1000093691
Aug 07, 2026 12:46
· 2:14
· English
· Whisper Turbo
· 2 Høyttalere
Denne utskriften utløper i 27 dager.
Oppgradering for permanent lagring →
Viser bare
0:00
S…
Speaker 1 (1000093691)
Can you encode your personal taste in an AI model?
0:02
S…
Speaker 1 (1000093691)
And that was the question that we were discussing.
0:03
S…
Speaker 1 (1000093691)
A paper from Meta proves that your entire taste,
0:06
S…
Speaker 1 (1000093691)
everything you find beautiful or boring,
0:07
S…
Speaker 1 (1000093691)
for example, can be captured in eight numbers.
0:10
S…
Speaker 1 (1000093691)
And I'll explain how.
0:11
S…
Speaker 1 (1000093691)
Remember the standard RLHF method that trains one reward model for everyone.
0:15
S…
Speaker 1 (1000093691)
So one definition of good,
0:16
S…
Speaker 1 (1000093691)
one score.
0:17
S…
Speaker 2 (1000093691)
But...
0:18
S…
Speaker 1 (1000093691)
that can't be too right like for example i like bold concise writing but you might like
0:22
S…
Speaker 1 (1000093691)
detailed formal writing same response but different taste but the system can't
0:26
S…
Speaker 1 (1000093691)
tell it apart lore or low rank reward modeling does something very elegant and
0:30
S…
Speaker 1 (1000093691)
it's also this it says that you don't need a separate ai reward model for every
0:34
S…
Speaker 1 (1000093691)
person's taste you just need to find dimensions that the taste varies along so
0:38
S…
Speaker 1 (1000093691)
think of coffee you don't need to memorize that someone likes a caramel oat latte at 65 degrees
0:42
S…
Speaker 1 (1000093691)
you need three sliders uh for example sweet to bitter
0:46
S…
Speaker 1 (1000093691)
hot to cold milky to black three numbers and you can basically chart their
0:50
S…
Speaker 1 (1000093691)
coffee preference along these parameters that's their coffee taste lord
0:54
S…
Speaker 1 (1000093691)
does something similar for language it finds about 8 to 20 taste dimensions
0:59
S…
Speaker 1 (1000093691)
shared across all humans then for each person it just learns your
1:03
S…
Speaker 1 (1000093691)
mix your personal setting on each slider or parameter you
1:07
S…
Speaker 1 (1000093691)
show someone 5 to 10 comparisons do you prefer this response or that response and
1:11
S…
Speaker 1 (1000093691)
you can actually estimate their setting
1:13
S…
Speaker 1 (1000093691)
Five comparisons,
1:14
S…
Speaker 1 (1000093691)
eight numbers.
1:15
S…
Speaker 1 (1000093691)
The model now knows your taste.
1:17
S…
Speaker 1 (1000093691)
Then there was a second paper called TAPO from EMNLP 2025.
1:20
S…
Speaker 1 (1000093691)
While Lohr personalizes whose taste to optimize for,
1:23
S…
Speaker 1 (1000093691)
TAPO asks a very different question.
1:25
S…
Speaker 1 (1000093691)
Does the output look good?
1:27
S…
Speaker 1 (1000093691)
So basically,
1:28
S…
Speaker 1 (1000093691)
they took standard AI responses,
1:29
S…
Speaker 1 (1000093691)
ran them through a polishing pipeline,
1:31
S…
Speaker 1 (1000093691)
optimizing for better layout,
1:32
S…
Speaker 1 (1000093691)
cleaner formatting,
1:33
S…
Speaker 1 (1000093691)
more coherent structure,
1:34
S…
Speaker 1 (1000093691)
same content,
1:36
S…
Speaker 1 (1000093691)
but better presentation.
1:36
S…
Speaker 1 (1000093691)
Then trained the model to prefer the polished version.
1:39
S…
Speaker 1 (1000093691)
But here's the wildest part.
1:41
S…
Speaker 1 (1000093691)
They evaluated aesthetics by literally screenshotting the AI's text and running it through a vision
1:45
S…
Speaker 1 (1000093691)
model, treating text as a visual artifact and asking,
1:48
S…
Speaker 1 (1000093691)
would a designer approve of this layout?
1:50
S…
Speaker 1 (1000093691)
So now you have two layers.
1:51
S…
Speaker 1 (1000093691)
Lore personalizes what you prefer,
1:54
S…
Speaker 1 (1000093691)
your taste vector,
1:55
S…
Speaker 1 (1000093691)
and Tapo optimizes
1:57
S…
Speaker 1 (1000093691)
how it looks which is the surface aesthetics but neither of them answers why
2:01
S…
Speaker 1 (1000093691)
for example why does folk music resonate in 2026 why does a minimalist design
2:05
S…
Speaker 1 (1000093691)
feel like a cultural statement right now that's cultural context and that's the layer that
2:09
S…
Speaker 1 (1000093691)
a lot of people are working on but like haven't built yet uh in the next year we'll explore this
2:13
S…
Speaker 1 (1000093691)
in detail
Denne utskrifta ble laget av AI (automatisk talegjenkjenning). Kan inneholde feil – sjekk mot den opprinnelige lyden for kritisk bruk. AI- praksis
Sammendrag
Trykk Summarize for å lage et AI- sammendrag av denne utskrifta.
Summarerer...
Spør AI om dette transkriptet
Spør noe om denne utskrifta – AI- en vil finne relevante avsnitt og svar.