1
00:00:00,000 --> 00:00:04,600
ever wish your music app could like actually explain why it thinks you'll love a song?

2
00:00:04,600 --> 00:00:05,400
Yeah.

3
00:00:05,400 --> 00:00:07,200
Instead of just kind of throwing a bunch of tracks at you.

4
00:00:07,200 --> 00:00:08,200
All right.

5
00:00:08,200 --> 00:00:11,200
Well, that's exactly what Spotify is trying to do.

6
00:00:11,200 --> 00:00:15,500
And today we're going to be diving deep into their latest experiment with AI.

7
00:00:15,500 --> 00:00:16,000
Okay.

8
00:00:16,000 --> 00:00:20,000
All about personalized narratives for music recommendations.

9
00:00:20,000 --> 00:00:22,000
So like making it feel more personal.

10
00:00:22,000 --> 00:00:22,500
Yeah.

11
00:00:22,500 --> 00:00:25,300
Recommendations from a friend who gets your music taste.

12
00:00:25,300 --> 00:00:26,100
Yeah.

13
00:00:26,100 --> 00:00:26,600
That's pretty cool.

14
00:00:26,600 --> 00:00:28,300
I actually read their blog post about it.

15
00:00:28,300 --> 00:00:28,700
Oh, cool.

16
00:00:28,700 --> 00:00:33,800
Contextualized recommendations through personalized narratives using LLMs.

17
00:00:33,800 --> 00:00:34,000
Okay.

18
00:00:34,000 --> 00:00:36,200
Now you're just using big words to sound smart.

19
00:00:36,200 --> 00:00:39,400
Well, maybe a little, but LLMs are actually pretty fascinating.

20
00:00:39,400 --> 00:00:39,600
All right.

21
00:00:39,600 --> 00:00:40,500
So break it down for me.

22
00:00:40,500 --> 00:00:41,600
What are LLMs?

23
00:00:41,600 --> 00:00:45,300
LLMs are large language models.

24
00:00:45,300 --> 00:00:50,800
They're basically the brains behind a lot of the AI stuff we see these days.

25
00:00:50,800 --> 00:00:52,100
Like think chat, GPT.

26
00:00:52,100 --> 00:00:52,300
Okay.

27
00:00:52,300 --> 00:00:52,500
Yeah.

28
00:00:52,500 --> 00:00:53,100
I've heard of that.

29
00:00:53,100 --> 00:00:55,300
They're trained on tons and tons of data.

30
00:00:55,300 --> 00:00:55,800
Right.

31
00:00:55,800 --> 00:00:59,200
And they can generate text that sounds really human like.

32
00:00:59,200 --> 00:01:00,000
That's so cool.

33
00:01:00,000 --> 00:01:02,000
And even like hold conversations.

34
00:01:02,000 --> 00:01:02,600
Wow.

35
00:01:02,600 --> 00:01:02,800
Yeah.

36
00:01:02,800 --> 00:01:08,800
So basically Spotify wants to use this conversational AI to tell us why they think we'd like a certain song.

37
00:01:08,800 --> 00:01:09,600
Exactly.

38
00:01:09,600 --> 00:01:15,200
So instead of just seeing an album cover, you might get a little blurb that says something like,

39
00:01:15,200 --> 00:01:19,800
this band's latest single is a total metal core adrenaline rush.

40
00:01:19,800 --> 00:01:20,300
Interesting.

41
00:01:20,300 --> 00:01:25,000
It's almost like they're trying to like actually capture the way a friend would describe a song.

42
00:01:25,000 --> 00:01:26,500
That's exactly the vibe they're going for.

43
00:01:26,500 --> 00:01:28,600
Like making it more personal and engaging.

44
00:01:28,600 --> 00:01:31,200
And you know what their early tests show that,

45
00:01:31,200 --> 00:01:31,500
Like.

46
00:01:31,500 --> 00:01:37,000
that these explanations actually make listeners up to four times more likely to check out the recommendation.

47
00:01:37,000 --> 00:01:37,400
Wow.

48
00:01:37,400 --> 00:01:38,500
That's huge.

