[flagged]
Would you like them all banned? All of their comments deleted?
Why is your response to disagreement to question why the other side is allowed to speak at all?
No, it isn't a perfect AI image gen algo. Much like spam blocking making me click a box or select the motorcycle isn't an ideal anti-spam measure.
Sometimes you have to implement something non-ideal to prevent abuse. It's as simple as that.
As far as I'm concerned, it makes sense that an LLM would have absolutely no clue about nearly-obsolete human melanin preferences, and it's surprising that as soon as machines (cue McBean and his Star-On machine) arrive that show us how silly we've been being, it'd be considered "unacceptable".
When I was a child, blue or green hair would've been considered unacceptable, and now it is (although largely limited to younger and older women); maybe I'll live long enough to see a world where the maître d' has blue skin and my banker has green and nobody except some old farts who are about to kick the bucket really cares?
MYOB. F.—I.W.
Is there some suggestion that these obviously wrong examples are intentionally ahistorical or intended to erase history? Is this a plot of some kind? How common are these issues relative to all the other images being produced? Apparently the tool seems to be reluctant to produce images of white people at all. What should the appropriate frequency of white people be? That last one is a serious question that should be answered with an actual number, because that is the central concern here (unless I'm mistaken).
What seems very likely to me, is that in an attempt to not make the tool embarrassingly only produce images of white people they overshot. Bear in mind for a second, that for anybody who isn't white and is trying to generate an image of a person doing a thing (with no historical context to infer a race for that person) it would possibly be deeply frustrating to only get images of white people.
It's not that. It would have been one thing that when you said "I want a picture of 4 doctors" for the tool not to generate any white doctors at all.
The issue here is that I've read numerous reports of people saying that when they _explicitly_ specified something to the effect of "I want a picture of 4 _white_ doctors" it would straight up refuse them on the spot.
The latter is _not_ a training weights issue.
This is the broader, deeper problem.
Hacker News is more of a library where people come to read stuff, and isn’t really a place for political battles like the one happening here. Usually a moderator will manually disable the flame war detector for a thread if there is some hope for it. But it is doing its job as far as I’m concerned, because this thread isn’t really the kind that motivates intellectual curiosity.
Are there issues ? Yes. Do they have soul searching to do? Yes Will they release a complely uncensored model ? No Will the exact prompts that showed problems be fixed ? Yes Will their be other problems ? Yes. Is this the end of Google? No