TRANSKRIPTIO 🔗
WEBVTT
1
00:00:00.150 --> 00:00:02.569
You know, there is a very specific feeling you get
2
00:00:02.629 --> 00:00:04.889
when a close friend hands you a mixtape or I
3
00:00:04.910 --> 00:00:08.619
guess these days text you a link to a track, right? Oh, totally.
4
00:00:08.679 --> 00:00:10.660
That feeling of like, you have to hear this.
5
00:00:10.900 --> 00:00:14.419
Yeah, exactly. And they say, this is so you. And
6
00:00:14.439 --> 00:00:16.899
when they actually get it right, it feels like magic,
7
00:00:17.219 --> 00:00:18.120
like they're reading your mind.
8
00:00:18.460 --> 00:00:20.379
Right. It's a deeply personal connection.
9
00:00:20.879 --> 00:00:24.730
But then in July 2015, something kind of strange happened.
10
00:00:25.550 --> 00:00:28.690
A machine started doing that. We all logged on and
11
00:00:28.730 --> 00:00:31.129
found this new feature, Discover Weekly. Yep.
12
00:00:31.449 --> 00:00:33.450
The legendary Monday morning drop.
13
00:00:33.810 --> 00:00:36.829
Right. And suddenly, this app was handing us 30 songs
14
00:00:36.869 --> 00:00:40.479
every week that felt, honestly, almost unsettlingly accurate.
15
00:00:41.100 --> 00:00:44.500
Uncanny, really. I mean, it was a total watershed moment
16
00:00:44.539 --> 00:00:47.539
for how we interact with media. We went from actively
17
00:00:47.640 --> 00:00:50.880
hunting for music in, you know, record stores or scrolling
18
00:00:50.939 --> 00:00:56.170
through very specific music blogs to having this incredibly intuitive curation,
19
00:00:56.229 --> 00:00:57.799
just served up on a platter.
20
00:00:58.079 --> 00:00:58.719
Served right to you.
21
00:00:58.880 --> 00:01:03.060
Exactly. It completely shifted the whole behavioral dynamic of music
22
00:01:03.100 --> 00:01:04.400
discovery for everyone.
23
00:01:04.420 --> 00:01:08.549
And, you know, it still feels like magic today. But
24
00:01:08.590 --> 00:01:11.150
looking through this massive stack of sources we pulled for
25
00:01:11.170 --> 00:01:15.150
this deep dive, and we've got technical briefs, research notes,
26
00:01:15.189 --> 00:01:18.439
the platform's own engineering history, you realize the magic is
27
00:01:18.609 --> 00:01:21.780
actually just a staggering amount of data and math. A
28
00:01:21.819 --> 00:01:24.379
lot of math. A lot of math. So our mission
29
00:01:24.420 --> 00:01:26.920
today is to Kind of strip away that illusion, right?
30
00:01:27.079 --> 00:01:31.700
We're going to demystify the actual machinery behind Spotify's recommendation algorithm.
31
00:01:31.879 --> 00:01:35.310
Because it's not magic. It's just very, very good prediction.
32
00:01:35.530 --> 00:01:38.590
Exactly. We want to figure out how a machine predicts
33
00:01:38.650 --> 00:01:41.810
what you, the listener, want to hear next. Because whether
34
00:01:41.849 --> 00:01:44.769
you're actively seeking out new underground artists or maybe you're
35
00:01:44.810 --> 00:01:47.609
just trying to protect your focus from, like, information overload,
36
00:01:48.269 --> 00:01:50.989
understanding the mechanics of this system turns you from a
37
00:01:51.030 --> 00:01:54.819
passive consumer into an active curator of your own audio world.
38
00:01:55.079 --> 00:01:56.140
You're taking back control.
39
00:01:56.319 --> 00:01:59.280
Yeah, you can actually train this machine. But to do that,
40
00:01:59.579 --> 00:02:01.920
we have to start at the beginning. Before the algorithm
41
00:02:02.019 --> 00:02:04.700
even looks at you, it has to categorize millions of songs.
42
00:02:04.879 --> 00:02:08.159
Right, and it categorizes all that music using three really
43
00:02:08.199 --> 00:02:12.500
distinct analytical pillars. And the first pillar is collaborative filtering.
44
00:02:12.520 --> 00:02:15.060
Okay, collaborative filtering. That sounds like.
45
00:02:15.740 --> 00:02:19.400
It's basically the logic behind people who bought this toaster
46
00:02:19.500 --> 00:02:22.000
also bought this bread knife. A lot of platforms use it.
47
00:02:22.139 --> 00:02:23.969
Oh, right. Like on Amazon or whatever.
48
00:02:24.169 --> 00:02:27.449
Exactly. But when you apply that to music at a
49
00:02:27.469 --> 00:02:31.449
global scale, it becomes incredibly complex. And the key thing
50
00:02:31.469 --> 00:02:35.000
here is the algorithm isn't actually listening to the audio
51
00:02:35.030 --> 00:02:35.560
for this part.
52
00:02:36.060 --> 00:02:36.659
Wait, it's not?
53
00:02:36.879 --> 00:02:40.680
No, not at all. It is exclusively analyzing human behavior.
54
00:02:41.180 --> 00:02:45.020
It uses what are called co-occurrence matrices. Co-occurrence.
55
00:02:45.099 --> 00:02:48.180
So meaning it's looking at how often two tracks just
56
00:02:48.300 --> 00:02:49.219
show up next to each other.
57
00:02:49.379 --> 00:02:53.939
Precisely. And primarily it's looking at user-generated playlists. So think
58
00:02:53.979 --> 00:02:57.539
about every time a listener creates, say, a Sunday morning
59
00:02:57.599 --> 00:02:58.620
coffee playlist, right?
60
00:02:58.800 --> 00:02:59.860
Oh, I have like three of those.
