I'm most interested in how they handle it for the users. It says potential "nudes" are blurred and can be "reported" or viewed. Will it tell the sender they're sending something that's going to be blurred? (which on one hand would let them resolve a false positive but also maybe let them find an adversarial image).
Really? I would have guessed most people would want to know who the "toxic" people are in their life so that they can take appropriate action instead of finding out in a worse way later.
Bumble is an app for meeting new people, not general-purpose socialization. I’d assume that if it had an auto-block feature for lewd images, it would remove the conversation from the inbox and there would most likely be no further contact.
We have been secretly, yet proudly drawing our members on stone walls since the dawn of time.
In Centralia, PA there use to be something called Grafitti Highway. It's an old highway road (they created a new road around it) that people would go and graffiti this old, unused highway. It might as been called Penis Highway as innately that's what was mostly drawn there. Too bad they shut it down lol
I'd say 95% of people on this website haven't received one and even Bumble said it's around 0.1% of users which is not comparable to spam levels at all
My recollection of Bumble is that all images were blurred (at least until you already sent pictures to each other?)
While I imagine having a good filter is important given the scale of online dating I would think all other filters would eliminate most of the problem (auto blurring images if people have only sent a few messages to each other, blur the first image sent, blur images by anyone who has been reported before, etc.)
Yes there is a length of conversation factor in visibility for normal photos too (I think).
But for "adult" photos they just blur.
I suspect people like this feature because they want to make sure they don't have people around them when they open that message (assuming it is a consensualy exchanged pic)
I'm almost certain your comment is tongue-in-cheek, but it did pique my curiosity - what do you think the actual ratio of unwanted explicit photos from each sex is? I have a hard time imagining it being less than 99:1.
There is a lot of DM spam on Twitter, TikTok, and Facebook/Instagram/Whatsapp involving explicit photos of female genitalia that would meet the criteria of obscene material under California's law[1] around sending unsolicited images by electronic means (SB 53).
Specially annoying when they use trending topics to tag their unrelated content. It never ocurred to me that I've seen more unsolicited female nudes in Twitter than everywhere else combined, including Reddit and its unsolicited chats.
I have never once received an unwanted nude from a woman but as spam whereas with men it has happened enough times for me to believe without statistical justification that unwanted images of male genitalia would be more common.
It does not matter who owns the body parts in the images. Unwanted genitalia is unwanted genitalia, no matter the sex or the purpose.
I think a lot of unsolicited female genitalia is sent for spam/scam proposes. A particularly pernicious sort is in the form of sextortion [1]. Images of female genitalia may be sent by a criminal to entice a (typically young or teenage) boy into believing he’s engaging in an online romantic relationship with a real girl. When he reciprocates with images of his own body (as requested to do so), the criminal responds by using those images to blackmail him into paying money with the threat of exposing him to friends and family. This sort of attack has also been used to target girls, though recently the targeting of boys has grown much faster. If I had to speculate, criminals likely see young boys as easier targets who can be enticed with images of the female body.
I suspect it's not just the occurrence rate which influences this feature being male-focussed. It's probably also the expected reception, which is itself a function of the occurrence rate.
If I opened an app one day and received an unsolicited picture of female genitalia, that would be strange. If it happened every day, then it's spam, even if I don't have a strong aversion to it.
The actual ratio will vary as a function of the attractiveness of the recipient. Your average joe probably gets none (or as close to none as statistically possible without actually being none). Your guy who looks like he just walked off a runway in Milan, OTOH, can probably count on getting at least a few a day, if only because they are the focus of so much attention.
I wonder if there is a way to use this to finetune an image generating AI to generate images that the AI would think is lewd but is not. (similar to what was done with a Yahoo model in 2016: https://news.ycombinator.com/item?id=12756462 )
More likely is people making adversarial dick pics to get past the filter, which is very quick to do when you have the parameters of a differentiable model. I noticed they were careful to state that they are releasing a model instead of the model they are using though.
Granted, I don't understand the psychology of people who send dick pics so I could be wrong about this, but I get the impression it's not a deeply considered thing conducive to spending lots of effort on. Adversarial dick pics are probably a rare threat just because most men won't be bothered.
Probably not much at all. They used a pretty straightforward and easy to run model, EfficientNetV2. Something that doesn't give out the best results, but will offer very quick inference and train time.
(Full disclosure; I work for a Bumble competitor. And I applaud Bumble for open sourcing this. It's great if we can work together as an industry to protect people from sexual abuse).
There's some questions I have about this release
the interest is in the model itself. They released the pre-trained model as a zip that you can download from their website but they do not mention the license for using this model anywhere. It is unclear from the documentation and license under what conditions you're allowed to use it. The zipfile itself doesn't contain any terms of use inside.
Though they give instructions on how to tweak the model with your own additional input images; again it's unclear from the license if you can create derivatives of the Model.
