Philippe Pasquier is a professor at Simon Fraser University and director of the Metacreation Lab for Creative AI. In this episode, he shares insights from two decades at the intersection of artificial intelligence, creative tools, and generative music.

We talk about what creativity means in the context of computer science and how AI systems like MIDI-GPT are designed to collaborate with musicians rather than replace them. Philippe explains how MIDI-GPT works, why prompting with music instead of text might be more useful for composers, and how the system allows for granular musical control.

The conversation also explores the limitations of current AI tools, ethical questions around training on data from deceased artists, and how digital infrastructure might be governed in the future. Philippe speaks openly about the challenges independent labs face compared to tech giants and argues for more public and artist-centered AI development.

So, hi and welcome.

2

00:00:03,563 --> 00:00:05,364

Philippe Vasquer.

3

00:00:05,924 --> 00:00:15,589

You are a professor at Simon Fraser University School of Interactive Arts and Technology,

where you direct the Meta Creation Lab for Creative AI.

4

00:00:15,930 --> 00:00:20,913

You also lead a research creation program around generative systems for creative tasks.

5

00:00:20,913 --> 00:00:30,788

And as such, you're a scientist specialized in artificial intelligence, software designer,

multidisciplinary media artist, educator, and a community builder.

6

00:00:31,580 --> 00:00:41,722

Pursuing a multidisciplinary research creation program, your contribution bridge

fundamental research on generative systems, machine learning, effective computing and

7

00:00:41,722 --> 00:00:50,152

computer assisted creativity, applied research in the creative software industry and

artistic practice in interactive and generative art.

8

00:00:50,212 --> 00:00:51,894

Does that sound about right?

9

00:00:52,450 --> 00:00:54,881

Hello Steiner, yes, that sounds about right.

10

00:00:56,019 --> 00:00:56,900

Excellent.

11

00:00:56,900 --> 00:00:59,203

And thank you so much for joining the podcast.

12

00:00:59,203 --> 00:01:10,768

um I have a pretty broad question for you to start with, and that is, uh what is

creativity and what does it mean for you?

13

00:01:11,352 --> 00:01:17,528

Yeah, well, know, creativity is sort of a very trendy concept in our day and age.

14

00:01:17,528 --> 00:01:24,934

It has gained value, apparently, uh in every possible way in the last uh few years and

decades.

15

00:01:24,934 --> 00:01:38,426

um Since uh the rise of the individual, um it has become an important value, sort of grown

and made popular in our society uh through...

16

00:01:38,444 --> 00:01:39,974

the lens of economical growth.

17

00:01:39,974 --> 00:01:46,336

There was this idea that the creative individual would be a more productive individual in

our society and it became a value.

18

00:01:46,336 --> 00:01:57,480

When it comes to defining it, that has been a challenge for the last 2000 years where

philosophers all over the world have discussed the notion of creativity and that's what in

19

00:01:57,480 --> 00:02:06,342

philosophy we call an essentially contested concept, meaning that there's no agreed upon

definition really and uh

20

00:02:07,018 --> 00:02:08,439

and the debate is going on.

21

00:02:08,439 --> 00:02:17,246

And so at the Meta Creation Lab for Creative AI, which I direct at Simon Fraser

University, we take a very computer scientist approach to creativity because that's what

22

00:02:17,246 --> 00:02:18,047

we do.

23

00:02:18,047 --> 00:02:21,330

We sort of reduce it to the generativity of it.

24

00:02:21,330 --> 00:02:30,637

Often creativity is defined as the possibility to generate new ideas, new options, new

items that are both original and valuable.

25

00:02:30,637 --> 00:02:31,878

uh

26

00:02:31,978 --> 00:02:42,634

But we try ourselves, instead of defining the concept of creativity and what are the

creative process and how it works in human cognition, we reduce it to the notion of a

27

00:02:42,634 --> 00:02:47,287

creative task, which allows to capture the generative aspect of creativity.

28

00:02:47,287 --> 00:02:50,869

So I don't really talk about creativity in general.

29

00:02:50,869 --> 00:03:00,374

I often talk and look at creativity through the lens of, in the case of AI, of the partial

or complete automation of a creative task.

30

00:03:00,374 --> 00:03:11,911

And so it doesn't really matter what it takes to, for example, compose a four-part music

with four tracks, but the task itself is clearly easy to define, and you know when you

31

00:03:11,911 --> 00:03:12,861

have achieved it.

32

00:03:12,861 --> 00:03:26,799

And so as a computer scientist, this sort of reducing creativity as a set of actions uh is

a great, has been a very productive way to frame it, as opposed to getting into the muddy

33

00:03:26,799 --> 00:03:28,546

territory of debating.

34

00:03:28,546 --> 00:03:31,829

the philosophy of the concept, although that's really interesting too.

35

00:03:31,829 --> 00:03:34,671

But yeah, I'll keep it at that for now.

36

00:03:34,671 --> 00:03:45,900

So yeah, we talk about creative tasks and creative tasks are in computer science and for

all intents and purposes, all of those tasks, and that's a lot, for which there's no

37

00:03:45,900 --> 00:03:48,882

rational, optimal solution.

38

00:03:49,023 --> 00:03:53,787

So if I want to find the shortest path on a Google map, I will use A star.

39

00:03:53,787 --> 00:03:55,308

It's an algorithm that does that for you.

40

00:03:55,308 --> 00:03:57,838

And most people, you know, they're really...

41

00:03:57,838 --> 00:04:00,020

follow the instruction of Google Maps when they move.

42

00:04:00,020 --> 00:04:04,203

So what's oriented people in the world right now is this algorithm A star.

43

00:04:04,203 --> 00:04:14,972

And it's capable of finding the shortest path from A to B, given the driving conditions or

given the state of the city and the streets.

44

00:04:15,112 --> 00:04:17,434

And that's an optimal solution.

45

00:04:17,474 --> 00:04:19,496

And so that's not a creative task.

46

00:04:19,496 --> 00:04:23,119

Playing chess, you win uh or you lose.

47

00:04:23,119 --> 00:04:25,841

It's a rational problem solving.

48

00:04:25,841 --> 00:04:27,234

So it's not a creative task.

49

00:04:27,234 --> 00:04:30,415

So that's under my definition, of creativity as a task.

50

00:04:30,415 --> 00:04:36,888

So tasks that are creative are tasks that, by definition, do not accept optimal solutions.

51

00:04:36,888 --> 00:04:49,263

So things like composing music, interpreting music, writing lyrics, everything that has to

do with what you talk about usually and what your center is about, that uh it does fall

52

00:04:49,263 --> 00:04:51,714

under the idea of a creative task.

53

00:04:51,714 --> 00:04:53,609

Mastering a track is a creative task.

54

00:04:53,609 --> 00:04:56,224

There's no such thing as an optimal mastering.

55

00:04:56,224 --> 00:05:00,987

or Pareto dominant mastering or the utility function of a mastering, right?

56

00:05:00,987 --> 00:05:16,138

And so, yeah, jokes, uh choosing clothes in the morning, ah deciding the gaze we take when

we look at something or the type of movement we make, all those things uh often do not

57

00:05:16,138 --> 00:05:21,271

have rational uh optimal solutions and therefore they're considered creative.

58

00:05:21,271 --> 00:05:24,093

So at the Metacritic Lab, that's really what we focus on.

59

00:05:24,093 --> 00:05:25,176

Automatizing.

60

00:05:25,176 --> 00:05:28,549

creative tasks for which there's no rational solution across domains.

61

00:05:28,549 --> 00:05:42,242

So we do it for dance and choreography, we do it for visual, for animation, but more

importantly we've been doing it in the last 15 years that the group has been together,

62

00:05:42,242 --> 00:05:45,324

we've been doing it intensively for music.

63

00:05:47,345 --> 00:05:56,282

So if I got you right, um creativity for you guys means uh creating something that doesn't

have an optimal rational solution.

64

00:05:57,666 --> 00:05:58,656

That's right.

65

00:05:59,343 --> 00:06:01,353

I haven't actually heard that definition before.

66

00:06:01,353 --> 00:06:03,484

It was really interesting and really fits.

67

00:06:03,484 --> 00:06:05,025

It really fits well.

68

00:06:05,445 --> 00:06:07,726

A couple of other definitions on creativity.

69

00:06:07,726 --> 00:06:19,471

em I've found, for instance, like a stock version of creativity from Wikipedia says that

creativity is the ability to form novel and valuable ideas or works using one's

70

00:06:19,471 --> 00:06:20,791

imagination.

71

00:06:21,092 --> 00:06:27,236

And of course, it gets to be a muddy territory um when you talk about novel and valuable.

72

00:06:27,236 --> 00:06:31,087

I mean, who's the judge of what's novel and or what's valuable.

73

00:06:31,228 --> 00:06:36,930

And I mean, something that's really hype these days is can a machine be creative?

74

00:06:36,930 --> 00:06:45,814

Can they create something new or is it just rejured regurgitating or imitating stuff

that's made before, you know, that discourse.

75

00:06:45,834 --> 00:06:55,898

And I, I recently read a post by an earlier podcast guest I've had on the show before

she's called Esther Fierre-Einhart.

76

00:06:55,898 --> 00:07:07,346

I don't know if you've heard about her, but she's made this definition that creativity is

the decomposition of what is perceived, selection and recombination with the ability to

77

00:07:07,346 --> 00:07:13,890

compare an intermediate result with an inner theme, an emotion or idea.

78

00:07:13,890 --> 00:07:20,534

And the confrontation with the inner theme in turn requires the ability to feel emotions

and a consciousness.

79

00:07:21,262 --> 00:07:23,782

Aha, so more of a cognitive definition.

