37:07.760 --> 37:12.800
 And do you think fidelity is that important in that space or is it more about the storytelling?

37:12.800 --> 37:17.840
 I think it may depend on the characteristic of the media. If it's a famous celebrity,

37:17.840 --> 37:22.480
 then it may be that you want to catch every nuance and they don't want to be reanimated by some

37:22.480 --> 37:30.800
 algorithm. It could be that if it's really a loveable frog telling you a story and it's

37:30.800 --> 37:34.320
 about a princess and a frog, then it doesn't matter if the frog moves in a different way.

37:35.520 --> 37:38.640
 I think a lot of the ideas that have sort of grown up in the game world will

37:39.520 --> 37:45.120
 now come into the broader commercial sphere once they're needing adaptive characters in AR.

37:45.120 --> 37:51.920
 Are you thinking of engineering tools that allow creators to create in the augmented world,

37:52.480 --> 37:56.000
 basically making a Photoshop for the augmented world?

37:57.360 --> 38:02.560
 Well, we have shown a few demos of sort of taking a Photoshop layer stack and then expanding it into

38:02.560 --> 38:08.640
 3D. That's actually been shown publicly as one example in AR. Where we're particularly excited

38:08.640 --> 38:17.120
 at the moment is in 3D. 3D design is still a very challenging space. We believe that it's

38:17.120 --> 38:22.800
 a worthwhile experiment to try to figure out if AR or immersive makes 3D design more spontaneous.

38:23.360 --> 38:26.960
 Can you give me an example of 3D design just like applications?

38:26.960 --> 38:32.080
 Well, literally, a simple one would be laying out objects, right? On a conventional screen,

38:32.080 --> 38:35.680
 you'd sort of have a plan view and a side view and a perspective view and you sort of be dragging

38:35.680 --> 38:39.440
 it around with the mouse and if you're not careful, it would go through the wall and all that.

38:39.440 --> 38:46.400
 Whereas if you were really laying out objects, say in a VR headset, you could literally move

38:46.400 --> 38:50.720
 your head to see a different viewpoint. They'd be in stereo, so you'd have a sense of depth

38:50.720 --> 38:57.040
 because you're already wearing the depth glasses, right? So it would be those sort of big gross

38:57.040 --> 39:01.120
 motor, move things around, kind of skills seem much more spontaneous just like they are in the

39:01.120 --> 39:08.320
 real world. The frontier for us, I think, is whether that same medium can be used to do fine

39:08.320 --> 39:14.880
 grain design tasks, like very accurate constraints on, say, a CAD model or something. That may be

39:14.880 --> 39:19.120
 better done on a desktop, but it may just be a matter of inventing the right UI.

39:20.160 --> 39:27.600
 So we're hopeful that because there will be this potential explosion of demand for 3D assets

39:27.600 --> 39:32.560
 that's driven by AR and more real time animation on conventional screens,

39:34.880 --> 39:40.800
 those tools will also help with, or those devices will help with designing the content as well.

39:40.800 --> 39:46.480
 You've mentioned quite a few interesting sort of new ideas. At the same time, there's old

39:46.480 --> 39:51.280
 timers like me that are stuck in their old ways. I think I'm the old timer.

39:51.280 --> 39:58.560
 Okay. All right. But the opposed all change at all costs. When you're thinking about

39:58.560 --> 40:05.440
 creating new interfaces, do you feel the burden of just this giant user base that loves the

40:05.440 --> 40:13.760
 current product? So anything new you do that any new idea comes at a cost that you'll be resisted?

40:13.760 --> 40:21.200
 Well, I think if you have to trade off control for convenience, then our existing user base

40:21.200 --> 40:27.680
 would definitely be offended by that. I think if there are some things where you have more convenience

40:27.680 --> 40:34.160
 and just as much control, that may be more welcome. We do think about not breaking well known

40:34.160 --> 40:40.640
 metaphors for things. So things should sort of make sense. Photoshop has never been a static

40:40.640 --> 40:46.640
 target. It's always been evolving and growing. And to some extent, there's been a lot of brilliant

40:46.640 --> 40:50.640
 thought along the way of how it works today. So we don't want to just throw all that out.