One might assume it falls under the Apache license that's included in the repository. But the model itself does not live in the repository (is not bundled with the source code) so it's not clear. Also Apache isn't really suited for assets and models. it's more meant for source code.
It would be nice if they would adopt a license that explicitly mentions how the model can be used like StableDiffusion does.
I wonder if marketing put the kibosh on the original title, since the URL slug differs substantially from the article's title
While doing a little view-source: spelunking, I also noticed they're straight-up lying about their alternate language link tags, which I guess is only slightly less interesting than if they had, in fact, translated that blog post into the 26 languages they pretend to support
Based on conversations with (male) friends, it seems to be the place for those who can't deal with Tinder anymore.
Personally I'm back to OkCupid - it may not have the number of users the others do, but at least I now get to know something about potential matches other than 5 pictures and that they enjoy traveling.
Dating apps are tailored to people in the top X% of attractiveness. Don't take it personally. Most people aren't successful on them, even the ones who get a lot of matches.
I'm relevantly good looking, no problem with getting a beautiful woman IRL, but on dating apps I get no matches. Or if I do its girls who are either average or tinder addicts who everyone should avoid.
My working theory is on a dating app woman have unlimited access, so tend to go torwards the top 99.9% of men.
Solution: Shadow ban hyper attractive people.
Solution 2: Only show people a similar level of attractiveness as yourself.
You are 100% correct, unsuccessful guys need to seek help and there are paid services out there. Or speak to a few women on these platforms (chicken-egg I know) and you will be told what others are doing that is turning them off. You need to treat this for what it is, marketing. You don't just put up your first attempt and say women aren't into me. You need to iterate and experiment. I know, it's hard work, 90% of guys won't put in the effort, they will just complain they aren't in the top 0.1% in attractiveness.
I disagree with that the top man can not date more than a few.
A single woman needs a romantic partner let's say ever three months, a man can have one every day. We see this in gay men having 500+ average sexual partners. Gay men can have sex as much as they want, because of access.
Top 0.1% of men can also have sex when ever they want, they are getting constant matches with 6,7,8 level woman. Well below their level, because its easier to get people below your level into bed without courting.
They have sex, he doesn't call her back, she is mad at men for another three months.
A few men can soak up all of the above average woman, considering they are on a bell curve and are rare them self. Only a subset of people are even on these dating apps.
Average looking people don't have this problem, because they are happy with matching other average woman who the top men are not interested in.
Summary: top men can date below their own level because a 7 woman, three below them is still very beautiful. But a 7 man can't date three below their own level because a 4 is not attractive and not worth the time using a app for.
Judging by how you are framing the issue and the way you've talked about women in this thread (which has been a red flag for me, at least) I suspect your experience has nothing to do with your physical traits.
If you'd read the thread, you'll find out that the poster's character is a part of the discussion (as he claims his character has nothing to do with his lack of success on dating sites).
The argument is, "I'm having trouble getting women to approve of me because of systemic issues with women."
I'm pointing out that the causal relationship the poster is making might be wrong. The issue could be the behavior if the poster, not some systemic issue with women.
A less dramatic version of your first solution is to show a profile progressively less the more matches are pending, and the more selective you are in rejecting pending matches.
Couple that with trying to predict likely matches and weight towards showing profiles which are likely to match and you would likely approach solution 2 without explicitly focusing only on attractiveness.
But at the same time you'll have killed a major revenue driver for a lot of dating sites: frustration over few matches.
How do dating apps make money? They were all prerevenue last time I used one.
I think trying to move away from images is a good idea, so people can showcase their worth to better compete with the physical gifted people who benefit from images.
I thought of a idea where you have to have a reel 10second video showcasing 5 traits
Let's say, you see the list of options and choose
Cooking
Humour
Adventure
Intellect
Strength
Say those are your traits, your reel will show case them. With the aim of creating relationships rather than hook ups.
Most of them make apps either by upsells or via ads, both of which gives them an incentive to prolong your usage of the site.
The typical user stays only for weeks or months at a time. Many users come back rarely enough that e.g. Bumble offers a lifetime premium membership for less than it costs you to renew their quarterly membership a third time ($149 for the lifetime membership I think), which should tell you something about how long they expect even the most committed users to hang around.
Which makes sense. After all, if they do their job, you'll meet someone and not need the app any more, and if they do nothing for you, you'll get tired of them and disappear eventually. Only a very narrow subset of users use the non-hookup focused sites for very long.
The challenge then is that this puts a very severe ceiling on your customer acquisition cost. Who knows what their average lifetime value of a user is, but it's going to be a small fraction of that $149 lifetime membership, with most users never paying anything at all.
If you can make one that is good enough that you don't need to advertise much, you might have a chance, but it's a very tough, very crowded market (for every Tinder and Bumble, there's are hundreds of dating sites which nobody remembers).
Most dating doesn't lead to permanent pairings, so the fact that there are roughly equal numbers of men and women isn't that relevant when people are free to date at different frequencies and durations.