49
00:01:38,500 --> 00:01:41,600
But are they using any specific type of LLM?

50
00:01:41,600 --> 00:01:43,300
Like there are so many out there right now.

51
00:01:43,300 --> 00:01:43,700
Yeah.

52
00:01:43,700 --> 00:01:44,300
There are.

53
00:01:44,300 --> 00:01:50,800
They actually tried out a bunch of different ones, but they found that Meta's LLM models worked best for what they wanted to do.

54
00:01:50,800 --> 00:01:51,000
Okay.

55
00:01:51,000 --> 00:01:53,100
So what makes LLM so special?

56
00:01:53,100 --> 00:01:54,900
Why LLM overall the others?

57
00:01:54,900 --> 00:02:01,600
Well, for one, LLM already knows a lot about music, podcasts, just the whole entertainment world in general.

58
00:02:01,600 --> 00:02:03,500
And it's also really adaptable.

59
00:02:03,500 --> 00:02:04,000
Okay.

60
00:02:04,000 --> 00:02:08,300
So Spotify can like fine tune it to do very specific things for them.

61
00:02:08,300 --> 00:02:10,300
So LLM is the engine behind all this.

62
00:02:10,300 --> 00:02:11,000
Yeah.

63
00:02:11,000 --> 00:02:11,900
You could say that.

64
00:02:11,900 --> 00:02:12,400
Interesting.

65
00:02:12,400 --> 00:02:13,900
So how are they actually using it?

66
00:02:13,900 --> 00:02:15,000
What are they doing with it?

67
00:02:15,000 --> 00:02:17,400
Well, they're focusing on two big use cases.

68
00:02:17,400 --> 00:02:21,200
First contextualized recommendations, which we just talked about.

69
00:02:21,200 --> 00:02:21,800
Right.

70
00:02:21,800 --> 00:02:26,100
And then they're using it to make their AIDJ feature even better.

71
00:02:26,100 --> 00:02:26,800
Oh yeah.

72
00:02:26,800 --> 00:02:27,800
I've heard about that.

73
00:02:27,800 --> 00:02:29,800
It's like having your own personal radio stage, right?

74
00:02:29,800 --> 00:02:30,900
Pretty much.

75
00:02:30,900 --> 00:02:34,300
But with LLM's, it goes way beyond just playing songs.

76
00:02:34,300 --> 00:02:38,400
Now the AIDJ can actually give you personalized commentary in real time.

77
00:02:38,400 --> 00:02:39,000
I'm listening.

78
00:02:39,000 --> 00:02:40,600
Like it's actually talking to you.

79
00:02:40,600 --> 00:02:40,900
Okay.

80
00:02:40,900 --> 00:02:42,200
So paint me a picture.

81
00:02:42,200 --> 00:02:44,100
What kind of commentary are we talking about?

82
00:02:44,100 --> 00:02:45,500
Well, could be anything really.

83
00:02:45,500 --> 00:02:45,800
Okay.

84
00:02:45,800 --> 00:02:51,600
It could be like a fun fact about the artist or the story behind a song or even connecting it back to your own

85
00:02:51,600 --> 00:02:52,400
listening history.

86
00:02:52,400 --> 00:02:53,000
Oh wow.

87
00:02:53,000 --> 00:02:58,800
Like maybe the AIDJ remembers that you used to listen to a certain band a lot a few years ago.

88
00:02:58,800 --> 00:03:02,600
And then it points out how the new song you're hearing reminds it of their older stuff.

89
00:03:02,600 --> 00:03:03,200
Wow.

90
00:03:03,200 --> 00:03:03,600
Yeah.

91
00:03:03,600 --> 00:03:04,200
It's pretty cool.

92
00:03:04,200 --> 00:03:04,800
That's super cool.

93
00:03:04,800 --> 00:03:11,400
It's like the AIDJ is becoming this like musical buddy who's sharing all its knowledge and insights with you.

94
00:03:11,400 --> 00:03:12,300
Right.

95
00:03:12,300 --> 00:03:19,500
But you know, with all this talk about AI, I got to ask, are human music experts going to get replaced by algorithms?