61
00:03:00.060 --> 00:03:02.599
Right, we all do. And if you put an acoustic
62
00:03:02.620 --> 00:03:05.490
folk song right after a soft jazz track, the algorithm
63
00:03:05.550 --> 00:03:08.870
registers that connection. It says, okay, these two belong together.
64
00:03:09.469 --> 00:03:11.530
So it's basically counting my choices as a vote.
65
00:03:11.770 --> 00:03:14.680
A human vote, exactly. And when you multiply that by
66
00:03:14.810 --> 00:03:18.919
hundreds of millions of users who are creating billions of playlists...
67
00:03:19.030 --> 00:03:22.990
This massive mathematical web is formed. The system can definitively
68
00:03:23.050 --> 00:03:26.129
say that song A and song B belong to the
69
00:03:26.169 --> 00:03:29.870
same behavioral cluster, simply because millions of humans voted with
70
00:03:29.889 --> 00:03:30.590
their track placement.
71
Load full transcript
00:03:31.090 --> 00:03:34.370
Oh. So we're basically doing all the heavy lifting for
72
00:03:34.389 --> 00:03:36.770
the algorithm just by organizing our own libraries.
73
00:03:36.870 --> 00:03:39.330
Pretty much. You're mapping the catalog for them.
74
00:03:39.789 --> 00:03:42.469
But, I mean, humans don't just group songs in playlists.
75
00:03:42.729 --> 00:03:45.550
We talk about them. We argue about them on Twitter
76
00:03:45.569 --> 00:03:48.280
or Reddit or whatever. I imagine the platform has a
77
00:03:48.289 --> 00:03:51.240
way of capturing the cultural conversation around the music too, right?
78
00:03:51.879 --> 00:03:54.680
Oh, absolutely. And that brings us to the second pillar,
79
00:03:54.780 --> 00:03:57.139
which is natural language processing or NLP.
80
00:03:57.939 --> 00:04:00.319
Okay. So it's reading text. Yes.
81
00:04:00.620 --> 00:04:04.719
The system essentially reads the internet. It constantly crawls music blogs,
82
00:04:04.810 --> 00:04:08.449
cultural articles, playlist descriptions, even social media to capture the
83
00:04:08.490 --> 00:04:09.870
vocabulary surrounding an artist.
84
00:04:10.689 --> 00:04:12.629
So it's looking for what people are saying about the music.
85
00:04:13.030 --> 00:04:15.849
Exactly. Okay. And this approach was heavily pioneered by a
86
00:04:15.870 --> 00:04:19.069
music intelligence company called the Echonest. They actually came out
87
00:04:19.110 --> 00:04:21.589
of the MIT Media Lab, and Spotify acquired them back
88
00:04:21.629 --> 00:04:22.009
in 2014.
89
00:04:22.009 --> 00:04:25.829
Oh, wow, 2014. So they've been doing this for a while. Oh, yeah.
90
00:04:26.449 --> 00:04:29.629
And the system is basically scanning for descriptive buzzwords. It's
91
00:04:29.689 --> 00:04:33.800
looking for terms like dreamy, lo-fi, synth-heavy, or, you know, melancholic.
92
00:04:34.620 --> 00:04:37.220
So it's not just noting that Pitchfork or some indie
93
00:04:37.259 --> 00:04:41.000
blog mentioned an artist. It's actually mapping the specific adjectives
94
00:04:41.040 --> 00:04:42.660
people use to describe the vibe of the song.
95
00:04:43.000 --> 00:04:46.829
Right. And it turns those adjectives into vectors, into math.
96
00:04:47.629 --> 00:04:50.310
So if an indie blog constantly uses the word reverb
97
00:04:50.449 --> 00:04:53.610
in the exact same paragraph as a specific band, the
98
00:04:53.730 --> 00:04:57.310
algorithm places that band in a very specific cultural neighborhood
99
00:04:57.389 --> 00:04:58.779
based purely on that text.
100
00:04:59.160 --> 00:04:59.699
That is wild.
101
00:04:59.939 --> 00:05:02.399
Yeah. The machine understands the cultural context of the music
102
00:05:02.740 --> 00:05:04.019
without ever hitting play on the track.
103
00:05:04.579 --> 00:05:06.439
But this brings up a huge blind spot, I think.
104
00:05:06.699 --> 00:05:06.800
Oh.
105
00:05:07.540 --> 00:05:09.459
What if an independent artist drops a brand new track
106
00:05:09.500 --> 00:05:12.199
at midnight? Nobody has added it to a playlist yet.
107
00:05:12.240 --> 00:05:14.100
So there is zero collaborative filtering data.
108
00:05:14.259 --> 00:05:14.439
Right.
109
00:05:14.519 --> 00:05:16.870
Right. And no blogs have written about it yet because
110
00:05:16.879 --> 00:05:19.670
it's brand new. So there is no natural language data.
111
00:05:20.730 --> 00:05:23.850
If it has literally zero historical data, how does the
112
00:05:23.889 --> 00:05:25.430
machine know who to push it to?
113
00:05:25.670 --> 00:05:28.610
That is the million dollar question. In computer science, that
114
00:05:28.670 --> 00:05:30.610
is known as the cold start problem.
115
00:05:30.769 --> 00:05:31.850
The cold start problem.
116
00:05:32.029 --> 00:05:35.069
Yeah. It is the most significant hurdle for any recommendation engine.
117
00:05:35.269 --> 00:05:38.379
Because if you only rely on past user behavior and
118
00:05:38.459 --> 00:05:41.680
text on the internet, new items are just completely invisible.
119
00:05:41.959 --> 00:05:43.620
They just fall into a black hole.
120
00:05:43.819 --> 00:05:47.060
Exactly. So to solve the cold start problem, the system
121
00:05:47.120 --> 00:05:51.360
utilizes its third pillar, which is raw audio analysis. This
122
00:05:51.480 --> 00:05:55.430
is where the machine finally actually listens to the music.
