Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
Craft matters in small ways, like how
0:02
coffee is made, or
0:04
how a wooden table is built piece by piece.
0:07
And in not so small ways, like how your money
0:09
is cared for. At UBS,
0:12
we elevate investing to a craft.
0:14
We deliver our services with passion,
0:16
expertise, and meticulous attention to detail.
0:19
This is what investing means to UBS. Not
0:22
just work, but a craft. Discover
0:24
more at ubs.com/craft.
0:27
The value of investments may fall as well as rise, and
0:30
you may not get back the amount originally invested. Welcome
0:41
to the New York Times Popcast. You
0:43
are language too vast for any training
0:45
model to grasp of music, news, and
0:48
criticism. I'm your host, John Karamanica. I
1:25
know for years you've been listening
1:27
to podcasts and thinking, these guys
1:29
need a theme song. Why
1:31
don't they have theme songs? Well, what
1:34
did I spend the last few days doing while playing
1:37
around with Suno and UDO, which
1:39
are two of the leading music
1:41
AI interfaces to approximate
1:44
what a popcast theme song might
1:46
sound like. I
1:48
put some prompts in and came up
1:50
with a bunch of different options, which I'm
1:52
going to be sprinkling through the show
1:55
this week, which is completely
1:58
devoted to contemporary. uses
2:00
of artificial intelligence in music, we
2:02
are going to be talking with
2:05
Rachel Metz, who covers artificial intelligence
2:07
more broadly at Bloomberg and Kristin
2:09
Robinson, who covers the music business,
2:12
including how the record industry is
2:14
trying to use AI for Billboard.
2:17
What you were hearing at the beginning of the show,
2:20
you were hearing, there's no name,
2:22
it was made with Suno. It's
2:24
the first track that
2:27
was generated with the prompt
2:30
of asking for a pop cast theme song
2:32
that sounded like 90s Midwest
2:35
emo. And I'm not
2:37
going to lie, they definitely got
2:39
that parts thing down. They definitely
2:41
got that sort of dry male
2:43
vocal thing down. It's a little
2:46
unsettling, but also kind of a
2:48
joyful experience. As I said,
2:50
throughout the show, we're going to be playing a
2:52
few other things that I'll be throwing to. But
2:55
for now, let's get into the
2:57
conversation. Rachel Metz is
2:59
here, Kristin Robinson is here.
3:01
Welcome to podcast. Thank you.
3:03
Everybody coming in virtually, but
3:05
hopefully authentically not artificially today,
3:07
fingers crossed. So I know
3:10
I've been keeping tabs on what's been happening with
3:14
AI music making intermittently
3:16
for the past, let's say six to 12
3:18
months. But both of you, this is a
3:20
very, very key part of your beats. Rachel,
3:23
you cover AI more broadly, music being one
3:25
segment of that. Kristin, you cover the music
3:27
business more broadly, AI being one segment of
3:29
that. And one of the reasons, the reason
3:32
I wanted to have both of you on
3:34
is I think you're approaching it through different
3:36
lenses, the same subject matter through different lenses
3:38
through your work. So I
3:41
thought it'd be interesting overlap and intersection
3:43
of perspectives. Can
3:45
we start with a thing that
3:47
happened recently that Felt
3:50
itchy to me? And Maybe we can start
3:52
with this as a fact as an also
3:54
an indicator of what's to come and maybe
3:56
then rewind and say, how do we get
3:59
to it? Really Travis. Really
4:01
Travis is and what has been
4:03
for many years. one of my
4:05
favorite country singers. Suffered a stroke
4:07
in twenty thirteen I believe and
4:10
and no longer Kim sang and
4:12
as a Randy Travis fans on
4:14
the grew up on the late
4:16
eighties early nineties hits emails of
4:18
a big loss but also I
4:21
currently music for a living. I
4:23
accept that not everything goldstein last
4:25
and that's okay as an older
4:27
son. Negative press release advertising a
4:29
new. Randy Travis songs and like
4:32
instantly I feel like the muscles
4:34
and my my max the top.
4:36
My soldiers defend a seal, the
4:38
discomforts and so what they've done.
4:40
This is a song called where
4:42
That came from and so let's
4:44
listen to Little Some the. Law.
4:56
So if you are fan or
4:58
any Travis, you're obviously you'll recognize
5:00
the voice of feals. mostly like
5:02
a rainy Travis voice. The inflections
5:04
feel like a good approximation of
5:06
Randy Travis, since, like some. What?
5:09
You're Actually Feeling isn't a i modeled
5:11
streamed on vocal stamps of about forty
5:13
Randy Travis songs overlaid on top of
5:16
a guide demo vocal sung by a
5:18
difference singer theoretically in the mood or
5:20
in service. The song itself is a
5:22
demo The the Asshole song was written
5:24
for any thrive as many years ago,
5:26
so it's not totally out of the
5:29
blue where they got these lyrics from,
5:31
or the idea from. But.
5:33
This is likely be the first
5:35
of what I imagine to be
5:37
many posthumous or posthumous asked recordings
5:39
for artist who know I can
5:42
make original recordings when when y'all
5:44
encountered this. Just. As a
5:46
fact did this seems. This.
5:48
Is normal? This is where it's going.
5:50
Or did anything about this see a
5:52
little bit Etsy. Maybe Rachel would start
5:54
with you and then we'll go first
5:56
and. For. Me is the first thing
5:58
I think is. I
6:00
have some questions about consent. Whose idea
6:02
was this? Did Randy
6:05
want to do this? How do we know that
6:07
he wants to do this? If
6:10
somebody has already died and they're a
6:12
state, they can make those decisions then.
6:15
And that seems totally fair to me. But if someone
6:17
is still alive, I think then it brings up
6:19
some really tricky questions. But aside from what does
6:21
it sound like and do we enjoy how it
6:23
sounds or is it creeping us out? And
6:27
I think the consent question is really interesting. There's
6:29
a CBS morning show or
6:31
Sunday morning show segment about this. And
6:33
he's in these rooms, but so
6:36
are cameras. I don't know who
6:38
was in those rooms when no cameras were
6:40
there. And also he can speak a little
6:42
bit, but not extensively. We
6:44
don't genuinely have a way
6:47
of knowing. One of the actual things that comes
6:49
up in the CBS Sunday morning segment, the
6:51
interviewer asks, he's there with Randy, he's there
6:54
with his wife. In
6:56
essence, you have thoughts, but you are
6:58
not able to express them. And
7:00
Randy sort of nods and said, yes. Do
7:03
we know how robust that process? We just
7:05
don't know. And I think the consent thing is
7:07
real. I have no idea if Randy Travis
7:09
wants this to happen. I've got to
7:12
assume his family wants it to happen.
7:14
His record label who has him under
7:16
contract wants it to happen. Bingo.
7:20
And it's striking that this happened in
7:22
Nashville, Nashville, Tennessee being the first state
7:24
that passed the law attempting to protect
7:26
the rights of artists to their voice
7:28
against supplications, the elder sides or whatever
7:30
that was passed a little while ago.
7:33
Kristen, how did this strike you both
7:35
as an individual data point, but also in the
7:37
broader conversation of how certain
7:39
artists are starting to, dare
7:42
I say, weaponize AI in their own favor?
7:44
I thought this was a really fascinating use
7:46
case. This is not the first time that
7:49
we've seen kind of something like this happen. Maybe
7:52
it was six months ago now, but
7:54
Edith Piaf, the old French singer, La
7:56
Vie en Rose, she passed away in
7:58
1966. So she
8:01
has been gone. And
8:03
her estate worked with Warner Music
8:05
Group and this film company, and
8:07
they're going to essentially resurrect her
8:09
voice using an AI voice model
8:11
and old recordings. And
8:14
they're going to use that to help narrate
8:16
and create this, I'm pretty sure it's like
8:18
an animated documentary about her life. And
8:20
they're also using AI to help recreate
8:23
her likeness. I find that
8:25
to be really fascinating because
8:27
when you think about it, I think that consensus
8:29
is at the core of whether or not
8:32
we feel like these things are okay. And
8:34
that comes back in Randy Trapp's instance as
8:36
well. But with someone who's truly
8:39
been deceased for 70 years, I don't know, she
8:41
could not have
8:44
conceived of the internet, let alone
8:46
AI, because she's been gone for
8:48
so long. So how do we
8:50
really know whether or not these
8:52
individuals would have ever been okay with this?
