1000093692

Aug 07, 2026 13:38 · 2:53 · English · Whisper Turbo · 3 Звучници

Овај транскрипт истекава 27 Дани. Надоградња за трајно складиштење →

Индекс лекције

0:00

S… Speaker 2 (1000093692)

Can you encode your personal taste into an AI model?

0:02

S… Speaker 2 (1000093692)

And this is the question that we've been discussing.

0:03

S… Speaker 1 (1000093692)

Quick recap,

0:04

S… Speaker 1 (1000093692)

in part one,

0:05

S… Speaker 2 (1000093692)

we saw that every reward function currently compresses all of human judgment into a single

0:09

S… Speaker 1 (1000093692)

number.

0:09

S… Speaker 2 (1000093692)

And that's why AI output feels slightly generic.

0:12

S… Speaker 1 (1000093692)

In part two,

0:13

S… Speaker 2 (1000093692)

we focused on the new kinds of reward models like Lore from Meta,

0:16

S… Speaker 2 (1000093692)

which showed that you can decompose personal taste into eight shared dimensions.

0:20

S… Speaker 1 (1000093692)

Now,

0:20

S… Speaker 2 (1000093692)

Lore solves what you prefer.

0:23

S… Speaker 2 (1000093692)

and tapo solves how it looks but neither solves why let's use music as

0:27

S… Speaker 2 (1000093692)

a lens here but this applies to any creative domain design writing fashion etc

0:31

S… Speaker 2 (1000093692)

taste has four dimensions current reward models and knowledge graphs capture

0:35

S… Speaker 2 (1000093692)

two of them dimension one is acoustic features which are the formal properties of the content

0:39

S… Speaker 2 (1000093692)

itself in music that's tempo timber energy standard reward models

0:43

S… Speaker 2 (1000093692)

currently are capturing this it's baked into the embeddings dimension two exposure

0:48

S… Speaker 2 (1000093692)

pathways which means that how the content has reached you friend sending you a song caddy

0:52

S… Speaker 1 (1000093692)

social

0:52

S… Speaker 2 (1000093692)

trust if you grew up with a particular song in your region it carries cultural inheritance

0:56

S… Speaker 2 (1000093692)

these are different pathways with different weights a process called collaborative filtering

1:00

S… Speaker 2 (1000093692)

which is like users you like also like this approach partially captures this spotify

1:04

S… Speaker 2 (1000093692)

uses it but it flattens the pathway into a single signal

1:08

S… Speaker 2 (1000093692)

Now let's talk about dimension 3,

1:09

S… Speaker 1 (1000093692)

which is Social Positioning.

1:10

S… Speaker 2 (1000093692)

This is Pierre Bourdieu's concept of Habitus from 1980s,

1:14

S… Speaker 2 (1000093692)

where he says that taste isn't just preference,

1:16

S… Speaker 2 (1000093692)

it's an identity signal.

1:17

S… Speaker 2 (1000093692)

It's cultural capital.

1:18

S… Speaker 2 (1000093692)

Liking obscure artists,

1:20

S… Speaker 2 (1000093692)

signal one kind of identity.

1:21

S… Speaker 2 (1000093692)

And no reward model captures this.

1:23

S… Speaker 2 (1000093692)

None of them have a representation for social signaling.

1:26

S… Speaker 2 (1000093692)

Dimension 4,

1:27

S… Speaker 2 (1000093692)

Temporal Context.

1:28

S… Speaker 2 (1000093692)

For example, why does this genre resonate at this moment?

1:31

S… Speaker 2 (1000093692)

The folk resurgence isn't because folk objectively sounds better in 2026.

1:35

S… Speaker 2 (1000093692)

It's actually because of the AI music background.

1:37

S… Speaker 2 (1000093692)

That's a cultural moment,

1:39

S… Speaker 2 (1000093692)

a temporal edge that didn't exist in 2023 and might not exist in 2028.

1:43

S… Speaker 2 (1000093692)

No reward model has a representation for time -dependent cultural context.

1:47

S… Speaker 2 (1000093692)

Current recommendation system treats preference as static.

1:50

S… Speaker 3 (1000093692)

It's not.

1:50

S… Speaker 2 (1000093692)

Current models capture acoustic features and exposure pathways.

1:54

S… Speaker 2 (1000093692)

It completely misses social positioning and temporal context and that's half the picture.

1:58

S… Speaker 2 (1000093692)

in a system that can capture all four dimensions a taste isn't a property of the user and

2:02

S… Speaker 2 (1000093692)

it's not a property of the song it's the pattern of connection across all four layers

2:06

S… Speaker 2 (1000093692)

for example two users might listen to the same folk artist same music but user a listens

2:11

S… Speaker 2 (1000093692)

to it because it signals authenticity in their community that's social positioning plus temporal

2:15

S… Speaker 2 (1000093692)

context user b listens to it because they grew up with it that's exposure pathway only

2:19

S… Speaker 2 (1000093692)

no cultural movement involved same behavior completely different taste a system

2:23

S… Speaker 2 (1000093692)

that maps all four dimensions can tell them apart but a flat preference data

2:27

S… Speaker 2 (1000093692)

can never hulu is a company that's been building for over a decade and they have what they call

2:31

S… Speaker 2 (1000093692)

cultural ai and they have something called a taste api but here's the thing they're

2:36

S… Speaker 2 (1000093692)

all solving different slices independently i think nobody has connected the three layers into one

2:40

S… Speaker 2 (1000093692)

stack integration can become a new product a taste api that doesn't just know your preferences

2:44

S… Speaker 2 (1000093692)

it knows why you have them the interesting thing is that the research is published the individual components

2:48

S… Speaker 2 (1000093692)

are proven whoever integrates them might end up building the infrastructure for computational taste

Овај транскрипт је створио АИ (автоматско препознавање говора). Може садржати грешке — проверити оригинални аудио за критичну употребу. Политика ВИ

❤️ Љубав STT.ai?

сажетак

The transcript discusses the limitations of current AI models in capturing personal taste, which is decomposed into four dimensions: acoustic features, exposure pathways, social positioning, and temporal context. Current models focus on acoustic features and exposure pathways, missing social positioning and temporal context. A system that can capture all four dimensions can differentiate between tastes based on the pattern of connection across these layers.

Питај ВИ о овом транкрипту

Питајте било шта о овом транскрипту - АИ ће наћи релевантне секције и одговор.

Тренкриптуј још један