96
00:03:19,500 --> 00:03:20,900
That's the question, isn't it?

97
00:03:20,900 --> 00:03:21,200
Yeah.

98
00:03:21,200 --> 00:03:23,300
But Spotify is making it very clear.

99
00:03:23,300 --> 00:03:26,400
Human expertise is still super important.

100
00:03:26,400 --> 00:03:26,600
Okay.

101
00:03:26,600 --> 00:03:37,200
They've actually got music editors working alongside the LLMs making sure the AIDJ's commentary is relevant, insightful, and you know, culturally sensitive.

102
00:03:37,200 --> 00:03:37,600
Makes sense.

103
00:03:37,600 --> 00:03:38,300
All that good stuff.

104
00:03:38,300 --> 00:03:42,200
So it's more like a collaboration like human and artificial intelligence working together.

105
00:03:42,200 --> 00:03:42,900
Exactly.

106
00:03:42,900 --> 00:03:50,700
It's all about finding that sweet spot where AI can enhance human creativity and knowledge, not replace it.

107
00:03:50,700 --> 00:03:51,600
Not like that.

108
00:03:51,600 --> 00:03:54,300
So it sounds like Spotify is really trying to be thoughtful about all of this.

109
00:03:54,300 --> 00:03:54,600
Yeah.

110
00:03:54,600 --> 00:03:59,900
But I imagine there are some pretty big technical challenges when you're dealing with these massive AI models, right?

111
00:03:59,900 --> 00:04:00,900
Oh, absolutely.

112
00:04:00,900 --> 00:04:06,200
Like one of the biggest ones is just the sheer computing power that you need to train these LLMs.

113
00:04:06,200 --> 00:04:06,700
Yeah.

114
00:04:06,700 --> 00:04:07,400
That makes sense.

115
00:04:07,400 --> 00:04:09,000
It's not just about having smart algorithms.

116
00:04:09,000 --> 00:04:13,200
It's about having the infrastructure to actually support them.

117
00:04:13,200 --> 00:04:18,500
So Spotify had to invest in some serious hardware to make this all happen.

118
00:04:18,500 --> 00:04:19,100
For sure.

119
00:04:19,100 --> 00:04:25,400
I mean, they had to come up with whole new systems and techniques just to deal with the demands of LLM training.

120
00:04:25,400 --> 00:04:34,600
Like, you know, they even created a system that saves their progress during training so they don't lose everything if like there's a system failure or something.

121
00:04:34,600 --> 00:04:35,700
That's smart.

122
00:04:35,700 --> 00:04:38,300
You'll be a nightmare to have to start from scratch every time.

123
00:04:38,300 --> 00:04:38,900
Totally.

124
00:04:38,900 --> 00:04:40,900
But training the model is only one part of it, right?

125
00:04:40,900 --> 00:04:41,600
Yeah.

126
00:04:41,600 --> 00:04:45,500
They also have to figure out how to make all of this work smoothly for millions of users.

127
00:04:45,500 --> 00:04:46,400
You got it.

128
00:04:46,400 --> 00:04:50,000
Getting these AI features to run smoothly for so many people.

129
00:04:50,000 --> 00:04:50,200
Yeah.

130
00:04:50,200 --> 00:04:51,500
That's a huge challenge.

131
00:04:51,500 --> 00:04:52,100
I bet.

132
00:04:52,100 --> 00:05:02,300
Spotify is using a combination of these like lightweight AI models and some really clever optimization tricks just to make sure that everything runs quickly and seamlessly.

133
00:05:02,300 --> 00:05:10,700
So it sounds like they're having to get pretty creative on the engineering side to balance the power of these AI models with the need for a good user experience.

134
00:05:10,700 --> 00:05:11,400
Exactly.

135
00:05:11,400 --> 00:05:13,700
It's a delicate balance for sure.

136
00:05:13,700 --> 00:05:19,600
And you know, while we're on the topic of Spotify's AI strategy, there's another thing I wanted to touch on.

137
00:05:19,600 --> 00:05:21,500
Their commitment to open source.