For that to work in practice implies a rigidly monogamous society where almost everyone participates in the same context, bonds for life and does so at an early stage of their lifecycle, as well as few mismatches between both relationship goals and pairing criteria. That's essentially the traditional non-polygamous agrarian civilization model if we ignore the fact that the 0.01% or so of nobles had multiple partners. On dating apps the dynamic is very different.
I agree that among the pool of people on these apps, there are lots of matches that would work out among average-looking people. So why doesn't that happen?
It's because a lot of attraction is based on things that photos and profiles can't convey. The medium of "digital app" is just very, very wrong for optimizing dating.
I've seen research that the most efficient way to handle dating is in-person speed dating.
> It's because a lot of attraction is based on things that photos and profiles can't convey.
Especially for men. The worst thing is the imbalance. Compatibility in a long-term relationship is certainly based on factors other than looks for both sexes but, like it or not, all relationships start with basic attraction. Women's basic attraction comes from their looks in the vast majority of cases. But for most men, it has to come from somewhere else. Usually status, power, ability, humour etc. So it works absolutely fine for women, but is a complete non-starter for the average man.
I'm not surprised speed-dating is more efficient. My experience with online dating is that even if I find a woman physically attractive I might not find her attractive in real life. We might simply not hit it off, or there can even be other factors like smell that are off-putting. Speed dating eliminates those immediately as well as allowing people to show off non-physical attractive qualities. I still don't think it's optimal for men, though. The best way is to get good at something and show it off in public. Like playing in a band, public speaking, leading and commanding respect of other men etc. That can make otherwise physically repulsive men attractive. There are countless examples.
Dating apps end up mostly being used for casual dating and hookups. The closest thing to that for average or below average guys is prostitution, which of course has been around a while and much money was and is being made.
Not convinced. There are obvious steps that'd be more "average guy" friendly, but would kill engagement and so revenue.
E.g. slow down showing profiles when they have more than X pending matches. An ex told me she'd paid for Gold on Tinder once out of curiosity (lets you see how many have swiped right on you). Her pending matches were in the 4 digits. These profiles "soak up" a lot of the time and attention of people on these platforms without creating a real chance of matches for users who aren't super aggressively matching everyone.
This again trains for a behaviour where men match a lot, and women can afford to be very selective.
But reducing the displays of the most popular profiles so their pending matches queue is never very deep won't do anything good for revenue.
Another option is to train models to predict likely matches, and sort displays accordingly, but that again won't do much for revenue, because it would mean average men would be more likely to be shown women closer to their attractiveness level which there is less competition for rather than women with 4 figure deep queues of waiting matches.
Remember, most dating sites make money from making it hard enough to get a match that people pay for premium services to increase their chance. They have a built in incentive to exaggerate the possibilities while maintaining friction in actually meeting someone. It must be hard for you to find someone for most of them to do well.
If you can convince people to pay for a less perversely incentivised model, maybe. E.g. a "pay per match" model maybe, if well protected against scammers. But suspect it'd be a very hard sell.
>Dating apps are tailored to people in the top X% of attractiveness.
They aren't. I'm in the top 1% for income and top 5% for looks. I used to get no matches on tinder. I asked a few female friends to come up with a profile they thought would be more attractive. After I saw what they came up with (I have a boat, I fly first class, I'm in a holiday in Europe) I deleted it and move them to acquaintances.
The type of woman who would find that sort of thing attractive is _not_ the type of woman who I'd want to sleep with let alone have a relationship with.
Something about your story doesn't add up. I would certainly consider myself NOT in the top %5 of looks, and get plenty of matches. And I'm not looking for trashy women who are only into material things either.
On looking at the average person I see on the street. Do you not have eyes?
>and get plenty of matches
Matches are the first step, then you have chatting, organizing a date and then a hookup/relationship. The attrition rate is abysmal at every level. It's just not time efficient to use online dating.
A close friend of mine kills it on tinder. He just has pics of himself looking good and smiling. He basically doesn't need to book a place to stay when going on holiday (lol). I would not put him in the top 5% (maybe top 10).
Your pics either sucked or you're not as attractive as you think you are. You don't need pics of yourself being rich. Or maybe you're older (he's 27).
In the show, Jian Yang sells the app to a social media company for precisely detecting this sort of phallic imagery, so I think Mike Judge knew exactly what he was doing.
OT--But Mike Judge was interviewed recently on Fly On The Wall, the podcast from David Spade and Dana Carvey (who doesn't have a podcast these days?)... and it was a great episode. There's a part where he explains where the Butthead's guffawing laugh comes from.. so good.
The podcast is focused on interviewing SNL alums.. but has recently branched out a bit. Anyway, I didn't know this, but Mike Judge did some Milton (the character who later appeared in Office Space) short for them.