8:55
That being said, in an estate situation,
8:58
legally, those decision-making powers have been
9:00
passed down to probably someone else
9:03
in her family. I'm not sure
9:05
exactly. With Randy, it's a little bit
9:07
different because he's still here. And
9:09
as long as he's of sound mind and
9:12
he's able to consent to it, then
9:14
I think that's okay. But it's been
9:16
really interesting. My mom is a very
9:19
big avid watcher of CBS Sunday morning.
9:22
And she called me after that came on because she
9:24
knows I cover AI and music. And
9:27
she was saying, I feel like
9:29
this isn't Randy Travis. To me,
9:31
this is not Randy Travis. And
9:33
I thought that was interesting because when I heard
9:35
it, I considered it in my
9:38
brain as Randy Travis. So when I
9:40
heard it, I was like, yeah, this is the new Randy
9:42
Travis song. It was using AI to assist it. But
9:44
she made a really good point. She was like, there was
9:46
a session singer who sang it.
9:48
And then they essentially used a voice filter
9:51
to map that voice, his timbre,
9:53
his tone over someone
9:55
else's voice who did a different performance.
9:58
That was not Randy. And I
10:00
don't think that either of us is right
10:02
and wrong. It's just interesting how people
10:05
can take away from this very
10:08
different things based on understanding the
10:10
process of how it's done. But
10:12
I do think that in
10:14
general, we feel differently about this. If the
10:16
artist is consenting to it versus if they're
10:18
not. And I also do think as
10:21
far as whether Randy Travis himself is
10:24
consenting truthfully because of
10:26
the nature of his injury, we actually,
10:28
I feel like we don't 100% totally
10:30
understand the depth of his consent, even
10:32
though he's consenting enough to appear on
10:34
CBS Sunday morning, or at least the
10:36
people who are advocating for him are
10:38
putting him on CBS Sunday morning. There's
10:40
also, I got a press release the
10:42
other day about a tour, a Randy
10:44
Travis tour, and I was like, what
10:47
is happening here? But what that, the tour
10:49
is, is this session singer, who
10:52
I guess I had not prior previously heard
10:54
of, but maybe sort of sings in a Randy
10:56
Travis motif, seeing the hits
10:58
of Randy Travis interspersed with like
11:00
biographical films and stuff, and Randy
11:03
Travis will be there, I guess.
11:05
And it's also, this is also maybe
11:08
not coincidentally, this is a Warner project
11:10
as well, so maybe this is an
11:12
indication that Warner as a label is
11:14
stepping out and taking more risks. It
11:17
reminds me a little bit, this is a forgive,
11:19
forgive this OG popcast take, which
11:21
I do not expect you guys to
11:23
know. But like when Luke Combs covered Fast
11:25
Car last year, like my running joke on
11:28
the show was like, it's a Luke Combs
11:30
AI cover. Like he didn't change a thing.
11:32
Like it's literally just Fast Car plugged
11:34
into a machine and someone pressed the Luke
11:37
Combs button and then they sang. And then,
11:39
and then, and then it's it. Nothing else
11:41
happened. And at the time I said, I
11:43
would not be surprised if record labels looking
11:45
for additional commercial angles
11:47
for their artists, especially artists
11:49
who maybe are in the mature phase
11:51
of their career, maybe post peak hit
11:53
making, if they would not say, Hey,
11:56
what are the 30 essential
11:58
songs of the eighties? fire
12:00
them up in the machine and put the Luke Combs filter
12:02
on and just see if anything sounds good. See if anything
12:04
we want to put out. That's
12:07
what I thought. I was like, that's the
12:09
worst case scenario. The worst case scenario is
12:11
we're going to get like a hundred different
12:13
versions of a guy like Luke Combs singing
12:15
a song like Fascar. But
12:17
actually the Randy Travis thing is
12:19
like lightly more insidious and Christian
12:21
to your point, what
12:24
part of that were maybe more to your mother's
12:27
point, what part of
12:29
this is Randy Travis, right? Like
12:31
the vocal is not the
12:34
songwriting. It's not. It's not written by Randy
12:36
Travis. It's written by songwriters for Randy Travis
12:38
X amount of years ago. The
12:41
vocal tone is computer
12:43
approximated. What
12:45
part of that is Randy Travis? I
12:48
truly don't know. And I wonder
12:50
you guys must be
12:52
encountering philosophical questions of this
12:54
nature over and over and over again. So
12:57
I wonder on the one hand
12:59
what part of that is Randy Travis. I'm curious
13:01
about both of your answers. And also the second
13:03
question is does it matter if
13:06
it is quote unquote Randy Travis.
13:08
It's Randy Travis in quotes. Does
13:11
any of the philosophical concern or
13:14
curiosity matter if it
13:16
somehow successfully lands between the quote mark?
13:19
I think that there's actually going to be
13:21
a shift in how we view this
13:23
over time. And I think there already has been.
13:26
If we look back to about a year ago,
13:28
that's when I think music related to AI had
13:30
its first big moment with Ghostwriter with
13:32
his fake Drake song. RIP Ghostwriter.
13:35
Yeah. You
13:37
had a great run. Big run. Big run. Big run. Got a
13:39
great run. And I think that was
13:41
a huge moment when it happened. It made
13:43
all the major news networks. A lot of
13:45
places didn't even know that there was music
13:48
related AI coming out and that people were
13:50
working on that. And when
13:52
that happened, I think people had
13:55
such an extreme reaction
13:57
to that. And I think that's very merited. It's the
13:59
first time we. never seen it. But fast
14:01
forward a year later and we have the Randy
14:04
Travis thing, we have Eadis Piaf, we have an
14:07
interview with Lauv six months ago. He
14:09
translated one of his songs into Korean
14:12
and then mapped his AI
14:14
voice model over it and released it. To
14:17
be fair, Lauv does sound like an AI
14:19
voice without any AI technology. Like Lauv voice
14:21
is just pretty much AI. You're trying to
14:23
get me in trouble. No,
14:26
no, no. Here's the thing
14:28
about popcast is the guests come on and be
14:30
experts and I get in trouble. That's how this
14:32
works. They know where to
14:34
find. Lauv's people know where to find me.
14:38
I'll expect an email from them later.