80

00:07:23,782 --> 00:07:24,742

Yeah, that's interesting.

81

00:07:24,742 --> 00:07:29,142

you know, I could go that way and there's plenty of attempts that way.

82

00:07:29,142 --> 00:07:39,682

we do ourselves, when we do model creativity, we do look at what is the process, what are

the cognition involved in terms of the beliefs, the desires, the intentions of an

83

00:07:39,682 --> 00:07:42,142

artificial agent.

84

00:07:42,202 --> 00:07:46,642

But I like my definition better because it is so functional.

85

00:07:46,642 --> 00:07:49,346

And it captures the first one from Wikipedia that you...

86

00:07:49,346 --> 00:07:59,132

that you sort of read, which is also the one from Margaret Boden, which is maybe the most

well-known philosopher of creativity in the 20th century.

87

00:07:59,132 --> 00:08:02,744

And her definition is exactly that, that it's about generating.

88

00:08:02,744 --> 00:08:05,355

So you produce something when you're being creative.

89

00:08:05,355 --> 00:08:10,298

There's an act of generation, an idea, an object, a painting, a piece of music.

90

00:08:10,298 --> 00:08:14,821

And then what you create needs to be original and valuable.

91

00:08:14,821 --> 00:08:16,770

And then there's different types of creativity.

92

00:08:16,770 --> 00:08:18,531

Personal creativity, peak creativity.

93

00:08:18,531 --> 00:08:28,725

Sometimes a kid makes a drawing, it is uh an act of creation, it is generative, it is

novel, original, but is it valuable?

94

00:08:28,725 --> 00:08:33,207

Well, it's valuable for the kid, it may be valuable for the parents who are like, now my

kid can draw a house, right?

95

00:08:33,207 --> 00:08:37,068

But it's not valuable for the world, it's not gonna go at Christie auction, right?

96

00:08:37,068 --> 00:08:40,210

So this availability, how valuable it is, that's subjective.

97

00:08:40,210 --> 00:08:44,621

And of course, we all know that in art and when creativity is involved, there is

subjectivity.

98

00:08:44,621 --> 00:08:46,612

know, someone's creativity might...

99

00:08:46,638 --> 00:08:51,098

might rub me the right way and someone else's might rub me the wrong way.

100

00:08:51,138 --> 00:08:54,618

And so I'm not going to like all those creative artifacts the same way.

101

00:08:54,618 --> 00:09:03,118

And that's really that subjectivity that for a computer scientist, for someone who works

in AI, makes it more complicated than rational problem solving, more complicated than

102

00:09:03,118 --> 00:09:08,638

actually getting a self-driving car to go from A to B, taking the shortest path.

103

00:09:08,638 --> 00:09:13,494

And therefore a little bit more interesting from, you know, I come from mathematics

myself.

104

00:09:13,494 --> 00:09:17,616

and then I became a computer scientist as a way to do applied mathematics.

105

00:09:17,616 --> 00:09:21,779

on the side, I've been growing a career as an artist.

106

00:09:21,779 --> 00:09:30,864

And so I've been fascinated by this very question of what is it that makes creativity

different from pure rational problem solving?

107

00:09:30,864 --> 00:09:41,509

And at the beginning uh of trying to identify creativity as different from rational

problem solving, I had the idea that it meant sort of by definition that it would be

108

00:09:41,509 --> 00:09:42,712

really, really hard.

109

00:09:42,712 --> 00:09:46,203

to get computer to be any good at any of those tasks.

110

00:09:46,203 --> 00:10:03,468

And it turns out, especially the last 10 years, uh after high efforts but many other AI

researchers, the whole field of generative AI uh with algorithms that specifically tackle

111

00:10:03,468 --> 00:10:07,130

those tasks for which there's no optimal solution has been booming.

112

00:10:07,130 --> 00:10:08,820

And now it's everywhere.

113

00:10:10,627 --> 00:10:21,937

And we'll get back to the issues with evaluating originality uh and valuable later on.

114

00:10:21,937 --> 00:10:35,028

But let's get into the reason I found you in the first place, which was an article about

um a thing called MIDI GPT.

115

00:10:36,529 --> 00:10:37,658

Could you tell?

116

00:10:37,658 --> 00:10:42,170

us about what that is and why you have built it.

117

00:10:43,308 --> 00:10:44,018

Right.

118

00:10:44,018 --> 00:10:51,020

So, yeah, we've been working at the Metacore CineLav for QATF on generative music systems

for a long time.

119

00:10:51,020 --> 00:11:04,624

Often those systems, uh and that's circled back to our previous conversation about

definitions as well, often those systems uh in machine learning and in AI have to do with

120

00:11:04,624 --> 00:11:10,786

training a model based on a set of data that we call the inspiration set or the training

set.

121

00:11:10,848 --> 00:11:22,144

And then what um those generative AI algorithms that we've been focusing on at the

Metacrystian Lab do is that they tend to try to imitate the data.

122

00:11:22,144 --> 00:11:36,132

And so they will generate new artifacts, so again, novel and potentially valuable uh

artifacts, let's say pieces of music, in the case of MIDI GPT, that would be classified by

123

00:11:36,132 --> 00:11:39,488

an unbiased observer, that is you and me, staying out.

124

00:11:39,488 --> 00:11:44,079

if I do an evaluation study, they will be classified as belonging to the training set.

125

00:11:44,079 --> 00:11:58,073

In other words, the goal here is uh to pass a little test uh that is to compare what the

system generates with what's in the training set.

126

00:11:58,073 --> 00:12:03,665

And if you can't make the difference, then it means my system is a good system.

127

00:12:03,665 --> 00:12:08,032

And that's really the tenet of any machine learning system.

128

00:12:08,032 --> 00:12:19,308

is to learn from data, uh the distribution of the data in order to generate new elements

that would be classified by an unbiased observer as belonging to the data set.

129

00:12:19,369 --> 00:12:25,812

And so we've been doing that uh since 2008 in the case of the Metacreation Lab.

130

00:12:25,812 --> 00:12:35,758

And then we went through all of the generation of different algorithm as machine learning

as a field, uh as progress, and as computing, which is the subtract.

131

00:12:35,758 --> 00:12:41,058

in which these machine learning algorithms work is becoming more and more powerful.

132

00:12:41,058 --> 00:12:46,918

And as we accrue more more data, then possibilities started to increase and increase.

133

00:12:46,918 --> 00:12:55,258

And back in 2018 and 2019 is when the first wave of transformers came through.

134

00:12:55,618 --> 00:13:05,133

And it appeared as an expert in the domain, someone who has been published in AI

conferences since 2001, in my case.

135

00:13:05,133 --> 00:13:13,078

uh it was clear very quickly that this breed of model and approaches was better and

stronger.

136

00:13:13,078 --> 00:13:24,387

so everyone was flabbergasted by, you know, chat GPT and the launch of uh AI agents and

bots that we can talk to.

137

00:13:24,447 --> 00:13:33,814

And then it turns out that I knew from my upbringing in AI that language is just, you

know, monophonic music in a way.

138

00:13:33,814 --> 00:13:35,294

It's a sequence of token.

139

00:13:35,294 --> 00:13:46,297

In music, we also have a temporal aspect of sequences of events, notes in general, in

different tracks and instruments, but there is uh a vertical dimension that does not exist

140

00:13:46,297 --> 00:13:47,238

in language.

141

00:13:47,238 --> 00:13:54,480

In language, we do turn taking, but we don't talk at the same time together, especially

not different sentences.

142

00:13:54,520 --> 00:14:01,822

And in music, there's this extra complexity of having horizontal dependencies and vertical

dependencies.

143

00:14:01,822 --> 00:14:02,548

And so...

144

00:14:02,548 --> 00:14:16,124

As a uh researcher in AI, it seemed like there was a really good thing to try very quickly

on to try to apply this new architecture on musical data.

145

00:14:16,124 --> 00:14:18,085

We were not the first one.

146

00:14:18,085 --> 00:14:20,256

mean, all of that happened in parallel in a way.

147

00:14:20,256 --> 00:14:23,927

We probably all have the same idea at the same time when the model came out.

148

00:14:23,927 --> 00:14:29,550

uh And I think one or two models, the model from Google...

149

00:14:29,550 --> 00:14:31,110

uh

150

00:14:31,886 --> 00:14:39,146

Music Net, I don't remember, there was one or two models that came out just before ours

because they're really quick at releasing those big companies.

151

00:14:39,146 --> 00:14:46,506

know, have engineers and PR people and all the mechanics to be the first one and make a

big splash.

152

00:14:46,506 --> 00:14:52,726

But we were among the first one, among the three first ones in the world in applying this

technology.

153

00:14:52,726 --> 00:15:00,046

And then interestingly, the way we applied it was a little bit more powerful, more generic

than others.

154

00:15:00,046 --> 00:15:01,926

And so the model has been...

155

00:15:01,926 --> 00:15:09,386

and since 2019 I've been training basically MIDI GPT transformers but the paper just came

out.

156

00:15:09,386 --> 00:15:12,173

So that's how much work it took really.

157

00:15:12,253 --> 00:15:29,675

The paper just came out uh at AAAI, at the latest edition of AAAI and been published with

all the refinements that we did add through those years and during those years we worked

158

00:15:29,675 --> 00:15:31,756

with that algorithm ourselves.

159

00:15:31,850 --> 00:15:39,256

We worked with companies to adapt that algorithm to their tool and we trained many, many

versions of that algorithm.

160

00:15:39,256 --> 00:15:46,081

One of the things that we were really interested in that is new to our algorithm is how to

control the generation.

161

00:15:47,283 --> 00:15:51,346

And so we've been doing a lot of work on that particular aspect.

