I think this is conflating the two problems. (Note, I think they are the ones conflating it, not you.)
They were trying to combat bias from training that has hurt them in the past. The infamous, "all doctors are evidently white american men." To do that, they tried messing with the prompts that users gave, and this led to the craziness they are currently facing.
They would like to somehow get the best results of both situations. I don't think they have any method that can do that, though. You would have to know the intent of the user to know when you can go for data accurate to the training set versus data that is more accurate to possibilities. (For a fun example, Planet of the Apes have a Lincoln Memorial that I would imagine would be very hard to get from Gemini, as is.)
The should just add a DEI slider with values from 0 to 100 and initialized to 50 and adjust the system prompt based on what the user selects. If someone screams to high heaven that gemini is too woke then reply with their DEI slider setting and say "well you had it turned up to 100". If someone screams Gemini is biased and unfair then the reply would be "you turned the DEI slider to 0".
Make the user be the one responsible for the behavior.
Problem is, these are neural nets and whatever instructions you give them affect any output, so censorship makes getting information out of the text chatbots even more unreliable than it is by default.
I wonder how deep in the model the bias is. If it's not a matter of adjusting a system prompt or minor fine tuning it could be a while before there's another release. I know one technique of de-censoring a model is basically fine tuning out the censorship so maybe that's what they're doing to Gemini.
I find it comical that they pretend not to know how to fix it, but just in case I’ll make it explicit:
Go into the text that you silently append to people’s prompts, and remove the EDI material.
We all know you didn’t train the model on images of POC Nazis and Vikings, and if you’re that concerned about bias then improve the training data. If there are no images available to train on for what you consider to be a problematic bias then leave it alone. It’s probably pretty close to reality and nobody is going to care except people who will simply find something else to complain about anyway.
> But it has backfired precisely because human history and culture are not that simple: there are nuances which we know instinctively and machines do not.
The reason it's hard to fix is not because machines don't know things we know, it's that people don't know things either. I would bet close to 100% of people complaining about 'woke' AI would not be able to produce historically accurate pictures either and not because of drawing ability but because of lack of historical understanding (again, not because of lack of access to facts/google/etc).
Here's an example: If you asked an image model to show you a picture of WW2 solders firing artillery in drag, you'd mock it as being 'woke', absurd, pushing "gender ideology" yadda yadda. But that did actually happen: https://www.reddit.com/r/interestingasfuck/comments/12cmtam/...
Any AI that does not return pictures of soldiers in drag firing artillery is biased and not historically accurate. Will DALL-E, SD, MJ, etc do that today without specific prompting? How many people here would draw such a picture if asked for pictures of ww2 soldiers? Most complaints about 'woke' AI are people complaining that AI doesn't re-affirm their pre-existing ideas. I'm highly skeptical that the people complaining about black popes have the same complaints about white depictions of jesus.
My comment on the other thread pointed out that these people have specific, political views that they want to force on everyone in every system and institution. They treat as evil anyone who is white, male, supports biological genders, straight, and so on. They then “fix” the “problem” by systematically discriminating against those groups by removing or replacing them. For OpenAI and Gemini, they are using both fine-tuning and system prompts to do this with Google’s being in the prompt.
That leads to the next problem: many media outlets are on board with the same stuff and lie about it, too. This outlet didn’t report that Google had people trying to force everyone to think a certain way, changed an easily-modified prompt to do it, etc. Instead, the author puts time into accusing everyone in the world of bias like racism, pretends Google was merely trying to counter some of it, and talks like what they did was really hard to avoid. They’re lying to back those in their camp.
Compare it to this short video by a conservative reporter:
While he’s biased and the ads are annoying, he quickly calls out the purpose of what they’re doing (eg CRT/DEI-style views), cites prior statements by a person in the program showing their politics are polar opposite of most users, showing the arrogance of the people behind these AI’s, and cites another expert who explains technically how the flaw happened (and was easily avoidable).