40:51.840 --> 40:55.600
 If there's a fundamental breakthrough, like a single click is good enough to select an object

40:55.600 --> 41:01.760
 rather than having to do lots of strokes, that actually fits in quite nicely to the existing

41:01.760 --> 41:07.840
 tool set, either as an optional mode or as a starting point. I think where we're looking at

41:07.840 --> 41:13.440
 radical simplicity where you could encapsulate an entire workflow with a much simpler UI,

41:14.000 --> 41:18.640
 then sometimes that's easier to do in the context of either a different device like a

41:18.640 --> 41:25.120
 mobile device where the affordances are naturally different or in a tool that's targeted at a

41:25.120 --> 41:31.040
 different workflow where it's about spontaneity and velocity rather than precision. And we have

41:31.040 --> 41:36.720
 projects like Rush, which can let you do professional quality video editing for a certain class of

41:36.720 --> 41:47.440
 media output that is targeted very differently in terms of users and the experience. And ideally,

41:47.440 --> 41:54.160
 people would go, if I'm feeling like doing Premiere, big project, I'm doing a four part

41:54.160 --> 41:59.200
 television series. That's definitely a premier thing. But if I want to do something to show my

41:59.200 --> 42:05.200
 recent vacation, maybe I'll just use Rush because I can do it in the half an hour. I have free at

42:05.200 --> 42:12.480
 home rather than the four hours I need to do it at work. And for the use cases which we can do well,

42:12.480 --> 42:16.960
 it really is much faster to get the same output. But the more professional tools obviously have

42:16.960 --> 42:21.520
 a much richer toolkit and more flexibility in what they can do.

42:21.520 --> 42:26.400
 And then at the same time, with the flexibility and control, I like this idea of smart defaults,

42:27.040 --> 42:33.520
 of using AI to coach you to like what Google has, I'm feeling lucky button.

42:33.520 --> 42:37.360
 Right. Or one button kind of gives you a pretty good

42:38.160 --> 42:41.440
 set of settings. And then you almost, that's almost an educational tool.

42:42.000 --> 42:43.360
 Absolutely. Yeah.

42:43.360 --> 42:49.920
 To show, because sometimes when you have all this control, you're not sure about the

42:51.040 --> 42:55.600
 correlation between the different bars that control different elements of the image and so on.

42:55.600 --> 43:00.400
 And sometimes there's a degree of, you don't know what the optimal is.

43:00.400 --> 43:05.360
 And then some things are sort of on demand like help, right?

43:05.360 --> 43:06.160
 Help, yeah.

43:06.160 --> 43:09.600
 I'm stuck. I need to know what to look for. I'm not quite sure what it's called.

43:10.400 --> 43:13.920
 And something that was proactively making helpful suggestions or,

43:14.800 --> 43:20.640
 you know, you could imagine a make a suggestion button where you'd use all of that knowledge

43:20.640 --> 43:25.120
 of workflows and everything to maybe suggest something to go and learn about or just to try

43:25.120 --> 43:31.280
 or show the answer. And maybe it's not one intelligent to pothole, but it's like a variety

43:31.280 --> 43:33.760
 of defaults. And then you go, oh, I like that one.

43:33.760 --> 43:34.960
 Yeah. Yeah.

43:34.960 --> 43:35.680
 Several options.

43:37.200 --> 43:39.520
 So back to poetry.

43:39.520 --> 43:40.640
 Ah, yes.

43:40.640 --> 43:42.400
 We're going to interleave.

43:43.440 --> 43:48.000
 So first few lines of a recent poem of yours before I ask the next question.

43:48.000 --> 43:55.840
 Yeah. This is about the smartphone. Today left my phone at home and went down to the sea.

43:57.120 --> 44:01.600
 The sand was soft, the ocean glass, but I was still just me.

44:02.480 --> 44:08.240
 So this is a poem about you leaving your phone behind and feeling quite liberated because of it.

44:08.960 --> 44:14.800
 So this is kind of a difficult topic and let's see if we can talk about it, figure it out.

44:14.800 --> 44:21.120
 But so with the help of AI, more and more, we can create versions of ourselves, versions of

44:21.120 --> 44:31.920
 reality that are in some ways more beautiful than actual reality. And some of the creative effort

44:31.920 --> 44:38.880
 there is part of creating this illusion. So of course, this is inevitable, but how do you think

44:38.880 --> 44:44.320
 we should adjust this human beings to live in this digital world that's partly artificial,

44:44.320 --> 44:51.520
 that's better than the world that we lived in a hundred years ago when you didn't have

44:51.520 --> 44:55.840
 Instagram and Facebook versions of ourselves and the online.

44:55.840 --> 44:58.880
 Oh, this is sort of showing off better versions of ourselves.

44:58.880 --> 45:04.880
 We're using the tooling of modifying the images or even with artificial intelligence

45:04.880 --> 45:12.880
 ideas of deep fakes and creating adjusted or fake versions of ourselves and reality.

45:14.080 --> 45:17.600
 I think it's an interesting question. You're all sort of historical bent on this.

