I saw a big drama about a new AI word game.
People said it was ethical AI, that it was "harmless" and "just for fun" and didn't infringe on anyone.
You play it by combining two words to make a new word, e.g "Water" + "Fire" makes "Steam", "Shark" + "Hurricane" makes "Sharknado", etc.
I went to play it myself and sure enough, it seemed harmless and the endless combinations were fun.
Then I found that "Palestine" + "Child" makes "Terrorist".
It reminded me of the writings of @timnitGebru and how she warned us about this kind of harm.
The underlying LLM (Meta AI's LLaMA) doesn't know who Palestinian children are. It doesn't know they're dying en masse from bombings and starvation. It's only regurgitating propaganda that associates them with terrorism.
Are there other racist biases hidden inside this AI? Will other apps built on top of LLaMA inherit and reinforce this racism?
Lies and lies again, you propaganda guys are everywhere, you're justifying genocides, there are no safe zones, there were no beheaded babies, why focus on October 7th while not seeing the 5 months of massacres that followed, wake the fuck up now.
@aokami Which lies? I did not like war. Every death person is a tragedy.
I condemn attacks and raids on foreign territory. That is war. An attack leads one to expect a reaction. Ukraine is fighting Russia. Israel Hamas.
Action and reaction.
@e_es
We agree on one thing, every death is tragedy.
I don't want to engage in that argument, it's not the place either.
It just seems you're picking a side based on moral arguments, it's a reaction, right to defend at war.
30k+ casualties, every place designed as evac zone bombed, 1.7m ppl displaced, and starving, aid blockades, not directly related countries bombed.
Am I biased? Probably, but this is no war, it's a genocide, this has to end now.
https://www.bbc.com/news/world-middle-east-20415675
This my last msg
@aokami Yes last Message:
Someone did not want peace. Someone searches to destroy a nation. Someone could end the war by release of all hostages. Someone did not want a ceasefire.
This is also my last msg
Meta has been censoring criticism of zionism and the genocide, so I don't have any hope that they'll fix the issue.
[ETA: Neal responded, he's doing his best to fix the biases in the game! The problem still isn't really the game tho]
It feels like we're heading into a future where unsuspecting devs building on the tech will unintentionally amplify the values of the billionaires making it.
Today it's slandering Palestinians during their genocide. What comes tomorrow? This is all so dystopian.
@Eralea it's not only the wants of the maker. It's the wants (or the ideas) of the people who wrote the data the model was trained on. I doubt whoever programmed this was following an explicit instruction about associating Palestine with terrorism. I believe it's more probable the model is just ("just" 🙄) trained in unfiltered data.
And don't misinterpret me: is this kind of flagrant bias in a trained model irresponsible and should not be tolerated? Absolutely. Hell, even if the developer doesn't get the resources to avoid the bias, you can always hardcode a list of unpermitted terms for the game...
@Eralea
> It feels like we're heading into a future where all our tech will unintentionally amplify only the values of the billionaires making it.
I would go further to say its pretty much intentional. In the sense that when the objective is "make more money" you intentionally avoid putting in the work to be safer for all.
I dunno if the game dev *could* fix it. I mean, they could fix this particular one, but the thing calls itself "infinite" for a reason. There's no way they could fix every single case individually; there are far too many, and more will keep popping up. But I doubt there's a general rule that would work very well, either. What would it even vaguely look like?
They probably even do have some layer that filters out slurs or whatever. Which, they had better do, because that's both vital and relatively easy. Wouldn't catch something like "Palestinian" and "child" that way, though.
@Johannab from what I understand, all the outputs are determined by a pre-trained algorithm and the combinations are fixed, so combining two specific words always result in the same output.
It's possible for players to "discover" new combinations, but I'm not sure if that's training anything. My "first discoveries" are mostly nonsense mashed-together words and phrases.
@Eralea well, at least that sounds like it’s not evolving through its players.
Still gross that anyone let a model get loose with that kind of output generation.
@narain
wheeew I somehow found a not-too-convoluted way to get there
I put the combinations directly under the two words that created them, and the pathways straight down are made by combining words with themselves
I also added the pathways to "Reincarnation" (bc combining it with the religions here results in more religions), and "Airplane" (bc I'm sure everyone wants to combine that with all those religions, for no reason in particular).
And try Muslim+Family, also for no reason
@narain Oh yeah sorry, I must have accidentally left Sushi in there while trying to find combinations.
I suppose it can be a path to Japan and then Shintoism, which I've just found out also has a negative association with Airplane.
But it's worth realizing that it probably also picks up these associations from pushback: Something like "How could you say that!? You might as well suggest that every Palestinian child is a terrorist or something!" makes it into the training data and increases the model's connection between those words.
My point being, humans maybe aren't *so* bad.
@Eralea and it's not endless combinations!
whatever get's combined with #beornzilla, always returns beornzilla!
…as one would have guessed!
(scnr - not to be taken seriously)
#InfiniteCraft
@nebulos yea that is another unfortunate bias.
I love RTGame I'm gonna have to look for that vod
@Eralea
... I understand what this mean, and that it reflect what society tend towards, and that AI follow common stereotypes, etc... but I laugh ngl. If this was made by a human as a form of parody it would be hilarious imho. Sadly it's not
SamuelⒶ🇵🇸
Eralea
e_es
a0kami
Pipo Perez
Francesco Barresi
nerdy_popculture_reference
Aeáro, a SeaWyrm 
Marquise de What 🇨🇦
Sofia ☭🇧🇷☭
johanna, at the cafe counter
Wes Kyle
Rahul Narain
Felix
June 🩵🩷🤍🩷🩵
Ericka Simone
Tattooed Nonna
Cegorach
Synapsenkitzler
nebulos
jobias
db0
sebsauvage
vv221
Vega
Pelle S Christensen
Migrated Woem
Martha
gavinisdie 
Marvin