123
00:05:55.670 --> 00:05:56.750
Okay, so it finally hits play.
124
00:05:57.170 --> 00:06:00.550
Essentially, yes. But it does so through convolutional neural networks,
125
00:06:00.709 --> 00:06:01.470
or CNNs.
126
00:06:01.810 --> 00:06:06.389
Wait, neural networks? I always associate convolutional neural networks with, like,
127
00:06:06.629 --> 00:06:08.110
image recognition, not audio.
128
00:06:08.370 --> 00:06:10.470
Yeah, that's usually where people hear about them.
129
00:06:10.629 --> 00:06:13.949
Like self-driving cars recognizing stop signs or facial recognition on
130
00:06:13.970 --> 00:06:14.389
your phone.
131
00:06:14.829 --> 00:06:17.750
You've hit on the exact mechanism. The neural network isn't
132
00:06:18.089 --> 00:06:20.209
listening to the song the way a human ear does.
133
00:06:20.430 --> 00:06:20.769
It's not?
134
00:06:21.089 --> 00:06:23.769
No. The system takes the raw audio file and it
135
00:06:23.930 --> 00:06:27.810
mathematically converts it into a spectrogram. which is a visual
136
00:06:27.879 --> 00:06:28.959
representation of the sound.
137
00:06:29.079 --> 00:06:29.500
Wait, really?
138
00:06:29.519 --> 00:06:30.120
Uh-huh.
139
00:06:30.600 --> 00:06:31.839
A visual representation?
140
00:06:32.019 --> 00:06:36.639
Yeah. Time is mapped on the x-axis, frequencies on the y-axis,
141
00:06:36.720 --> 00:06:40.399
and the amplitude, or the volume of those frequencies, is
142
00:06:40.540 --> 00:06:42.399
represented by color intensity.
143
00:06:42.839 --> 00:06:46.740
Oh, wow. So the algorithm is literally looking at a
144
00:06:46.839 -->
00:06:47.699
heat map of the song.
145
00:06:47.959 --> 00:06:48.180
Yes.
146
00:06:49.060 --> 00:06:51.540
It turns the audio into a picture and analyzes the shapes.
147
00:06:52.259 --> 00:06:53.060
That is fascinating.
148
00:06:53.100 --> 00:06:56.519
It really is. The CNN analyzes that image to extract
149
00:06:56.579 --> 00:07:00.160
what they call a sonic fingerprint. It reads the tempo,
150
00:07:00.220 --> 00:07:03.139
the loudness, the musical key, the energy, and even the
151
00:07:03.220 --> 00:07:04.259
timbre of the instruments.
152
00:07:04.480 --> 00:07:05.459
Just by looking at the picture?
153
00:07:05.800 --> 00:07:08.750
Just by the visual patterns on the spectrogram. It can
154
00:07:08.759 --> 00:07:11.370
detect if a track has a heavy bass line or
155
00:07:11.449 --> 00:07:14.709
a distorted guitar, all from the image. And because of this,
156
00:07:15.269 --> 00:07:18.230
an entirely new song with zero streams can be mathematically
157
00:07:18.290 --> 00:07:22.949
matched to older, popular songs that have identical sonic fingerprints.
158
00:07:23.350 --> 00:07:27.259
Oh, I see. So it bypasses the cold start problem entirely.
159
00:07:27.319 --> 00:07:29.850
Exactly. It just pushes the new track to listeners who
160
00:07:30.009 --> 00:07:32.949
already favor that specific sonic architecture.
161
00:07:33.290 --> 00:07:37.370
Okay, so we have playlists mapping human behavior, web crawlers
162
00:07:37.430 --> 00:07:42.310
mapping cultural vocabulary, and neural networks visually analyzing sound waves.
163
00:07:43.029 --> 00:07:45.100
That covers how the machine understands the music.
164
00:07:45.180 --> 00:07:45.439
Right.
165
00:07:45.779 --> 00:07:47.740
But how does it map those billions of data points
166
00:07:47.800 --> 00:07:51.560
onto my personal habits? Because, I mean, my listening habits
167
00:07:51.600 --> 00:07:53.480
are chaotic. I don't just listen to one genre.
168
00:07:54.160 --> 00:07:57.379
Most people don't. And the transition from analyzing the catalog
169
00:07:57.579 --> 00:08:01.879
to analyzing you, the listener, centers on your taste profile.
170
00:08:01.899 --> 00:08:02.800
The taste profile.
171
00:08:03.120 --> 00:08:06.120
Yeah. The system logs every single interaction you have within
172
00:08:06.139 --> 00:08:10.079
the app, and it assigns different mathematical weights to different actions.
173
00:08:10.259 --> 00:08:13.029
So a quick skip, like hitting next within the first
174
00:08:13.350 --> 00:08:16.029
few seconds of a track, that carries a really heavy
175
00:08:16.170 --> 00:08:16.730
negative weight.
176
00:08:16.910 --> 00:08:19.069
It's like saying, absolutely not. Right.
177
00:08:19.589 --> 00:08:21.800
But saving a song... Or adding it to your own
178
00:08:21.839 --> 00:08:25.579
playlist carries a heavy positive weight. But it goes a
179
00:08:25.600 --> 00:08:28.540
lot deeper than just positive and negative. It categorizes your
180
00:08:28.660 --> 00:08:30.259
actions by context.
181
00:08:30.759 --> 00:08:34.059
Context. Meaning like the device I'm using or the time
182
00:08:34.100 --> 00:08:34.320
of day?
183
00:08:35.059 --> 00:08:38.330
Down to the minute. The system differentiates between the upbeat
184
00:08:38.399 --> 00:08:40.659
pop you cast to your smart speaker on a Saturday
185
00:08:40.710 --> 00:08:44.649
night and the ambient instrumental music you play on your
186
00:08:44.669 --> 00:08:46.269
phone during a Tuesday morning commute.