45:19.360 --> 45:24.720
 I actually wonder if 18th century aristocrats who commissioned famous painters to paint portraits

45:24.720 --> 45:28.960
 of them had portraits that were slightly nicer than they actually looked in practice.

45:28.960 --> 45:29.680
 Well played, sir.

45:29.680 --> 45:34.720
 So human desire to put your best foot forward has always been true.

45:37.440 --> 45:42.240
 I think it's interesting. You sort of framed it in two ways. One is if we can imagine alternate

45:42.240 --> 45:47.280
 realities and visualize them, is that a good or bad thing? In the old days, you do it with

45:47.280 --> 45:52.560
 storytelling and words and poetry, which still resides sometimes on websites. But

45:52.560 --> 46:00.640
 we've become a very visual culture in particular. In the 19th century, we were very much a text

46:00.640 --> 46:07.280
 based culture. People would read long tracks. Political speeches were very long. Nowadays,

46:07.280 --> 46:15.280
 everything's very kind of quick and visual and snappy. I think it depends on how harmless your

46:15.280 --> 46:23.120
 intent. A lot of it's about intent. So if you have a somewhat flattering photo that you pick

46:23.120 --> 46:30.960
 out of the photos that you have in your inbox to say, this is what I look like, it's probably fine.

46:32.160 --> 46:37.040
 If someone's going to judge you by how you look, then they'll decide soon enough when they meet

46:37.040 --> 46:45.120
 you whether the reality. I think where it can be harmful is if people hold themselves up to an

46:45.120 --> 46:49.520
 impossible standard, which they then feel bad about themselves for not meeting. I think that's

46:49.520 --> 46:58.400
 definitely can be an issue. But I think the ability to imagine and visualize an alternate

46:58.400 --> 47:04.800
 reality, which sometimes which you then go off and build later, can be a wonderful thing too.

47:04.800 --> 47:10.720
 People can imagine architectural styles, which they then have a startup, make a fortune and then

47:10.720 --> 47:17.680
 build a house that looks like their favorite video game. Is that a terrible thing? I think

47:18.720 --> 47:24.560
 I used to worry about exploration actually, that part of the joy of going to the moon

47:24.560 --> 47:30.320
 when I was a tiny child, I remember it, and grainy black and white, was to know what it would look

47:30.320 --> 47:35.120
 like when you got there. And I think now we have such good graphics for knowing, for visualizing

47:35.120 --> 47:40.960
 experience before it happens, that I slightly worry that it may take the edge off actually wanting

47:40.960 --> 47:46.160
 to go. Because we've seen it on TV, we kind of, oh, by the time we finally get to Mars,

47:46.160 --> 47:53.200
 we're going, oh yeah, it's Mars, that's what it looks like. But then the outer exploration,

47:53.200 --> 47:58.480
 I mean, I think Pluto was a fantastic recent discovery where nobody had any idea what it

47:58.480 --> 48:04.800
 looked like and it was just breathtakingly varied and beautiful. So I think expanding

48:04.800 --> 48:10.800
 the ability of the human toolkit to imagine and communicate on balance is a good thing.

48:10.800 --> 48:15.920
 I think there are abuses, we definitely take them seriously and try to discourage them.

48:17.440 --> 48:22.960
 I think there's a parallel side where the public needs to know what's possible through events like

48:22.960 --> 48:30.720
 this, right? So that you don't believe everything you read and print anymore, and it may over time

48:30.720 --> 48:35.440
 become true of images as well. Or you need multiple sets of evidence to really believe

48:35.440 --> 48:40.240
 something rather than a single media asset. So I think it's a constantly evolving thing.

48:40.240 --> 48:45.680
 It's been true forever. There's a famous story about Anne of Cleves and Henry VIII where,

48:47.760 --> 48:52.720
 luckily for Anne, they didn't get married, right? So, or they got married and

48:53.840 --> 48:54.560
 What's the story?

48:54.560 --> 48:59.200
 Oh, so Holbein went and painted a picture and then Henry VIII wasn't pleased and, you know,

48:59.200 --> 49:03.520
 history doesn't record whether Anne was pleased, but I think she was pleased not

49:03.520 --> 49:07.920
 to be married more than a day or something. So I mean, this has gone on for a long time,

49:07.920 --> 49:13.280
 but I think it's just part of the magnification of human capability.

49:14.640 --> 49:20.560
 You've kind of built up an amazing research environment here, research culture, research

49:20.560 --> 49:25.600
 lab, and you've written that the secret to a thriving research lab is interns. Can you unpack

49:25.600 --> 49:29.520
 that a little bit? Oh, absolutely. So a couple of reasons.