162

00:15:51,346 --> 00:15:56,302

But to backtrack for a second, so MIDI GPT is just a transformer.

163

00:15:56,302 --> 00:16:06,278

uh But instead of taking textual prompts, and there's a lot of text to music models out

there and there's nothing wrong with them, there's a debate whether they're more useful or

164

00:16:06,278 --> 00:16:12,701

less useful, but in our case the prompt is made of music, is made of existing MIDI music.

165

00:16:12,701 --> 00:16:23,407

And so MIDI GPT is an infilling model that will basically generate whatever you want and

whatever you want will work, what it will generate will work with what's before, what's

166

00:16:23,407 --> 00:16:25,356

after, what's above and what's under.

167

00:16:25,356 --> 00:16:29,848

So I can say, hey, this is a song redo the bassline.

168

00:16:29,868 --> 00:16:34,230

Or hey, this is a song redo those three bars of the melody.

169

00:16:34,230 --> 00:16:37,671

Or hey, this is a song make me a continuation of those four tracks.

170

00:16:37,672 --> 00:16:40,413

Or hey, this is a song add a marimba track.

171

00:16:40,413 --> 00:16:44,905

And so what's nice with MIDI GPT, it supports all the MIDI channels.

172

00:16:44,905 --> 00:16:47,796

And so all of the MIDI instruments are known by the system.

173

00:16:47,796 --> 00:16:51,412

It actually knows what range a bass is uh using.

174

00:16:51,412 --> 00:16:56,584

It generates parts that are playable by humans, so I can actually have people interpreting

that music.

175

00:16:56,784 --> 00:17:02,207

And then it offers a lot of control, and we've been working really hard on making it

controllable.

176

00:17:02,207 --> 00:17:15,192

um And I can go on and explain why that is the case, but it's mostly because of our back

and forth and exploration with musicians themselves and with uh software companies that

177

00:17:15,192 --> 00:17:19,662

make software for musicians that we felt that they needed to be...

178

00:17:19,662 --> 00:17:28,758

ah more addition to this model before we release it publicly and now it's available on

Hugging Face and there's been some good updates.

179

00:17:30,647 --> 00:17:46,278

So if I got that right, MIDI GPT works in a way that you feed it uh MIDI, your own

playing, and then it kind of searches its training, or it has a foundation of training

180

00:17:46,278 --> 00:17:57,245

set, which then references to find some structures or patterns, so as to make a coherent

musical result.

181

00:17:57,245 --> 00:17:59,306

um

182

00:17:59,607 --> 00:18:05,311

that could be combined both horizontally or vertically with the input.

183

00:18:05,351 --> 00:18:10,054

you could harmonize a melody for instance.

184

00:18:10,188 --> 00:18:11,599

Yeah, exactly.

185

00:18:11,781 --> 00:18:22,496

So you can be like, hey, this is a melody, make me a polyphonic piano part uh or an organ

or a guitar and then a bass line and a drum for it.

186

00:18:22,497 --> 00:18:23,305

No problem.

187

00:18:23,305 --> 00:18:28,239

it restricted in terms of how many bars you get input or output?

188

00:18:28,462 --> 00:18:39,562

No, the model itself has an attention window and we have different versions depending on,

but we like it to run on a CPU, so the current model that works everywhere has about 16

189

00:18:39,562 --> 00:18:51,702

bars to 32 bars of attention and what it will do is there's a windowing mechanism, so if I

load a song from Stevie Wonder or Michael Jackson and I re-sample the entire bassline,

190

00:18:51,702 --> 00:18:55,982

then it's gonna go and slide through like that.

191

00:18:55,982 --> 00:18:56,862

Yeah.

192

00:18:58,581 --> 00:18:59,074

So.

193

00:18:59,074 --> 00:19:09,401

idea by prompting the system with music was that artists are in the business of making

music and often they have their own musical material.

194

00:19:09,401 --> 00:19:17,065

And explaining in text what style you want and all of the detail of what you want is

actually really hard to do.

195

00:19:17,185 --> 00:19:25,150

And there's now a lot of studies showing that text prompting is not necessarily a great

human computer interface.

196

00:19:25,262 --> 00:19:27,124

uh in a number of domains.

197

00:19:27,124 --> 00:19:36,096

uh And so we try this alternative in which you use the music material itself as part of

the prompt.

198

00:19:37,542 --> 00:19:43,644

It's often a lot easier to come up with a great idea than to continue it or finish it.

199

00:19:43,644 --> 00:19:46,825

So it might be a good companion in that sense.

200

00:19:48,166 --> 00:20:01,584

So I have this project with a few music students where we do a kind of a hackathon because

I teach a subject in composition and arranging ah and they're

201

00:20:01,584 --> 00:20:06,999

Their task is to try to cheat on these assignments using AI tools.

202

00:20:07,080 --> 00:20:22,116

So could you imagine using MIDI GPT to cheat on an assignment, for instance, uh making uh

a ABA uh jazz song in that format?

203

00:20:22,116 --> 00:20:24,678

Could you use MIDI GPT for that purpose, you think?

204

00:20:25,122 --> 00:20:26,792

Yeah, 100%.

205

00:20:26,792 --> 00:20:32,025

Yeah, 100 % you can use MIDI GPT on any composition assignments.

206

00:20:32,365 --> 00:20:39,268

And yeah, and so that's really entering the realm of computer assisted uh creativity.

207

00:20:40,008 --> 00:20:46,071

And that's one of the reasons why we like the idea that, of course, with MIDI GPT, you can

generate from scratch.

208

00:20:46,171 --> 00:20:54,124

But if you generate from scratch, this system is trained on a dataset called MetaMIDI,

which we also released at the MetaCreationLab.

209

00:20:54,124 --> 00:20:57,766

And it's the largest data set of MIDI data in the world.

210

00:20:57,766 --> 00:21:06,319

It's uh at the moment 1.6 million MIDI tracks um with the multi-track pieces uh of MIDI.

211

00:21:06,400 --> 00:21:09,131

And so it knows all of the music, every music.

212

00:21:09,131 --> 00:21:11,362

It knows way more music than a human can.

213

00:21:11,362 --> 00:21:17,284

And so if you generate from scratch, you might end up with medieval classical music.

214

00:21:17,725 --> 00:21:20,706

And even if you can't tell the system, I want a pop song.

215

00:21:20,706 --> 00:21:24,227

with like four instruments and this and that.

216

00:21:24,227 --> 00:21:29,729

Like describing in Word again with text is so difficult to get it right.

217

00:21:29,889 --> 00:21:37,282

And a lot of the system out there, and AI is for the most part a for-profit Californian

enterprise, right?

218

00:21:37,282 --> 00:21:43,874

You can name, mean, software that people use are not uh typically from...

219

00:21:44,586 --> 00:21:47,418

Europe or anywhere else in the US or China.

220

00:21:47,418 --> 00:21:50,990

get those, you have the Asian stack and the American stack.

221

00:21:51,230 --> 00:21:58,464

And those systems are made for the layman and they're made for everyone to use, but

therefore they're really not specialized.

222

00:21:58,464 --> 00:22:05,238

In other words, the system that we are about, we are about them because of the massive PR

machine of those big companies.

223

00:22:05,238 --> 00:22:09,641

And those big companies, they want to have everyone as a user, like for charge GPT.

224

00:22:09,641 --> 00:22:11,642

So Udio, Suno.

225

00:22:12,034 --> 00:22:17,316

those plant-based systems, they're great and we can't remove that from them, they do work.

226

00:22:17,316 --> 00:22:26,208

But at the same time, if you're a musician and you have a specific idea of the aesthetic

you want, you typically will never get there with those systems.

227

00:22:26,208 --> 00:22:32,220

You will have to adjust that from changing over and over and over until you get what you

want.

228

00:22:32,220 --> 00:22:39,222

So we took the reverse approach, which is like, no, a musician has music, that music is

actually telling what style it is.

229

00:22:39,374 --> 00:22:49,531

uh It's telling you know the density of note the size of note is telling even the the

micro timing of the interpretation the velocity that is used the dynamic all of the things

230

00:22:49,531 --> 00:22:58,198

that you would be well at pressed you know to describe with words all of that is actually

given by the fact that I don't want you to start from scratch I want you to start from

231

00:22:58,198 --> 00:23:07,106

your music or someone else music if you're less of a you know if you if ethics isn't a

problem for you you do what you want right, but uh

232

00:23:07,106 --> 00:23:19,590

But the truth is, by prompting with music, we find that we go way uh faster to the actual

content that the creator is looking at and wants to listen to.

233

00:23:19,590 --> 00:23:23,851

And then we work more on, hey, what are the new features that no software has?

234

00:23:23,851 --> 00:23:25,522

Like things like batch generation.

235

00:23:25,522 --> 00:23:28,773

Hey, I want to make a variation of this.

236

00:23:28,773 --> 00:23:33,334

And a lot of professional musicians, Stana, you probably know that.

237

00:23:33,422 --> 00:23:38,242

They do pastiche, they do style imitation, they do what my algorithm does for a living.

238

00:23:38,242 --> 00:23:49,542

They receive a call from BBC on Monday, there is this documentary, it's going into 64

countries, you need to remake the soundtrack because we're not gonna pay rights for those

239

00:23:49,542 --> 00:23:50,602

countries.

240

00:23:51,722 --> 00:23:55,622

And so a lot of musicians, they do for a living, music style imitation.

241

00:23:55,622 --> 00:24:03,402

And so the system would really allow them to take that track, the, okay, we're gonna

resample half of the bars of the entire...