187
00:08:46.789 --> 00:08:49.289
It knows the difference between Saturday night me and Tuesday
188
00:08:49.330 --> 00:08:50.389
morning me. Exactly.
189
00:08:51.129 --> 00:08:54.750
It attempts to isolate your core persistent tastes from your
190
00:08:54.820 --> 00:08:56.549
temporary situational moods.
191
00:08:57.039 --> 00:08:59.100
Okay, but if it knows my core tastes so well,
192
00:08:59.500 --> 00:09:01.700
and it knows the context of my Tuesday morning commute,
193
00:09:02.720 --> 00:09:03.960
I have to ask about the curveballs.
194
00:09:04.460 --> 00:09:05.120
The curveballs?
195
00:09:05.179 --> 00:09:07.720
Yeah. Because every once in a while, it will throw
196
00:09:07.759 --> 00:09:10.220
a song into my mix that feels completely out of
197
00:09:10.259 --> 00:09:12.799
left field. It's a genre I don't usually listen to
198
00:09:12.840 --> 00:09:15.940
at all. Is that a glitch in my taste profile,
199
00:09:16.159 --> 00:09:18.070
or is it doing that on purpose?
200
00:09:18.590 --> 00:09:20.149
It is entirely intentional.
201
00:09:20.450 --> 00:09:21.129
Really? Yeah.
202
00:09:21.629 --> 00:09:25.090
This touches on a foundational concept in machine learning known
203
00:09:25.210 --> 00:09:28.149
as the explore versus exploit dilemma.
204
00:09:28.389 --> 00:09:29.789
Explore versus exploit.
205
00:09:30.029 --> 00:09:34.320
Right. Think about it. If an algorithm only ever exploits
206
00:09:34.360 --> 00:09:37.500
your known preferences by giving you your absolute favorite genres
207
00:09:37.860 --> 00:09:41.019
over and over, the experience quickly becomes a stale echo
208
00:09:41.059 --> 00:09:42.120
chamber and you lose interest.
209
00:09:42.340 --> 00:09:42.799
That makes sense.
210
00:09:43.149 --> 00:09:47.419
However, if it explores too heavily and constantly serves you
211
00:09:47.580 --> 00:09:51.820
unfamiliar genres, the cognitive load is too high, you get frustrated,
212
00:09:51.840 --> 00:09:52.919
and you just close the app.
213
00:09:53.320 --> 00:09:55.009
Oh, I see. It's like a local restaurant. If the
214
00:09:55.049 --> 00:09:57.809
chef only ever exploits my love for my usual order,
215
00:09:57.830 --> 00:10:01.620
I'm going to get bored eventually. they explore by forcing
216
00:10:01.659 --> 00:10:03.519
a weird new special on me every time I walk
217
00:10:03.539 --> 00:10:05.620
in the door, I'll probably just stop going.
218
00:10:05.879 --> 00:10:08.500
That's a perfect analogy. The chef has to
balance it.
219
00:10:09.019 --> 00:10:10.529
So how does the machine balance it?
220
00:10:10.840 --> 00:10:14.299
To manage that balance, the platform relies on algorithms modeled
221
00:10:14.379 --> 00:10:18.440
after multi-armed bandits. In the engineering research we looked at,
222
00:10:18.519 --> 00:10:21.919
this specific system is sometimes referred to as BART, which
223
00:10:21.980 --> 00:10:24.799
stands for Bandits for Recommendations as Treatments.
224
00:10:25.080 --> 00:10:28.759
Wait, multi-armed bandit. That's a casino term, right? It refers
225
00:10:28.799 --> 00:10:29.779
to a row of slot machines?
226
00:10:30.139 --> 00:10:33.830
Yes, exactly. So imagine you're standing in front of 10
227
00:10:33.830 --> 00:10:36.309
slot machines. You know one of them pays out frequently
228
00:10:36.429 --> 00:10:39.110
because you've played it before. That's your core musical taste.
229
00:10:39.789 --> 00:10:42.429
You can exploit that machine for a guaranteed but very
230
00:10:42.509 --> 00:10:44.950
predictable return. Or you can pull the lever on a
231
00:10:44.970 --> 00:10:47.820
completely unknown machine to explore its payout rate.
232
00:10:48.399 --> 00:10:49.120
I see where this is going.
233
00:10:50.200 --> 00:10:53.500
The algorithm is constantly running probability calculations on when to
234
00:10:53.519 --> 00:10:56.019
play it safe and when to pull an unknown lever.
235
00:10:56.940 --> 00:10:59.200
So when a strange song pops into your feed, the
236
00:10:59.240 --> 00:11:02.570
system is briefly shifting into explore mode. It is tossing
237
00:11:02.610 --> 00:11:04.970
you a curveball to gather fresh data on your boundaries.
238
00:11:05.190 --> 00:11:06.090
Wow, it's testing me.
239
00:11:06.850 --> 00:11:09.769
It's watching your real-time reaction. Do you skip it immediately?
240
00:11:09.789 --> 00:11:12.190
Do you let it play halfway through? Do you save it?
241
00:11:12.879 --> 00:11:15.559
your reaction updates the probability math for the next time
242
00:11:15.659 --> 00:11:16.600
it decides to explore.
243
00:11:16.879 --> 00:11:19.519
So it's actively testing the fences of my taste profile.
244
00:11:20.240 --> 00:11:22.870
That completely explains the random surprises in the daily mixes,
245
00:11:23.330 --> 00:11:26.570
which actually that brings us to the user interface. Because
246
00:11:26.610 --> 00:11:30.129
all of this underlying math, the three pillars, the multi-armed bandits,
247
00:11:30.490 --> 00:11:32.190
it has to surface somewhere, right?
248
00:11:32.509 --> 00:11:34.590
Right. The listener has to actually interact with it.
249
00:11:35.169 --> 00:11:39.379
And the platform has several different avenues for delivering these recommendations. Yeah.
