2 ms·

Maybe someone should do a solver for this CAPTCHA and put the code on GitHub. It does look really easy to solve using Machine Learning.

sailorganymede 5y ago
If I was a betting man, I’d wager the reason why they give these weirder types of challenges is because solving CAPTCHAs means you’re labelling them, which gives who ever is hosting it labelled data to train their own models against. that’s why CAPTCHAs sometimes ask ppl to spot all the zebra crossings and stuff etc - no machine can solve it yet, so ppl can do it, and by doing it and solving it, you produce a labelled dataset for autonomous cars.
- jamesy0ung 5y ago
  That's how hCaptcha does it https://www.hcaptcha.com/labeling https://www.hcaptcha.com/labeling You play to get a dataset labelled and trained.
  - londons_explore 5y ago
    Except there are now enough labelled datasets of those things, even public freely available ones. If you wanted to increase the precision of the model further, you'd end up using very ambiguous images which wouldn't be suitable for a captcha.
- gambiting 5y ago
  Yes, which is why every time you are given one of those captchas you deliberately pick at least one incorrect image, and it still lets you through. Back in the day when google captcha had you typing in words, it was super clear which one was generated and which one was scanned from a book or whatever - you'd always type in the scanned word completely randomly and it would still accept it.
noduerme 5y ago
If CAPTCHA was only about trying to keep bots out of websites, it wouldn't be that hard. CAPTCHA is a profitable enterprise, because every user response is used to train an AI. It is presented as a free service for website owners, but actually it uses your users to be trainers. A truly free open-source captcha wouldn't be very hard to implement or to keep updated, but no one does it because it would be very expensive to host the server.
captn3m0 5y ago
You can subscribe here for whenever it happens: https://github.com/dessant/buster/issues/320 https://github.com/dessant/buster/issues/320

sailorganymede 5y ago

If I was a betting man, I’d wager the reason why they give these weirder types of challenges is because solving CAPTCHAs means you’re labelling them, which gives who ever is hosting it labelled data to train their own models against. that’s why CAPTCHAs sometimes ask ppl to spot all the zebra crossings and stuff etc - no machine can solve it yet, so ppl can do it, and by doing it and solving it, you produce a labelled dataset for autonomous cars.

jamesy0ung 5y ago
That's how hCaptcha does it https://www.hcaptcha.com/labeling https://www.hcaptcha.com/labeling You play to get a dataset labelled and trained.
- londons_explore 5y ago
  Except there are now enough labelled datasets of those things, even public freely available ones. If you wanted to increase the precision of the model further, you'd end up using very ambiguous images which wouldn't be suitable for a captcha.
gambiting 5y ago
Yes, which is why every time you are given one of those captchas you deliberately pick at least one incorrect image, and it still lets you through. Back in the day when google captcha had you typing in words, it was super clear which one was generated and which one was scanned from a book or whatever - you'd always type in the scanned word completely randomly and it would still accept it.

noduerme 5y ago

If CAPTCHA was only about trying to keep bots out of websites, it wouldn't be that hard. CAPTCHA is a profitable enterprise, because every user response is used to train an AI. It is presented as a free service for website owners, but actually it uses your users to be trainers. A truly free open-source captcha wouldn't be very hard to implement or to keep updated, but no one does it because it would be very expensive to host the server.

captn3m0 5y ago

You can subscribe here for whenever it happens: https://github.com/dessant/buster/issues/320 https://github.com/dessant/buster/issues/320