14:41
But yeah, the Lauv thing happens. And then
14:43
of course we have Drake and
14:45
I'm sure we'll get to this. The
14:47
Drake Kendrick beef and Drake using Tupac
14:49
and Snoop Dogg's voices. And
14:52
that was done without consent. But
14:54
think about how much less we
14:56
all reacted to when that happened
14:58
versus a year ago with Ghost
15:00
Rider taking Drake actually. It's ironic
15:03
that Drake is the one now
15:05
out here, quote unquote, stealing other
15:07
people's voices when he first had
15:09
his stolen. So I find
15:12
it quite interesting that the goalpost is
15:14
going to move. Also I
15:16
remember feeling with the Ghost Rider moment
15:18
that quibble the big conversation amongst me and
15:20
other hip hop fans and critics
15:22
that I was talking to were very
15:24
preoccupied with all
15:27
of the little things that Drake does
15:29
in songs that you'd never see it
15:31
on like the notation. You'd never
15:33
see it on a lyric sheet, but when you hear
15:36
it, when you feel it, when you hear it and
15:39
identify the places in the Ghost
15:41
Rider song where Ghost Rider failed
15:43
at that. Now, obviously it's
15:45
an accomplishment. It's a big accomplishment. But I remember
15:47
most of my friends were kind of like, ah,
15:49
the way that this word goes into this word,
15:51
like Drake would never do that. He would leave
15:53
a pause between A and B. That's
15:55
how we thought about it. But I think what
15:57
you just said is right, which is That
16:00
was because it was such a novel use case now
16:03
like when Drake
16:05
did and we'll get to the to the
16:07
Kendra beef and but when he did the
16:09
pot snoop thing Yes, I
16:11
could identify all that stuff where he
16:13
failed at like sounding like snoop But
16:15
I no longer expected it it like
16:17
something the goalpost had shift and I
16:20
wonder so Rachel with that in mind
16:22
Do these philosophical concerns matter at
16:24
all or have? consumers
16:26
and listeners Already established
16:28
almost like a parallel category
16:30
like there is drape music
16:32
and then there is quote-unquote
16:35
drape music oh
16:38
That's a good question. I feel like for
16:40
most people this stuff is still so new
16:42
for for one second. Let's step back AI
16:45
generated music or music that has been has used
16:48
AI in some way is Has
16:51
been on for a long time and it is
16:54
it is true that in the past couple
16:56
years It's actually developed quite rapidly in terms of
16:58
being able to sort of like really well
17:00
mimic Human-made music especially like vocals and just layering
17:02
of different things on top of each other
17:04
I feel like the Suno and Yudio in particular
17:07
They really showed the average person like
17:09
how far that stuff has developed in
17:11
the last couple years, but I
17:13
feel like most people Don't
17:16
aren't super familiar with the idea
17:18
of AI generated music and so
17:20
I do think that these questions of what is
17:22
real What is not real? What is artistry? What
17:24
is human? I feel like those things matter and At
17:28
least they matter to me And
17:30
I as a listener. I still want to
17:32
know what's going on Something
17:34
is using AI simply as
17:37
like a tool to build on recording to
17:39
edit something That's a little different from if
17:41
you're wholly generating something and going back to
17:43
this Randy example You said what was it
17:46
40 tracks? Or
17:48
I think the 40 vocals vocal
17:50
stems of 40 Randy Travis
17:52
songs Okay, that seems pretty
17:55
small for a data set to me. I feel
17:57
like 40 of anything does not a large
18:00
that make. So if you had more, if you want to
18:02
make it sound better, add more.
18:05
You mentioned that obviously people have been using
18:07
AI in various forms for a while. Maybe
18:09
can we step back from this current moment
18:11
and go back prior before a year ago?
18:14
What are some of the
18:16
earlier examples of people attempting to bring AI
18:18
technology into the music space before we get
18:20
to this glutt? I think we're like at
18:22
a glutt moment, even if most people don't
18:25
realize we're at a glutt moment, but I
18:27
feel like we're there. What
18:29
was happening two years ago, four years ago,
18:31
dare I say six, eight, 10 years ago
18:33
that set the table for this? I
18:35
dare. Yeah. So I think
18:37
the first time that I as
18:39
a technology journalist became aware
18:41
of using AI
18:44
to wholly generate music would have been
18:46
around 2016 when Google was working on
18:48
what was called Project
18:50
Magenta. And it was a research project and
18:52
it was more like at the time they
18:54
were doing like noodling on a piano. And
18:56
it kind of blows my mind when I
18:58
think about what you can do now very
19:00
easily and cheaply versus what was then state
19:02
of the art. And at that time they
19:05
couldn't get AI to form a coherent piece
19:07
of music. It never ended. It was just
19:09
kind of like tickling the ivories, which seemed
19:11
really cool at the time. It was a
19:13
jam band? Google was it? It was just
19:15
a jam band? It's a jam band.
19:17
It's just a piano. Yeah, basically. It was just kind of keep
19:19
going. And I kept being like, Oh, I hope they release it,
19:21
but they would, they never released it. And then it
19:24
kind of there are a few developments
19:26
over the years. OpenAI actually had a
19:28
project called Jukebox. I feel like nobody
19:30
noticed it because it came out at
19:32
the very beginning of the pandemic. It was like,
19:35
I think it was like mid March 2020. I
19:38
was on maternity leave. So I obviously was like,
19:40
whatever, I don't care. But they were trying to,
19:42
they were using like major artists names and trying
19:45
to generate music at the time. It sounds really
19:47
weird. If you want to, if you go online
19:49
and look up OpenAI Jukebox, you'll find like a
19:51
bunch of samples and it's like, Whoa, what is
19:53
this? But yeah, over the past few
19:56
years with the rise of different
19:58
kinds of generative models, AI, generated
20:00
voices, getting better and better. It's kind of like a
20:03
few different pieces of technology coming
20:05
together to create what we
20:07
have now. People like the
20:09
UDO and Suno teams kind of putting all
20:11
these pieces together in a really smart way.
20:14
Before we take that any further, we're gonna
20:16
take a break real fast. Going into this
20:18
break, hear something from
20:20
UDO. UDO, this is a song, it
20:22
comes with song titles.
20:25
The song is City Sounds Unbound. This
20:27
is from a prompt asking for a
20:29
pop cast theme song in the style
20:31
of 90s New York hip-hop. I will
20:33
say this is a little bit more
20:35
salt and pepper than 90s New York
20:38
hip-hop. It's giving 1987 to 1989,
20:42
maybe more, but it does have that spare
20:46
crisp percussion. It does
20:49
have that sassy vocal
20:51
presentation. Again, it's great
20:54
lulz. So this is City Sounds Unbound pop
20:56
cast theme song. We'll be back in a
20:58
minute. I'm
21:20
always something new. Craft
21:29
matters in small ways like how coffee is
21:31
made. Or how
21:33
a wooden table is built piece by piece. And
21:37
in not so small ways like how your money is
21:39
cared for. At UBS, we
21:41
elevate investing to a craft. We
21:43
deliver our services with passion, expertise,
21:45
and meticulous attention to detail. This
21:48
is what investing means to UBS.
21:50
Not just work, but a craft. Discover
21:54
more at ubs.com/craft.
21:57
The value of investments may fall as well as rise and you may not
21:59
get back the amount of originally invested. I
22:05
use the New York Times Games app
22:07
every single day. I love playing Connections.
22:10
With Connections I need to twist my brain to
22:12
see the different categories. I think I know this
22:14
connection. Look, Bath is a city in
22:16
England, Sandwich is a city in England, Reading
22:18
is a city in England, and I'm going to
22:21
guess Derby is a city in England. I started
22:23
Wordle 194 days ago and I haven't missed a
22:25
day. The New York Times Games app has
22:27
all the games right there. I absolutely
22:29
love spelling bee. I always
22:31
have to get genius. I've
22:33
seen you yell at it and say that. That
22:36
should be a word. Totally should be a
22:38
word. Sudoku is kind of my version of
22:40
lifting heavy weights at the gym. At this point
22:42
I'm probably more consistent with doing the crossword than
22:44
brushing my teeth. When I can finish
22:46
a hard puzzle without hints, I feel like
22:48
the smartest person in the world. When I
22:50
have to look up a clue to help me, I'm
22:53
learning something new. It gives me joy
22:55
every single day. Start playing in the New
22:57
York Times Games app. You can
23:00
download it at nytimes.com/games app.
23:06
We've mentioned Yudio and Suno a couple times.
23:08
These are, it seems like, the two industry
23:10
leaders at the moment in terms of, and
23:12
for those who are listening who have not
23:14
had a chance to play around with it,
23:17
I do recommend just toying around.
23:19
You can make a bunch of songs, at least for
23:21
free. If you're curious about the technology, the
23:23
Popcast theme song at the beginning
23:26
was made with Suno.