250
00:11:39.419 --> 00:11:43.809
The interface segments these algorithms into distinct products. So Discover Weekly,
251
00:11:43.870 --> 00:11:47.799
which arrives on Mondays, leans heavily on collaborative filtering and
252
00:11:47.840 -->
00:11:51.899
the explore side of the bandit algorithm to surface new-to-you tracks.
253
00:11:52.200 --> 00:11:53.820
Monday is for exploring. Exactly.
254
00:11:53.860 --> 00:11:56.120
Then you have Release Radar, which drops on Fridays, and
255
00:11:56.159 --> 00:11:59.950
that leans heavily on the exploit side. It focuses strictly
256
00:12:00.070 --> 00:12:03.470
on new releases from artists you already demonstrate a very
257
00:12:03.529 --> 00:12:04.870
high affinity for. Okay.
258
00:12:05.629 --> 00:12:07.610
And then you have the AI DJ, which they just
259
00:12:07.649 --> 00:12:11.519
launched in 2023. That one is really interesting to me
260
00:12:11.559 --> 00:12:14.419
because it takes all the back-end personalization data we just
261
00:12:14.440 --> 00:12:17.960
talked about and wraps it in a generative voice model.
262
00:12:18.080 --> 00:12:20.440
Yeah, creating a synthetic radio host.
263
00:12:20.600 --> 00:12:24.019
Right, a fake DJ to actually explain to you why
264
00:12:24.059 --> 00:12:26.759
it's playing certain clusters of songs. Like, hey, I noticed
265
00:12:26.799 --> 00:12:28.519
you were listening to a lot of this last summer.
266
00:12:28.840 --> 00:12:31.659
It's a very clever way to humanize the data. And
267
00:12:31.700 --> 00:12:35.779
beyond the DJ, the platform also utilizes autoplay and radio features.
268
00:12:36.070 --> 00:12:38.570
That's where the algorithm takes the baton the moment an
269
00:12:38.669 --> 00:12:40.909
album or playlist ends, and it just tries to maintain
270
00:12:40.950 --> 00:12:43.269
the current sonic fingerprint indefinitely.
271
00:12:43.490 --> 00:12:44.289
Right, just keeps going.
272
00:12:45.190 --> 00:12:47.850
And honestly, the entire home screen itself is a dynamic
273
00:12:47.889 --> 00:12:51.210
grid ranked by a predicted engagement. The system is calculating
274
00:12:51.250 --> 00:12:54.870
the exact probability of you clicking a specific podcast or
275
00:12:54.950 --> 00:12:57.850
daily mix based on your historical behavior at that exact
276
00:12:57.870 --> 00:12:58.309
time of day.
277
00:12:59.000 --> 00:13:01.639
So my home screen looks totally different from yours at, say,
278
00:13:01.639 --> 00:13:02.940
2 p.m. on a Thursday.
279
00:13:03.379 --> 00:13:04.090
Entirely different.
280
00:13:04.659 --> 00:13:08.309
Now, with all this incredibly sophisticated automation, I mean, neural networks,
281
00:13:08.370 --> 00:13:12.309
reading spectrograms, slot machine probability algorithms running in real time,
282
00:13:13.269 --> 00:13:16.950
it really begs the question of human involvement. Are human
283
00:13:17.009 --> 00:13:20.889
curators essentially obsolete in this ecosystem? Like, has the machine
284
00:13:20.990 --> 00:13:23.470
entirely replaced the program director at a radio station?
285
00:13:23.970 --> 00:13:27.490
It's a fair question, but no. Human curation is far
286
00:13:27.529 --> 00:13:31.289
from obsolete. In fact, it is structurally vital
to the platform.
287
00:13:31.720 --> 00:13:35.159
The company employs what they call an algotorial strategy.
288
00:13:35.519 --> 00:13:41.659
Algotorial. That's a portmanteau of algorithmic and editorial, right? Precisely.
289
00:13:42.019 --> 00:13:45.519
It is a hybrid model because the massive flagship playlists,
290
00:13:45.860 --> 00:13:49.039
the ones with millions of followers, they are not generated
291
00:13:49.090 --> 00:13:50.350
from scratch by cold math.
292
00:13:50.529 --> 00:13:50.870
They aren't.
293
00:13:51.409 --> 00:13:55.110
No. Human editors who are deep experts in specific cultural
294
00:13:55.169 --> 00:13:58.889
scenes or genres curate the foundational pool of tracks. An
295
00:13:58.970 --> 00:14:01.750
editor might decide that a specific list of, say, 100
296
00:14:01.750 --> 00:14:04.809
songs represents the absolute best of a particular genre for
297
00:14:04.850 --> 00:14:05.450
that given week.
298
00:14:06.090 --> 00:14:09.500
Oh, I see. And then the algorithm takes that human-curated
299
00:14:09.580 --> 00:14:12.899
pool and filters it for the individual listener. Exactly.
300
00:14:13.340 --> 00:14:16.840
The algorithm scores those 100 tracks against your personal taste profile.
301
00:14:17.340 --> 00:14:20.340
And then it selects a customized subset, maybe 30 songs,
302
00:14:20.440 --> 00:14:23.289
and personalizes the order in which they appear on your
303
00:14:23.370 --> 00:14:24.330
specific screen.
304
00:14:25.029 --> 00:14:27.809
So the human defines the boundaries of the sandbox, right?
305
00:14:28.269 --> 00:14:30.370
And the machine organizes the sand for you.
306
00:14:30.769 --> 00:14:33.750
That's a great way to put it. Furthermore, independent artists
307
00:14:33.830 --> 00:14:36.590
and labels can actually utilize a portal to pitch their
308
00:14:36.730 --> 00:14:39.120
unreleased music directly to these human editors.
309
00:14:39.539 --> 00:14:41.519
Oh, so they can bypass the cold start problem by
310
00:14:41.539 --> 00:14:42.559
just pitching to a human.