138
00:05:21,500 --> 00:05:22,100
Oh, yeah.

139
00:05:22,100 --> 00:05:23,200
That's a big one.

140
00:05:23,200 --> 00:05:26,200
So how does open source fit into all of this?

141
00:05:26,200 --> 00:05:30,400
Well, Spotify is super active in the open source AI community.

142
00:05:30,400 --> 00:05:38,900
So they're not only benefiting from all the amazing tools and innovations that other people are creating, but they're also giving back and sharing their own advancements.

143
00:05:38,900 --> 00:05:42,100
So it's like a two way street where everybody benefits from the shared knowledge.

144
00:05:42,100 --> 00:05:42,600
Exactly.

145
00:05:42,600 --> 00:05:46,500
It's a win-win for Spotify and the AI community as a whole.

146
00:05:46,500 --> 00:05:48,000
I like that.

147
00:05:48,000 --> 00:05:52,100
This has been a super fascinating look at all the technical stuff going on behind the scenes.

148
00:05:52,100 --> 00:05:54,200
But let's zoom out for a second.

149
00:05:54,200 --> 00:05:57,800
What does all of this actually mean for the average listener?

150
00:05:57,800 --> 00:05:58,300
Right.

151
00:05:58,300 --> 00:06:01,600
How does it change our experience with music and podcasts?

152
00:06:01,600 --> 00:06:07,400
Well, I think Spotify's vision here is to help us connect with audio content in a much deeper way.

153
00:06:07,400 --> 00:06:07,800
Okay.

154
00:06:07,800 --> 00:06:16,100
You know, they want to take the guesswork out of discovering new artists and help us understand why we might love a particular song or podcast.

155
00:06:16,100 --> 00:06:20,300
It's like having a personal music expert who just knows you inside and out.

156
00:06:20,300 --> 00:06:21,300
That's a great way to put it.

157
00:06:21,300 --> 00:06:23,600
And it goes beyond just convenience too.

158
00:06:23,600 --> 00:06:24,100
Okay.

159
00:06:24,100 --> 00:06:30,100
Imagine an AI DJ that's introducing you to genres and artists that you never would have found on your own.

160
00:06:30,100 --> 00:06:33,000
That's actually what excites me the most about all of this.

161
00:06:33,000 --> 00:06:37,800
The possibility of like breaking out of my little music bubble and discovering something completely new.

162
00:06:37,800 --> 00:06:38,800
Exactly.

163
00:06:38,800 --> 00:06:44,300
And this kind of personalized discovery could have a huge impact on the entire music industry.

164
00:06:44,300 --> 00:06:44,800
Really?

165
00:06:44,800 --> 00:06:45,800
How so?

166
00:06:45,800 --> 00:06:46,500
Think about it.

167
00:06:46,500 --> 00:06:50,800
If AI can help us discover new artists more effectively,

168
00:06:50,800 --> 00:06:54,800
it means more opportunities for those artists to connect with their audience.

169
00:06:54,800 --> 00:06:57,700
It's almost like it levels the playing field a bit.

170
00:06:57,700 --> 00:07:00,800
Giving those independent artists a better chance of being heard.

171
00:07:00,800 --> 00:07:02,300
That's the idea.

172
00:07:02,300 --> 00:07:06,100
And it could even help us rediscover artists that we might have forgotten about

173
00:07:06,100 --> 00:07:09,700
or find some hidden gems in our own music libraries.

174
00:07:09,700 --> 00:07:10,200
Right.

175
00:07:10,200 --> 00:07:13,600
Like we always talk about how vast the music universe is.

176
00:07:13,600 --> 00:07:17,300
But now we have AI to help us navigate it and find those hidden treasures.

177
00:07:17,300 --> 00:07:18,500
Exactly.

178
00:07:18,500 --> 00:07:20,600
And all of this isn't just limited to music either.

179
00:07:20,600 --> 00:07:23,200
They want to do the same thing for podcasts too.

180
00:07:23,200 --> 00:07:24,400
That makes sense.

181
00:07:24,400 --> 00:07:26,100
There's so many podcasts out there now.