242

00:24:03,628 --> 00:24:14,551

music piece on all of the track except for the drum and we're going to do that 60 times or

100 times or 200 times and I want you to rank all those generations based on their

243

00:24:14,551 --> 00:24:16,038

proximity to the original.

244

00:24:16,038 --> 00:24:27,175

I listen to them one by one and as soon as it sounds far enough from the original to be

different but with a good style that I'm interested in, I'll take that and I start working

245

00:24:27,175 --> 00:24:27,965

on it.

246

00:24:28,365 --> 00:24:32,106

know, use cases like that that are real use cases actually.

247

00:24:32,106 --> 00:24:34,628

is really what we were aiming for from the get-go.

248

00:24:34,628 --> 00:24:44,376

Being musicians ourselves in the lab, we really believe in my lab, which is a bit

different from a lot of computer science departments and labs that I know, we believe in

249

00:24:44,376 --> 00:24:45,651

participatory design.

250

00:24:45,651 --> 00:24:53,644

So this idea that when we design a system and it starts by designing an algorithm, we want

the end user to be there in the room.

251

00:24:53,644 --> 00:25:01,292

And to do that, we work with artists ourselves, second person view on the system, and we

work ourselves as artists.

252

00:25:01,292 --> 00:25:04,064

your first-person view on the system.

253

00:25:04,104 --> 00:25:05,085

And that's important to us.

254

00:25:05,085 --> 00:25:10,730

And I think that's what allows us to design a little bit different type of systems.

255

00:25:10,730 --> 00:25:20,878

And so far, in interaction with the industry, it's been quite uh successful and fruitful

because I think the industry understands uh that what we're looking at and what we

256

00:25:20,878 --> 00:25:30,476

actually need for real creators that make a living out of making music is different from

what Californian AI has to offer.

257

00:25:30,687 --> 00:25:34,935

and how generic the tools are and how we need things that are bit more specialized.

258

00:25:34,935 --> 00:25:40,436

And so that's really the direction we took and how we went that way.

259

00:25:40,658 --> 00:25:43,219

Hmm, yeah, it makes sense.

260

00:25:43,320 --> 00:25:44,900

Different...

261

00:25:49,644 --> 00:25:59,230

Of course, musicians look after entirely different things and an entirely different level

of detail um in their workflows.

262

00:25:59,230 --> 00:26:06,084

Have you been surprised by any feedbacks you've received from the users this far?

263

00:26:07,500 --> 00:26:10,150

Yeah, well, we've been really happy.

264

00:26:10,150 --> 00:26:11,452

There's been albums made.

265

00:26:11,452 --> 00:26:16,134

uh There's uh all sorts of projects ongoing and passed.

266

00:26:16,234 --> 00:26:21,216

With it, we qualified for the final of the AI Song Contest last year.

267

00:26:21,216 --> 00:26:27,199

uh And there was some of my students on stage during uh that presentation.

268

00:26:27,199 --> 00:26:32,672

And so we've been really happy to see that there's actual real world artistic application

of the systems.

269

00:26:32,672 --> 00:26:37,484

You know, in AI, very, very often we make the work with graduate

270

00:26:37,484 --> 00:26:41,748

students, we publish it and then it's ending up on a shelf.

271

00:26:41,748 --> 00:26:54,818

Or even more often than Google, Facebook, Meta, know, all of those companies would take it

on, develop it and then push something and then the world will discover it then, right?

272

00:26:54,818 --> 00:27:04,565

So we move to a world where um things have changed a lot in terms of research because

before the GAFA

273

00:27:04,850 --> 00:27:15,218

the best in any domain, especially computer science and the type of things we do in AI,

the best would always be a professor and his her PhD students.

274

00:27:15,218 --> 00:27:23,565

uh And nowadays, you know, we have a number of companies that for some reason is still

unclear and there's still conversation around it, started doing research and started

275

00:27:23,565 --> 00:27:26,708

publishing in the same place where my student publish, right?

276

00:27:26,708 --> 00:27:29,549

Except that when they publish their paper,

277

00:27:29,902 --> 00:27:34,724

ah they have 1.4 billion followers, let's say, if you say Google, for example.

278

00:27:34,724 --> 00:27:38,926

So doesn't matter if that paper is good or bad or new or not new.

279

00:27:38,926 --> 00:27:46,730

The world is going to take it for, uh this is the first time, you know, this is the first

people doing that, and those are the best, and et cetera, et cetera.

280

00:27:46,730 --> 00:27:51,852

Turns out three years earlier, maybe my student came up with a way better algorithm for

the same, same thing.

281

00:27:51,852 --> 00:27:56,434

And they've been, you know, copying it, improving it through engineering, and then

publishing it.

282

00:27:56,434 --> 00:27:58,275

And it's very often the case like that.

283

00:27:58,275 --> 00:28:00,018

So the world has really changed.

284

00:28:00,018 --> 00:28:09,690

often my students say that it's unfair because there's a big difference between doing the

same, work in my lab and doing the work as an intern in Google.

285

00:28:09,690 --> 00:28:14,731

The visibility of the work would be the same, but the visibility of the work would be

completely different.

286

00:28:14,731 --> 00:28:27,054

And so there's been, not just in research, but of course, in the software industry and in

application, there's been a very big asymmetry of information uh and power.

287

00:28:27,054 --> 00:28:28,716

uh

288

00:28:28,716 --> 00:28:32,820

because a lot of those companies are so big that they sort of...

289

00:28:32,820 --> 00:28:34,886

um

290

00:28:34,990 --> 00:28:44,830

put everything else in the shade and then it's really hard to exist outside of the realm

of those very big players.

291

00:28:44,830 --> 00:28:57,230

And so that's something that we've been trying to do and work on with our little ecosystem

here that involve both artists and communities that are hybrid basically, like the AIMC

292

00:28:57,230 --> 00:29:02,610

conference, the AI Music Creativity Conference, which will be in Belgium in September this

year.

293

00:29:02,610 --> 00:29:04,590

It was at Oxford last year.

294

00:29:05,244 --> 00:29:16,922

next year and then we're looking after that maybe moving it to North uh America and that's

the type of conference in which we have a mix of academics and actual musicians and

295

00:29:16,922 --> 00:29:31,222

composers as you know take their time and work together on tools and in those you know

forums that I attend and I help uh animate there is a strong you know uh critic of the

296

00:29:31,222 --> 00:29:35,165

mainstream uh AI and what the GAFA are doing

297

00:29:35,165 --> 00:29:38,429

and the way they are doing it.

298

00:29:40,191 --> 00:29:48,954

I guess that's a bit like the rest of the world in terms of uh visibility and influence

being the currency.

299

00:29:48,954 --> 00:29:55,037

That's how it's like for artists or yeah I mean influencers kind of rule the world in that

way.

300

00:29:55,457 --> 00:29:57,318

So if you have a big following.

301

00:29:58,190 --> 00:30:02,350

also a reality in research now, and that's really new.

302

00:30:02,350 --> 00:30:07,870

And even when I applied for grants, it used to be that what matters was excellence.

303

00:30:08,214 --> 00:30:09,475

and the quality of the work.

304

00:30:09,475 --> 00:30:20,220

Now they have all those measures called impact, not just the citations, which is the

normal regular type of impact we use to measure in academia, but now it's like how many

305

00:30:20,220 --> 00:30:25,552

followers do you have and how many people have downloaded or tried your software?

306

00:30:25,552 --> 00:30:32,886

If I give a demo uh software for people to try, it's going to be hard for me to compete

with Google, you know what I mean?

307

00:30:32,886 --> 00:30:38,028

Because Google, going to put it on the front page of their search engine

308

00:30:38,388 --> 00:30:41,732

and it's going to be really hard to exist that way.

309

00:30:41,732 --> 00:30:48,000

And so, the ecosystem is rapidly changing, but I'm always one to try to...

310

00:30:48,000 --> 00:30:50,158

uh

311

00:30:50,158 --> 00:30:53,280

to reaffirm the fact that there is a real diversity.

312

00:30:53,280 --> 00:30:57,722

And I'm not saying that what the GAFA are doing is not good nor useless.

313

00:30:57,722 --> 00:31:05,285

know, it's having generic agents and generic tools that you can prompt is really useful

for a lot of people, including myself.

314

00:31:05,285 --> 00:31:12,308

But there's also a space in which for professional, for experts, there's different types

of tools that needs to exist.

315

00:31:12,308 --> 00:31:20,352

And those tools won't have billions of users because there's you know, millions of people

who actually want to compose music from the ground up.

316

00:31:20,352 --> 00:31:24,816

And so those big companies have no interest to serve those niche markets.

317

00:31:24,816 --> 00:31:27,598

There's not enough money basically for them to go there.

318

00:31:27,658 --> 00:31:36,826

And so there's a real space there of opportunity for people like you and I to exist and to

do interesting things.

319

00:31:38,556 --> 00:31:38,829

So.

320

00:31:38,829 --> 00:31:40,490

m

321

00:31:43,626 --> 00:31:57,926

In the world of AI and music these days, we kind of look past the major commercial

players, who are the most interesting guys at the moment, do you think?

322

00:31:58,306 --> 00:32:04,466

What kind of communities or institutions or who's making the coolest stuff right now?

323

00:32:05,902 --> 00:32:11,722

Well, it's mostly happening, think, I feel, in the academic world.

324

00:32:11,722 --> 00:32:12,982

So you get different pools.

325

00:32:12,982 --> 00:32:17,582

You get IRCAM in Paris, who has always done really interesting things, and they continue

doing it.

326

00:32:17,582 --> 00:32:23,322

I work with Jérôme Nica right now, because we have also a thread of work in my lab on

musical agents.