311
00:14:42.840 --> 00:14:46.700
Right. It ensures that human intuition remains a primary entry
312
00:14:46.740 --> 00:14:48.740
point into the algorithmic ecosystem.
313
00:14:49.039 --> 00:14:50.559
You know, that makes the system feel a lot less
314
00:14:50.600 --> 00:14:51.519
dystopian to me.
315
00:14:51.860 --> 00:14:52.610
I agree.
316
00:14:53.230 --> 00:14:56.389
But, you know, blending human curation with global machine prediction
317
00:14:56.409 --> 00:15:00.580
is Inevitably creates friction. When you have a single algorithm
318
00:15:00.620 --> 00:15:03.799
steering the attention of hundreds of millions of people, you
319
00:15:03.840 --> 00:15:06.000
are quite literally shaping culture itself.
320
00:15:06.360 --> 00:15:07.919
Absolutely. The stakes are huge.
321
00:15:08.940 --> 00:15:11.990
Looking through the research notes, there are some really fierce
322
00:15:12.029 --> 00:15:15.590
debates surrounding the
consequences of this system. We need to
323
00:15:15.610 --> 00:15:16.470
look at those impartially.
324
00:15:16.850 --> 00:15:19.730
Yeah, these debates are essential for understanding the environment you're
325
00:15:19.769 --> 00:15:22.629
navigating as a listener. And one of the most heavily
326
00:15:22.730 --> 00:15:26.210
researched topics in the sources is the concept of filter bubbles.
327
00:15:26.340 --> 00:15:26.879
Filter bubbles.
328
00:15:27.080 --> 00:15:31.700
Right. The concern is that hyper-personalization narrows a listener's cultural diet.
329
00:15:32.340 --> 00:15:34.519
If the machine's primary goal is to predict what you
330
00:15:34.639 --> 00:15:37.539
already want, it creates a feedback loop that traps you
331
00:15:37.590 --> 00:15:40.269
in an echo chamber of your own aesthetic preferences.
332
00:15:40.590 --> 00:15:44.049
Right. I mean, if the algorithm only feeds me lo-fi
333
00:15:44.129 --> 00:15:46.929
indie rock because it knows I definitely won't skip it,
334
00:15:47.870 --> 00:15:51.149
then I'll never discover jazz or classical or hip-hop. My
335
00:15:51.190 --> 00:15:52.730
taste will just completely calcify.
336
00:15:53.049 --> 00:15:55.730
That is exactly the fear. But it's important to note
337
00:15:55.789 --> 00:15:59.149
that the academic research on music recommendation engines is actually
338
00:15:59.190 --> 00:16:00.049
quite mixed on this.
339
00:16:00.220 --> 00:16:00.429
It is.
340
00:16:00.740 --> 00:16:04.360
Yeah, some studies definitely support the echo chamber theory. However,
341
00:16:04.620 --> 00:16:08.659
other data sets indicate that algorithmic recommendation can actually broaden
342
00:16:08.799 --> 00:16:12.490
a user's horizons far more effectively than traditional terrestrial radio
343
00:16:12.590 --> 00:16:13.049
ever did.
344
00:16:13.110 --> 00:16:16.789
Because they're only playing, because traditional radio is just playing
345
00:16:16.809 --> 00:16:17.909
the top 40.
346
00:16:17.909 --> 00:16:22.129
Exactly. It's geographically and temporally limited. It plays a very
347
00:16:22.190 --> 00:16:24.629
narrow band of top hits to appeal to the broadest
348
00:16:24.710 --> 00:16:29.240
possible demographic. The algorithm, conversely, will readily expose a listener
349
00:16:29.279 --> 00:16:32.899
in Chicago to an obscure microgenre from Tokyo, as long
350
00:16:32.940 --> 00:16:34.340
as the sonic fingerprints align.
351
00:16:34.580 --> 00:16:38.820
That makes sense. It's genre agnostic, as long as the
352
00:16:39.120 --> 00:16:40.580
math works. Right. Yeah.
353
00:16:41.379 --> 00:16:43.850
The degree to which it narrows or broadens your taste
354
00:16:43.990 --> 00:16:46.629
often just depends on how the individual user responds to
355
00:16:46.669 --> 00:16:49.389
those responses. explore tests we discussed earlier.
356
00:16:49.990 --> 00:16:53.090
Which perfectly transitions to the industry side of the debate.
357
00:16:53.649 --> 00:16:57.779
Because if the
algorithm determines what gets discovered, it ultimately
358
00:16:57.820 --> 00:16:59.019
determines who gets paid, right?
359
00:16:59.299 --> 00:16:59.600
Oh, yes.
360
00:16:59.799 --> 00:16:59.980
Yeah.
361
00:17:00.299 --> 00:17:01.190
That is the core of it.
362
00:17:01.659 --> 00:17:04.759
And there is significant friction between artists and the platform
363
00:17:05.319 --> 00:17:08.119
regarding how this automated discovery is monetized.
364
00:17:08.619 --> 00:17:11.160
A major flashpoint we saw in the sources is a
365
00:17:11.200 --> 00:17:15.069
program that was introduced in 2020 called Discovery Mode. discovery
366
00:17:15.109 --> 00:17:17.930
mode yeah this system allows artists and rights holders to
367
00:17:17.970 --> 00:17:21.539
flag specific tracks which signals to the algorithm that they
368
00:17:21.579 --> 00:17:25.859
want those songs prioritized in personalized listening contexts like autoplay
369
00:17:25.940 --> 00:17:26.880
and radio.
370
00:17:26.779 --> 00:17:29.640
So they can ask for an algorithmic boost.
371
00:17:29.660 --> 00:17:33.980
Yes but the catch is the economic trade-off in exchange
372
00:17:34.019 --> 00:17:37.319
for that algorithmic boost the rights holders agree to accept
373
00:17:37.500 --> 00:17:40.460
a lower royalty rate for any streams generated in those
374
00:17:40.519 --> 00:17:43.039
specific discovery contexts.
