Can hCaptcha be used to replace ReCAPTCHA at BitcoinTalk?

[Lucius]

Which one of you convinced Google that the mailbox is a parking meter?
Welcome the new world order and the tyranny of the AI.

reCaptcha can sometimes be really weird to say the least, but their AI obviously still has some basic issues like recognizing a parking meter from a mailbox. This is one of the problems that AI has from the very beginning, some seemingly simple things are still complicated for them, and there seems to be no progress in the last 5 years -> Simple Pictures That State-of-the-Art AI Still Can't Recognize.

And when it comes to hCaptcha, lately (at least in my case) it all comes down to mostly a lot of clicking - because if, say, they ask to mark all the bikes, then it looks exactly like this.

[libert19]

As they always give images cropped from different positions, doesn't that mean they could give us 9 squares and depending on which one we click, the AI understands which part exactly of the picture contains the target object? Some squares contain only a very small part of the object or none.

There were times when they asked me to select a "bicycle" when the whole canvas showed a motorcycle instead.

It's deliberate afaik, it says somewhere that if no matching found, click skip button and you will get access to whatever you were trying to access.

[PrimeNumber7]

Which one of you convinced Google that the mailbox is a parking meter?
Welcome the new world order and the tyranny of the AI.

reCaptcha can sometimes be really weird to say the least, but their AI obviously still has some basic issues like recognizing a parking meter from a mailbox. This is one of the problems that AI has from the very beginning, some seemingly simple things are still complicated for them, and there seems to be no progress in the last 5 years -> Simple Pictures That State-of-the-Art AI Still Can't Recognize.

captchas are used to prevent people from using automation to solve the 'puzzle'. This includes preventing an adversary from using their own AI from solving puzzles.

An important aspect of training AI is to allow a model to attempt to solve a puzzle (make a prediction), and to subsequently give the model feedback as to how it performed. If you give a model bad/inaccurate feedback, its best possible performance is bad. If you pass an image of a mailbox into a model, and tell the model it is wrong when it correctly predicts the image is not of a parking meter, in the future the model will make less accurate predictions.

If the puzzle that SM was asked to solve was presented again, the person SM might not choose the parking meter to be a mailbox, while AI might classify it to be a mailbox after receiving feedback that it is. Or a future puzzle might ask SM to choose all the mailboxes, and present the same picture, Sm the person would choose that picture, while a model might not choose that picture after receiving feedback that it is of a parking meter. Someone who makes either of these 'mistakes' might be presented an additional challenge.

In other words, the "mistake" that SM highlighted was not a mistake.