|
ymgve posted:how do you quick fix bias in your model? just add words like "black" and "woman" at the end of users' prompts at random! "We accidentally did a racism with our AI, but don't worry, we 'fixed' it"
|
# ? Jul 21, 2022 18:14 |
|
|
# ? May 29, 2024 14:31 |
|
It's weird that would need teased out, you'd think part of the academic or research team documents would explain how they intended to handle that kind of bias in the sample set.
|
# ? Jul 21, 2022 18:22 |
|
A Strange Aeon posted:It's weird that would need teased out, you'd think part of the academic or research team documents would explain how they intended to handle that kind of bias in the sample set. It's a quick "fix" that was introduced during the last few days, in the research papers it's just "we're aware of the problem and are thinking hard about solutions"
|
# ? Jul 21, 2022 18:33 |
|
Someone was just tired of getting emails while the people in charge of datasets were trying to sort it out. Training takes a long time it's not something you can quickly change, at least not from my understanding. It works, but should have been implemented as an option you can turn off since it's changing your prompt and been up front about how it's working.
|
# ? Jul 21, 2022 18:46 |
|
pixaal posted:Someone was just tired of getting emails while the people in charge of datasets were trying to sort it out. Training takes a long time it's not something you can quickly change, at least not from my understanding. It wouldn't have been that much work to filter for prompts that include words that represent people and then check to see if there's any racial keywords before inserting one. I mean, it would have been more work than just randomly inserting words into prompts, but I could build a fairly intelligent system to do this in a couple days. The way they seem to be doing it is hilariously incompetent.
|
# ? Jul 21, 2022 18:58 |
|
|
# ? Jul 21, 2022 19:09 |
|
A Strange Aeon posted:It's weird that would need teased out, you'd think part of the academic or research team documents would explain how they intended to handle that kind of bias in the sample set. https://openai.com/blog/dall-e-2-pre-training-mitigations/ They did try that before training the model, but its a difficult and ill-defined problem.
|
# ? Jul 21, 2022 19:13 |
|
TIP posted:It wouldn't have been that much work to filter for prompts that include words that represent people and then check to see if there's any racial keywords before inserting one. I didn't even think about making the work around better, your right that could be vastly improved and probably should be communicated properly. Fixing it at the training level is probably a pipe dream anyway. History has been white washed already and without a time machine that's hard to restore.
|
# ? Jul 21, 2022 19:13 |
|
AI be morbin'
|
# ? Jul 21, 2022 19:19 |
|
https://twitter.com/JanelleCShane/status/1550187189610455040
|
# ? Jul 21, 2022 19:34 |
|
Guess who just got a DALL-E invite (and also learned that "fleshlight" breaks their ToS...)
|
# ? Jul 21, 2022 20:41 |
|
A Strange Aeon posted:It's weird that would need teased out, you'd think part of the academic or research team documents would explain how they intended to handle that kind of bias in the sample set. I don't think they expected as much bias as was found. I think humanity is far more biased than they expected and since the training set was from the internet it was exponentially more biased then humanity.
|
# ? Jul 21, 2022 21:19 |
|
For some reason, this is the only one of these pictures that featured an actual giant and not just a big waterpark. Those are some really freaky rear end giants. Almost reminds me of the live action "Where The Wild Things Are".
|
# ? Jul 21, 2022 22:16 |
|
TIP fucked around with this message at 23:17 on Jul 21, 2022 |
# ? Jul 21, 2022 23:13 |
|
|
# ? Jul 21, 2022 23:23 |
|
Foreverially tied up and loving it
|
# ? Jul 21, 2022 23:36 |
|
|
# ? Jul 21, 2022 23:47 |
|
See the text that it makes? The AI's all seem to make this type of text across different implementations. It terrifies me. What does it say? I cannot get it to translate to english. I read some where that it's just random noise. Then why is it so similar? Same dataset?
|
# ? Jul 22, 2022 00:16 |
|
It will not give me a translation. Any ideas would be great!
|
# ? Jul 22, 2022 00:27 |
|
|
# ? Jul 22, 2022 00:35 |
|
pikachu as a prehistoric fertility idol, clay sculpt, 5000BC, British museum openAI Dall-E invite came! ohhh I'm down to 8 credits after goofing around with it all afternoon.