23:28
The songs that you'll hear kind of
23:30
sprinkled throughout were made with either Suno
23:32
or Yudio. First of all, on
23:35
a baseline level, you go
23:38
to the interface and you say, I would like a
23:40
theme song for Popcast, the New York Times podcast in
23:43
a style of blank or
23:46
whatever. The lyrics are about this. Or maybe
23:48
in one case, in Yudio's case, you can
23:50
actually write some lyrics better than the song
23:52
will be built around. But
23:55
can we talk a little bit about what's
23:57
underneath the hood? Because obviously, this is something
23:59
that... neither company I
24:01
think has said there is
24:03
a large model that this is trained
24:05
on that these programs are trained on.
24:08
What do we know about those models? We
24:10
don't know that much about the training process
24:13
of these. They've been very tight-lipped about it.
24:15
Of course there's quite a few models kind
24:17
of looking at AI more broadly that have
24:19
been fairly open about the fact that their
24:22
data sets include copyrighted works. I
24:24
think in music because the entire music business
24:26
is based off of exploiting
24:29
copyrights and making money off
24:31
of private ownership of songs.
24:33
This is an extremely
24:36
important thing for music
24:39
lovers as well as musicians and
24:41
labels and publishers to know about and
24:43
they've been pretty quiet about it so
24:46
I think that definitely has rubbed some
24:48
people the wrong way. We don't know
24:50
if it includes copyrighted works. There is
24:52
a former stability AI
24:55
VP of audio who left
24:57
the company and he's founded his
24:59
own nonprofit. It's called Fairly Trained. His name
25:01
is Ed Newton-Rex and basically he's kind of
25:03
made it his mission to talk
25:06
about the training process of AI and music
25:08
and he did do I think it was
25:10
an op-ed for Music
25:13
Business Worldwide where he talked about how
25:15
he was able to
25:17
generate things that sounded
25:19
extremely close to true
25:21
copyrights songs that we all
25:24
know and love by prompting Suno and
25:26
the Udeo enough that they could get
25:28
around to getting these copyrighted works pretty
25:30
explicitly. So it does seem like that
25:32
is a very good possibility to be
25:35
part of the training set but their
25:37
terms and conditions at least for Udeo
25:39
which I checked out recently does put
25:42
the onus on users. So if the user
25:44
wants to release a song let's say it
25:46
has a melody in it that's copyright from
25:48
a copyrighted song and it's bar
25:50
for bar or note for note rather and they
25:53
put it up there. Do a little soldier
25:55
boy joke for all the ads. On
26:01
Spotify and something were to happen and you
26:03
were to get sued, I do believe that
26:05
Onus would be on you, given the terms
26:07
and conditions. So that's definitely something to keep
26:09
in mind when using these things. Also
26:12
one thing that if I think this came
26:14
up in maybe was it
26:16
Brian Hyatt's Rolling Stone story about Suno
26:19
interviewing one of the investors and if it's wrong, if
26:21
I'm wrong, tell me, but I think interviewing one of
26:23
the investors who essentially said, I number
26:26
one, if this company already had
26:28
deals with record labels, I wouldn't
26:30
invest. Like on
26:32
the one hand, I don't want to know
26:34
exactly what the training model is and also
26:37
it's better if we're not in bed with
26:39
the record labels and the publishers at this
26:41
point. But it does seem clear,
26:44
much like Spotify itself or
26:46
TikTok or eventually these programs, they
26:49
will be in business with the
26:51
record labels. That's what it feels
26:53
like inevitability. These kinds
26:56
of relationships will be woven into artist
26:58
contracts, be woven into songwriter contracts or
27:00
producer contracts. Is it your understanding that
27:02
that's something that's actively in motion or
27:04
is that still feel one, two, three,
27:06
five years on the road? That
27:08
is happening. There are
27:10
definitely talks going on. We may also use lawsuits,
27:12
but there are definitely some talks going on. I
27:16
think that the ideal situation for the
27:18
majors right now is I
27:20
think what it seems to be
27:22
their strategy is picking and choosing
27:24
certain music, AI companies to anoint
27:27
as the ones that they want to succeed. Given
27:30
they own the vast majority
27:32
of important copyrights in music history,
27:34
which is honestly for recorded
27:36
music, a relatively short history, I think
27:39
they're trying to control the outcome
27:41
and veer it in a certain
27:43
direction. They certainly want
27:45
to participate in getting licenses
27:48
for their works if that's
27:50
what's needed, but we'll
27:53
see what happens. I do think there
27:55
is a possibility of a future where
27:57
Suno and Udeo become so
27:59
entrenched. in the creative process and
28:01
become so popular that it could force
28:03
the hands of music companies
28:06
to come to the table and work with them
28:08
even though they might right
28:10
now not be interested in that.
28:13
But in the near term, I do think that kind
28:15
of these more quote unquote fairly trained models
28:18
that use licensed works or
28:20
they make their training data
28:22
in-house even by hiring musicians to make
28:24
a bunch of samples for them. I
28:27
think that ideally that's what they're looking
28:29
for. Is that
28:31
what's happening? Do you think that they're
28:34
essentially rather than if they're not training
28:36
on publicly available or commercially available music,
28:38
they're essentially hiring session musicians to be
28:41
like, play me a hardcore song,
28:43
play me an emo song, play me this? A
28:46
lot of different models trained in different
28:48
ways. Like there's an AI
28:50
music company called Soundful and they train
28:52
on one shot. So that's
28:54
one type of thing instead of a loop. A
28:56
lot of people train on loops or I'm also
28:58
seeing a lot of people
29:01
are turning to production libraries. So people
29:03
who make music that's cheap
29:05
to be licensed for TV, film, YouTube
29:07
videos, et cetera. And they're
29:10
licensing those catalogs and that
29:12
can contain up to a million works, which
29:14
it really puts a dent in things when
29:16
you're trying to get there. And but yeah,
29:18
there are people who are hiring musicians just
29:21
to feed the machine right now, which I
29:23
think is a very interesting idea. As
29:25
a musician, do you feel like you're training
29:28
models to replace you in certain
29:32
things? If these AI models get good
29:34
enough, do they replace production libraries? Are
29:36
the production libraries and these musicians contributing
29:38
to speeding that up or is this
29:40
all inevitable? And it's better
29:42
to just get the money in the short term
29:45
and take the bet on the long term. I've
29:48
got to assume those musicians are under NDAs
29:50
if they're doing that. But if not, or
29:52
if you want to break them, email anybody
29:54
on this show. We'd love to hear more
29:56
about that process. OK,
29:59
let me play. doubles advocate for a second
30:01
because one thing that happened after the Ghost Rider song,
30:03
or at least
30:06
one thing I was thinking about, was let's
30:08
take Drake. Drake is incredibly
30:11
famous, incredibly successful,
30:13
and someone who
30:16
genuinely engages in the
30:18
meme ecosystem about Drake,
30:21
both positive memes, negative memes. Drake
30:24
understands how to make
30:26
songs, how to create visual
30:29
content, and also how to
30:31
react to the inevitable kind
30:33
of like effluvia on the
30:35
internet that is made non-consensually
30:37
about you. If I'm
30:39
Drake, why don't I just
30:41
make the Drake app which
30:43
can have varying levels of AI
30:45
technology. The cheapest level is
30:47
just Drake vocal filter. It's like we can run
30:50
through this podcast through whatever that
30:52
vocal filter is, and it just sounds like three Drake's talking
30:54
to each other. The second level may
30:56
be a higher level of training
30:58
with song construction and more
31:00
idiosyncrasies. Maybe the highest level is some
31:03
combination of that and then actually get
31:05
to interacting with people who make OVO
31:07
records to make it sound exactly like
31:10
a Drake song. Why
31:12
shouldn't that be a monetizable
31:15
income stream that a
31:17
smart artist would take advantage of?
31:19
What's the what's the downside of
31:21
that? Oh
31:24
man, there's a lot to unpack there.
