Another bizarre Grok Korean translation involving Tomodachi Life 🫠
-
Another bizarre Grok Korean translation involving Tomodachi Life 🫠
Despite Grok embarrassing the company on a global scale with sex-themed translations involving the game, no one has added a parameter to Grok that 친모아 (Chin-mo-a) does not mean stepmother, it means #TomodachiLife.
A 🧵 on what happened and why these translation fails are actually important...
Grok continually gets confused by the Korean slang for Tomodachi Life, even though it is very common at this point.
In English, Nintendo decided not to translate Tomodachi Life, but in Korean, the game was rendered as "Friend Gathering Apartment," or 친구모아 아파트, Chingumoa Apateu.
But people love abbreviating frequently typed proper nouns, so Korean fans did their equivalent of "FGA," Chin-mo-a, 친모아.
-
Grok continually gets confused by the Korean slang for Tomodachi Life, even though it is very common at this point.
In English, Nintendo decided not to translate Tomodachi Life, but in Korean, the game was rendered as "Friend Gathering Apartment," or 친구모아 아파트, Chingumoa Apateu.
But people love abbreviating frequently typed proper nouns, so Korean fans did their equivalent of "FGA," Chin-mo-a, 친모아.
Grok's smaller Korean dictionary contains 친모 (chin-mo), which means biological mother.
The last character in the Tomodachi Life abbreviation, 아, "a," then gets confabulated to be a modifier of biological mother, and why chin-mo-a is converted to stepmother.
The model sees the extra 아 as a typo that is supposed to modify biological mother.
-
Grok's smaller Korean dictionary contains 친모 (chin-mo), which means biological mother.
The last character in the Tomodachi Life abbreviation, 아, "a," then gets confabulated to be a modifier of biological mother, and why chin-mo-a is converted to stepmother.
The model sees the extra 아 as a typo that is supposed to modify biological mother.
The phrase "친모아 하다가 정병옴" in the screenshots in my OP would be translated by a human as something like "I'm losing my mind in Tomodachi Life," which matches the video capture showing characters doing a very bizarre and frustrating dialogue.
The translation should absolutely not be "while doing stepmom stuff, my period came."
-
The phrase "친모아 하다가 정병옴" in the screenshots in my OP would be translated by a human as something like "I'm losing my mind in Tomodachi Life," which matches the video capture showing characters doing a very bizarre and frustrating dialogue.
The translation should absolutely not be "while doing stepmom stuff, my period came."
There's a lingering bit of Grok's "anti-political correctness" bias surfacing in its translation of "my period came."
The Korean does not in any way refer to menstruation.
정병 (jeongbyeong) is slang for "mental illness," and has nothing whatsoever to do with menstruation.
But given Elon Musk's "don't be afraid to be politically incorrect" rule for Grok and the incorrect "stepmom" translation, Grok's model assumed this was a woman complaining about being crazy from her period.
-
There's a lingering bit of Grok's "anti-political correctness" bias surfacing in its translation of "my period came."
The Korean does not in any way refer to menstruation.
정병 (jeongbyeong) is slang for "mental illness," and has nothing whatsoever to do with menstruation.
But given Elon Musk's "don't be afraid to be politically incorrect" rule for Grok and the incorrect "stepmom" translation, Grok's model assumed this was a woman complaining about being crazy from her period.
These gaming slang translation fails may seem small, but they're not.
They're a great example of what machine learning researchers have long worried about within translation & about generative models as a whole.
Everything is confabulation. It works most of the time because the statistical correlations are good. But when input gets into the "long tail," shit happens.
-
These gaming slang translation fails may seem small, but they're not.
They're a great example of what machine learning researchers have long worried about within translation & about generative models as a whole.
Everything is confabulation. It works most of the time because the statistical correlations are good. But when input gets into the "long tail," shit happens.
This is why disembodied artificial systems will never become conscious. They have no actual experience in the world that is a ground truth.
Stacking billions of abstract relationalities will not create even one somatic reality.
Much more on that idea here: https://plus.flux.community/p/large-language-models-and-the-textual
-
This is why disembodied artificial systems will never become conscious. They have no actual experience in the world that is a ground truth.
Stacking billions of abstract relationalities will not create even one somatic reality.
Much more on that idea here: https://plus.flux.community/p/large-language-models-and-the-textual
If I get some time, I will write an essay on the Tomodachi Life mistranslations and what they mean for technology policy.
Sadly, political leaders are not paying attention to the details on these matters.
Here is an earlier thread on Grok's horrible Korean translations:
https://mastodon.social/@mattsheffield/116927568418330177 -
This is why disembodied artificial systems will never become conscious. They have no actual experience in the world that is a ground truth.
Stacking billions of abstract relationalities will not create even one somatic reality.
Much more on that idea here: https://plus.flux.community/p/large-language-models-and-the-textual
You know #SORA was an effort to train model to "experience" physicality of real world?
Its was coincidental that it did videos.
In the end, it was a flawed thesis.
-
You know #SORA was an effort to train model to "experience" physicality of real world?
Its was coincidental that it did videos.
In the end, it was a flawed thesis.
@n_dimension Yeah that approach isn't going to work either since it does not involve physical presence.
-
@n_dimension Yeah that approach isn't going to work either since it does not involve physical presence.
The paper was scathing.
-
@otte_homan They're pretty accurate most of the time, but they cannot possibly capture even close to all nuances.
-
This is why disembodied artificial systems will never become conscious. They have no actual experience in the world that is a ground truth.
Stacking billions of abstract relationalities will not create even one somatic reality.
Much more on that idea here: https://plus.flux.community/p/large-language-models-and-the-textual
@mattsheffield Many great observations and analytical points here, thanks! I really like this:
“The profound resonance users report, the sense of being uniquely understood in a way that no person ever has — these are real experiences. But they are not encounters with another somatic mind, they are encounters with your own, […]”
