|
eXXon posted:Breaking news: PlaygroundAI user “N-word Balls” makes a pretty convincing goatse. Tried some along those lines, still has to include the thread title and alas every iteration cut far too much of the context. There's a trick to squeezing a few more characters in that I can't remember off the top of my head and it's too late to go digging, plus I don't think it would be enough even then. I'll try tomorrow!
|
#
?
Mar 17, 2023 06:18
|
|
|
|
| # ? Jun 1, 2024 08:02 |
|
|
StratGoatCom posted:One of the few upshots here is that this may provoke reactions from the security establishment mind. I'm gonna go ahead and predict here: - Democrats or left leaning people seeing likely fakes of anyone will dismiss them. - Republicans will dismiss any that show their party in a bad light, and utterly believe the rest. Which is basically where we are now for the rest of it so I'm not sure what you're expecting is going to be the tipping point here.
|
|
#
?
Mar 17, 2023 07:30
|
|
|
Do we reach a point soon where just about any video or photographic evidence of an event or a person doing something is automatically questionable, refutable, disputable and essentially worthless as a means of proof?
|
|
#
?
Mar 17, 2023 12:13
|
|
|
BiggerBoat posted:Do we reach a point soon where just about any video or photographic evidence of an event or a person doing something is automatically questionable, refutable, disputable and essentially worthless as a means of proof? Always has been.
|
|
#
?
Mar 17, 2023 12:42
|
|
|
There were some rumblings about needing a common video watermarking standard all recording devices would use in 2018 (and earlier too, probably, the issue keeps popping up since at least Amiga Paint), but it turns out we don't really care.
|
|
#
?
Mar 17, 2023 13:15
|
|
|
yeah, obfuscation and watermarking technology generally seem to work pretty terribly. i don't even know what corporations use to keep stuff from leaking and getting uploaded some streamers will just get their own background music that i guess they hope can be used either for thwarting other people taking their stuff, or maybe to make it harder for automated copyright systems to detect something on their own streams with background music
|
|
#
?
Mar 17, 2023 13:22
|
|
|
I think Europe already has laws the advertising images featuring manipulated images of people need to explicitly say they've been altered. I think the goal there is fighting body image issues. Might be an effective solution to expand rules like that so images need to explicitly say "Image not altered" or "This image has been digitally altered", and just have it apply to anything publicly shared. If cameras and snapchat can add the appropriate marking, which is easy, it wouldn't even be a burden to the public. It won't happen, because federal government and half of America is dysfunctional, but I think it is the type of solution needed
|
|
#
?
Mar 17, 2023 15:39
|
|
|
StumblyWumbly posted:I think Europe already has laws the advertising images featuring manipulated images of people need to explicitly say they've been altered. I think the goal there is fighting body image issues. Yeah it will never happen but that sounds like the best case. Or to even make it "this image has been altered" by default and you have to click a basic thing where you [legal fine print that absolves the social media platform] say that the content is unaltered.
|
|
#
?
Mar 17, 2023 15:40
|
|
|
How is it determined whether or not a photo has been digitally altered? Because literally all digital photography has been digitally altered in some way. Be it something simple like color adjustment, to using computational photography as the default photo taking mode on all modern smartphones.
|
|
#
?
Mar 17, 2023 15:50
|
|
|
StumblyWumbly posted:I think Europe already has laws the advertising images featuring manipulated images of people need to explicitly say they've been altered. I think the goal there is fighting body image issues. It's a bit more involved than that. Note that this is 2021 publication. On this subject specifically, there's the code of practice on disinformation explicitly bans deep fakes. For big tech, this is tied through DSA, meaning fines of up to 6% global turnover. cinci zoo sniper fucked around with this message at 16:02 on Mar 17, 2023 |
|
#
?
Mar 17, 2023 15:51
|
|
|
Boris Galerkin posted:How is it determined whether or not a photo has been digitally altered? Because literally all digital photography has been digitally altered in some way. Be it something simple like color adjustment, to using computational photography as the default photo taking mode on all modern smartphones. Different options, starting from photo DRM to running ML models that are trained to detect forgery. On the latter, one specific example I know of some video footage analysis models that are trained on identifying changes in blood flow under the skin.
|
|
#
?
Mar 17, 2023 15:57
|
|
|
kliras posted:yeah, obfuscation and watermarking technology generally seem to work pretty terribly. i don't even know what corporations use to keep stuff from leaking and getting uploaded Movies use https://en.wikipedia.org/wiki/Coded_anti-piracy It doesn't prevent leaks, it just lets them trace who did it and sue them. Money uses https://en.wikipedia.org/wiki/EURion_constellation and Photoshop, Xerox, etc are encouraged to block anything they detect with it.
|
|
#
?
Mar 17, 2023 17:04
|
|
|
duz posted:Movies use https://en.wikipedia.org/wiki/Coded_anti-piracy
|
|
#
?