31:28
If you're an artist, I've been thinking about this a
31:30
lot lately because I've been covering OpenAI and Scarlett
31:33
Johansson's anger about OpenAI having a voice
31:35
for chat TPT that sounds similar to
31:37
her voice or that she believes firmly
31:39
sound similar to her and a lot
31:41
of people also sound similar to her.
31:44
And I just keep thinking about how valuable
31:47
your voice is when you are various
31:49
types of artists if you're an actor.
31:51
Obviously if you are a singer, your
31:53
voice is valuable to you and also in
31:55
some cases like Drake, I think we can argue
31:58
that it's iconic, right? People can recognize it. They
32:00
can hear it and go, oh, that's obviously
32:02
this artist. So when you get
32:04
to that level, which select a few people
32:06
are, I feel like
32:08
it must be a very complicated thought
32:11
process of do I want to monetize? Like we are
32:13
getting to a point where you can totally build an
32:15
app like that. It would be still quite a bit
32:17
of work to make it sound really good, but you could totally
32:19
do it. But
32:23
to Kristen's earlier point about do
32:25
we participate now? Should I get the
32:27
money for this now or not? I
32:29
feel like it must be very tricky to decide what
32:31
to do there. But do you really want to
32:33
not have control over your voice? I feel like
32:35
a lot of people would say, at least for
32:37
now, no. But won't that
32:40
happen anyway? Because
32:42
how far away is the technology where someone
32:45
could just do that and not have to
32:47
pay anybody to sound like Drake? Like
32:49
why wouldn't an artist proactively, it's
32:52
already there basically. So
32:54
why wouldn't an artist get out? Like in
32:56
my mind, one of the
32:58
downsides is obviously you have a bad actor, right?
33:00
You have someone who uses it to make
33:02
Drake say something objectionable. Like think about
33:04
cameo and people on cameo being told
33:06
to be like, say something and then
33:08
it turns into white supremacist dog whistle
33:10
or something. And obviously you don't want
33:12
the AI version of that. But
33:15
barring that, assuming there's some customer,
33:17
there's some intervention between you creating the
33:19
thing and being able to share it
33:21
publicly, why wouldn't this
33:24
be a good idea for someone who's
33:26
incredibly famous and is already going to
33:28
get meme slash AI'd? Kristen, what
33:30
do you think? I mean, I think
33:32
that there are some artists who are
33:34
already leaning in and trying
33:36
to find controlled ways to do this.
33:39
But I think that it's a whole spectrum. And I
33:41
think at the end of the spectrum of being very fast
33:43
and loose about this is like grinds. Making
33:46
her own AI model last year, she was
33:48
very, very early on this. I think she
33:50
was inspired by what happened with Ghost Rider.
33:52
I know she was tweeting about that at
33:54
the time. And then shortly thereafter, released her
33:56
own voice model and said basically, I don't
33:58
care what y'all do with it. You can even
34:00
release it commercially if you want, but I want 50% of
34:03
whatever royalties you make from it. I've
34:05
since talked to some other AI voice
34:07
companies that are hoping to do kind
34:09
of official partnership type stuff as
34:12
their business model. And
34:14
some of them have thought that there's not a
34:17
ton of commercial value so far that we've
34:19
seen from voice models. So
34:21
like none of those grimes AI songs that her
34:23
fans made, I haven't heard of
34:25
any of them making good money.
34:27
There's one by Keto, K-I-T-O, using
34:30
the AI grimes model. That's the only one
34:32
in the streaming economy. I don't even know how
34:35
much that made. But
34:37
some people are thinking about monetizing it at
34:39
the root. So whenever your voice is
34:41
used by a fan and that fan pays
34:43
a subscription fee to the service, then you
34:45
get paid on that conversion.
34:48
People are thinking of all sorts of ways to try to
34:51
make this work for artists. I think
34:53
the challenge is being a really
34:55
good artist is curating yourself and
34:57
selling a brand version of yourself. And I
34:59
feel like that's never been more true than
35:02
an age of social media and
35:04
where we're at today. I think people really
35:07
want to feel like their
35:09
artist is someone who's like a guiding light.
35:11
They know exactly who they are as a
35:13
person and as an art
35:16
maker. So I think
35:18
that this does get into slippery slope
35:20
territory of when you start
35:22
letting your fans pretend to be you. Do
35:25
we really trust them to mark and say, this
35:27
is AI? Do we really trust them to
35:29
follow the rules? But yeah, you
35:31
could even argue, though, that we lost control of
35:33
who we are on the internet a long time
35:35
ago. That's sort of
35:37
my feeling. My feeling is, no, I don't trust fans
35:39
any more than I trust trolls
35:41
or trust meme makers. Everybody
35:45
is fundamentally untrust. Everybody who's not
35:47
the artist themselves is fundamentally untrustworthy.
35:49
And I wonder if the next
35:51
phase of this
35:54
is going in artistry broadly,
35:56
not simply in music across
35:58
disciplines is essential. essentially, how do
36:00
we monetize the mistrust that's
36:03
already baked into online presentation of
36:05
art? So that's on the one hand. The
36:08
lesson goal side of me says, let's
36:11
say it's not a full Drake app like what I
36:13
was describing, but let's say it's essentially
36:15
an audition app. It's
36:18
I have a song that I think Drake should
36:20
sample or I have a vocal pattern that I
36:22
think he might find useful. This I can see
36:24
working really well in Nashville, which is very heavily
36:26
a songwriter culture and demos are
36:29
pitched around constantly. Why
36:32
not have a company and says, Hey, I'm
36:34
pitching this to Carrie Underwood. Here's the demo
36:36
in Carrie Underwood voice rather than hiring some
36:38
random demo singer to sing it how they
36:40
think Carrie Underwood might sing it. That
36:43
feels like a semi valid use case
36:45
for some of this technology. It's
36:48
already being done that way. There's
36:50
a lot of songwriters that are
36:52
using AI voice filters privately, which
36:55
I feel a lot better about knowing that they're
36:57
not being commercially released. It's kind
36:59
of not a ton of harm. The
37:01
only harm is potentially upsetting the artist by them
37:03
receiving a demo and they're like, what the hell?
37:06
Why am I on this track? I never recorded. I
37:09
sound terrible. Yeah. So
37:11
yeah. I mean, like, who knows the session vocalist might have
37:13
not nailed the performance in
37:16
the style of you, but yeah, you're
37:18
already seeing a lot of songwriters. I
37:20
think I reported on this last summer. So it's been about
37:23
a year of people doing this. And so
37:25
it's a way to land demos. It is
37:27
really, really hard to pitch a record
37:29
and actually get it to land. And
37:32
for a lot of songwriters, they'll do anything to try to
37:34
make it better. Before AI,
37:36
during my reporting of this, I found
37:39
one songwriter who told me that he has
37:41
this go to guy that whenever he's pitching
37:43
to Adam Levine, Maroon 5, he
37:46
has an Adam Levine sound like session
37:48
singer that he'll get his demos recorded
37:50
with. And that's the darkest thing I've
37:53
ever been said on Popcast. The
38:00
poor person out there is an
38:02
entire metier as Adam Levine sound-alike
38:04
demos. Man, phrase up for you, bro.
38:07
This guy also has a go-to Justin Bieber
38:09
vocalist. Like nothing said, slightly.
38:15
But at least the argument would be like, well,
38:17
at least in that instance, you're hiring a person
38:19
to do a job where they get paid. But
38:22
these guys are saying that session vocalists only make $100 to $250
38:24
for... That's
38:27
kind of negligible side money for these people usually.
38:30
So yeah, now people are using ad voice
38:33
filters to try to better target pitch records.
38:35
I've seen with some people they'll include like
38:37
a folder to the NR that represents the
38:39
artist and in the folder will be the
38:42
AI version and the clean just regular demo
38:44
version so they can pick based
38:46
on what the artist prefers,
38:48
which I think is probably the smartest way
38:50
to go about it without freaking someone out.