A lot of their copy seems to indicate the priority is detecting dick picks. I realize those are the vast, vast majority of "unsolicited lewd images", but I do wonder how capable their model is at detecting pussy pics, if at all.
Does this mean it can also detect "send nudes" to avoid false positives on solicited lewd images? Or does the UI have a on/off toggle for currently-accepting-lewd-images?
Many years ago (ca. 2009?) at a previous employer we had an enterprise email scanner which included a content filtering feature which could block "pornographic images". We ended up turning it off as it had too many false positives since it was an insurance company that dealt with injuries so a lot of legitimate injury photos would get classed as porn and blocked. I think it was basically looking at how much of an image was skin tones.
The more interesting thing about that feature was that it A) only blocked nude images of white people, and B) would aggressively block photos of golden retrievers much to the annoyance of someone in the claims team who bred them and occasionally received photos to her work email.
And it really sticks out as one of the few paradigm shifts we’ve had in the last 10 years. In 2009, it really was an impossible task to do that kind of image recognition. But by 2012, it was available in a library for anyone.
Better Off Ted had a great premise, but it didn't survive past two seasons because it didn't "click". It was the right idea, but just like with successful shows, certain magic that not even the creators understand has to happen. And it didn't happen here, which is a shame.
* During COVID, a medical school professor complains his class livestream would get cut off constantly without warning because he is teaching anatomy and showing pictures.
* Civics class online quiz. Multiple choice question. A: *; B: *; C: *; D: * - apparently all four choices are censored because they are names of Chinese leader.
* A police officer complains his files related to a rape case he is investigating was deleted by the cloud service provider
Google docs spell/grammar check shuts off if the word "anal" appears anywhere in a sentence. https://imgur.com/dTJmsTt
For some reason "anal" and the F-bomb are the only two words I can find where it does this.
The craziest part is if you type something like "analss gland" it figures out you're trying to type "anal gland" and shuts down spellcheck, even though anal gland (dogs) is a perfectly g-rated use of the word anal.
Even the word fuck can be perfectly acceptable in plenty of contexts. Weird that someone went to the trouble of building in weird kill-switch exceptions in a spell checker.
“This feature remains severely disliked by users, with feedback such as ‘Fuck this moronic new button!’ being an example of the milder responses.”
As an Australian, I’m fucking livid these fucks think they have the right to fuck around with my use of the word fuck.
That is to say… as a native speaker of the lovely little English variant where the word fuck or words derived from it are valid in almost every single linguistic context the English language possesses… I take offence at this petty US cultural imperialism.
Sometimes I think there's a specific person at Google with some kind of vocabulary-related mental illness. Just try and get Google Voice Typing to type the word "o'clock". You can stand there saying it over and over, and Google won't stop listening indicating it's hearing and interpreting the word, but it just refuses to type it.
why is it even a 'feature'? someone had to go in add a switch for the word anal to turn spell check off. it isnt like they just dont want to be involved with swear words. this is a feature that makes the app worse?!?
plus anal isnt even a swear word. how do you want me to describe the hole where the shit comes out? this is just wrong on so many levels.
I used to work at a place that blocked a significant fraction of incoming Dutch emails. Eventually we figured out it rejected any message with "kunt" ("he/she can"), which is not exactly uncommon. Apparently they were flagged as deliberate attempts to circumvent the profanity filter.
Do you want to block attempts to skirt the profanity filter? How to you avoid reintroducing the Scunthorpe problem?
Jeez, darn are all mild forms of expletives, should they be blocked? What about bloody? That many anglophones would go to if you couldn't swear?
Surely if you can't use those, people are just going to say gosh instead, or Smeg or frel. Should they also be banned. Swear words exist for a reason, you can perhaps get rid of the harshest forms, you are never going to remove the need or desire for people to swear though.
Around 2007, there was an image board type of site that was struggling to build a system that could reliably detect these sort of images.
Their solution ended up being to run it through face detection first -- which was pretty reliable even back then.
It turns out, people who sent unsolicited lewd photos tended to do so without their face in frame. So if a face was detected, it was significantly less likely to be lewd.
Just an amusing anecdote. Wish I could find the blog post describing it.
Also a great example of effective feature engineering.
The success of gradient boosting and deep learning lies largely in the ability of those models to "learn" sophisticated high-order features from raw data. Their particular success in working with images, audio, and video lie in their ability to construct more (and more sophisticated) features, beyond what a team of smart humans with domain knowledge could have constructed by hand given huge resources. In a sense, the only difference between model architectures is the feature space that they embed the raw data into.
That said, sometimes you just need to give the model a good starting place, and sometimes you can obviate A sophisticated model largely or entirely, if you can come up with highly "explanatory" features on your own.
To this day, I think the ability to come up with high-impact features is one of the differentiating factors of really good industry data scientist.
There should be a feature to warp and bend the naughty bits so they're tiny and shriveled and encrusted with warts and scabs and spiders, and post a copy back to the offender, to discourage them.