Mar 17, 2023 17:05
|
|
|
There Bias Two posted:Wait how the gently caress was this available back in 1962?! This has completely blown my mind. Watch this one too from 1968: https://www.youtube.com/watch?v=yJDv-zdhzMY
|
|
#
?
Mar 17, 2023 17:16
|
|
|
SaTaMaS posted:It seems like Facebook could use all their social media data to generate a conversational LLM. They do have a significant AI effort though it's hard to figure out what they've accomplished. Quoting from a while ago but lol, I’d like to see GPT-4 trained on an archive of all Facebook posts. You know; garbage in, garbage out…
|
|
#
?
Mar 17, 2023 17:27
|
|
|
Snackmar posted:Watch this one too from 1968: https://www.youtube.com/watch?v=yJDv-zdhzMY https://en.wikipedia.org/wiki/PARC_(company) https://en.wikipedia.org/wiki/The_Mother_of_All_Demos
|
|
#
?
Mar 17, 2023 17:28
|
|
|
F4rt5 posted:Quoting from a while ago but lol, I’d like to see GPT-4 trained on an archive of all Facebook posts. https://www.youtube.com/watch?v=xslW5sQOkC8 It seems to work better going the other way, at Stanford they took one of Facebook's open source LLaMa models and improved it by training it with output from GPT 3.5. Apparently they were able to create an impressive model for a fraction of the training cost.
|
|
#
?
Mar 17, 2023 17:34
|
|
|
cinci zoo sniper posted:Different options, starting from photo DRM to running ML models that are trained to detect forgery. …
|
|
#
?
Mar 17, 2023 17:41
|
|
|
Ruffian Price posted:There were some rumblings about needing a common video watermarking standard all recording devices would use in 2018 (and earlier too, probably, the issue keeps popping up since at least Amiga Paint), but it turns out we don't really care. What would prevent that watermark from being added to deepfakes and rendering the safeguard worthless if not a further detriment since it'd have a watermark showing it is "real" video/pictures?
|
|
#
?
Mar 17, 2023 18:09
|
|
|
Evil Fluffy posted:What would prevent that watermark from being added to deepfakes and rendering the safeguard worthless if not a further detriment since it'd have a watermark showing it is "real" video/pictures? You could achieve that by doing essentially the same thing as HDCP: Have "approved" devices contain an encryption key that is used to cryptographically sign the image upon capture. But, same as with HDCP, sooner or later the keys will leak, making it a temporary measure.
|
|
#
?
Mar 17, 2023 18:17
|
|
|
Yup, it would absolutely mean a court-mandated monopoly. Want to have your recordings admissible as evidence, gotta buy this approved brand of portable recorder.
|
|
#
?
Mar 17, 2023 19:30
|
|
|
StumblyWumbly posted:I think Europe already has laws the advertising images featuring manipulated images of people need to explicitly say they've been altered. I think the goal there is fighting body image issues. Except any disclaimer like that will be a paragraph in 4 point type that's on the screen for like 3 seconds like any credit card or card ad is now. More than likely, what's going to happen is any politician will just claim "deepfake" and anyone who can afford a decent lawyer in court (rich people) will be able to cast doubt about the authenticity of any recording that shows them clearly committing the crime they're charged with.
|
|
#
?
Mar 17, 2023 23:53
|
|
|
Ruffian Price posted:Yup, it would absolutely mean a court-mandated monopoly. Want to have your recordings admissible as evidence, gotta buy this approved brand of portable recorder. Whoopsie doodle, someone stole the signing keys used to "identify" a device! Well shucks! This ignores things like de-lidding the DRM chip, the person at the fab making the thing quietly making duplicates, and so on. At best, you can say that a specific ID was available to at least that specific person, rather than saying that they themselves made it.
|
|
#
?
Mar 18, 2023 03:48
|
|
|
Volmarias posted:Whoopsie doodle, someone stole the signing keys used to "identify" a device! Well shucks! I still use an RSA SecurID as a keychain for the pure memory of horror when I accidentally dropped something heavy on it and broke the LCD...that was a FUN time to get access back into that account. Also instilled a need to have multiple different 2FA devices just incase I go swimming with my phone.
|
|
#
?
Mar 18, 2023 13:17
|
|
|
SaTaMaS posted:https://www.youtube.com/watch?v=xslW5sQOkC8 This is kind of nuts. Yes it means you need a 'better' model to train it against. But for OpenAi this is a disaster, spend months of work and hundreds of millions training a model to only have it essentially copied within days of release. We might get into an almost mexican standoff between the major players all not wanting to release their model first.
|
|
#
?
Mar 21, 2023 10:51
|
|
|
So will training models using older models be the hot new stuff now? Seems clever, like a bootstrapping situation.
|
|
#
?
Mar 21, 2023 12:37
|
|
|
His Divine Shadow posted:So will training models using older models be the hot new stuff now? Seems clever, like a bootstrapping situation. They actually tried that but it was too much garbage in garbage out. And it was inadvertent because people kept posting the awful responses from the first models online.
|
|
#
?