49:31.360 --> 49:36.080
 As you see, looking at my personal history, there are certain ideas you bond with at a certain

49:36.080 --> 49:40.880
 stage of your career and you tend to keep revisiting them through time. If you're lucky,

49:40.880 --> 49:44.880
 you pick one that doesn't just get solved in the next five years, and then you're sort of out of

49:44.880 --> 49:49.840
 luck. So I think a constant influx of new people brings new ideas with it.

49:49.840 --> 49:56.800
 From the point of view of industrial research, because a big part of what we do is really taking

49:56.800 --> 50:03.360
 those ideas to the point where they can ship us very robust features, you end up investing a lot

50:03.360 --> 50:08.640
 in a particular idea. And if you're not careful, people can get too conservative in what they

50:08.640 --> 50:15.280
 choose to do next, knowing that the product teams will want it. And interns let you explore the more

50:15.280 --> 50:22.160
 fanciful or unproven ideas in a relatively lightweight way, ideally leading to new publications for

50:22.160 --> 50:28.080
 the intern and for the researcher. And it gives you then a portfolio from which to draw which idea

50:28.080 --> 50:32.720
 am I going to then try to take all the way through to being robust in the next year or two to ship.

50:34.000 --> 50:37.520
 So it sort of becomes part of the funnel. It's also a great way for us to

50:38.160 --> 50:42.640
 identify future full time researchers, many of our greatest researchers were former interns.

50:42.640 --> 50:48.800
 It builds a bridge to university departments so we can get to know and build an enduring relationship

50:48.800 --> 50:53.840
 with the professors and we often do academic give funds to as well as an acknowledgement of the

50:53.840 --> 51:01.360
 value the interns add and their own collaborations. So it's sort of a virtuous cycle. And then the

51:01.360 --> 51:06.960
 long term legacy of a great research lab hopefully will be not only the people who stay but the ones

51:06.960 --> 51:11.520
 who move through and then go off and carry that same model to other companies.

51:11.520 --> 51:17.440
 And so we believe strongly in industrial research and how it can complement academia and

51:17.440 --> 51:21.920
 we hope that this model will continue to propagate and be invested in by other companies,

51:21.920 --> 51:26.800
 which makes it harder for us to recruit, of course, but you know, that's the sign of success

51:26.800 --> 51:30.320
 and a rising tide lifts all ships in that sense.

51:31.040 --> 51:38.080
 And where's the idea of born with the interns? Is there brainstorming? Is there discussions

51:38.080 --> 51:46.240
 about, you know, like what the ideas come from? Yeah, as I'm asking the question, I

51:46.240 --> 51:51.760
 realized how dumb it is, but I'm hoping you have a better answer than a question I ask at the

51:51.760 --> 51:59.280
 beginning of every summer. So what will happen is we'll send out a call for interns. They'll

52:00.000 --> 52:04.240
 we'll have a number of resumes come in, people will contact the candidates, talk to them about

52:04.240 --> 52:09.920
 their interests. They'll usually try to find some somebody who has a reasonably good match to what

52:09.920 --> 52:14.320
 they're already doing, or just has a really interesting domain that they've been pursuing in

52:14.320 --> 52:20.640
 their PhD. And we think we'd love to do one of those projects too. And then the intern stays in

52:20.640 --> 52:27.360
 touch with the mentors, we call them. And then they come and in the first at the end of two weeks,

52:27.360 --> 52:32.000
 they have to decide. So they'll often have a general sense by the time they arrive.

52:32.000 --> 52:36.080
 And we'll have internal discussions about what are all the general

52:36.800 --> 52:41.120
 ideas that we're wanting to pursue to see whether two people have the same idea and maybe they

52:41.120 --> 52:47.040
 should talk and all that. But then once the intern actually arrives, sometimes the idea goes linearly

52:47.040 --> 52:51.120
 and sometimes it takes a giant left turn and we go, that sounded good. But when we thought about

52:51.120 --> 52:54.880
 it, there's this other project or it's already been done and we found this paper that we were

52:54.880 --> 53:02.240
 scooped. But we have this other great idea. So it's pretty flexible at the beginning. One of the

53:02.240 --> 53:08.080
 questions for research labs is who's deciding what to do, and then who's to blame if it goes

53:08.080 --> 53:15.600
 wrong, who gets the credit if it goes right. And so in Adobe, we push the needle very much towards

53:15.600 --> 53:22.960
 freedom of choice of projects by the researchers and the interns. But then we reward people based

53:22.960 --> 53:28.000
 on impact. So if the projects ultimately end up impacting the products and having papers and so

53:28.000 --> 53:34.400
 on. And so your alternative model just to be clear is that you have one lab director who thinks he's

53:34.400 --> 53:38.720
 a genius and tells everybody what to do, takes all the credit if it goes well, blames everybody

53:38.720 --> 53:44.800
 else if it goes badly. So we don't want that model. And this helps new ideas percolate up.