Funnily enough one of my SFW pictures got detected recently on bumble and my working theory is that they detected my phone as a bulge. It might be fun to try to test what type of content it misidentifies.
20 years ago when candy bar phones were popular i designed a dildo-shaped phone case – a sort of “push up bra” for men. Not sure why my product never took off.
This looks like the detector does not run on the phone but I could be wrong about that. If it does not, that implies all messages on this dating app are not private. These conversations and pictures could be very sensitive.
our scale allows us to collect a best-in-the-industry dataset of both lewd and non-lewd images
For another, it seems like they are saying they collect all these images of people's private parts and use them for their training set.
Though it's possible they're sourcing the photos in the above way, in general I'd assume that the messages are not private. Unless messages are end-to-end encrypted on a platform, you have to assume the platform can access them.
This isn't some huge revelation. E2E encryption isn't something that Bumble advertises, and their privacy policy [1] plainly states:
> We review the content of messages sent in the App to identify topics, sentiments, and trends across our Users. We will take steps to remove personally identifying information from such messages, prior to reviewing them. We will not share the content of User messages or information we derive from them with any third party.
edit: and for the record if your messages on a dating app are not E2E encrypted then you are making a terrible mistake. I can't believe anyone would voluntarily make this decision.
> edit: and for the record if your messages on a dating app are not E2E encrypted then you are making a terrible mistake. I can't believe anyone would voluntarily make this decision.
Reality goes something like this: You roll out your E2EE dating app, it gets successful, you're winning bigly. Then cops show up with an order from a judge demanding you hand over information regarding user #890202 because they are suspected of using your platform to entice minors into sexual activity. This is usually where the E2EE worship falls flat, for legitimate societal reason.
If you want E2EE you need to use a dedicated app specifically ONLY for that (Signal, WhatsApp, etc). Any time you bolt on other business or incentives and then use E2EE as the messaging platform you will discover a world of pain in my opinion.
it's always that isn't it? They could go raid the guy's home perhaps. Alternatively you sent embarrassing messages to a woman on a dating app, you have a weird fetish you discussed with someone on there, you're cheating on your wife, or maybe some nation-state actor wants some blackmail on you. A hacker or disgruntled employee has dumped a database of all the messages and journalists and Twitter users are going through it right now. Or maybe you don't trust Bumble. They seem to have an axe to grind anyway.
Not that the issue of minors isn't a problem but I don't think that should always override people's privacy. And FWIW there is plenty of space for people to share compromising messages on the dating app before they can manage to convince a stranger to download Signal and continue the conversation there. That's if you're lucky enough that she doesn't immediately think you are a scammer of some kind.
I am on the side of privacy extremists but a lot of people find that to be a nonstarter so to each their own. I wouldn't dare go on a dating app that didn't have E2EE though or at least a way to anonymize yourself so no phone # required. ;)
I think privacy-first is the the correct mental model to have. I've had to explain to some people that the usage of a certain platform by drug dealers, pedophiles, and organized violent crime isn't a defect, but really a feature in theory. If the tech or platform sucked/was-compromised, they wouldn't gravitate towards it. It is a weird conversation, but most people do seem to grok it in that sense.
>Traversing the trade-offs between state-of-the-art performance and the ability to serve our user base at scale, we implemented (in its latest iteration) an EfficientNetv2-based binary classifier: a convolutional network that has faster training speed and overall better parameters efficiency. It uses a combination of better designed architecture and scaling, with layers like MBConv (that utilizes 1×1 convolutions to wide up the space and depth-wise convolutions for reducing the number of overall parameters) and FusedMBConv (that merges some steps of the vanilla MBConv above for faster execution), to jointly optimize training speed and parameter efficiency. The model has been trained leveraging our GPU powered data centers in a continuous exercise of dataset, network and hyperparameters (the settings used to speed up or improve the training performance) optimization.
This is a lot of words to say, "we used an off-the-shelf architecture". Like it's literally just Google's public implementation copied over.
253 comments
[ 3.6 ms ] story [ 288 ms ] threadI suspect given the option, most people would turn on an option for "block dick pics silently" and leave it on.
Similar to how I never actually check my spam folder anymore.
Considering what Bumble/Tinder are maybe this makes sense?
In Centralia, PA there use to be something called Grafitti Highway. It's an old highway road (they created a new road around it) that people would go and graffiti this old, unused highway. It might as been called Penis Highway as innately that's what was mostly drawn there. Too bad they shut it down lol
(Which I think is still mostly guys doing it, at least historically...)
I'd note that the title here is slightly misleading. It doesn't do anything to check if the pics are solicited or not.
While I imagine having a good filter is important given the scale of online dating I would think all other filters would eliminate most of the problem (auto blurring images if people have only sent a few messages to each other, blur the first image sent, blur images by anyone who has been reported before, etc.)
But for "adult" photos they just blur.