38:52
But people are expecting this now. So I
38:55
think it's less shocking
38:57
than even a year ago. Rachel,
39:00
can we talk a little bit about legislation? Can
39:02
we talk about what's happened in Tennessee? But
39:05
also how legislation
39:07
is actually going to interact with some of
39:09
these philosophical questions because I may be
39:13
overly cynical in the sense that I
39:15
think I accept that there is some
39:17
disaggregation of the self that happens when
39:19
you present creative things in an online
39:21
space and famous people
39:23
who are discussed and memed
39:26
intensely, probably somewhere
39:29
in their spirit understand that what's
39:31
happening on the internet is 70% of them
39:34
or 30% or 90% or
39:37
at any given moment it's shifting. But
39:39
obviously legislation is not that nuanced. And
39:42
I do think it's important that some of
39:44
these things are enshrined in law. I wonder
39:46
if you could talk about what happened in Tennessee
39:49
and then maybe some of the other things
39:51
that are starting to happen about how to
39:53
address protecting an artist's
39:55
voice legally. There's
39:57
a lot of interesting stuff
39:59
happening. legally like all across
40:01
generative AI. With the music
40:04
stuff, as you mentioned earlier, there's the Elvis
40:06
Act in Tennessee. Gotta love an act named
40:08
after the king. I just thought that was
40:10
really funny. You have to have that heavy
40:13
draw when you talk about it. I apologize.
40:17
But so like that in particular
40:19
I thought was very interesting. The
40:21
aim of that is to protect
40:23
musicians' rights. They want to protect
40:25
them from AI impersonations in particular
40:27
in a place like Tennessee that
40:30
has such a heavy music, huge
40:32
music scene, huge music history, obviously
40:34
very focused on the humans that
40:36
are making the music, either singing
40:39
or just instrumenting. And so it
40:41
does make sense that a place like Tennessee would come
40:43
up with legislation like this. I think it's the only,
40:45
there are very, like in general, there are very few
40:47
laws in the United States related
40:50
to AI in
40:52
general. There are a handful that
40:54
are related to like specific applications
40:56
of technology. In Illinois, they
40:59
have a law called BIPA, which is
41:01
aimed at protecting people from unauthorized use
41:03
of biometric data, for instance. But like
41:05
you don't see much. And so this Elvis
41:07
rule I thought was very interesting. What I'm wondering
41:09
about is how they're going to enforce it. Like
41:12
how, what is considered an artist under
41:14
the law or someone with a recognizable
41:16
voice? I mean, they're saying this rule
41:18
is meant to protect people from unauthorized
41:20
AI, like reproductions of your voice.
41:22
If you have a recognized voice, you don't want
41:25
someone to just be able to make a new
41:27
one. So how do you decide whose
41:29
voice is, like is there a threshold for like
41:31
who's recognizable as an artist? I'm still a little
41:33
unclear on that. And then also how do you,
41:36
what are you going to do when someone complains?
41:38
And where do both offending parties have to
41:40
be in Tennessee? So like, I think it's
41:42
an interesting idea. I have a lot of
41:44
questions. I think that what
41:47
we're starting to see overall is a
41:49
bunch of efforts, both at the federal
41:51
level, at the state level, of people
41:53
trying to regulate different types
41:55
of generative AI. So AI
41:58
that's generating text, generating. sounds
42:00
such as music, generating images, and a
42:02
lot of the stuff is aimed at
42:04
the training data. So should the
42:07
prevailing law regarding training
42:09
data scraped from the internet, which we
42:11
have traditionally considered fair use, should that
42:13
stay or should that change? We might
42:16
see that changing by medium. That's going
42:18
to be interesting to see. So it's
42:20
possible that a court would say companies
42:22
like Suno and Yudio can continue to
42:24
scrape the internet forever for free to
42:26
do whatever, but maybe something
42:28
that's doing generative AI for images cannot.
42:30
Like I've talked to some legal experts
42:32
about that and so it's very
42:35
sticky right now. What's gonna happen with the
42:37
inputs, what's gonna happen with the outputs, and I think
42:39
it's gonna take years for this to be resolved in
42:41
any way. We're gonna take a quick break right now.
42:43
Here's another popcast theme song.
42:46
This one is a Suno generation. This
42:48
is the second generated song
42:50
when I asked a very big
42:53
picture prompt for a popcast theme song,
42:55
not specifying genre, but I think implying
42:57
pop. This is sort of a light
43:00
electro pop. Feels a little
43:02
bit European. It's giving Cascada.
43:05
It's giving Gigi
43:09
D'Agostino. So anyway, here's
43:11
this pop two from Suno pop
43:13
casting song. Families
44:12
have a lot going on. Let
44:15
Ollie help manage the mental load with
44:17
new cognitive health supplements for everyone 4
44:19
and up like delicious lollie focus pops
44:22
or lollymellopops for kids. And
44:24
for parents, try 3 new brainy juice to
44:26
help you focus, chill out, or get energized.
44:29
Find these cognitive health buddies for the whole
44:31
fam at ollie.com. These
44:34
statements have not been evaluated by the Food and
44:36
Drug Administration. This product is not intended to diagnose,
44:38
treat, store, or prevent any disease. We
44:44
have the beginning of the
44:46
legal conversation about what's protectable, who and
44:48
how it's enforced. On some level, I've
44:50
got to assume that the AI industry
44:53
is also going to go on, if
44:55
it hasn't already, something of
44:57
a marketing blitz. And
44:59
when I was alluding to the tears of AI,
45:01
like, is there good AI and bad AI? Is
45:07
there funny AI and serious AI?
45:10
With Ghostwriter and there was
45:12
that website or the channel that I emailed
45:14
you guys there I ruined it, or it's
45:16
the Beach Boys covering 99 problems. Oh,
45:19
oh, oh. On some level, is
45:22
the AI industry going to use
45:24
simple, accessible, winky, winky stuff
45:26
as like, I see AI as your
45:28
friend. It's just fun what we're
45:30
doing here. While underneath,
45:34
much more sinister and disruptive things
45:36
are happening. Because that's, I feel
45:38
like we're already seeing the tearing
45:40
of AI between kind of like
45:43
publicly legible uses of AI,
45:46
and then far more abstruse
45:48
and powerful uses of AI
45:50
underneath. Is that just Suno
45:52
and UDO marketing, essentially? It's being like, here's a
45:54
fun AI song, you guys should love it. And
45:57
then you guys pay attention to that while we
45:59
do the video. dirty work? In
46:01
general, kind of. I think it's more a little
46:04
more complicated than that. But yeah, if you think
46:06
back a couple years when there was a
46:08
big backlash against facial recognition technology,
46:10
you saw a lot of companies
46:13
saying, please regulate us. We want to
46:15
be regulated, yet nobody is coming up
46:18
with the rules. And I think you're
46:20
seeing a similar thing, and we have seen over
46:22
the past year or so, a lot of
46:24
it is like application based, right? It's not AI
46:26
generally because it's become so vast and there's so
46:29
many different applications of what in the past we
46:31
would have just thought of as like machine learning.