375
00:17:42.599 --> 00:17:45.640
Whoa That sounds like a highly contentious tradeoff.
376
00:17:45.779 --> 00:17:49.640
It is fiercely debated. Critics within the industry have drawn
377
00:17:49.700 --> 00:17:53.460
parallels to payola, which is the illegal practice of paying
378
00:17:53.579 --> 00:17:57.359
radio stations to play specific songs. Right. They argue that
379
00:17:57.400 --> 00:18:01.140
discovery mode forces independent artists into a race to the bottom,
380
00:18:01.460 --> 00:18:04.180
basically requiring them to take a pay cut just to
381
00:18:04.220 --> 00:18:07.859
remain visible in an algorithmically dominated ecosystem.
382
00:18:08.240 --> 00:18:11.109
That's a strong criticism. But what does the platform say?
383
00:18:11.509 --> 00:18:14.730
Well, the platform, conversely, argues that it functions as a
384
00:18:14.809 --> 00:18:19.609
highly accessible marketing tool. Unlike traditional promotional campaigns that require
385
00:18:19.690 --> 00:18:23.950
massive upfront capital, discovery mode requires zero upfront budget.
386
00:18:23.990 --> 00:18:25.509
So you don't need a huge marketing team.
387
00:18:25.750 --> 00:18:29.069
Exactly. They argue it allows independent artists to gain global
388
00:18:29.130 --> 00:18:30.970
reach that they otherwise just couldn't afford.
389
00:18:30.990 --> 00:18:34.880
And it fundamentally alters the economics of streaming either way
390
00:18:34.890 --> 00:18:37.319
you look at it. And the sources point out that
391
00:18:37.359 --> 00:18:40.460
this is really just one piece of a broader economic puzzle.
392
00:18:41.079 --> 00:18:45.579
alongside debates over base royalty rates, the pooling of subscription revenue,
393
00:18:46.059 --> 00:18:48.900
and even controversies over so-called ghost artists.
394
00:18:49.160 --> 00:18:49.980
Oh, the ghost artists.
395
00:18:50.000 --> 00:18:53.890
Yeah, these fake artists racking up algorithmic streams on ambient
396
00:18:53.970 --> 00:18:55.410
noise playlists. It's wild.
397
00:18:55.950 --> 00:19:00.529
The streaming economy is incredibly complex, and it is actively contested.
398
00:19:00.950 --> 00:19:03.890
The critical takeaway for the listener here is just awareness.
399
00:19:04.500 --> 00:19:08.059
The seamless, highly intuitive listening experience provided by the app
400
00:19:08.119 --> 00:19:10.680
is built on top of a very dense library often
401
00:19:10.740 --> 00:19:12.759
fraught economic reality for the creators.
402
00:19:13.400 --> 00:19:17.049
Absolutely. So given all of these mechanics, the sonic analysis,
403
00:19:17.109 --> 00:19:20.430
the probability tests, the economic levers, how do we take
404
00:19:20.470 --> 00:19:24.109
the wheel? I want to actively shape my profile, not
405
00:19:24.130 --> 00:19:26.759
just drift along the algorithm's currents.
406
00:19:26.920 --> 00:19:28.359
Right. You want to train it. Yeah.
407
00:19:28.920 --> 00:19:32.359
What are the concrete practical levers a user can pull?
408
00:19:32.759 --> 00:19:34.900
And to be clear, we are not talking about secret
409
00:19:34.920 --> 00:19:37.480
developer hacks here, just the actual built-in tools.
410
00:19:38.029 --> 00:19:41.450
Well, the most effective lever is intentional feedback. You have
411
00:19:41.490 --> 00:19:45.319
to remember that the algorithm interprets passivity as acceptance.
412
00:19:45.480 --> 00:19:46.700
Passivity is acceptance.
413
00:19:47.049 --> 00:19:49.619
Yes. If a track appears on Discover Weekly that you
414
00:19:49.779 --> 00:19:52.579
actively dislike, do not just skip it or let it
415
00:19:52.640 --> 00:19:56.079
play out of apathy. Use the interface to hide the song.
416
00:19:56.599 --> 00:19:59.039
Select the options to exclude the artist.
417
00:19:59.440 --> 00:20:00.619
So I have to be proactive.
418
00:20:00.859 --> 00:20:04.380
Exactly. Conversely, if you want the algorithm to learn a
419
00:20:04.420 --> 00:20:07.119
new preference, you have to deliberately seek it out and
420
00:20:07.180 --> 00:20:10.680
engage with it by saving tracks and following artists. Your
421
00:20:10.859 --> 00:20:12.660
actions dictate the probability models.
422
00:20:13.140 --> 00:20:15.799
So it's basically like training a puppy. A short skip
423
00:20:15.900 --> 00:20:18.160
is a harsh no. A save is a treat.
424
00:20:18.440 --> 00:20:19.539
That's a great way to think about it.
425
00:20:20.200 --> 00:20:22.420
But wait, what about contextual pollution?
426
00:20:22.599 --> 00:20:23.000
What do you mean?
427
00:20:23.640 --> 00:20:26.259
Well, a lot of people fall
asleep to ambient rain sounds, right?
428
00:20:26.700 --> 00:20:29.140
Or they share an account with their toddler. If I
429
00:20:29.190 --> 00:20:32.049
play eight hours of white noise or thunderstorm sounds every
430
00:20:32.089 --> 00:20:35.750
single night, will the neural network assume my absolute favorite
431
00:20:35.769 --> 00:20:39.789
genre is thunderstorm and just destroy my music recommendations?
432
00:20:40.130 --> 00:20:43.069
Oh, that is a very common issue, and the platform
433
00:20:43.109 --> 00:20:47.150
actually provides two specific tools to mitigate it. First, you
434
00:20:47.170 --> 00:20:50.849
can utilize the private session feature. Activating this before you
435
00:20:50.890 --> 00:20:55.069
play your sleep sounds effectively pauses the taste profile data collection.