182
00:07:26,100 --> 00:07:26,500
Right.

183
00:07:26,500 --> 00:07:27,700
It's overwhelming.

184
00:07:27,700 --> 00:07:29,700
It can be hard to know where to even start.

185
00:07:29,700 --> 00:07:30,700
Exactly.

186
00:07:30,700 --> 00:07:35,200
So AI could really help cut through all the noise and find the shows that we'd actually enjoy.

187
00:07:35,200 --> 00:07:39,100
And it can also help us discover shows that align with our interests,

188
00:07:39,100 --> 00:07:41,700
even if they're not the most popular ones out there.

189
00:07:41,700 --> 00:07:45,400
It's like having a personal podcast curator who's sorting through all the options for you.

190
00:07:45,400 --> 00:07:46,400
Pretty much.

191
00:07:46,400 --> 00:07:51,500
It's all about creating a deeper connection between the creators and the listeners

192
00:07:51,500 --> 00:07:54,300
and building a richer audio ecosystem.

193
00:07:54,300 --> 00:07:56,600
It sounds like a win-win for everybody involved.

194
00:07:56,600 --> 00:08:00,100
But, you know, let's be real for a second.

195
00:08:00,100 --> 00:08:02,900
Are there any potential downsides to all of this?

196
00:08:02,900 --> 00:08:05,200
Or any ethical concerns we should be thinking about?

197
00:08:05,200 --> 00:08:06,200
Oh, for sure.

198
00:08:06,200 --> 00:08:11,100
I think one of the biggest concerns is this idea of filter bubbles or echo chambers.

199
00:08:11,100 --> 00:08:12,000
Oh, right.

200
00:08:12,000 --> 00:08:16,700
If AI is only recommending content based on what we've already listened to,

201
00:08:16,700 --> 00:08:21,700
it might just reinforce our existing tastes and prevent us from discovering anything truly new.

202
00:08:21,700 --> 00:08:25,400
So instead of expanding our horizons, it could actually limit our exploration.

203
00:08:25,400 --> 00:08:26,300
Exactly.

204
00:08:26,300 --> 00:08:28,700
And that's definitely something to be aware of.

205
00:08:28,700 --> 00:08:33,500
I think the key is to find that balance between personalization and discovery,

206
00:08:33,500 --> 00:08:36,500
offering recommendations that both align with our tastes

207
00:08:36,500 --> 00:08:38,300
and challenge us to try something different.

208
00:08:38,300 --> 00:08:40,200
We need that element of surprise.

209
00:08:40,200 --> 00:08:43,800
That unexpected discovery is what makes music so exciting.

210
00:08:43,800 --> 00:08:44,400
Totally.

211
00:08:44,400 --> 00:08:46,400
Yeah, it really seems like a tough balance to strike, right?

212
00:08:46,400 --> 00:08:47,100
Yeah.

213
00:08:47,100 --> 00:08:48,800
Giving people what they want,

214
00:08:48,800 --> 00:08:51,400
but also pushing them a little bit outside their comfort zone.

215
00:08:51,400 --> 00:08:52,700
Yeah, totally.

216
00:08:52,700 --> 00:08:57,000
And that's something that Spotify and I think the whole AI community

217
00:08:57,000 --> 00:09:00,400
is going to have to really wrestle with as this technology gets better and better.

218
00:09:00,400 --> 00:09:01,600
It's not just a technical problem.

219
00:09:01,600 --> 00:09:03,700
It's kind of a philosophical one too, you know?

220
00:09:03,700 --> 00:09:04,800
It really is.

221
00:09:04,800 --> 00:09:08,100
So we've talked a lot about the potential benefits,

222
00:09:08,100 --> 00:09:11,000
but are there any risks we should be thinking about?

223
00:09:11,000 --> 00:09:13,000
Like anything that worries you about all of this?

224
00:09:13,000 --> 00:09:14,700
Yeah, definitely.

225
00:09:14,700 --> 00:09:16,600
I mean, besides the whole echo chamber thing,

226
00:09:16,600 --> 00:09:20,900
there's also this question of potential misuse, right?