So, colleges, tech companies, and liberal outlets keep supporting systematic discrimination built on Marxist-like theories of equity. They often lie about what they are doing. If they get caught, they lie to make each other look better. We should never trust such people or their policies with our future.
Whereas, Jesus said we should be righteous and loving and just. That the truth will set us free. We’re honest about what we do, where it came from, what good it’s done over time when people actually followed God’s Word (and design), and that we’re better off going back to what worked. Then, fixing problems we see in honest, godly ways in humility. Not the deception or hubris these program directors seem to have.
14 comments
[ 2.5 ms ] story [ 41.7 ms ] threadNo. The explanation for why this has happened is because the AI was specifically instructed to do so.
They were trying to combat bias from training that has hurt them in the past. The infamous, "all doctors are evidently white american men." To do that, they tried messing with the prompts that users gave, and this led to the craziness they are currently facing.
They would like to somehow get the best results of both situations. I don't think they have any method that can do that, though. You would have to know the intent of the user to know when you can go for data accurate to the training set versus data that is more accurate to possibilities. (For a fun example, Planet of the Apes have a Lincoln Memorial that I would imagine would be very hard to get from Gemini, as is.)
Make the user be the one responsible for the behavior.
Google probably has the same giant prompt but dialed to over 9000.
Problem is, these are neural nets and whatever instructions you give them affect any output, so censorship makes getting information out of the text chatbots even more unreliable than it is by default.
Go into the text that you silently append to people’s prompts, and remove the EDI material.
We all know you didn’t train the model on images of POC Nazis and Vikings, and if you’re that concerned about bias then improve the training data. If there are no images available to train on for what you consider to be a problematic bias then leave it alone. It’s probably pretty close to reality and nobody is going to care except people who will simply find something else to complain about anyway.
The reason it's hard to fix is not because machines don't know things we know, it's that people don't know things either. I would bet close to 100% of people complaining about 'woke' AI would not be able to produce historically accurate pictures either and not because of drawing ability but because of lack of historical understanding (again, not because of lack of access to facts/google/etc).
Here's an example: If you asked an image model to show you a picture of WW2 solders firing artillery in drag, you'd mock it as being 'woke', absurd, pushing "gender ideology" yadda yadda. But that did actually happen: https://www.reddit.com/r/interestingasfuck/comments/12cmtam/...
Any AI that does not return pictures of soldiers in drag firing artillery is biased and not historically accurate. Will DALL-E, SD, MJ, etc do that today without specific prompting? How many people here would draw such a picture if asked for pictures of ww2 soldiers? Most complaints about 'woke' AI are people complaining that AI doesn't re-affirm their pre-existing ideas. I'm highly skeptical that the people complaining about black popes have the same complaints about white depictions of jesus.
That leads to the next problem: many media outlets are on board with the same stuff and lie about it, too. This outlet didn’t report that Google had people trying to force everyone to think a certain way, changed an easily-modified prompt to do it, etc. Instead, the author puts time into accusing everyone in the world of bias like racism, pretends Google was merely trying to counter some of it, and talks like what they did was really hard to avoid. They’re lying to back those in their camp.
Compare it to this short video by a conservative reporter:
https://youtu.be/RGOtLLEVKEM?si=lPmY9t2VcG-CYDQJ
While he’s biased and the ads are annoying, he quickly calls out the purpose of what they’re doing (eg CRT/DEI-style views), cites prior statements by a person in the program showing their politics are polar opposite of most users, showing the arrogance of the people behind these AI’s, and cites another expert who explains technically how the flaw happened (and was easily avoidable).
So, colleges, tech companies, and liberal outlets keep supporting systematic discrimination built on Marxist-like theories of equity. They often lie about what they are doing. If they get caught, they lie to make each other look better. We should never trust such people or their policies with our future.
Whereas, Jesus said we should be righteous and loving and just. That the truth will set us free. We’re honest about what we do, where it came from, what good it’s done over time when people actually followed God’s Word (and design), and that we’re better off going back to what worked. Then, fixing problems we see in honest, godly ways in humility. Not the deception or hubris these program directors seem to have.