 ugh, this is pretty clearly a CAPTCHA being used to tune a machine learning algorithm. these are definitely ML-generated pictures and I imagine what's going on is they are adding some noise to a query for "cake" and then asking me to tag the cake-looking stuff. BUT! It means that it doesn't actually matter what I select. I picked a random assortment of stuff and it let me through just fine. Take THAT, training model!! :AngelDevil: 39 comments
@darius how is that even supposed to work as a captcha then?? it's like they were so focused on the "using free labor to train our ML" part that they forgot the original purpose was to block bots @prehensile I think you actually have to pick stuff that is close-enough-to-true but not quite accurate in order for it to pass you (basically they are checking if your answer deviates *too far* from how wrong they *think* they are, it's an epistemic mess) @darius I got a similar one the other day that I failed because I was interpreting #9 as a "strawberry cake"  @darius Know how that goes. accessibility nightmares alone with those... it's nice to know that they are difficult for even those with sight.
[DATA EXPUNGED]
@darius Weird! I thought the usual CAPTCHA approach was to provide a random mix of known and unknown images, then use clever statistics and aggregation to get certainty about unknown labels. Sounds like this particular implementation is buggy. @adrian yes, I was simplifying when I said "it doesn't matter what I select" -- I suspect I can select stuff that is wrong but also close enough within a threshold of expected-wrongness for it to pass. OTOH in the next captcha I told them that an airplane was a house and they were like "sure thing man"  @darius I guess that would help with the reverse image search problem. But it doesn't make me feel great about the strength of hCaptcha.  @darius wait, wasn't captcha always used for training? I remember back when Google used words they were training their text recognition algorithms. There used to be two words, one was generated and the other was a scan from a book or newspaper. They knew the solution to the generated one and wanted us to solve the scan for them. I sometimes had fun trying to spot the scanned word and "solve" it with random text.    @darius i got the same captcha once but it was brutal. i had to go through three of these before i was let in.
[DATA EXPUNGED]
 @darius I’ve seen this a few times, but I’ve also been under the assumption that we’ve been training an ML model for Alphabet’s self driving cars for years. re this:  The OG recaptcha was smarter: there were two words, one to check you, and another one it wants you to recognize. It was usually easy to tell which is which, the unknown word had more noise. So I would enter "fuck" instead of that word. Worked every time 👍 @darius this is something ive noticed too! you know what would be funny. if we all just clicked the wrong stuff and messed up the ML :thounking: "Dog? that sure looks like a cake to me!" @darius @darius Yeah, with hCaptcha, the secret is that the users are still the product, you're just using a different middleman. From their website: "When you use hCaptcha, companies bid on the work your users do as they prove their humanity. You get the rewards." Google mostly uses street view stuff in their captchas, but hCaptcha literally sells a data labelling service in which your users are the employees. |
Gregory's Server
@darius having fun determining the most useless data to provide 💅🏻