46:33
But I think with the music stuff, I
46:36
think it looks like the companies are looking more
46:38
for partnerships. But yeah, I
46:40
think there definitely is, like the companies will say, please
46:42
regulate us. But what they mean is, let us help
46:44
you come up with the rules, because of course then
46:46
the rules, it's sort of a Fox and the Hennihaus
46:49
kind of situation. Let us help you come up with
46:51
the rules, and those rules will necessarily
46:53
be rules that we're okay
46:55
with. And they will probably benefit us,
46:57
perhaps like a startup that's working on
46:59
closed source models, wants rules that help
47:02
regulate open source models, for instance. We've
47:05
talked a lot about AI as a song
47:07
making tool or as a tool to help
47:09
a person who makes songs make songs. AI
47:12
is being used in a bunch of other
47:14
ways in the music business, it's being used
47:16
in play listing, other kinds of
47:18
large data model training. Can you talk a little
47:20
bit about where's AI happening
47:22
that we're not seeing? Everybody,
47:25
again, we know about BBL
47:27
Jersey, we know about Drake rapping as Tupac
47:29
and Snoop. Everybody knows about
47:31
that. And I think, to your point,
47:33
Rachel, the AI industry would be very
47:35
happy if that's where public understanding of
47:38
AI music stopped. What's
47:40
happening underneath that we should be
47:42
mindful of? I think
47:44
that there's a lot of different applications of AI in
47:46
the music business today that we don't talk about that
47:49
often. The stuff that's not generative,
47:51
the stuff that's not about music creation,
47:53
that stuff is also happening
47:55
in a variety of ways. I think one
47:58
of the most interesting and... potentially
48:00
helpful use cases for AI
48:02
that we haven't talked about yet
48:05
is royalty collection and processing. That's
48:07
just notoriously a huge pain in
48:09
the ass. And there's,
48:11
given there's so many different
48:14
little microtransactions every single time you stream
48:16
a song, and then there's in a
48:18
case of certain rap songs, it
48:20
could be with a bunch of samples. It
48:22
could be up to like 30 songwriters. So
48:25
then you have to divide that all up
48:27
between all the right songwriters, get the money
48:29
to them, get the money to their publisher.
48:31
It's a mess. And
48:34
I think a lot of people think we could be doing better
48:36
in the music industry, getting
48:38
correct metadata and kind of making sure
48:41
to clean that stuff up. But
48:43
in general, I do think that given there's
48:45
just such a large, there's trillions
48:47
of lines of transactions to process,
48:50
AI will certainly be helpful in
48:52
that case. And also, because it's
48:54
so challenging, and makes it really
48:56
hard for artists and songwriters to
48:58
audit their record labels and music
49:01
publishers. And so I've heard
49:03
from some business managers that they're
49:05
like really hoping that this is going to be
49:07
a way to put
49:09
some checks and balances on that and really sort
49:11
through what their royalty statements are
49:13
looking like. And in addition,
49:16
you can see on a place like
49:18
Spotify, they've integrated the AI DJ feature,
49:20
which is pretty fun. I could see
49:22
it being interesting in the future. In
49:25
addition, they also have this beta
49:27
AI play
49:29
listing tool where you can type in kind of
49:32
a vibe, and then it will generate
49:34
a list of tracks. And I think that's fun
49:36
and playful. For now,
49:39
I don't really think that it's that
49:41
exciting right now. But maybe it could be in
49:44
general with AI that interacts with
49:46
the music business and interactive music
49:48
creation. A lot of this stuff,
49:50
we're still figuring out what the actual use cases
49:52
will be long term. Are some of these things
49:54
just gimmicks for kids that are
49:57
just having fun on the internet? Or are there actual
49:59
applications? that will be commercially
50:01
valuable, we don't know
50:03
yet. I think that we're trying
50:06
every single time that I'm talking to people,
50:08
I'm always asking them what ways are they
50:11
using it that's actually going to be helpful
50:13
for them or helpful for their fans. But
50:15
yeah, there are quite a few things
50:17
including this AI DJ, AI playlisting
50:20
features that Spotify have that are just kind of
50:22
playful and fun and we'll just see where it
50:24
goes from there. I haven't tried
50:26
that although it does make me want to see if
50:29
it could figure out how to recreate
50:32
what I played on my college radio show in 1995.
50:34
If I just give it a few problems and be
50:36
like this is what my radio show was like and
50:38
it was about can you sort of like
50:40
loosely estimate what a playlist was like. John,
50:43
you're too unique for that. I don't
50:45
think it could do it. That's exactly
50:47
right. They weren't gonna put the quiet
50:50
storm in between the backpack wrap. They
50:52
weren't gonna do that but I did
50:54
that. That's right. Rachel,
50:56
what about you? Do you
50:58
think there are either
51:01
good AI or nefarious use cases that
51:04
are lurking underneath the like high-five look
51:06
at the joke we just made AI
51:08
that people seem to be talking about?
51:11
I think there are always, this is like
51:13
a really unsatisfying answer, but I feel like
51:15
there are always good and bad uses for
51:17
pretty much any technology and AI is absolutely
51:20
no exception. You can't really say this is
51:22
good, this is bad categorically because different people
51:24
have different ideas. I think that what
51:26
we're seeing with its use in the music
51:28
industry is as Kristen said there's still a
51:30
lot of figuring it out but at the
51:32
same time I think it's interesting
51:35
that like one of the
51:37
uses of AI that the average person
51:39
is most familiar with just is related
51:41
to like song selection on Spotify. If
51:44
it's just creating like
51:46
a radio station that's sort of tuned
51:48
to you, I still find that kind
51:50
of like weirdly magical. Using that on
51:53
Spotify or like Netflix I'm like how
51:55
do they always know what I want? Well
51:57
obviously it's been trained on a ton of
51:59
data. And every time I use it,
52:01
I am like giving it yet more information about
52:03
me that it can then mine to keep me
52:05
glued to it in the future. That's
52:07
a good thing and a bad thing, right? So a
52:10
lot of these uses of AI are going
52:12
to be a combination of both. And I
52:14
feel like a lot of what we can
52:16
do as consumers, we're journalists, but we also
52:18
use these things. I think it's just important
52:20
to be really mindful of what is going
52:22
on behind the scenes as much as we
52:24
can be. And then obviously as journalists, we
52:26
have to interrogate them and tell everybody about
52:28
it. As we go
52:30
out, is there for either of you a
52:32
particular, whether it's a song or
52:34
a particular moment that you've encountered in your
52:37
reporting that you think people should know about
52:39
that maybe they don't know, but that's not
52:41
one of the kind of like 10 layering,
52:43
red alarm screening AI
52:45
use cases the last 12 months. Is there
52:47
anything you want to point people to to
52:50
say, this is interesting, fascinating or
52:52
clever that people don't know? I
52:55
have one. Okay. I
52:57
think, and maybe
52:59
this might have just inspired me to actually write
53:01
a story on it, but I have
53:04
been talking with a bunch of artist managers
53:06
who have said that the worst thing that
53:08
you can ever have to get your artist
53:10
to do is radio liners. It's like that,
53:12
Hey, this is Kristen and you're listening to
53:14
KISS FM. They
53:17
hate that. It's like pulling teeth.
53:19
And I very briefly worked at a record label where I was sent
53:21
out to do that. And it was
53:23
actually horrible. So I have heard
53:26
that some managers have tried using
53:28
AI to turn in AI radio
53:31
liners. Wait, did I read that in one of yours?
53:33
I feel like I read about that in someone's story.
53:35
Was it one of your stories? So
53:38
I like made a very quick mention of it
53:40
as like a bullet point in a newsletter that
53:42
I just started. Follow it. Machine
53:44
learning, billboard.com. So yeah,
53:47
so, but I haven't done a full story on that. And I feel like
53:49
now I should now that we're talking about
53:51
it, but I do think that that's really
53:54
a fascinating way to streamline the process of
53:56
music promotion, make your artists more
53:58
available. That being said, if I was. radio
54:00
station and I figured out that almost
54:02
the liners I was getting were faked,
54:05
I'd be pissed. So I don't know. I actually, Loki,
54:07
don't know if I'd be pissed. Also, they might want
54:09
to make one for podcasts. Like by all means, send
54:11
one over. I don't care if it's AI. We'll
54:14
start this... It would be hilarious if it was
54:16
probably funnier if it was AI. Can
54:19
we start this podcast with Adam Levine, AI? No,
54:24
just ask the guy, the real session musician.