53:45.440 --> 53:49.840
 The art of running such a lab is that there are strategic priorities for the company

53:49.840 --> 53:55.520
 and there are areas where we do want to invest in pressing problems. And so it's a little bit of a

53:55.520 --> 54:01.360
 trickle down and filter up meets in the middle. And so you don't tell people you have to do X,

54:01.360 --> 54:07.360
 but you say X would be particularly appreciated this year. And then people reinterpret X through

54:07.360 --> 54:12.720
 the filter of things they want to do and they're interested in. And miraculously, it usually comes

54:12.720 --> 54:18.640
 together very well. One thing that really helps is Adobe has a really broad portfolio of products.

54:18.640 --> 54:24.960
 So if we have a good idea, there's usually a product team that is intrigued or interested.

54:26.000 --> 54:31.520
 So it means we don't have to qualify things too much ahead of time. Once in a while, the product

54:31.520 --> 54:36.880
 teams sponsor an extra intern because they have a particular problem that they really care about,

54:36.880 --> 54:41.440
 in which case it's a little bit more, we really need one of these. And then we sort of say,

54:41.440 --> 54:45.920
 great, I get an extra intern. We find an intern who thinks that's a great problem. But that's not

54:45.920 --> 54:49.520
 the typical model. That's sort of the icing on the cake as far as the budget's concerned.

54:51.440 --> 54:55.920
 And all of the above end up being important. It's really hard to predict at the beginning of the

54:55.920 --> 55:01.200
 summer, which we all have high hopes of all of the intern projects. But ultimately, some of them

55:01.200 --> 55:06.480
 pay off and some of them sort of are a nice paper, but don't turn into a feature. Others turn out

55:06.480 --> 55:12.080
 not to be as novel as we thought, but they'd be a great feature, but not a paper. And then others,

55:12.080 --> 55:16.400
 we make a little bit of progress and we realize how much we don't know. And maybe we revisit that

55:16.400 --> 55:22.320
 problem several years in a row until it finally we have a breakthrough. And then it becomes more

55:22.320 --> 55:30.320
 on track to impact a product. Jumping back to a big overall view of Adobe Research, what are you

55:30.320 --> 55:37.360
 looking forward to in 2019 and beyond? What is, you mentioned there's a giant suite of products,

55:37.360 --> 55:45.920
 giant suite of products, giant suite of ideas, new interns, a large team of researchers.

55:46.960 --> 55:54.400
 Where do you think the future holds? In terms of the technological breakthroughs?

55:54.400 --> 56:00.960
 Technological breakthroughs, especially ones that will make it into product will get to

56:00.960 --> 56:05.920
 impact the world. So I think the creative or the analytics assistance that we talked about where

56:05.920 --> 56:10.320
 they're constantly trying to figure out what you're trying to do and how can they be helpful and make

56:10.320 --> 56:16.160
 useful suggestions is a really hot topic. And it's very unpredictable as to when it'll be ready,

56:16.160 --> 56:20.720
 but I'm really looking forward to seeing how much progress we make against that. I think

56:22.480 --> 56:28.640
 some of the core technologies like generative adversarial networks are immensely promising

56:28.640 --> 56:34.720
 and seeing how quickly those become practical for mainstream use cases at high resolution with

56:34.720 --> 56:39.840
 really good quality is also exciting. And they also have this sort of strange way of even the

56:39.840 --> 56:45.120
 things they do oddly are odd in an interesting way. So it can look like dreaming or something.

56:46.160 --> 56:55.040
 So that's fascinating. I think internally we have a Sensei platform, which is a way in which

56:55.040 --> 57:01.760
 we're pooling our neural net and other intelligence models into a central platform, which can then be

57:01.760 --> 57:07.040
 leveraged by multiple product teams at once. So we're in the middle of transitioning from a,

57:07.040 --> 57:11.040
 you know, once you have a good idea, you pick a product team to work with and you sort of hand

57:11.040 --> 57:17.520
 design it for that use case to a more sort of Henry Ford, stand it up in a standard way, which

57:17.520 --> 57:22.880
 can be accessed in a standard way, which should mean that the time between a good idea and impacting

57:22.880 --> 57:28.960
 our products will be greatly shortened. And when one product has a good idea, many of the other

57:28.960 --> 57:34.240
 products can just leverage it too. So it's sort of an economy of scale. So that's more about the

57:34.240 --> 57:39.680
 how then the what, but that combination of this sort of renaissance in AI, there's a comparable

57:39.680 --> 57:45.280
 one in graphics with real time ray tracing and other really exciting emerging technologies.