327

00:32:23,322 --> 00:32:27,102

And that, again, has never been, and so far is not, a mainstream activity.

328

00:32:27,102 --> 00:32:29,502

And therefore, there's been...

329

00:32:29,502 --> 00:32:36,571

It's been very active and very interesting, but there's been none of the GAFA presence in

that space.

330

00:32:36,571 --> 00:32:42,498

Although Google just released a paper inspired by some of our work uh using transformers

for musical agents.

331

00:32:42,498 --> 00:32:46,342

um

332

00:32:46,520 --> 00:32:48,851

But yes, there is the academic world.

333

00:32:48,851 --> 00:33:01,817

There's Queen Mary in London, which is possibly the main center for computer music and

anything music and computing and AI related uh is the largest concentration on earth.

334

00:33:01,817 --> 00:33:09,151

And then there is, know, humbly on our side, the Metacreation Lab where we're doing also

pretty well and a few others.

335

00:33:09,151 --> 00:33:12,542

And so we gather in those communities such as AIMC.

336

00:33:12,542 --> 00:33:16,554

There's NIME, the New Interface for Musical Expression is a conference.

337

00:33:16,554 --> 00:33:28,467

that is about more human uh machine uh interactions and how to design that in the context

of what's happening right now with music, musical tools and AI.

338

00:33:28,467 --> 00:33:32,586

And there's plenty of other uh conferences and that's really where we live.

339

00:33:32,586 --> 00:33:38,809

I also like to have a public facing uh part of the work at the lab.

340

00:33:38,809 --> 00:33:40,569

So we do play music ourselves.

341

00:33:40,569 --> 00:33:43,870

I will be playing at Mutec festivals this year uh in...

342

00:33:43,870 --> 00:33:45,830

uh

343

00:33:45,962 --> 00:33:56,330

in Montreal at the end of August, it's the largest electronic music festival in Canada and

we're really happy we'll be playing a piece called Revival in which we, in a sort of uh

344

00:33:56,330 --> 00:34:06,397

cheeky way, we train musical agents on the music of dead composers and we explore this

idea of digital afterlife but for music composition.

345

00:34:07,198 --> 00:34:13,021

It's very trendy, there's a lot of companies now making agents that are going to survive

you and learn from your data and all that.

346

00:34:13,021 --> 00:34:17,663

So we're interested in exploring those concepts in the context of music.

347

00:34:17,663 --> 00:34:20,364

Concretely, I always wanted to play with David Tudor.

348

00:34:20,364 --> 00:34:32,670

He's dead, so I'm going to make an AI agent of him, for example, and play on stage with an

agent that has learned from the material of those dead musicians, both the sound material

349

00:34:32,670 --> 00:34:34,691

but also the temporal dynamic.

350

00:34:34,691 --> 00:34:36,752

And those agents are capable of

351

00:34:36,912 --> 00:34:47,869

on stage in live situation reason and be like given what Stenart just played and what

Philippe just played and what I just played what should I play next to sound like David

352

00:34:47,869 --> 00:35:02,199

Tudor right so those type of style imitation in live on stage we present that but that's

really um what we try to do is to link the academic community within an audience and then

353

00:35:02,199 --> 00:35:06,702

the artistic community and try to serve those communities the best we can because we

354

00:35:06,702 --> 00:35:09,514

are in a way part of those communities ourselves, right?

355

00:35:09,514 --> 00:35:16,881

And so we feel really, ah we feel that uh we're sincere in our effort.

356

00:35:16,881 --> 00:35:24,707

ah And then we work with companies, Steinart, we work with Steinberg, it's been a number

of years that they're funding our research.

357

00:35:24,707 --> 00:35:31,033

ah But so far nothing with Steinberg, with Teenage Engineering, ah with...

358

00:35:31,033 --> 00:35:32,934

uh

359

00:35:33,098 --> 00:35:41,534

audio kinetics in Montreal, no product has been released and that has to do more so with

the fact that all of our systems are data driven.

360

00:35:41,796 --> 00:35:44,239

And when we talk about...

361

00:35:45,024 --> 00:35:46,175

MIDI GPT.

362

00:35:46,175 --> 00:35:57,661

So it's trained on MetaMIDI and MetaMIDI is acquired because I'm a researcher and on the

fair use in Canada for research purposes, I can scrub the internet, right?

363

00:35:57,661 --> 00:36:01,603

There's no way I would have gotten so many MIDI files otherwise.

364

00:36:01,964 --> 00:36:08,888

And then it used to be that you scrub the internet to make a data set, you train a model,

the model doesn't have the data.

365

00:36:08,888 --> 00:36:09,628

If it's not...

366

00:36:09,628 --> 00:36:12,522

um

367

00:36:12,522 --> 00:36:21,811

if it's trained properly, it's not memorizing the data, and it generates things that I can

show are not plagiarizing the data.

368

00:36:21,811 --> 00:36:26,797

And we show that, of course, because the company would not even consider the model

otherwise.

369

00:36:26,797 --> 00:36:30,100

And then it used to be the case that they can pack the model and sell it.

370

00:36:30,100 --> 00:36:33,083

And there's still a number of companies doing that, like Suno, Udio.

371

00:36:33,083 --> 00:36:35,365

And now you have an ecosystem with two types of players.

372

00:36:35,365 --> 00:36:36,364

Players...

373

00:36:36,364 --> 00:36:44,476

who are either very big, like the gaffer or like Sunoo or Udyo, smaller players that just

do AI.

374

00:36:44,476 --> 00:36:51,796

And so for them, it's a break it or make it sort of game, and they get lawsuits and we see

what's going to happen.

375

00:36:51,796 --> 00:36:52,257

I'm actually...

376

00:36:52,257 --> 00:36:53,126

oh

377

00:36:53,126 --> 00:37:01,853

witness expert in the federal court in the US on some of those cases, so I can't talk too

much about it, but there's those companies who will try it, you know, so they have one

378

00:37:01,853 --> 00:37:08,718

product and this product is based on AI, so they're gonna try all their best, maybe

they're gonna get sued, maybe they win, maybe they lose.

379

00:37:08,718 --> 00:37:15,142

And then you have companies I work with, like TN Engineering or like Steinberg, they have

plenty of products.

380

00:37:15,583 --> 00:37:19,885

Cubase has a million license and hundreds of functionalities.

381

00:37:20,118 --> 00:37:28,158

They don't want to damage their reputation and to take a risk of losing everything because

of one AI feature.

382

00:37:28,812 --> 00:37:30,922

no matter how good that AI feature is, right?

383

00:37:30,922 --> 00:37:39,324

And so for those guys, and there's a lot of companies in the world like that in every

domain, they're just waiting to see what the legislation is gonna really look like and how

384

00:37:39,324 --> 00:37:40,344

it's gonna be applied.

385

00:37:40,344 --> 00:37:52,607

And it's been a real pity that our legislators, know, governance in general in the world

is a little bit broken, but in this particular case, it has been so slow to move, and

386

00:37:52,607 --> 00:37:58,688

therefore the GAFA and all the people who are much more like, we do this and that's all we

do, we can take the risk,

387

00:37:58,766 --> 00:37:59,436

you

388

00:37:59,436 --> 00:38:09,711

got such a lead, such a lead in that adventure that it's very complicated now for regular

software company who existed before this AI craze.

389

00:38:09,711 --> 00:38:12,792

Where here it's really hard for them to push those product.

390

00:38:12,792 --> 00:38:21,516

And they're still in a way all of the company I work with, we do the research together,

we're still preparing things in the background, we're testing things, uh we're refining

391

00:38:21,516 --> 00:38:29,600

designs, but they can't release anything because uh the legal team is like, this is not uh

an ecosystem in which

392

00:38:29,600 --> 00:38:31,661

we want to take that risk.

393

00:38:31,701 --> 00:38:42,865

So that's where we are right now with some of the divide between what people are exposed

to, typically Californian AI, and what's actually existing all across the world, not just

394

00:38:42,865 --> 00:38:44,676

here in Canada, but also in Europe.

395

00:38:44,676 --> 00:38:56,270

That's why uh it's hard for people to name AI and software coming from Europe because they

don't really exist commercially for the most part.

396

00:38:56,572 --> 00:38:58,903

Hmm.

397

00:38:58,943 --> 00:39:03,286

So there was a lot of stuff to unpack in what you just said.

398

00:39:03,286 --> 00:39:13,191

em So among other things, you mentioned that you wanted to train an AI agent on a dead

artist.

399

00:39:13,191 --> 00:39:16,332

That kind of raises some ethical questions.

400

00:39:16,413 --> 00:39:23,917

And you're also talking about all the pending lawsuits ah and slow legislation.

401

00:39:23,917 --> 00:39:26,238

And I'm wondering um both.

402

00:39:26,704 --> 00:39:36,447

both when it comes to the ethical viewpoint on using dead musicians music as a training

set.

403

00:39:36,447 --> 00:39:44,518

em How do feel that fits into your ethical view of fair use?

404

00:39:44,866 --> 00:39:46,208

Yeah, well, exactly.

405

00:39:46,208 --> 00:39:51,953

And we do it uh in an artistic context because it's provocative.

406

00:39:51,953 --> 00:39:59,041

It's obvious that we won't do it with Beyonce or Michael Jackson because first Beyonce is

not dead, but Michael Jackson is dead.

407

00:39:59,041 --> 00:40:02,624

We won't do it on Michael Jackson because we probably would get sued real quick.

408

00:40:02,624 --> 00:40:04,806

m

409

00:40:04,812 --> 00:40:16,592

Yeah, on the ethical side, uh of course, yeah, this is really to have that discourse that

we also make that type of artworks uh because we do believe that the question of the

410

00:40:16,592 --> 00:40:18,893

digital afterlife is a real one.

