4 ms·

We 'messed up' with black Nazi blunder, Google co-founder admits

8 points by paulcjh 3y ago

raxxorraxor 3y ago
So is the mistake in this particular example or that the AI forgoes reality in favor of a more socially engineered answers?
- jevoten 3y ago
  The mistake is they got caught. It will be more subtle next time.
082349872349872 3y ago
Could someone please ELI5 how an AI not being "aware" how salient high- vs low-melanin phenotypes were for particular humans is supposed to be a blunder? https://news.ycombinator.com/item?id=39535423 https://news.ycombinator.com/item?id=39535423 Are we intending to go (with only a brief detour) from a demon-haunted world almost directly into an "AI"-haunted one?
- rasmus-kirk 3y ago
  Well, because the AI already knows how the Germans looked in WW2, the "problem" is the censoring in place. If you removed the censoring from the algorithm, then it would generate the accurate images you asked for. Not advocating removing all censorship on these models, but if you actually need WW2 images for some project, say a Youtube video, getting a bunch of images of black people with Stahlhelm's isn't really ideal. Just imagine a serious-styled amateur Youtube WW2 documentary using AI-images here and there, but almost all of the German army consists of minorities, it really sets the wrong tone! There is a sweet spot for this sort of thing.
  - aebtebeten 3y ago
    > because the AI already knows how the Germans looked in WW2 Does it? I had thought it was supposed to be generating these, not retrieving them? (in which case I would find it unremarkable if it had encoded Picture(Nazi) = head(human) + hat(Stalhelm) instead of going all the way to Picture(Nazi) = head(human.melanin(low)) + hat(Stahlhelm); see cousin comment about 6-fingered hands vs. Hox conservation) But maybe I'm biased because I've mainly tried to get AI to do coding and math, and my usual results have been exactly along these lines: the general shape might be almost there, but the details never match up.
- raxxorraxor 3y ago
  If your proposition was true, Google would not have tried to modify the melanin phenotypes in the images that were created.
  - 082349872349872 3y ago
    Is it true that Google modified those images? I would like to know why we all seem to think a bundle of weights ought to possess highly contingent historical knowledge? (especially when said bundles of weights are well known to generate six-fingered hands, even though Hox genes have been conserved for the last 300 million years or so?)
chrisjj 3y ago
> Mr Brin said the system was not thoroughly tested before being released to the public. Is thorough test of "AI" even a thing?
chrisjj 3y ago
https://archive.ph/UBDSb https://archive.ph/UBDSb

8 points by paulcjh 3y ago

So is the mistake in this particular example or that the AI forgoes reality in favor of a more socially engineered answers?

jevoten 3y ago
The mistake is they got caught. It will be more subtle next time.

082349872349872 3y ago

Could someone please ELI5 how an AI not being "aware" how salient high- vs low-melanin phenotypes were for particular humans is supposed to be a blunder? https://news.ycombinator.com/item?id=39535423 https://news.ycombinator.com/item?id=39535423 Are we intending to go (with only a brief detour) from a demon-haunted world almost directly into an "AI"-haunted one?

rasmus-kirk 3y ago
Well, because the AI already knows how the Germans looked in WW2, the "problem" is the censoring in place. If you removed the censoring from the algorithm, then it would generate the accurate images you asked for. Not advocating removing all censorship on these models, but if you actually need WW2 images for some project, say a Youtube video, getting a bunch of images of black people with Stahlhelm's isn't really ideal. Just imagine a serious-styled amateur Youtube WW2 documentary using AI-images here and there, but almost all of the German army consists of minorities, it really sets the wrong tone! There is a sweet spot for this sort of thing.
- aebtebeten 3y ago
  > because the AI already knows how the Germans looked in WW2 Does it? I had thought it was supposed to be generating these, not retrieving them? (in which case I would find it unremarkable if it had encoded Picture(Nazi) = head(human) + hat(Stalhelm) instead of going all the way to Picture(Nazi) = head(human.melanin(low)) + hat(Stahlhelm); see cousin comment about 6-fingered hands vs. Hox conservation) But maybe I'm biased because I've mainly tried to get AI to do coding and math, and my usual results have been exactly along these lines: the general shape might be almost there, but the details never match up.
raxxorraxor 3y ago
If your proposition was true, Google would not have tried to modify the melanin phenotypes in the images that were created.
- 082349872349872 3y ago
  Is it true that Google modified those images? I would like to know why we all seem to think a bundle of weights ought to possess highly contingent historical knowledge? (especially when said bundles of weights are well known to generate six-fingered hands, even though Hox genes have been conserved for the last 300 million years or so?)

chrisjj 3y ago

> Mr Brin said the system was not thoroughly tested before being released to the public. Is thorough test of "AI" even a thing?

chrisjj 3y ago

https://archive.ph/UBDSb https://archive.ph/UBDSb