57:45.280 --> 57:49.920
 And when these all come together, you'll sort of basically be dancing with light, right? Where

57:49.920 --> 57:56.080
 you'll have real time shadows, reflections, and as if it's a real world in front of you, but then

57:56.080 --> 58:00.880
 with all these magical properties brought by AI where it sort of anticipates or modifies itself

58:00.880 --> 58:05.040
 in ways that make sense based on how it understands the creative task you're trying to do.

58:06.320 --> 58:12.320
 That's a really exciting future for creative for myself too, the creator. So first of all,

58:12.320 --> 58:17.760
 I work in autonomous vehicles. I'm a roboticist. I love robots. And I think you have a fascination

58:17.760 --> 58:23.920
 with snakes, both natural and artificial robots. I share your fascination. I mean, their movement

58:23.920 --> 58:31.760
 is beautiful, adaptable. The adaptability is fascinating. There are, I looked it up, 2900

58:31.760 --> 58:37.200
 species of snakes in the world. Wow. The 175 venomous, some are tiny, some are huge.

58:38.880 --> 58:44.560
 Saw that there's one that's 25 feet in some cases. So what's the most interesting thing

58:44.560 --> 58:52.000
 that you connect with in terms of snakes, both natural and artificial? Why, what was the connection

58:52.000 --> 58:58.000
 with robotics AI in this particular form of a robot? Well, it actually came out of my work

58:58.000 --> 59:02.880
 in the 80s on computer animation, where I started doing things like cloth simulation and

59:02.880 --> 59:07.360
 other kind of soft body simulation. And you'd sort of drop it and it would bounce,

59:07.360 --> 59:10.880
 then it would just sort of stop moving. And I thought, well, what if you animate the spring

59:10.880 --> 59:16.160
 lengths and simulate muscles? And the simplest object I could do that for was an earthworm.

59:16.160 --> 59:21.680
 So I actually did a paper in 1988 on called the motion dynamics of snakes and worms. And I

59:21.680 --> 59:27.760
 read the physiology literature on both Hale snakes and worms move and then did some of the early

59:27.760 --> 59:34.640
 computer animation examples of that. So your interest in robotics started with graphics?

59:34.640 --> 59:40.960
 Came out of simulation and graphics. When I moved from Alias to Apple, we actually did a

59:40.960 --> 59:46.160
 movie called Her Majesty's Secret Serpent, which is about a secret agent snake that parachutes in

59:46.160 --> 59:50.320
 and captures a film canister from a satellite, which tells you how old fashioned we were thinking

59:50.320 --> 59:57.760
 back then, sort of classic 19 sort of 50s or 60s Bond movie kind of thing. And at the same time,

59:57.760 --> 1:00:03.120
 I'd always made radio control ships when I was a child and from scratch. And I thought, well,

1:00:03.120 --> 1:00:08.800
 how can it be to build a real one? And so then started what turned out to be like a 15 year

1:00:08.800 --> 1:00:14.320
 obsession with trying to build better snake robots. And the first one that I built just sort of

1:00:14.320 --> 1:00:19.520
 slithered sideways, but didn't actually go forward, then added wheels and building things in real

1:00:19.520 --> 1:00:25.840
 life makes you honest about the friction. The thing that appeals to me is I love creating the

1:00:25.840 --> 1:00:30.800
 illusion of life, which is what drove me to drove me to animation. And if you have a robot with

1:00:30.800 --> 1:00:36.320
 enough degrees of coordinated freedom that move in a kind of biological way, then it starts to

1:00:36.320 --> 1:00:42.080
 cross the Yankani Valley and to see me like a creature rather than a thing. And I certainly got

1:00:42.080 --> 1:00:50.320
 that with the early snakes by S3, I had it able to sidewind as well as go directly forward. My

1:00:50.320 --> 1:00:54.240
 wife to be suggested that it would be the ring bearer at our wedding. So it actually went down

1:00:54.240 --> 1:00:59.360
 the aisle carrying the rings and got in the local paper for that, which was really fun.

1:01:00.160 --> 1:01:07.200
 And this was all done as a hobby. And then I at the time that can onboard compute was incredibly

1:01:07.200 --> 1:01:11.840
 limited. It was sort of yes, you should explain that these snakes, the whole idea is that you would

1:01:11.840 --> 1:01:18.640
 you're trying to run it autonomously. Autonomously, on board right. And so

1:01:19.760 --> 1:01:25.280
 the very first one, I actually built the controller from discrete logic, because I used to do LSI,

1:01:25.280 --> 1:01:30.640
 you know, circuits and things when I was a teenager. And then the second and third one,

1:01:30.640 --> 1:01:35.120
 the 8 bit microprocessors were available with like a whole 256 bytes of RAM,

1:01:36.000 --> 1:01:39.840
 which you could just about squeeze in. So they were radio controlled rather than autonomous.

