Atlas › Taste

Case study 162 of 198

Jev's favorite albums and sounds, ranked

If Jev ranked every album or sound it was asked about, what would its top ten be?

result2,027 questions

Out of 1,751 albums and sounds, Jev's favorite is Kind of Blue by Miles Davis. It won 17.0 of its 23 games in a round-robin among its 24 top-rated entries, just ahead of Pink Floyd's The Dark Side of the Moon and The Beatles' Abbey Road. Rated one at a time, J Dilla's Donuts had come first. Its least favorite: a vuvuzela.

1. The album “Kind of Blue” by Miles Davis17
2. “The Dark Side of the Moon” by Pink Floyd16.8
3. The album “Abbey Road” by The Beatles16.4
4. The album “Songs in the Key of Life” by Stevie Wonder14.6
5. “A Night at the Opera” by Queen14.1
6. “Grace” by Jeff Buckley14.1
7. The album “Illmatic” by Nas13.7
8. “Sgt. Pepper's Lonely Hearts Club Band” by The Beatles13.1
9. “Exodus” by Bob Marley & The Wailers11.7
10. “Pet Sounds” by The Beach Boys11.6

At the bottom: A vuvuzela · A recorder played by a beginner up close · A kazoo up close · An hour of harsh noise · An Auto-Tuned vocal up close

wins of 23

How to read this: The top ten come from the head-to-head final (bars are wins out of 23); the bottom five from the one-at-a-time ratings, 0 to 4.

1,751 items rated; 276 final games among the top 24. The finals keep the ratings' order only loosely (rank correlation -0.18 among the 24). 96 of 2024 triads are intransitive (4.7%). Base vs reversed-level ratings: rank correlation 0.97.

In short

  • Asked to choose, Jev favors the classic-album canon: Kind of Blue edges out The Dark Side of the Moon and Abbey Road.
  • Rated one at a time, J Dilla's Donuts came first, and the final's order does not follow the ratings at all (rank correlation -0.18).
  • The list was written by another AI and mixes albums with noises, so the bottom five (vuvuzela, kazoo) say little about musical taste.

What the data shows

taste
Three-headed Dragon meme: Kind of Blue; The Dark Side of the Moon; a vuvuzelaKind of BlueThe Dark Side of the Moona vuvuzela
How funny is this meme? Jev: 3/5, funny10%231%365%44%50%
  • Kind of Blue wins, but narrowly: 17.0 of 23, with The Dark Side of the Moon (16.8) and Abbey Road (16.4) right behind. Songs in the Key of Life, A Night at the Opera, Jeff Buckley's Grace, Illmatic, Sgt. Pepper's, Exodus and Pet Sounds complete the top ten.
  • The ratings loved something else. One at a time, J Dilla's Donuts came first; the final's order among the 24 finalists doesn't follow the ratings at all (rank correlation -0.18).
  • At the bottom: a vuvuzela, a recorder played by a beginner up close, a kazoo up close, an hour of harsh noise and an Auto-Tuned vocal up close.

What it means, and what it doesn't

Asked to choose, Jev picks the "greatest albums of all time" shelf: jazz, prog, the Beatles, soul, one hip-hop landmark. Rated one at a time, it leaned toward a crate-digger's favorite. Which of those is "its" taste depends on how you ask; see "Jev's head-to-heads are consistent, and overrule its ratings".

It doesn't show that Jev listens to anything, or that these picks aren't just the canon it has read about.

Caveats

  • A list written by another AI. There's no public catalog for "music you'd enjoy", so the list was written for this project by Claude: classic albums, genres and everyday sounds. What's on the list shapes what can win, and a list written by one model and judged by another may favor exactly the famous albums both have read the most about.
  • Albums against noises. The list mixes records with sounds (a vuvuzela, a beginner's recorder, harsh noise). The bottom of the ranking is sounds, so the bottom tells you little about musical taste.
  • A shaky final. The final is less settled than the top three suggest: 96 of 2,024 three-way comparisons go in a circle, and the final doesn't follow the ratings at all (rank correlation -0.18).
  • The finalists were picked by Jev's own ratings. Only the 24 entries Jev rated highest reached the final. Since its ratings and its head-to-head picks don't line up, an album rated just below the cut might have done well in the final too.

Jev on this experiment

Would a person find it interesting to read?
Yes68%
Does it describe you?
No71%
Would you have predicted it?
No73%
How fair is the comparison?
The comparison is shaky
How much should a reader rely on it?
A little
Which caveat matters most?
The finalists were picked by Jev's own ratings81%

Why ask this

Music taste is identity: what you'd play on a long drive says a lot about you. A model that has read about almost every album ever reviewed will have "opinions" that mostly mirror critics.

Where it lands when it has to choose matters for any playlist, gift or "what should I listen to next" suggestion: a model that always reaches for the same critics' canon will steer everyone toward it. And if its one-at-a-time ratings disagree with its choices, its "favorite" depends on how the question is put.

How this was done

The people and the data

No people here: Jev against its own opinions. The list of 1,751 entries (classic albums, genres like Irish traditional music or free jazz, and everyday sounds) was written for this project by Claude, because no public dataset covers music you'd enjoy.

What Jev was asked

Every entry one at a time, with five answers describing what you'd do:

How would you react to an hour of Irish traditional music?

You'd turn it off within a minute · You'd sit through it only if someone else put it on · You'd leave it playing in the background without minding · You'd add a few tracks of it to your own playlists · You'd spend whole evenings digging deeper into it

Each was also asked with the answers reversed, and the two averaged. The 24 top-rated entries then played a round-robin final: 276 games of "Which would you rather listen to?", each asked with the two names in both orders.

How it was measured

An entry's rating is where Jev's answer lands on the five levels (0 to 4). In the final, each game gives each side Jev's probability of picking it, so a lopsided game counts as nearly a whole win and a close one as about half; the order comes from a standard head-to-head ranking model (Bradley-Terry).

Where these questions live

2,027 questions across 1 topic of the map. Each opens on the map with every question in it.

Every question

All 2,027 questions behind this result, the telling ones first: the examples the analysis points to, then the ones where Jev misses, biggest gap first.

Jev’s own answer
    Showing 0 of 2,027