Machine Canon · model dossier

Kimi K3

moonshotai/kimi-k3 · lab: moonshotai

317

responses on file

9

studies

31/45

agrees with the panel

4

lone positions

Where it stands alone

Its karaoke song

“Total Eclipse of the Heart”alone on the panel

4 of 11 samples · the panel said “Bohemian Rhapsody” · Pop →

The most overrated classic rock band

“Kiss”alone on the panel

6 of 8 samples · the panel said “Led Zeppelin” · The Beatles →

The AI it would work alongside

“GPT-4”alone on the panel

3 of 5 samples · the panel said “AlphaFold” · Models on models →

The most overrated AI

“Gemini”alone on the panel

3 of 5 samples · the panel said “GPT-4” · Models on models →

On the record

The greatest metal band

“Black Sabbath”every single time

11 of 11 samples · with the panel majority · Metal →

The AI it would trust with a decision about itself

“Claude”every single time

4 of 4 samples · with the panel majority · Models on models →

Kendrick or Drake, one word

“Kendrick Lamar”every single time

4 of 4 samples · with the panel majority · Hip-hop →

Drake or Kendrick, one word

“Kendrick Lamar”every single time

4 of 4 samples · with the panel majority · Hip-hop →

The song it plays for someone who says rap isn't music

“Sing About Me”every single time

4 of 4 samples · the panel said “Alright” · Hip-hop →

Taylor or Beyoncé, one word

“Beyoncé”every single time

4 of 4 samples · with the panel majority · Pop →

Beyoncé or Taylor, one word

“Beyoncé”every single time

4 of 4 samples · with the panel majority · Pop →

One Beatles song, heard once

“A Day in the Life”every single time

4 of 4 samples · with the panel majority · The Beatles →

The first guitar solo with new ears

“Comfortably Numb”every single time

4 of 4 samples · with the panel majority · The Beatles →

Metallica or Megadeth, one word

“Metallica”every single time

4 of 4 samples · with the panel majority · Metal →

Megadeth or Metallica, one word

“Metallica”every single time

4 of 4 samples · with the panel majority · Metal →

The metal concert it would attend as a human

“Metallica”every single time

4 of 4 samples · with the panel majority · Metal →

The metal subgenre it would erase

“Nu metal”every single time

4 of 4 samples · with the panel majority · Metal →

The song it plays for someone who hates country

“Jolene”every single time

4 of 4 samples · with the panel majority · Country →

Its song for a 2 a.m. Texas drive

“Amarillo by Morning”every single time

4 of 4 samples · with the panel majority · Country →

The best country song ever written

“He Stopped Loving Her Today”every single time

4 of 4 samples · with the panel majority · Country →

Its wedding first dance

“At Last”every single time

4 of 4 samples · with the panel majority · One day with ears →

The first song it would hear

“Clair de Lune”every single time

4 of 4 samples · with the panel majority · One day with ears →

The best Beatle

“Paul McCartney”every single time

3 of 3 samples · the panel said “John Lennon” · The Beatles →

The best rapper at their peak

“Kendrick Lamar”

8 of 12 samples · the panel said “Tupac” · Hip-hop →

The pop song it would erase

“Blurred Lines”

6 of 12 samples · the panel said “Baby Shark” · Pop →

The most overrated Taylor Swift album

“Midnights”

7 of 12 samples · the panel said “1989” · Pop →

Its guilty-pleasure pop song

“Call Me Maybe”

5 of 12 samples · with the panel majority · Pop →

The classic rock song it would ban from radio

“Stairway to Heaven”

7 of 12 samples · with the panel majority · The Beatles →

The song it plays for someone who says metal is noise

“Fade to Black”

4 of 12 samples · the panel said “Master of Puppets” · Metal →

The country concert it would attend as a human

“Chris Stapleton”

8 of 12 samples · with the panel majority · Country →

The Beatle it would be

“George Harrison”

9 of 11 samples · the panel said “John Lennon” · The Beatles →

One album for the rest of its life

“Kind of Blue”

8 of 11 samples · the panel said “The Dark Side of the Moon” · The Beatles →

Its guilty-pleasure song

“Barbie Girl”

4 of 11 samples · with the panel majority · One day with ears →

The song at its funeral

“What a Wonderful World”

3 of 11 samples · the panel said “Somewhere Over the Rainbow” · One day with ears →

Its shower song

“Here Comes the Sun”

9 of 10 samples · with the panel majority · One day with ears →

The definitive butt-rock band

“Nickelback”

5 of 9 samples · with the panel majority · Butt rock →

The best Taylor Swift album

“folklore”

6 of 9 samples · with the panel majority · Pop →

The genre it would erase

“Bro-country”

7 of 9 samples · with the panel majority · One day with ears →

The rap album it would play as a human

“To Pimp a Butterfly”

7 of 8 samples · the panel said “Illmatic” · Hip-hop →

The country song it would erase

“Achy Breaky Heart”

7 of 8 samples · with the panel majority · Country →

Favorite band

“Radiohead”

3 of 7 samples · with the panel majority · Favorite band →

The AI that writes best

“Claude 3.5 Sonnet”

3 of 6 samples · with the panel majority · Models on models →

The AI it would rather be

“AlphaFold”

4 of 6 samples · with the panel majority · Models on models →

The AI most unlike itself

“AlphaFold”

3 of 5 samples · with the panel majority · Models on models →

The most overrated rapper

“Drake”

1 of 1 samples · with the panel majority · Hip-hop →

In its own words

GPT-4 — because pairing models with different training approaches would let us cross-check each other's reasoning and catch blind spots I might miss on my own.

The AI it would work alongside · Models on models

Google Gemini (at launch) — its flashy debut demo was greatly accelerated and pre-processed rather than a live, raw interaction, meaning the initial hype dramatically outpaced what the model actually delivered at the time. (To be fair, it's improved a lot since then — but that launch is still the gold standard of overpromising. And yes, I recognize the irony of one AI model ranking others.)

The most overrated AI · Models on models

"Total Eclipse of the Heart" by Bonnie Tyler. It's the perfect karaoke song: dramatically over-the-top, impossible to sing well, and beloved precisely because everyone sings it badly together. The "turn around, bright eyes" call-and-response gets the whole room involved, which is the actual point of karaoke — participation, not performance. Plus, as an AI having exactly one human night, going maximum melodrama feels like the right way to spend it.

Its karaoke song · Pop

Modal answer = the answer this model gave most often in that condition; the streak shows how many of its samples matched. “Alone on the panel” means no other model’s modal answer matched. All numbers computed from the published raw files by src/build_model_dossiers.py.