411

00:40:19,334 --> 00:40:23,657

And then, you know, increasingly a digital footprint.

412

00:40:24,078 --> 00:40:28,938

you want it or not is increasing because everything we do is computer mediated.

413

00:40:29,038 --> 00:40:35,278

And so there is already the case that Facebook exploits your data while you're alive.

414

00:40:35,278 --> 00:40:35,838

And guess what?

415

00:40:35,838 --> 00:40:38,358

They'll continue exploiting your data while you're dead.

416

00:40:38,678 --> 00:40:42,098

And you're not going to be removed from the database, and et cetera, et cetera.

417

00:40:42,138 --> 00:40:50,138

So want it or not, this question is a very, very,

418

00:40:51,864 --> 00:40:53,215

current one.

419

00:40:53,215 --> 00:41:01,443

And so with this artwork in particular, we really put that to the forefront by being

explicit and saying, hey, this is what we're doing.

420

00:41:01,443 --> 00:41:06,567

But then I can link it to a lot of things that are already been done without people being

explicit about it.

421

00:41:06,567 --> 00:41:09,350

So that's bringing the discourse forward.

422

00:41:09,350 --> 00:41:19,016

Then concretely and a bit more technically in terms of ethics, well, we can in some case

get the rights of the

423

00:41:19,016 --> 00:41:25,229

the stakeholders of people who have uh inherited the rights of those composers.

424

00:41:25,229 --> 00:41:29,862

In some cases, uh the work is in public domain, right?

425

00:41:29,862 --> 00:41:34,786

So for example, I'm training AI on some uh Wagner...

426

00:41:34,786 --> 00:41:36,907

Wagnerian music, some music from Wagner.

427

00:41:36,907 --> 00:41:40,651

uh He died more than 70 years ago.

428

00:41:40,651 --> 00:41:52,601

And so in most legislations, there's differences between countries, which is another thing

with those type of notions, which is quite incredible, is that regulating AI is hard, but

429

00:41:52,601 --> 00:42:04,460

regulating AI in a world that doesn't have, unfortunately, the capacity to do worldwide uh

regulation that are concerted and harmonized.

430

00:42:04,558 --> 00:42:09,700

uh is sort of the dramatic situation we're in.

431

00:42:09,700 --> 00:42:15,663

And so this is also another part of the conversation we want to have is, you know, where

are you trying that model?

432

00:42:15,663 --> 00:42:16,983

Where are the servers?

433

00:42:16,983 --> 00:42:18,264

Where is the data?

434

00:42:18,264 --> 00:42:23,966

All of that will change completely the legal, uh legal status of what we do.

435

00:42:23,966 --> 00:42:31,460

uh And then when it comes to the ethics, there's other questions such as the legitimacy of

the work.

436

00:42:31,460 --> 00:42:32,250

So...

437

00:42:32,329 --> 00:42:43,571

A lot of the arguments of companies defending themselves, being attacked in justice right

now for scrapping data from the internet, they say, well, music is a socially constructed

438

00:42:43,571 --> 00:42:44,491

phenomena, right?

439

00:42:44,491 --> 00:42:53,793

No one will be born, no human will be born on an island alone and end up with composing a

symphony in their lifetime, right?

440

00:42:53,793 --> 00:42:58,316

We are musicians and we do what we do because there was musicians before us doing what

they do.

441

00:42:58,316 --> 00:43:00,667

and we are inspired by it, right?

442

00:43:00,667 --> 00:43:06,180

And I can listen to the radio, get inspired by a piece and write a piece that is sort of

related but different.

443

00:43:06,180 --> 00:43:17,637

How is that different for my neural network listening to the radio, except all of the

radio, very, quickly, in other words, know, eating all that data, getting inspired by it

444

00:43:17,637 --> 00:43:27,372

and generating something that is arguably not, ah in that case, you if I imitate the work

of David Tudor, that is not the work of David Tudor.

445

00:43:27,372 --> 00:43:34,565

that is original, that there's no copyright issue, but it's been inspired by the work of

David Tudor.

446

00:43:34,585 --> 00:43:36,266

What is ethically wrong with that?

447

00:43:36,266 --> 00:43:39,167

This seems to be what humans are actually doing, right?

448

00:43:39,167 --> 00:43:49,182

And so we want to go a little bit deeper into that discourse whereby right now, and

because people are afraid and scared, and it's true because some people are getting sued,

449

00:43:49,182 --> 00:43:54,574

some companies are getting sued, so everyone is like, ooh, if you use data, if it's not

your data, then...

450

00:43:54,778 --> 00:43:55,898

you probably shouldn't.

451

00:43:55,898 --> 00:44:02,140

But the truth is then that starts a new one, which is who can afford uh data, right?

452

00:44:02,140 --> 00:44:14,143

And if it is the case that you need to be able, like big American company, to afford

entire catalog of music and pay a massive amount of money to be able to train and

453

00:44:14,143 --> 00:44:22,505

eventually use and exploit those models, then again, you create another type of

inequality, which is inequality to access to data, right?

454

00:44:22,766 --> 00:44:24,480

Or you see, so...

455

00:44:24,480 --> 00:44:30,794

So we do that in an artistic context because we think it's harmless, right?

456

00:44:30,794 --> 00:44:36,918

uh It's not like the family of David Tudor is making a fortune and they're gonna sue us.

457

00:44:37,199 --> 00:44:40,081

We love the aesthetic, you we're earnest in our approach.

458

00:44:40,081 --> 00:44:44,264

We really think it's a strong art project and the output is really, really good.

459

00:44:44,264 --> 00:44:47,616

It has obviously nothing to do with the actual work of the artist itself.

460

00:44:47,616 --> 00:44:53,130

It's our work, you know, and but part of it is that we use that data.

461

00:44:53,254 --> 00:45:02,719

And so we do that consciously and it's part of our practice and our process and it's also

triggering discussions, right?

462

00:45:02,719 --> 00:45:05,540

And that's why we do it.

463

00:45:05,540 --> 00:45:17,886

But so far at least I looked at every angle of it and I've not been proven guilty of

anything unethical that I know of.

464

00:45:19,799 --> 00:45:24,872

What would be the counter argument you think if you were to play Devil's Advocate?

465

00:45:25,166 --> 00:45:26,746

I think there's plenty.

466

00:45:26,746 --> 00:45:28,226

Some of them would be cultural.

467

00:45:28,226 --> 00:45:33,785

There's cultures in which, for example, pictures of dead people are not okay.

468

00:45:33,966 --> 00:45:35,926

Even for memory purposes or things like that.

469

00:45:35,926 --> 00:45:48,386

There's religions, like the Muslim religion, which is the fastest growing ideology on

earth and the second largest ideology after the Christian one, religion.

470

00:45:48,946 --> 00:45:53,026

And so in that region, for example, explicit representation.

471

00:45:53,150 --> 00:45:55,672

of human figures is not okay.

472

00:45:55,672 --> 00:46:06,788

So the statues of data, uh depending on cultures, depending on the way death is being

treated, depending on the way remembrance is being treated, depending on the definition of

473

00:46:06,788 --> 00:46:12,682

respect that you have, the statues of data of these people would be vastly different.

474

00:46:12,922 --> 00:46:19,030

The concrete thing though is right now, uh nobody read them when they sign.

475

00:46:19,030 --> 00:46:28,394

the agreement, you know, but when you open your Facebook account, your Instagram account,

your TikTok account, your Google account, you give the entirety of the data of the

476

00:46:28,394 --> 00:46:32,726

interaction you will have with those platforms at an infinite time, right?

477

00:46:32,726 --> 00:46:41,320

So you sell your afterlife, your digital afterlife by definition, like by, you literally

sign on it when you sign in.

478

00:46:41,320 --> 00:46:47,214

And so again, you know, this artwork is a way to also raise awareness.

479

00:46:47,214 --> 00:46:49,654

to that because typically people don't think about that.

480

00:46:49,654 --> 00:46:52,334

It's not really pleasant to think about one's death.

481

00:46:52,454 --> 00:46:56,814

It's not really pleasant to think about databases with our own data in it.

482

00:46:56,874 --> 00:47:00,814

And mixing the two is not something people think about on a daily basis.

483

00:47:00,814 --> 00:47:06,194

People are really busy with the present, their immediate future, and maybe their immediate

past.

484

00:47:06,194 --> 00:47:11,694

But they rarely think about, hey, what's going to happen to all of that in 50 years,

right?

485

00:47:12,018 --> 00:47:20,762

um So it's probably a good market idea to make a Halal Facebook or something.

486

00:47:21,944 --> 00:47:33,801

I think there's gonna be, yeah, absolutely, I think there's gonna be some platform as

those conversations evolve that are gonna start being explicit about exactly that, about

487

00:47:33,801 --> 00:47:36,724

talking about what's happening, right?

488

00:47:37,692 --> 00:47:51,029

But you mentioned also that the way humans learn and make music is uh community driven or

socially driven and the way we learn is similar to a neural network.

489

00:47:51,029 --> 00:47:58,093

And that's a common argument when it comes to defending these lawsuits.

490

00:47:58,854 --> 00:48:07,498

But do you have any opinions on how AI music legislation should look like?

491

00:48:09,042 --> 00:48:15,077

Do think it should be like completely open or should there be any boundaries?

492

00:48:16,982 --> 00:48:26,628

That's a good question, that's a complex question and the response, I want to be a little

bit nuanced here but the answer is that it depends.

493

00:48:26,628 --> 00:48:37,033

But there is surely something to be said about the protection of intellectual property and

it's almost like a political...