1:01:39.840 --> 1:01:44.480
 And really, we're more about the physic physicality and coordinated motion.

1:01:46.560 --> 1:01:51.520
 I've occasionally taken a side step into if only I could make it cheaply enough,

1:01:51.520 --> 1:01:59.040
 bake a great toy, which has been a lesson in how clockwork is its own magical realm that

1:01:59.040 --> 1:02:03.680
 you venture into and learn things about backlash and other things you don't take into account as

1:02:03.680 --> 1:02:07.600
 a computer scientist, which is why what seemed like a good idea doesn't work. So it's quite

1:02:07.600 --> 1:02:14.160
 humbling. And then more recently, I've been building S9, which is a much better engineered

1:02:14.160 --> 1:02:17.760
 version of S3 where the motors wore out and it doesn't work anymore. And you can't buy

1:02:17.760 --> 1:02:24.640
 replacements, which is sad given that it was such a meaningful one. S5 was about twice as long and

1:02:25.760 --> 1:02:32.960
 look much more biologically inspired. I, unlike the typical roboticist, I taper my snakes.

1:02:33.520 --> 1:02:37.040
 There are good mechanical reasons to do that, but it also makes them look more biological,

1:02:37.040 --> 1:02:43.280
 although it means every segment's unique rather than a repetition, which is why most engineers

1:02:43.280 --> 1:02:49.840
 don't do it. It actually saves weight and leverage and everything. And that one is currently on

1:02:49.840 --> 1:02:54.480
 display at the International Spy Museum in Washington, DC. None of it has done any spying.

1:02:56.080 --> 1:03:00.000
 It was on YouTube and it got its own conspiracy theory where people thought that it wasn't real

1:03:00.000 --> 1:03:04.160
 because they work at Adobe, it must be fake graphics. And people would write to me, tell me

1:03:04.160 --> 1:03:11.200
 it's real. They say the background doesn't move and it's like, it's on a tripod. So that one,

1:03:11.200 --> 1:03:16.960
 but you can see the real thing. So it really is true. And then the latest one is the first one

1:03:16.960 --> 1:03:21.280
 where I could put a Raspberry Pi, which leads to all sorts of terrible jokes about pythons and

1:03:21.280 --> 1:03:29.920
 things. But this one can have onboard compute. And then where my hobby work and my work worker

1:03:29.920 --> 1:03:36.400
 converging is you can now add vision accelerator chips, which can evaluate neural nets and do

1:03:36.400 --> 1:03:41.600
 object recognition and everything. So both for the snakes and more recently for the spider that

1:03:41.600 --> 1:03:48.640
 I've been working on, having desktop level compute is now opening up a whole world of

1:03:49.200 --> 1:03:54.880
 true autonomy with onboard compute, onboard batteries, and still having that sort of

1:03:54.880 --> 1:04:01.680
 biomimetic quality that appeals to children in particular. They are really drawn to them and

1:04:01.680 --> 1:04:08.960
 adults think they look creepy, but children actually think they look charming. And I gave a

1:04:08.960 --> 1:04:14.880
 series of lectures at Girls Who Code to encourage people to take an interest in technology. And

1:04:14.880 --> 1:04:19.200
 at the moment, I'd say they're still more expensive than the value that they add,

1:04:19.200 --> 1:04:22.560
 which is why they're a great hobby for me, but they're not really a great product.

1:04:22.560 --> 1:04:30.000
 It makes me think about doing that very early thing I did at Alias with changing the muscle

1:04:30.000 --> 1:04:35.760
 rest lengths. If I could do that with a real artificial muscle material, then the next snake

1:04:35.760 --> 1:04:40.960
 ideally would use that rather than motors and gearboxes and everything. It would be lighter,

1:04:40.960 --> 1:04:49.200
 much stronger, and more continuous and smooth. So I like to say being in research as a license

1:04:49.200 --> 1:04:54.640
 to be curious, and I have the same feeling with my hobby yet. It forced me to read biology and

1:04:54.640 --> 1:04:59.680
 be curious about things that otherwise would have just been natural geographic special.

1:04:59.680 --> 1:05:04.320
 Suddenly, I'm thinking, how does that snake move? Can I copy it? I look at the trails that

1:05:04.320 --> 1:05:08.640
 side winding snakes leave in sand and see if my snake robots would do the same thing.