Mar 21, 2023 12:59
|
|
|
HootTheOwl posted:They actually tried that but it was too much garbage in garbage out. What will they train on when the internet turns into an ouroboros of AI poo poo?
|
|
#
?
Mar 21, 2023 13:05
|
|
|
This completely screws the investment incentive. Why invest in OpenAi when any rando company can just come along with a cheap model and train it using yours for a fraction of the cost. I'm sure open ai will scream this breaks their tos (which it definitely does) but they broke an absolutely ton of TOS' when they originally trained their models, claiming fair use. I find it hard to believe it would be classed as fair use for scraping people's stuff without permission to train their model, but not fair use to effectively scrape their model to train some other companies model. Especially when it currently stands that the text and images that AIs output cannot be copyrighted. Mega Comrade fucked around with this message at 13:35 on Mar 21, 2023 |
|
#
?
Mar 21, 2023 13:27
|
|
|
live by the scrape, die by the scrape
|
|
#
?
Mar 21, 2023 13:31
|
|
|
Someone I work with said that he doesn't think any of these LLMs are going to be anything beyond a common commodity before too long. All the secret sauce is how you build on top of it.
|
|
#
?
Mar 21, 2023 13:46
|
|
|
sinky posted:What will they train on when the internet turns into an ouroboros of AI poo poo? They were already complaining that the golden age of big data learning this way is over because of it. I think the art bots especially were vulnrable. Their new tactic is taking the user uploaded models made from their successful arts. Which is going to lead to a whole lot of porn.
|
|
#
?
Mar 21, 2023 13:52
|
|
|
shoeberto posted:Someone I work with said that he doesn't think any of these LLMs are going to be anything beyond a common commodity before too long. All the secret sauce is how you build on top of it. Before long? You can, today, download the fully trained weights for a model surprisingly close to GPT3.5 and run it at a perfectly reasonable speed on a garden variety CPU. The only thing holding the floodgates closed is licensing liability concerns.
|
|
#
?
Mar 21, 2023 13:56
|
|
|
Aramis posted:Before long? You can, today, download the fully trained weights for a model surprisingly close to GPT3.5 and run it at a perfectly reasonable speed on a garden variety CPU. The only thing holding the floodgates closed is licensing liability concerns. Wait which one is this?
|
|
#
?
Mar 21, 2023 14:31
|
|
|
Aramis posted:Before long? You can, today, download the fully trained weights for a model surprisingly close to GPT3.5 and run it at a perfectly reasonable speed on a garden variety CPU. The only thing holding the floodgates closed is licensing liability concerns. Liability concerns, and also just ignorance. ChatGPT is "AI" for an enormous amount of people who think it's cool but don't actually understand the tech. As the hype cycle cools they're going to lose that monopoly on being The AI for the public.
|
|
#
?
Mar 21, 2023 14:31
|
|
|
tractor fanatic posted:Wait which one is this?
|
|
#
?
Mar 21, 2023 14:43
|
|
|
shoeberto posted:Someone I work with said that he doesn't think any of these LLMs are going to be anything beyond a common commodity before too long. All the secret sauce is how you build on top of it. OpenAI's secret sauce is their approach to the alignment problem, apparently it's easy to train a LLM on a bunch of text, but getting it to know how to avoid being goaded into sounding like a nazi or giving bomb making instructions is much more difficult.
|
|
#
?
Mar 21, 2023 15:05
|
|
|
SaTaMaS posted:OpenAI's secret sauce is their approach to the alignment problem, apparently it's easy to train a LLM on a bunch of text, but getting it to know how to avoid being goaded into sounding like a nazi or giving bomb making instructions is much more difficult. Well it was. Now you just need one company to pay for that and everyone else can just point their own models at it to gain most of that custom tweaking for close to nothing.
|
|
#
?
Mar 21, 2023 15:10
|
|
|
SaTaMaS posted:OpenAI's secret sauce is their approach to the alignment problem, apparently it's easy to train a LLM on a bunch of text, but getting it to know how to avoid being goaded into sounding like a nazi or giving bomb making instructions is much more difficult. That, and also that it's still spectacularly expensive to train a LLM from scratch. However, what's happening right now is that it's becoming increasingly obvious that this is a price that only has to be paid once. It's going to be interesting when the next "generation" of architecture comes around. Whoever pays to train it first will be 100% guaranteed to be a sucker who's expenditure will benefit everyone else with little to no reward.
|
|
#
?
Mar 21, 2023 15:15
|
|
|
|
| # ? Jun 1, 2024 08:02 |
|
|
Okay maybe the funniest possible outcome is specifically because of treating image and text scrapings as bypassing copyright, it makes it effortless to train your own AI on scraping another AI's output and then sell yours at a fraction of the price, but then your competitor pays that once and trains their AI on it...
|
|
#
?
Mar 21, 2023 15:50
|
|