436
00:20:55.609 --> 00:20:58.130
It blinds the algorithm to your activity for that session.
437
00:20:58.480 --> 00:21:00.319
Okay, that's great for one-off situations.
438
00:21:01.099 --> 00:21:04.220
It is, but for recurring habits, there is a better lever.
439
00:21:04.849 --> 00:21:07.849
You can navigate to any specific playlist, whether it's your
440
00:21:07.950 --> 00:21:11.430
white noise or a playlist of children's music. Open the
441
00:21:11.470 --> 00:21:14.470
menu and select exclude from your taste profile.
442
00:21:14.789 --> 00:21:16.690
Oh, I didn't even know that existed.
443
00:21:16.890 --> 00:21:20.400
It's a lifesaver. This tells the algorithm to permanently ignore
444
00:21:20.460 --> 00:21:24.240
all streaming data originating from that specific playlist. It basically
445
00:21:24.279 --> 00:21:28.059
quarantines the data so it doesn't cross-contaminate your primary music recommendations.
446
00:21:28.400 --> 00:21:31.440
That is a crucial feature for anyone with kids. Wow.
447
00:21:32.490 --> 00:21:34.930
It really comes down to managing the data you feed
448
00:21:34.970 --> 00:21:35.450
the machine.
449
00:21:35.789 --> 00:21:36.309
It really does.
450
00:21:37.349 --> 00:21:40.730
So to tie all of this together, we started by
451
00:21:40.769 --> 00:21:43.869
wondering how a machine mimics the intuition of a friend
452
00:21:43.910 --> 00:21:47.210
handing you a mixtape. And the reality is that it
453
00:21:47.250 --> 00:21:49.670
requires a staggering synthesis of data.
454
00:21:49.930 --> 00:21:50.450
Staggering.
455
00:21:51.049 --> 00:21:54.079
It uses collaborative filtering to map the behavior of millions
456
00:21:54.119 --> 00:21:58.019
of strangers. It uses natural language processing to read the
457
00:21:58.059 --> 00:22:02.720
cultural vocabulary of the Internet. It uses convolutional neural networks
458
00:22:02.779 --> 00:22:05.700
to visually map the raw sound waves of the audio itself.
459
00:22:06.160 --> 00:22:06.980
All working together.
460
00:22:07.220 --> 00:22:09.940
Yeah. And it filters all of that through a probability
461
00:22:09.980 --> 00:22:13.500
matrix that constantly calculates whether
to serve you your favorites
462
00:22:13.960 --> 00:22:17.779
or test your boundaries, all while blending in human editorial taste.
463
00:22:17.960 --> 00:22:19.960
It's an incredible piece of engineering.
464
00:22:20.289 --> 00:22:23.809
It is. Every time you open a Monday playlist, you're
465
00:22:23.849 --> 00:22:27.430
looking at the result of a massive, silent conversation between
466
00:22:27.970 --> 00:22:31.059
your past behavior... the habits of millions of other people,
467
00:22:31.640 --> 00:22:34.799
and a machine trying to calculate your future desires. And
468
00:22:35.000 --> 00:22:37.799
the core lesson from the sources is clear. The more
469
00:22:37.859 --> 00:22:42.079
deliberately and honestly you listen, the more accurately the machine responds.
470
00:22:42.640 --> 00:22:45.279
It really just functions as a digital mirror of your
471
00:22:45.339 --> 00:22:48.200
auditory habits. But, you know, understanding how it works today
472
00:22:48.279 --> 00:22:51.960
raises a really compelling question about where this technology is
473
00:22:51.980 --> 00:22:52.940
heading tomorrow.
474
00:22:53.000 --> 00:22:54.440
Tomorrow? What do you mean?
475
00:22:54.740 --> 00:22:57.859
Well, currently the algorithm is entirely dependent on the physical
476
00:22:57.900 --> 00:23:00.660
inputs you provide, your searches, your saves, your skips, right?
477
00:23:01.339 --> 00:23:06.859
But biometric technology, particularly in smartwatches and wearables, is advancing rapidly.
478
00:23:07.220 --> 00:23:10.019
Oh, connecting the audio data to physical data.
479
00:23:10.200 --> 00:23:14.180
Exactly. Imagine a scenario where the recommendation engine no longer
480
00:23:14.269 --> 00:23:16.569
needs you to press a button. What happens when the
481
00:23:16.640 --> 00:23:18.990
algorithm integrates with your biometric sensors?
482
00:23:19.170 --> 00:23:19.289
Wow.
483
00:23:19.920 --> 00:23:22.720
It could monitor your heart rate, your blood pressure, or
484
00:23:22.740 --> 00:23:25.869
your galvanic skin response in real time. If it detects
485
00:23:25.910 --> 00:23:28.990
your stress levels rising during a commute, it could automatically
486
00:23:29.049 --> 00:23:32.309
cross-reference that data with sonic profiles known to lower your
487
00:23:32.349 --> 00:23:35.470
heart rate, seamlessly fading in a specific ambient track.
488
00:23:35.549 --> 00:23:37.109
Without you even asking for it. Right.
489
00:23:37.789 --> 00:23:40.410
It wouldn't just be reacting to your musical preferences anymore.
490
00:23:40.450 --> 00:23:42.759
It would be reacting to your physiological state.
491
00:23:43.440 --> 00:23:46.819
Which leaves us with a wild thought to ponder. If
492
00:23:46.859 --> 00:23:50.549
an algorithm is monitoring your biometrics... And serving you audio
493
00:23:50.569 --> 00:23:54.700
specifically designed to alter your physiological state in real time.
494
00:23:55.880 --> 00:24:00.089
At that point, who is really regulating your emotions? You
495
00:24:00.549 --> 00:24:01.049
or the machine?