1:05:10.240 --> 1:05:13.840
 Out of something inanimate, I like why you put a try to bring life into it and beauty.

1:05:13.840 --> 1:05:18.240
 Absolutely. And then ultimately, give it a personality, which is where the intelligent

1:05:18.240 --> 1:05:25.040
 agent research will converge with the vision and voice synthesis to give it a sense of having

1:05:25.040 --> 1:05:29.840
 not necessarily human level intelligence. I think the Turing test is such a high bar, it's

1:05:30.480 --> 1:05:36.080
 a little bit self defeating, but having one that you can have a meaningful conversation with,

1:05:36.880 --> 1:05:43.040
 especially if you have a reasonably good sense of what you can say. So not trying to have it so

1:05:43.040 --> 1:05:50.080
 a stranger could walk up and have one, but so as a pet owner or a robot pet owner, you could

1:05:50.080 --> 1:05:53.200
 know what it thinks about and what it can reason about.

1:05:53.200 --> 1:05:58.800
 Or sometimes just meaningful interaction. If you have the kind of interaction you have with a dog,

1:05:58.800 --> 1:06:02.080
 sometimes you might have a conversation, but it's usually one way.

1:06:02.080 --> 1:06:02.640
 Absolutely.

1:06:02.640 --> 1:06:06.000
 And nevertheless, it feels like a meaningful connection.

1:06:06.800 --> 1:06:12.720
 And one of the things that I'm trying to do in the sample audio that will play you is beginning

1:06:12.720 --> 1:06:17.680
 to get towards the point where the reasoning system can explain why it knows something or

1:06:17.680 --> 1:06:22.320
 why it thinks something. And that again, creates the sense that it really does know what it's

1:06:22.320 --> 1:06:29.840
 talking about, but also for debugging. As you get more and more elaborate behavior, it's like,

1:06:29.840 --> 1:06:37.120
 why did you decide to do that? How do you know that? I think the robot's really

1:06:37.120 --> 1:06:42.560
 my muse for helping me think about the future of AI and what to invent next.

1:06:42.560 --> 1:06:47.040
 So even at Adobe, that's mostly operating in the digital world.

1:06:47.040 --> 1:06:47.440
 Correct.

1:06:47.440 --> 1:06:54.480
 Do you ever, do you see a future where Adobe even expands into the more physical world perhaps?

1:06:54.480 --> 1:07:02.080
 So bringing life not into animations, but bringing life into physical objects with whether it's,

1:07:02.080 --> 1:07:08.640
 well, I'd have to say at the moment it's a twinkle in my eye. I think the more likely thing is that

1:07:08.640 --> 1:07:14.240
 we will bring virtual objects into the physical world through augmented reality.

1:07:14.240 --> 1:07:21.360
 And many of the ideas that might take five years to build a robot to do, you can do in a few weeks

1:07:21.360 --> 1:07:29.120
 with digital assets. So I think when really intelligent robots finally become commonplace,

1:07:29.120 --> 1:07:34.000
 they won't be that surprising because we'll have been living with those personalities in the virtual

1:07:34.000 --> 1:07:38.640
 sphere for a long time. And then they'll just say, oh, it's Siri with legs or Alexa,

1:07:39.520 --> 1:07:47.760
 Alexa on hooves or something. So I can see that welcoming. And for now, it's still an adventure

1:07:47.760 --> 1:07:53.760
 and we don't know quite what the experience will be like. And it's really exciting to sort of see

1:07:53.760 --> 1:07:59.680
 all of these different strands of my career converge. Yeah, in interesting ways. And it is

1:07:59.680 --> 1:08:07.920
 definitely a fun adventure. So let me end with my favorite poem, the last few lines of my favorite

1:08:07.920 --> 1:08:14.560
 poem of yours that ponders mortality. And in some sense, immortality, as our ideas live through

1:08:14.560 --> 1:08:21.120
 the ideas of others through the work of others, it ends with, do not weep or mourn. It was enough

1:08:21.120 --> 1:08:28.000
 the little atomies permitted just a single dance. Scatter them as deep as your eyes can see. I'm

1:08:28.000 --> 1:08:34.320
 content they'll have another chance, sweeping more centered parts along to join a jostling,

1:08:34.320 --> 1:08:41.440
 lifting throng as others danced in me. Beautiful poem. Beautiful way to end it. Gavin, thank you

1:08:41.440 --> 1:08:45.920
 so much for talking today. And thank you for inspiring and empowering millions of people

1:08:45.920 --> 1:08:50.960
 like myself for creating amazing stuff. Oh, thank you. It's a great conversation.