|
# ? Jul 22, 2022 07:14 |
nazca posted:
these seem similar but are actually really different. Bottom right is very obviously some kinda white blocktext font and imo is only superficially similar to say, top or middle left. Check out this set of glyphs, which probably will look like poo poo in night mode: clearly all of those represent the same thing, ofc. I don't see poo poo in the bottom left that even remotely seems like a repeating glyph of the same type. That one is i think supposed to be the same font but at best you have an extremely blocky rectangle I. The rest of the characters in that one are too low detail to be usable, but have some differences. I think it probably is just noise and it's a mix of our brains trying to extract info from something that has no info, and the fake brain trying to create info and coming up with something that seems good enough. either that or whatever info is actually in there is encoded in a way that doesn't make sense at all, like such and such region has an overall value of x, and the next one has something else, etc. Where the actual lines don't represent it. ProfessorMarvel posted:Guess who just got a DALL-E invite (and also learned that "fleshlight" breaks their ToS...) like the text here, a human would see those two Gs as the same thing, there's no indication the network assigns anything to them, or even considers them as similar units. I think you'd have to figure out a way to quantify similarities or something that is without bias, or like... analyze information in a way that's hard for me to articulate. Because none of these are repeating glyphs, with maybe the bullet points being an exception, you'd have to imagine some situation like we got a 2d bitmap image from aliens and have to figure out the writing depicted in it. There's probably a way this could be done but idk. Cause you can't assume the neural net is using discrete glyphs as units of info. I think this is just a rehash of the yard sale / fake room conversation that came up earlier: looks like a yard sale would look at that resolution, but there's nothing actually in there, there's no identifiable objects or anything. It just looks similar to how words would look if you took a lovely pic. We know real words would have clear glyphs, how's the network supposed to get it if it's just pattern matching text inputs to image training sets to image outputs?
|
|
# ? Jul 22, 2022 07:42 |
|
I received my Dall-E invitation. Comin' In Hot! Back to Art School!
|
# ? Jul 22, 2022 08:17 |
|
3 Toed Sloth.
|
# ? Jul 22, 2022 09:40 |
|
I swear Dall-E sometimes throws in a completely unrelated pic.
|
# ? Jul 22, 2022 14:40 |
|
Click the image and use the report function in the lower left - the report function is also for nonsense generations, not just offensive stuff Those of you that got in after the payment switch got flipped, when did you apply for the beta? Wondering if they still do filtering or if they just take people in the order people applied
|
# ? Jul 22, 2022 17:37 |
|
|
# ? Jul 22, 2022 21:04 |
|
|
# ? Jul 23, 2022 00:24 |
|
|
# ? Jul 23, 2022 00:48 |
|
|
# ? Jul 23, 2022 00:59 |
|
|
# ? Jul 23, 2022 01:01 |
|
zalgo comes!
|
# ? Jul 23, 2022 01:02 |
|
|
# ? Jul 23, 2022 01:09 |
|
gandalf and saruman embarrassed at meeting each other while dancing at a gay ent club webcomic style
|
# ? Jul 23, 2022 01:11 |
|
doctorfrog posted:gandalf and saruman embarrassed at meeting each other while dancing at a gay ent club webcomic style
|
# ? Jul 23, 2022 01:17 |
|
Apparently Dalle LOVES William-Adolphe Bouguereau, and using his name makes anything betterdalle discord posted:other example where you can very clearly see how adding the painter changes the result: dalle discord posted:Anime Key visual of a young witch holding two beakers full of liquid, mad scientist's lab background
|
# ? Jul 23, 2022 01:20 |
|
Tunicate posted:Apparently Dalle LOVES William-Adolphe Bouguereau, and using his name makes anything better "you can very clearly see how adding the painter changes the result" isn't that expected, and true of many artists?
|
# ? Jul 23, 2022 01:25 |
|
Yeah but he's way more stable than a lot of others, for some reason
|
# ? Jul 23, 2022 01:27 |
|
|
# ? Jul 23, 2022 02:59 |
|
|
# ? May 29, 2024 14:31 |
|
Requested a rat innkeeper holding up a key Got a rat innkeeper with a messed up enormous right hand. Erased the right hand and asked for the same. Fixed. Erased the key and asked for a candle instead. Numerous variations later I ended up with a goat holding a candle.
|
# ? Jul 23, 2022 03:27 |