I suspect people like this feature because they want to make sure they don't have people around them when they open that message (assuming it is a consensualy exchanged pic)
1. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtm...
I have never once received an unwanted nude from a woman but as spam whereas with men it has happened enough times for me to believe without statistical justification that unwanted images of male genitalia would be more common.
I am not sure how this is relevant to building a classifier, however. Or relevant to California law.
I think a lot of unsolicited female genitalia is sent for spam/scam proposes. A particularly pernicious sort is in the form of sextortion [1]. Images of female genitalia may be sent by a criminal to entice a (typically young or teenage) boy into believing he’s engaging in an online romantic relationship with a real girl. When he reciprocates with images of his own body (as requested to do so), the criminal responds by using those images to blackmail him into paying money with the threat of exposing him to friends and family. This sort of attack has also been used to target girls, though recently the targeting of boys has grown much faster. If I had to speculate, criminals likely see young boys as easier targets who can be enticed with images of the female body.
[1] https://en.wikipedia.org/wiki/Sextortion
If I opened an app one day and received an unsolicited picture of female genitalia, that would be strange. If it happened every day, then it's spam, even if I don't have a strong aversion to it.
IIRC, in something like 2000 interviews there were 4 male flashers and zero female.
Four more than expected when we're paying for their time.
2000:4:0
Archive of the original post (NSFW?): https://web.archive.org/web/20161022003102/https://open_nsfw...
Two minute papers - 1 pixel attack:
https://www.youtube.com/watch?v=AOZw1tgD8dA
Decent papers:
https://arxiv.org/abs/1812.00181
https://arxiv.org/pdf/1807.00051
Well, I wasn’t planning on encountering that phrase today, perhaps ever to be honest
There's some questions I have about this release
the interest is in the model itself. They released the pre-trained model as a zip that you can download from their website but they do not mention the license for using this model anywhere. It is unclear from the documentation and license under what conditions you're allowed to use it. The zipfile itself doesn't contain any terms of use inside.
Though they give instructions on how to tweak the model with your own additional input images; again it's unclear from the license if you can create derivatives of the Model.
One might assume it falls under the Apache license that's included in the repository. But the model itself does not live in the repository (is not bundled with the source code) so it's not clear. Also Apache isn't really suited for assets and models. it's more meant for source code.
It would be nice if they would adopt a license that explicitly mentions how the model can be used like StableDiffusion does.
While doing a little view-source: spelunking, I also noticed they're straight-up lying about their alternate language link tags, which I guess is only slightly less interesting than if they had, in fact, translated that blog post into the 26 languages they pretend to support
Personally I'm back to OkCupid - it may not have the number of users the others do, but at least I now get to know something about potential matches other than 5 pictures and that they enjoy traveling.
My working theory is on a dating app woman have unlimited access, so tend to go torwards the top 99.9% of men.
Solution: Shadow ban hyper attractive people.
Solution 2: Only show people a similar level of attractiveness as yourself.
Much more likely you don't put enough effort into your profile or you have something on, that is a red flag for most women.
I disagree with that the top man can not date more than a few.
A single woman needs a romantic partner let's say ever three months, a man can have one every day. We see this in gay men having 500+ average sexual partners. Gay men can have sex as much as they want, because of access.
Top 0.1% of men can also have sex when ever they want, they are getting constant matches with 6,7,8 level woman. Well below their level, because its easier to get people below your level into bed without courting.
They have sex, he doesn't call her back, she is mad at men for another three months.
A few men can soak up all of the above average woman, considering they are on a bell curve and are rare them self. Only a subset of people are even on these dating apps.
Average looking people don't have this problem, because they are happy with matching other average woman who the top men are not interested in.
Summary: top men can date below their own level because a 7 woman, three below them is still very beautiful. But a 7 man can't date three below their own level because a 4 is not attractive and not worth the time using a app for.
I'm pointing out that the causal relationship the poster is making might be wrong. The issue could be the behavior if the poster, not some systemic issue with women.
Couple that with trying to predict likely matches and weight towards showing profiles which are likely to match and you would likely approach solution 2 without explicitly focusing only on attractiveness.
But at the same time you'll have killed a major revenue driver for a lot of dating sites: frustration over few matches.
I think trying to move away from images is a good idea, so people can showcase their worth to better compete with the physical gifted people who benefit from images.
I thought of a idea where you have to have a reel 10second video showcasing 5 traits
Let's say, you see the list of options and choose
Cooking Humour Adventure Intellect Strength
Say those are your traits, your reel will show case them. With the aim of creating relationships rather than hook ups.
The typical user stays only for weeks or months at a time. Many users come back rarely enough that e.g. Bumble offers a lifetime premium membership for less than it costs you to renew their quarterly membership a third time ($149 for the lifetime membership I think), which should tell you something about how long they expect even the most committed users to hang around.