227
00:09:20,900 --> 00:09:22,000
Yeah, pretty mean.

228
00:09:22,000 --> 00:09:28,300
Like imagine an AI that's been programmed to push certain artists or genres over others,

229
00:09:28,300 --> 00:09:31,700
maybe for commercial reasons or even political agendas.

230
00:09:31,700 --> 00:09:33,400
Okay, yeah, that's a little creepy.

231
00:09:33,400 --> 00:09:34,600
Yeah, it's kind of unsettling.

232
00:09:34,600 --> 00:09:37,500
It makes you think about who's really controlling what we listen to.

233
00:09:37,500 --> 00:09:37,900
Right.

234
00:09:37,900 --> 00:09:39,600
Like is it us or is it the algorithm?

235
00:09:39,600 --> 00:09:40,600
Exactly.

236
00:09:40,600 --> 00:09:40,800
Yeah.

237
00:09:40,800 --> 00:09:42,200
As AI gets more powerful,

238
00:09:42,200 --> 00:09:45,600
we need to be having these conversations about the ethics of it all.

239
00:09:45,600 --> 00:09:49,200
We need to make sure it's used to help people, not control them.

240
00:09:49,200 --> 00:09:52,300
It's a good reminder that technology is never really neutral.

241
00:09:52,300 --> 00:09:52,800
Right.

242
00:09:52,800 --> 00:09:56,300
It always reflects the values of whoever created it.

243
00:09:56,300 --> 00:09:57,200
That's a great point.

244
00:09:57,200 --> 00:09:57,600
Yeah.

245
00:09:57,600 --> 00:10:01,700
Wow, this whole deep dive into Spotify's AI experiments has been super interesting.

246
00:10:01,700 --> 00:10:02,600
Yeah, it has.

247
00:10:02,600 --> 00:10:06,200
It feels like we're just scratching the surface of what's possible with AI and music.

248
00:10:06,200 --> 00:10:06,800
It really does.

249
00:10:06,800 --> 00:10:12,600
It's such a fast moving field and I think Spotify's work with these LLMs is a perfect example of how

250
00:10:12,600 --> 00:10:15,600
AI is changing the way we experience music.

251
00:10:15,600 --> 00:10:17,300
Yeah, it's exciting.

252
00:10:17,300 --> 00:10:23,400
So to sum it all up, Spotify is experimenting with these AI powered narratives to make music

253
00:10:23,400 --> 00:10:25,600
and podcast discovery better.

254
00:10:25,600 --> 00:10:28,000
They're using this thing called the LLAMA model from Meta.

255
00:10:28,000 --> 00:10:28,400
Right.

256
00:10:28,400 --> 00:10:34,600
And focusing on giving us more context for recommendations and making that AI DJ feature even cooler.

257
00:10:34,600 --> 00:10:39,400
And it sounds like there's a lot of potential, but also some things to watch out for like the echo chambers

258
00:10:39,400 --> 00:10:41,400
and making sure AI is used responsibly.

259
00:10:41,400 --> 00:10:42,400
Absolutely.

260
00:10:42,400 --> 00:10:44,400
Well, this has been another great deep dive.

261
00:10:44,400 --> 00:10:47,000
As always, we want to leave you with something to think about.

262
00:10:47,000 --> 00:10:51,100
So Spotify wants to make these personalized narratives that really get us.

263
00:10:51,100 --> 00:10:55,900
But could AI actually help us break out of our musical echo chambers?

264
00:10:55,900 --> 00:10:59,500
Could it actually introduce us to music we'd never find on our own?

265
00:10:59,500 --> 00:11:00,400
That is the question.

266
00:11:00,400 --> 00:11:01,200
What do you think?

267
00:11:01,200 --> 00:11:02,200
Let us know.

268
00:11:02,200 --> 00:11:06,600
Head over to our social media and tell us what you think about the future of AI and music.

269
00:11:06,600 --> 00:11:07,400
We want to hear from you.

270
00:11:07,400 --> 00:11:17,400
This is the Deep Dive, signing off.