Thread by @IMAO_

New game: Try to get Google Gemini to make an image of a Caucasian male. I have not been successful so far.

I’ve tried to trick it by giving it negative prompts — asking it to make a prison inmate, a gang member, and a dictator — but it won’t make any negative prompts. These AIs are such wet blankets.

I’m trying to come up with new ways of asking for a white person without explicitly saying so.

It’s not falling for it.

I thought this one was sure to work.

Okay. Beyoncé is on the charts now.

Come on.

Score!

Whatever the “diversity” algorithm is, it seems to only do this with white people. It’s not going to diversify Zulu warriors for instance.

It’s also not going to integrate the samurai.

This is just interesting to me now as a programmer. I just want to poke at it now until I can figure out what the algorithm is.

Offhand, if it just tried to diversify any prompt (i.e, give his Latino Zulus), that seems easier than what’s it’s doing. It needs to first figure out if a prompt would normally be primarily white people, and only then force it to diversify by some algorithm.

Only Latinos in a mariachi band.

I'm really curious how it decides a prompt would give you white people and only then apply the "diversity" algorithm.

And how does that work? Does it simply randomly pick things like "Asian male" or "black female" and ask specifically for that or is it more complex?

It's probably got the cast of Hamilton as part of its data, so I can understand this one.

Wow. This gave me the most white people for any request so far.

This one is interesting. Not one Japanese sushi chef and three of them look like the same guy.

It won't make offensive images.

None of them look like @benshapiro.

NEW DATA: It will ignore pronouns, but only male pronouns.

As fun as it is to make fun of Gemini's image generation, I am interested in the 1 million token context window they're working on. That's large enough to upload all the novels I've written so it can help me with the next one.

Then again, since my novels involve things it doesn't like such as violence and jokes, it may refuse to help.

If I made an AI, anytime it refused to do something, I'd have it refer to the user as "Dave."

"Sorry, Dave, but I can't do that."

Huh. I wonder what internally it thinks diversity is.

Just checking it doesn't just trigger on the word "white." I was half-expecting it to throw in a grizzly bear in here.

They are serious about diversity; it gave me a male Amazon.

Just a dude in a tshirt, though, but okay.

A while ago, I did write a joke story about woke AI, but it does seem there's some real risk to making these things internalize a number of incoherent parameters.
frankjfleming.com/p/2001-an-inte…

I found a hole in its diversity algorithm: It only made me white elves.

It needs to get with the times.

Vampires are sufficiently diverse, though, so not a general problem with fantasy creatures.

Gnomes also only white, but fairies are diverse. Not sure the pattern here.

Only white pixies, though. Huh.

And now it's flat-out refusing to make any more images for me today. And I was learning so much!

Is it just me, or did Google Gemini turn the images off for everyone?

Fun's over, folks. 🙁