Google Gemini caused quite a stir this week due to the tool’s apparent hesitance when asked to generate images depicting White people. Of course for folks who are more familiar with Google’s history on certain issues, these events are less of a surprise.
>"Rather than punishing people who advocate for ideological diversity, it is my hope that Google will, in the future, embrace them."
Unlikely. Google has a long history of making terrible, tyrannical decisions and ignoring the ensuing public outrage. Just look at what they did with YouTube.
Try asking it what a boy is, the knots it twists itself into are hilarious and just get worse as you drill down on the obvious logical inconsistencies in its answers
I appreciate your input, assessment and perception on this situation. Amazing that at my age I’d be reading this from a rabbit and a few others on here and X . Legacy media is dead . The only in-depth intellectual honesty I can find is from people like you . Thanks Rabbit . ☮️
The vast majority of engineers at Google would agree that the behavior of Gemini here was wildly stupid. It just so happens that there's no incentive by anyone to push back on poorly-thought out diversity initiatives because it would be idiotic from a career standpoint.
I’ve been using GPT-2, 3, 3.5, 4, 4 turbo and my own homebrew for long -form writing (now full novels) mostly Midjourney for illustrations. GPT has simultaneous bias issues - “the Black man was driving his kids down the street” or gay, or Asian woman, or blond woman would get fascinating, repeatable corrosive responses in early versions. However, mention gay or trans and responses inevitably include flowery inclusiveness statements. You must be incredibly careful in what you ask for. When I generated gay porn, it took some time to evolve prompts (iterate GPT-4 conversing with itself (i only use API’s) to generate guardrail evasion) which didn’t constantly flip gay into queer or LGBTQ+ unreadable junk. Nobody goes into a hot back room at a LGBTQ+ leather bar in 1978.
Midjourney is much more complex a problem. It generally will never spontaneously generate images with noneuropean characteristics unless you use the keyword diverse, or a specific prompt including for instance “African American”. God forbid you say “African”. It is very difficult to engineer a prompt which results in multiple races or ages in an image for men.
The most successful way to get great images with different skin types is to choose an artist - Kehinde Wiley will get a radically different response than Horst P. Horst as target. I have A superb AI autobiography of Muhammad Ali illustrated by Kehinde Wiley, and a superb AI autobiography of Greta Garbo illustrated by Horst P. Horst. You work within the constraints.
All other image systems just aren’t anywhere near Midjourney, all other text systems aren’t anywhere near ChatGPT.
I have thousands of gay hardcore novels I’m waiting to train into Mistral locally (when I have an extra year), and hundreds of thousands of gay male “physique” images to train a local image system when they’re adequate, therefore probably not soon.
I’m sure anyone who really uses these develops test vectors to manage bias and evolves prompts which compensate.
Thanks for this much-needed reminder of the insanity of Google's management. Let us not forget that the very top Google execs have no qualms about doing whatever is necessary to flip the result of a presidential election: Google TGIF Following 2016 Election: https://www.breitbart.com/tech/2018/09/12/leaked-video-google-leaderships-dismayed-reaction-to-trump-election/
>"Rather than punishing people who advocate for ideological diversity, it is my hope that Google will, in the future, embrace them."
Unlikely. Google has a long history of making terrible, tyrannical decisions and ignoring the ensuing public outrage. Just look at what they did with YouTube.
Try asking it what a boy is, the knots it twists itself into are hilarious and just get worse as you drill down on the obvious logical inconsistencies in its answers
I appreciate your input, assessment and perception on this situation. Amazing that at my age I’d be reading this from a rabbit and a few others on here and X . Legacy media is dead . The only in-depth intellectual honesty I can find is from people like you . Thanks Rabbit . ☮️
The vast majority of engineers at Google would agree that the behavior of Gemini here was wildly stupid. It just so happens that there's no incentive by anyone to push back on poorly-thought out diversity initiatives because it would be idiotic from a career standpoint.
When I saw that even Jeff Dean had pronouns in his bio, I realized how deep is the problem
I think at this point we have to assume the Deep State owns Google. They control it, they tell it what to do.
I’ve been using GPT-2, 3, 3.5, 4, 4 turbo and my own homebrew for long -form writing (now full novels) mostly Midjourney for illustrations. GPT has simultaneous bias issues - “the Black man was driving his kids down the street” or gay, or Asian woman, or blond woman would get fascinating, repeatable corrosive responses in early versions. However, mention gay or trans and responses inevitably include flowery inclusiveness statements. You must be incredibly careful in what you ask for. When I generated gay porn, it took some time to evolve prompts (iterate GPT-4 conversing with itself (i only use API’s) to generate guardrail evasion) which didn’t constantly flip gay into queer or LGBTQ+ unreadable junk. Nobody goes into a hot back room at a LGBTQ+ leather bar in 1978.
Midjourney is much more complex a problem. It generally will never spontaneously generate images with noneuropean characteristics unless you use the keyword diverse, or a specific prompt including for instance “African American”. God forbid you say “African”. It is very difficult to engineer a prompt which results in multiple races or ages in an image for men.
The most successful way to get great images with different skin types is to choose an artist - Kehinde Wiley will get a radically different response than Horst P. Horst as target. I have A superb AI autobiography of Muhammad Ali illustrated by Kehinde Wiley, and a superb AI autobiography of Greta Garbo illustrated by Horst P. Horst. You work within the constraints.
All other image systems just aren’t anywhere near Midjourney, all other text systems aren’t anywhere near ChatGPT.
I have thousands of gay hardcore novels I’m waiting to train into Mistral locally (when I have an extra year), and hundreds of thousands of gay male “physique” images to train a local image system when they’re adequate, therefore probably not soon.
I’m sure anyone who really uses these develops test vectors to manage bias and evolves prompts which compensate.
For public consumption? Hilarious.
Isn’t it funny how big of a deal this was at the time, and how quaint it all seems now?