54:27
No, I'm a higher than Adam
54:29
Levine guy. Yo, that's going to
54:32
be my ultimate anniversary, like my
54:34
10 years on Popcast whenever
54:36
that happens. Totally so nice. I'm a higher
54:38
than Adam Levine guy to sing a podcast
54:40
theme song in the voice of Adam Levine,
54:42
one of the artists I have been absolutely
54:45
the cruelest to in my critical career. But
54:47
anyway, if anyone wants to say AI,
54:49
shout out, hilarious. I'll probably play it.
54:51
Rachel, anything that you want to draw
54:53
folks through? One thing
54:55
that I did when I was working on a
54:58
feature about AI music recently and about Suno and
55:00
UDO in particular was near the end of the
55:02
process after playing around with them, I went on
55:04
to Spotify because I was just curious if these
55:07
things were starting to trickle in there and just
55:09
search for the word Suno. I would
55:11
recommend that people do that and just kind of
55:13
see... You can't guarantee that the stuff you're
55:15
finding is made using Suno, but a lot of
55:17
it sure sounds like it. There's still sort
55:19
of a certain tone to these things and
55:22
like kind of a slight weirdness to the
55:24
song. The way that the songs
55:26
progress, you're like, that's not exactly what one
55:28
would expect. But it is a nice way to
55:30
get a sense for what people happen to be
55:32
doing with it. That's sort of like percolating on
55:34
Spotify. So that would be my suggestion besides trying
55:36
these out for yourself is just go on Spotify
55:38
and search for Suno or maybe search for UDO.
55:40
There's probably stuff from there on there too at
55:42
this point. One thing that we
55:44
didn't actually explicitly get to are differences between
55:47
Suno and UDO. And I do think it's
55:49
worth saying. So I, in making a pop
55:51
cast theme song, which we've been hearing throughout
55:53
the episode, I put the same prompts into
55:55
Suno and UDO, basically
55:58
Cotton Bass. Maybe I left. out some
56:00
punctuation, but essentially the same prompts. One
56:03
thing I felt with UDO is
56:05
it returned over and over again to
56:08
the same lyrical themes. I didn't give
56:10
any lyric prompts. I only gave topic
56:12
prompts. There was Beats and
56:14
Rhymes came up a lot in the
56:17
UDO ones, maybe three different ones
56:20
in UDO used that or a
56:22
version of that phrasing, whereas I
56:24
thought the Suno lyric generation felt
56:27
a little bit more natural. Also,
56:30
on the whole, I felt like
56:33
the styles that it approximated
56:35
better were styles that I
56:37
already, maybe it's just my
56:39
association, already associate with a
56:41
kind of degree of digital
56:43
texture, electro-pop or certain strains
56:45
of hip-hop, whereas things that
56:47
are a little bit more guitar driven or
56:50
fundamentally melodic driven, I felt
56:52
they struggled more with. I wonder,
56:54
do you guys ultimately,
56:56
I think Suno was more
56:58
effective than UDO for me in my
57:01
task? Before we go, do you guys
57:03
see the difference or feel the difference between the
57:06
two and where do you land? I definitely noticed
57:08
some big differences between the two when I was
57:10
playing around with them. For me, Suno felt a
57:12
little bit more intuitive to use. One
57:15
thing that I would note is Suno
57:17
uses chat GPT to generate
57:19
lyrics. So it's working with open AI technology.
57:21
UDO off the top of my head, I'm
57:23
not sure, but I do know that Suno
57:26
uses chat GPT. So if you
57:28
like the feel of chat GPT, I think we
57:30
can all agree it has a style. That's
57:33
where they're pulling from and then it brings
57:35
back the song lyrics and also song
57:37
titles, I think, are all coming back.
57:39
And then they're sort of like musicifying
57:41
it. UDO, meanwhile, I think also the
57:44
way they build the songs in UDO,
57:46
in the kinds of chunks that they
57:48
build, I found it a little bit
57:50
more challenging to use. But I could
57:52
also see how someone who has a
57:54
greater understanding of music and song structure
57:56
might actually find that more appealing. And
57:59
the stuff that I've made I guess I've found
58:01
a few more things in Suno that I've
58:03
liked, but usually I'll create my own lyrics
58:06
and just see where that takes me. So
58:08
I don't know if that's made a difference, but
58:11
I think they're both extremely advanced. I think
58:14
one thing I don't think we explicitly
58:16
said so far is that
58:19
most music-related generative AI models that
58:21
are out there today, either
58:24
they do the voice or
58:26
they do the track, not both. And
58:30
to see that Suno and Udo are
58:32
doing both is really remarkable. And something
58:34
I have only seen probably
58:36
in the last, I don't know, since the start of 2024-ish.
58:40
But I think they're definitely neck and
58:42
neck. That being said, something that should
58:44
be noted about Suno is they just
58:46
raised $125 million in their latest funding
58:48
round. And
58:52
Udo, I'm pretty sure, as late as funding round,
58:54
raised $10 million. That being
58:56
said, who knows what their
58:59
financials look like, why Suno wanted to raise so much,
59:01
all that stuff. I'm not sure. But
59:03
I am really curious to see what Suno
59:05
comes up with with all the money that
59:08
they have just gotten from investors. So
59:10
I'm really curious to see if in the six months
59:12
from now we're going to be talking about
59:14
them being neck and neck, or
59:16
if one of them will pull away. But I
59:18
guess it should be noted that also BBL Drizzy
59:20
was made on Udo. And that
59:23
turned out to be quite good and
59:25
is something that I feel like I
59:27
hear people humming and singing all
59:29
the time. Yeah, the most commercially
59:32
successful or publicly recognizably successful AI
59:34
generated the track thus far. And
59:37
to my point earlier about Drake, he just found a way
59:39
to rap on it. So I'm just
59:41
saying, any free ideas out here on
59:43
Popcat. Rachel Metz, Krista Robinson, thanks for
59:45
going into the brave new world with
59:48
us. I'm confident we're going to be
59:50
back talking about this again probably sooner
59:52
than we're all comfortable with. Thank
59:55
you so much for having us. And that
59:57
is our show, What a Ride.
59:59
It has been every Popcat. ever
1:00:01
is at nytimes.com/popcast. Popcast
1:00:03
Deluxe with me and Joe is
1:00:05
at YouTube, youtube.com/popcast and subscribe all
1:00:08
across the board. Email
1:00:10
us, popcast at nytimes.com. Get
1:00:13
in the discord, get in
1:00:15
the Facebook group, tinyurl.com/popcast discord
1:00:17
or slash popcast Facebook. The
1:00:20
stickers are at itsthepopcast.myshopify.com.
1:00:23
There's a Zazzle shop. I put
1:00:25
some mugs in there, some buttons, some baby
1:00:28
onesies. Go nuts, knock yourself out. We're gonna
1:00:30
have more on the merch stuff soon. We
1:00:32
will let you know. As
1:00:34
always, our producers Pedro Rosado from Head
1:00:36
Stepper Media. We'll be back
1:00:38
next week. Let's go out with
1:00:41
another UDO generated popcast theme song.
1:00:43
This is also made
1:00:45
to the general prompt for a theme song, sort
1:00:47
of in a pop style. Somehow
1:00:50
this is extremely British. This
1:00:52
is very dry 90s,
1:00:54
light, a little bit sour
1:00:56
even. British-esque
1:00:58
presentation of a pop song. This
1:01:01
is Tune In, Feel
1:01:03
Good, popcast theme song. ♪
1:01:06
I'm a young girl from P.T.C.
1:01:09
Ukraine and the Polkies. ♪ ♪
1:01:13
Ain't been feeling any good. ♪ ♪
1:01:16
So tune in to P.T.C. for
1:01:19
this little group. ♪ ♪
1:01:21
I'm a little bit like you should.
1:01:23
♪ ♪
1:01:27
I'm a little bit like you should. ♪
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More