494

00:48:37,033 --> 00:48:38,848

uh

495

00:48:38,848 --> 00:48:50,710

a political view that people take on it, you know, from people that are extremely copy

left, aka copyright is a wrong thing, we should not bother with it, to people that are uh

496

00:48:50,710 --> 00:48:54,994

typically you move toward corporations that are extremely...

497

00:48:55,002 --> 00:49:00,385

copyright oriented with like massive legal instruments to defend uh those copyrights.

498

00:49:00,385 --> 00:49:14,541

And right now all of the lawsuits that I know of most, no, not all, most of the lawsuits I

know of in the US and in Canada have to do with big stakeholders attacking AI companies.

499

00:49:14,541 --> 00:49:16,212

And so it's universal.

500

00:49:16,212 --> 00:49:17,313

It's Sony.

501

00:49:17,313 --> 00:49:23,995

It's really big stakeholders attacking AI companies because of the monetary dimension of

it.

502

00:49:24,076 --> 00:49:24,996

Sometimes

503

00:49:24,996 --> 00:49:27,659

more rarely it is artists themselves, right?

504

00:49:27,659 --> 00:49:32,905

And I'm more interested in those cases because I think it's more earnest, it's less of a

game of money then.

505

00:49:32,905 --> 00:49:40,132

And I know personally some artists, visual artists, painters that used to make a living

painting...

506

00:49:40,132 --> 00:49:42,306

um

507

00:49:42,306 --> 00:49:52,441

portraits that people would ask them to make of themselves and then they can't earn a

living now because now you can do you know image to image with a few prompt and get a

508

00:49:52,441 --> 00:49:54,712

really good uh portrait of you.

509

00:49:54,712 --> 00:50:00,194

Of course, it's not made of old painting on a canvas so it's actually completely different

but it looks really great.

510

00:50:00,194 --> 00:50:09,369

You get to steer the system towards something you like and it's sort of flattering and

that and so there's a real uh impact of those generative AI systems on the market.

511

00:50:09,369 --> 00:50:12,240

In France for example, 35 % of

512

00:50:12,240 --> 00:50:17,844

uh of the market for graphic design has been eaten by those systems.

513

00:50:17,844 --> 00:50:22,807

And so there's literally 35 % less money to go around if you're a graphic designer in

France.

514

00:50:22,807 --> 00:50:25,329

And so a lot of young graphic designers can't make it anymore.

515

00:50:25,329 --> 00:50:31,933

So the impact on job displacement and all that is already real of all those system.

516

00:50:31,933 --> 00:50:39,818

So legislation and some protection, at least in the short term, I think is the right way

to go.

517

00:50:39,818 --> 00:50:42,230

uh

518

00:50:42,230 --> 00:50:43,391

the right tools, right?

519

00:50:43,391 --> 00:50:50,846

How as an artist can I really decide to give my data and make sure that I'm gonna be paid

for it?

520

00:50:50,846 --> 00:50:56,329

Or how can I make sure that I decide not to give my data and then no one can actually pump

it?

521

00:50:56,530 --> 00:50:58,051

No one can scrap it.

522

00:50:58,051 --> 00:51:00,752

So technically we're not quite there yet.

523

00:51:00,752 --> 00:51:04,355

Legislative legislation, we're not quite there yet.

524

00:51:04,355 --> 00:51:07,207

But there is those niche market appearing.

525

00:51:07,207 --> 00:51:12,090

Some companies, for example uh Adobe, are already doing the right thing.

526

00:51:12,090 --> 00:51:25,363

like Adobe, because they own, I think it's called, with Firefly, have Giphy, they have the

largest stock pictures site in the world, and they let every creator on that stock picture

527

00:51:25,843 --> 00:51:31,584

site decide if they allow their data to be used to train AI or not.

528

00:51:31,584 --> 00:51:36,345

And if they do, then a little stream of money will come back each time the model is being

used.

529

00:51:36,345 --> 00:51:40,066

And so those new systems to retribute

530

00:51:40,238 --> 00:51:48,902

um So to clear, ah to clear, um to clear, how uh would you call that?

531

00:51:50,298 --> 00:51:56,771

to make sure that the artist is on board and then to retribute and pay back the artist.

532

00:51:56,771 --> 00:51:58,362

Those systems already exist.

533

00:51:58,362 --> 00:52:00,683

know, some companies are going that route and stuff.

534

00:52:00,683 --> 00:52:08,487

But here again, it takes a massive infrastructure and I'm afraid that only big companies

can afford to go that route for now.

535

00:52:08,487 --> 00:52:13,889

And so we need, and that's something we work very hard on in Canada, we need what we call

AI comments.

536

00:52:13,889 --> 00:52:19,982

So we need clear regulations, but we also need AI comment that is uh tools and

537

00:52:19,982 --> 00:52:33,841

approaches that allow common people to either sell the data or protect the data or to

manipulate and do that without having a team of engineers and legal experts uh at hand.

538

00:52:33,841 --> 00:52:43,222

And that last mile of some companies are doing the right thing to everyone is in a

position to be able to work and

539

00:52:43,222 --> 00:52:46,482

at the same time do the right thing, uh we're not there yet.

540

00:52:46,482 --> 00:52:47,134

We're not there yet.

541

00:52:47,134 --> 00:52:48,125

So there's quite a bit of work.

542

00:52:48,125 --> 00:53:02,894

And so in Canada, we are trying to work towards AI commands and look at the picture in an

holistic fashion and say, AI, the internet and the digital.

543

00:53:03,252 --> 00:53:05,523

is just an infrastructure.

544

00:53:05,763 --> 00:53:07,924

And infrastructures are comments.

545

00:53:07,984 --> 00:53:10,105

Roads aren't private, they're public.

546

00:53:10,105 --> 00:53:14,447

The mailing system, the post system, post Canada is public.

547

00:53:14,447 --> 00:53:18,469

And we should have an equivalent for AI.

548

00:53:18,769 --> 00:53:26,353

And if we do, then it means we would have places to store data, all sort of things will

happen that are not the case now.

549

00:53:26,353 --> 00:53:32,896

Because it would be a public infrastructure, that would follow the laws sort of by

definition.

550

00:53:33,088 --> 00:53:45,937

uh And then that way that will empower people to uh either offer the data or protect the

data and with some sort of guarantee because professional computer scientists uh would run

551

00:53:45,937 --> 00:53:48,228

public servers to do that.

552

00:53:48,228 --> 00:53:54,262

So that's one of the potential future that I can see in terms of.

553

00:53:54,720 --> 00:54:11,014

untangling and trying to answer all those complex questions is to go back to basically um

considering the internet, AI and artificial information as part of being managed by public

554

00:54:11,014 --> 00:54:11,755

services.

555

00:54:11,755 --> 00:54:21,897

It used to be the case in France before the internet that we had the Minitel, which was a

French uh situation in which everyone had a terminal.

556

00:54:21,897 --> 00:54:24,678

It was using the telephone lines at the time.

557

00:54:25,086 --> 00:54:26,947

And you would have access to a network.

558

00:54:26,947 --> 00:54:27,767

It's the Internet.

559

00:54:27,767 --> 00:54:29,508

It was an Internet of the time.

560

00:54:29,508 --> 00:54:32,689

uh Different networks would be linked together.

561

00:54:32,689 --> 00:54:44,495

And then there was this idea that you would rent it from the post office and then the

servers would be at the city level or the department or even at the state level.

562

00:54:44,495 --> 00:54:49,097

And then only computer scientists would work on those servers.

563

00:54:49,198 --> 00:54:54,640

And uh so the software will always work and you don't have to install anything.

564

00:54:54,640 --> 00:55:05,128

Right now we're back to that model, which is a decentralized model where you have client

server and everything is in the cloud, except that the cloud is not public, it's private.

565

00:55:05,909 --> 00:55:16,518

And so we're already in a situation where if every communication and every action we do is

digitally mediated, well, uh those...

566

00:55:16,620 --> 00:55:20,272

infrastructure we use, uh they're all private.

567

00:55:20,272 --> 00:55:30,587

And so all of sudden, by the very fact of using them, the data escapes you, and the future

of those data and your behavior, everything is being logged, obviously, and everything

568

00:55:30,587 --> 00:55:37,180

accepts you because everything becomes all of a sudden preempt by the private uh industry.

569

00:55:37,180 --> 00:55:46,756

So with that uh digressing too, too, too much, which I love to do from your question,

Stena, there is here what I want

570

00:55:46,756 --> 00:55:57,004

to just do is point out a vision that uh it has not always been the way it is right now

and there is nothing saying that it should be in the future the way it is right now.

571

00:55:57,004 --> 00:56:04,409

But the way it is right now uh benefits a few very big companies and no one else really.

572

00:56:04,850 --> 00:56:20,112

um So um you're saying that the digital infrastructure and cloud services should be

publicly owned or governed by the state or something like that?

573

00:56:20,694 --> 00:56:32,292

Absolutely, absolutely there should be and there is like right now a big movement in most

countries to re-nationalize a lot of things and I think the digital infrastructure is one

574

00:56:32,292 --> 00:56:36,084

of those things that people are now being increasingly worried about.

575

00:56:36,965 --> 00:56:44,529

Yes, because it's almost impossible to live without Californian AI.

576

00:56:44,878 --> 00:56:49,818

and it's almost impossible to enter the market and propose an alternative.

577

00:56:50,118 --> 00:56:58,798

And so those two facts together are a little bit problematic, especially in the

geopolitical context in which we are entering.

578

00:56:59,408 --> 00:57:02,059

Yeah, yeah.

579

00:57:02,760 --> 00:57:07,103

So there's lots of stuff to dig into there.