Which makes sense. After all, if they do their job, you'll meet someone and not need the app any more, and if they do nothing for you, you'll get tired of them and disappear eventually. Only a very narrow subset of users use the non-hookup focused sites for very long.
The challenge then is that this puts a very severe ceiling on your customer acquisition cost. Who knows what their average lifetime value of a user is, but it's going to be a small fraction of that $149 lifetime membership, with most users never paying anything at all.
If you can make one that is good enough that you don't need to advertise much, you might have a chance, but it's a very tough, very crowded market (for every Tinder and Bumble, there's are hundreds of dating sites which nobody remembers).
I agree that among the pool of people on these apps, there are lots of matches that would work out among average-looking people. So why doesn't that happen?
It's because a lot of attraction is based on things that photos and profiles can't convey. The medium of "digital app" is just very, very wrong for optimizing dating.
I've seen research that the most efficient way to handle dating is in-person speed dating.
Especially for men. The worst thing is the imbalance. Compatibility in a long-term relationship is certainly based on factors other than looks for both sexes but, like it or not, all relationships start with basic attraction. Women's basic attraction comes from their looks in the vast majority of cases. But for most men, it has to come from somewhere else. Usually status, power, ability, humour etc. So it works absolutely fine for women, but is a complete non-starter for the average man.
I'm not surprised speed-dating is more efficient. My experience with online dating is that even if I find a woman physically attractive I might not find her attractive in real life. We might simply not hit it off, or there can even be other factors like smell that are off-putting. Speed dating eliminates those immediately as well as allowing people to show off non-physical attractive qualities. I still don't think it's optimal for men, though. The best way is to get good at something and show it off in public. Like playing in a band, public speaking, leading and commanding respect of other men etc. That can make otherwise physically repulsive men attractive. There are countless examples.
E.g. slow down showing profiles when they have more than X pending matches. An ex told me she'd paid for Gold on Tinder once out of curiosity (lets you see how many have swiped right on you). Her pending matches were in the 4 digits. These profiles "soak up" a lot of the time and attention of people on these platforms without creating a real chance of matches for users who aren't super aggressively matching everyone.
This again trains for a behaviour where men match a lot, and women can afford to be very selective.
But reducing the displays of the most popular profiles so their pending matches queue is never very deep won't do anything good for revenue.
Another option is to train models to predict likely matches, and sort displays accordingly, but that again won't do much for revenue, because it would mean average men would be more likely to be shown women closer to their attractiveness level which there is less competition for rather than women with 4 figure deep queues of waiting matches.
Remember, most dating sites make money from making it hard enough to get a match that people pay for premium services to increase their chance. They have a built in incentive to exaggerate the possibilities while maintaining friction in actually meeting someone. It must be hard for you to find someone for most of them to do well.
If you can convince people to pay for a less perversely incentivised model, maybe. E.g. a "pay per match" model maybe, if well protected against scammers. But suspect it'd be a very hard sell.
They aren't. I'm in the top 1% for income and top 5% for looks. I used to get no matches on tinder. I asked a few female friends to come up with a profile they thought would be more attractive. After I saw what they came up with (I have a boat, I fly first class, I'm in a holiday in Europe) I deleted it and move them to acquaintances.
The type of woman who would find that sort of thing attractive is _not_ the type of woman who I'd want to sleep with let alone have a relationship with.
Something about your story doesn't add up. I would certainly consider myself NOT in the top %5 of looks, and get plenty of matches. And I'm not looking for trashy women who are only into material things either.
On looking at the average person I see on the street. Do you not have eyes?
>and get plenty of matches
Matches are the first step, then you have chatting, organizing a date and then a hookup/relationship. The attrition rate is abysmal at every level. It's just not time efficient to use online dating.
Your pics either sucked or you're not as attractive as you think you are. You don't need pics of yourself being rich. Or maybe you're older (he's 27).
https://tsdr.uspto.gov/#caseNumber=79343576&caseType=SERIAL_...
https://www.quora.com/How-did-the-term-private-dick-for-a-pr...
The podcast is focused on interviewing SNL alums.. but has recently branched out a bit. Anyway, I didn't know this, but Mike Judge did some Milton (the character who later appeared in Office Space) short for them.
I recommend it.
Does this mean it can also detect "send nudes" to avoid false positives on solicited lewd images? Or does the UI have a on/off toggle for currently-accepting-lewd-images?
The more interesting thing about that feature was that it A) only blocked nude images of white people, and B) would aggressively block photos of golden retrievers much to the annoyance of someone in the claims team who bred them and occasionally received photos to her work email.
That gave me a good laugh
And it really sticks out as one of the few paradigm shifts we’ve had in the last 10 years. In 2009, it really was an impossible task to do that kind of image recognition. But by 2012, it was available in a library for anyone.
AlexNet was September 2012.
FlickR did their XKCD1425 bird detector in Oct 2014: https://code.flickr.net/2014/10/20/introducing-flickr-park-o...
