We 'messed up' with black Nazi blunder, Google co-founder admits | Hacker News Reader

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

telegraph.co.uk

8 pointspaulcjh9 comments

> Mr Brin said the system was not thoroughly tested before being released to the public.

Is thorough test of "AI" even a thing?

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

The mistake is they got caught. It will be more subtle next time.

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

Are we intending to go (with only a brief detour) from a demon-haunted world almost directly into an "AI"-haunted one?

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.

> 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.

If your proposition was true, Google would not have tried to modify the melanin phenotypes in the images that were created.

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?)

https://archive.ph/UBDSb

Reply on news.ycombinator.com