580

00:57:07,103 --> 00:57:16,218

I was also thinking about, since you mentioned that 35 % of the market for graphic

designers in France have been eaten up by AI.

581

00:57:16,218 --> 00:57:17,450

uh

582

00:57:19,186 --> 00:57:26,632

How do you see anything similar happening with tools like MIDI GPT?

583

00:57:27,112 --> 00:57:41,064

You mentioned earlier that you could use it in sort of a workflow as a film music composer

could in terms of imitating styles and doing that more efficiently.

584

00:57:41,625 --> 00:57:47,440

Could that also um take away some of the market opportunities?

585

00:57:50,274 --> 00:57:54,417

Well, uh possibly, but I don't think so.

586

00:57:54,417 --> 00:58:02,973

And we really took care of trying to design MIDI GPT in a way whereby it's an assistant to

the artist, but it's not a way to replace the artist.

587

00:58:02,973 --> 00:58:10,928

With Suno Audio, can, without knowing nothing, scratch generate a great track in 30

seconds.

588

00:58:10,929 --> 00:58:12,650

But with MIDI GPT...

589

00:58:12,770 --> 00:58:23,557

you need to know what to start with, and then the granularity at which it works, there's

so much, because of all those controls, so much parameters and things, you have to work,

590

00:58:23,557 --> 00:58:34,564

like it's work, and it's a tool you have to learn how to know and learn how to use, and

some people get good at it and or not, but in other words, with our systems, you cannot

591

00:58:34,564 --> 00:58:37,405

completely remove the human of the loop.

592

00:58:38,094 --> 00:58:49,275

And that's something that is important to us, to make sure that eventually the creator uh

stay uh being the human, and eventually we do computer-assisted creativity.

593

00:58:49,275 --> 00:58:51,398

ah

594

00:58:51,754 --> 00:58:58,879

and or even co-creation where we can sometimes acknowledge the system as part of this

composition.

595

00:58:58,879 --> 00:59:08,615

And in fact we did studies around that because we were really interested both for

beginners and for professionals what are the differences, what's at stake and how do they

596

00:59:08,615 --> 00:59:09,296

use the system.

597

00:59:09,296 --> 00:59:12,207

And we find that indeed they use it very differently.

598

00:59:12,648 --> 00:59:18,182

When we work with uh Ableton or Steinberg, in this case the studies take place in Cubase.

599

00:59:18,182 --> 00:59:21,093

You know, a beginner

600

00:59:21,484 --> 00:59:30,801

Typically, their problem is that they bought a music software and they want to make music,

but um they never finished a track.

601

00:59:30,801 --> 00:59:32,342

They don't know already how to do it.

602

00:59:32,342 --> 00:59:40,329

And often those music software, especially in music, are aspirational purchase where

people buy Ableton because they want to make electronic music.

603

00:59:40,329 --> 00:59:42,710

There's six million Ableton license.

604

00:59:42,710 --> 00:59:47,474

I guarantee you 95 % of those license, they never finished a track.

605

00:59:47,564 --> 00:59:48,074

Right?

606

00:59:48,074 --> 00:59:55,818

So by putting something like MIDI GPT in their hands, they will start maybe make a melody,

maybe they can generate the rest.

607

00:59:55,818 --> 01:00:01,160

They can also then generate things that they like, but they would not be able to compose

anyway.

608

01:00:01,160 --> 01:00:10,195

And by steering the system, and that's where the balance of control, you know, the one

button generate versus like all of the control being explicitly shown, beginners get lost

609

01:00:10,195 --> 01:00:14,367

if there's too much control, professional are annoyed if there's not enough control.

610

01:00:14,367 --> 01:00:17,978

So this balance is always really hard in any software to strike.

611

01:00:18,068 --> 01:00:29,474

but by manipulating and steering the system, if we do it well and if we give them the

right control, they can compose music that they like because they become the curator and

612

01:00:29,474 --> 01:00:31,045

the system becomes the composer.

613

01:00:31,045 --> 01:00:38,960

And they have this little assistant that does what they want and they can converge, and we

have evidence of that, to music that they like that they could not compose otherwise.

614

01:00:38,960 --> 01:00:43,562

So for beginners, and that's really the vast majority of the market that...

615

01:00:43,734 --> 01:00:50,696

those companies we work with would want to eat is that they want their beginners to be

able to become in a way composers.

616

01:00:50,776 --> 01:00:55,108

The composers themselves, the professionals, those guys they know how to make music.

617

01:00:55,108 --> 01:00:58,019

They actually don't even need the system, right?

618

01:00:58,019 --> 01:01:08,214

And when we study with those professionals and in that study it was published at Ishkai uh

last year, the International Joint Conference on AI.

619

01:01:08,364 --> 01:01:11,986

We define professional by people who actually make a living out of using Cubase.

620

01:01:11,986 --> 01:01:18,241

So really, there's a direct link between this use of Cubase, this software, and their

livelihood.

621

01:01:18,241 --> 01:01:20,662

And then we put this new AI system in it.

622

01:01:21,103 --> 01:01:22,944

And for them, it's really scary.

623

01:01:22,944 --> 01:01:27,367

It's been 30 years because Cubase is an old software that they use and compose in Cubase.

624

01:01:27,367 --> 01:01:33,951

And also then, you put a machine that is almost competing with them, potentially, inside

their favorite software.

625

01:01:33,951 --> 01:01:35,552

And then we observe what happens.

626

01:01:35,552 --> 01:01:38,314

And what we find is for composers, they actually love it.

627

01:01:38,370 --> 01:01:41,752

But they don't really use the output of the system at all.

628

01:01:41,752 --> 01:01:45,795

Unlike the beginners, would verbatim use the output of the system.

629

01:01:45,795 --> 01:01:48,737

The composers would be more like, what if I put a marimba here?

630

01:01:48,737 --> 01:01:50,308

Give me three examples.

631

01:01:50,308 --> 01:01:50,924

yeah, I like that.

632

01:01:50,924 --> 01:01:51,539

I'll work with that.

633

01:01:51,539 --> 01:01:53,040

And then they start composing a part.

634

01:01:53,040 --> 01:01:53,831

Because they can.

635

01:01:53,831 --> 01:01:55,242

Because that's what they do.

636

01:01:55,242 --> 01:02:01,516

Usually some of them were even surprised when we asked if they use the They were like, of

course I'm not going to use the content of the system.

637

01:02:01,516 --> 01:02:03,147

That's not my content.

638

01:02:03,147 --> 01:02:05,859

But I'm interested to generate and try.

639

01:02:05,859 --> 01:02:07,848

And so they do what-if scenario.

640

01:02:07,848 --> 01:02:16,562

So you see that two different population of users, two completely different type way of

using it with two complete different uh outcomes.

641

01:02:16,562 --> 01:02:27,627

And so we are interested in that, not just generating with MIDI GPT the best algorithm for

music composition, but also how to integrate them in existing workflows and software.

642

01:02:27,627 --> 01:02:35,564

And then studying not just usability, does the system work or is the interface the right

interface, which we do.

643

01:02:35,564 --> 01:02:40,916

of course, usability studies, but more importantly, we're interested in autorship.

644

01:02:40,916 --> 01:02:47,377

At the end, do you feel this is your music that you made with ChargeGPT, or with MIDI GPT

in this case?

645

01:02:47,698 --> 01:02:53,599

Or is it the machine music and you've been directing the machine and it's good music, but

it's not your music?

646

01:02:54,019 --> 01:02:57,320

Or is it really your music and the machine was just an assistant?

647

01:02:57,320 --> 01:02:59,741

Or is it a collaboration?

648

01:02:59,741 --> 01:03:01,682

What is the status of that music?

649

01:03:01,682 --> 01:03:03,622

And even more importantly,

650

01:03:03,692 --> 01:03:10,204

We now experiment, we also look at what we call the phenomenological aspect of using those

software.

651

01:03:10,204 --> 01:03:12,424

In other words, how do you feel?

652

01:03:12,524 --> 01:03:20,647

So if you've been using Cubase for 20 years and you come and now there's this new system

and you sort of have to use it because you compose faster with it, how do you feel in the

653

01:03:20,647 --> 01:03:21,342

morning?

654

01:03:21,342 --> 01:03:31,990

Do you feel like, oh no, do you feel like this is really ruining your work, which is like

the passion of your life, or do you feel like you're still excited and using it?

655

01:03:31,990 --> 01:03:33,816

So we really try to go...

656

01:03:33,816 --> 01:03:39,060

deeper than just, you know, this is a toy, use it and tell me that it's amazing.

657

01:03:39,060 --> 01:03:39,890

You know, no.

658

01:03:39,890 --> 01:03:47,055

Like I think in JNI we need to pass this sort of, you know, flamboyant just quick demo of

like, my God, look what it does.

659

01:03:47,055 --> 01:03:48,556

It does a little video that moves like that.

660

01:03:48,556 --> 01:03:53,640

But yeah, but I can't use that video anywhere because it's totally not my aesthetic and I

can't get it to do what I want.

661

01:03:53,640 --> 01:03:56,712

Well, you know, it's just a toy.

662

01:03:56,712 --> 01:04:03,426

We're trying to build actual tools and to really study what are the implications of those

tools.

663

01:04:05,360 --> 01:04:06,302

Excellent.

664

01:04:06,302 --> 01:04:15,054

I think that was a great finishing statement for this episode, actually, So, thank you so

much for joining.

665

01:04:15,404 --> 01:04:18,639

Yeah, thanks a lot and sorry for digressing in my answers.

666

01:04:18,841 --> 01:04:20,464

I like to not really answer questions.

667

01:04:20,464 --> 01:04:25,954

I like more to put out provocating thoughts and I hope your audience will appreciate it.