(If anyone here hasn't watched the series, do yourself a favor and do so.)
That said, still a very entertaining show.
It turned out the untextured 3D pig's head in my image was tripping Discord's porno filter. I can only assume it was the shape of its snout.
* During COVID, a medical school professor complains his class livestream would get cut off constantly without warning because he is teaching anatomy and showing pictures.
* Civics class online quiz. Multiple choice question. A: *; B: *; C: *; D: * - apparently all four choices are censored because they are names of Chinese leader.
* A police officer complains his files related to a rape case he is investigating was deleted by the cloud service provider
For some reason "anal" and the F-bomb are the only two words I can find where it does this.
The craziest part is if you type something like "analss gland" it figures out you're trying to type "anal gland" and shuts down spellcheck, even though anal gland (dogs) is a perfectly g-rated use of the word anal.
https://imgur.com/KjWcQJm
“This feature remains severely disliked by users, with feedback such as ‘Fuck this moronic new button!’ being an example of the milder responses.”
That is to say… as a native speaker of the lovely little English variant where the word fuck or words derived from it are valid in almost every single linguistic context the English language possesses… I take offence at this petty US cultural imperialism.
Just looking at the sheer flexibility of the word on the wiktionary reveals it’s impressive breadth of use. https://en.m.wiktionary.org/wiki/fuck#English
https://github.com/ltworf/international_code_of_conduct/blob...
"my anal ysis shows..."
that doesnt get flagged as a spelling error?
why is it even a 'feature'? someone had to go in add a switch for the word anal to turn spell check off. it isnt like they just dont want to be involved with swear words. this is a feature that makes the app worse?!?
plus anal isnt even a swear word. how do you want me to describe the hole where the shit comes out? this is just wrong on so many levels.
Jeez, darn are all mild forms of expletives, should they be blocked? What about bloody? That many anglophones would go to if you couldn't swear?
Surely if you can't use those, people are just going to say gosh instead, or Smeg or frel. Should they also be banned. Swear words exist for a reason, you can perhaps get rid of the harshest forms, you are never going to remove the need or desire for people to swear though.
Their solution ended up being to run it through face detection first -- which was pretty reliable even back then.
It turns out, people who sent unsolicited lewd photos tended to do so without their face in frame. So if a face was detected, it was significantly less likely to be lewd.
Just an amusing anecdote. Wish I could find the blog post describing it.
The success of gradient boosting and deep learning lies largely in the ability of those models to "learn" sophisticated high-order features from raw data. Their particular success in working with images, audio, and video lie in their ability to construct more (and more sophisticated) features, beyond what a team of smart humans with domain knowledge could have constructed by hand given huge resources. In a sense, the only difference between model architectures is the feature space that they embed the raw data into.
That said, sometimes you just need to give the model a good starting place, and sometimes you can obviate A sophisticated model largely or entirely, if you can come up with highly "explanatory" features on your own.
To this day, I think the ability to come up with high-impact features is one of the differentiating factors of really good industry data scientist.
Edit: missing "to"
Reported profile photos: lewd
Reported message photos: lewd
Though it's possible they're sourcing the photos in the above way, in general I'd assume that the messages are not private. Unless messages are end-to-end encrypted on a platform, you have to assume the platform can access them.
> We review the content of messages sent in the App to identify topics, sentiments, and trends across our Users. We will take steps to remove personally identifying information from such messages, prior to reviewing them. We will not share the content of User messages or information we derive from them with any third party.
[1] https://bumble.com/en-us/privacy
Isn't that what the SavedModel is that they shared on googleapis.com right here? https://storage.googleapis.com/private_detector/private_dete...
edit: and for the record if your messages on a dating app are not E2E encrypted then you are making a terrible mistake. I can't believe anyone would voluntarily make this decision.
Reality goes something like this: You roll out your E2EE dating app, it gets successful, you're winning bigly. Then cops show up with an order from a judge demanding you hand over information regarding user #890202 because they are suspected of using your platform to entice minors into sexual activity. This is usually where the E2EE worship falls flat, for legitimate societal reason.
If you want E2EE you need to use a dedicated app specifically ONLY for that (Signal, WhatsApp, etc). Any time you bolt on other business or incentives and then use E2EE as the messaging platform you will discover a world of pain in my opinion.
Not that the issue of minors isn't a problem but I don't think that should always override people's privacy. And FWIW there is plenty of space for people to share compromising messages on the dating app before they can manage to convince a stranger to download Signal and continue the conversation there. That's if you're lucky enough that she doesn't immediately think you are a scammer of some kind.
I am on the side of privacy extremists but a lot of people find that to be a nonstarter so to each their own. I wouldn't dare go on a dating app that didn't have E2EE though or at least a way to anonymize yourself so no phone # required. ;)
This is a lot of words to say, "we used an off-the-shelf architecture". Like it's literally just Google's public implementation copied over.