TINAC

i added audio captcha…. well i tried but its way too hard

ok because of hackatime update delay this devlog covers a lot of things:

1. I added verification support
2. Made it an image of text not just text, it is compressed using zlib and its 1-bit
3. Improved the noise generation, it now uses a more advanced 2D markov chain

In the video, you can notice:

1. The noise has more “structure” to it than before. This is what the upgraded 2D markov chain is doing, and it makes it harder for ML models to recognize the decoys
2. The text is actually an image of text, notice the border around it and the pixelation.
3. I can click “submit” now

Made the demo for the captcha.
This is incredibly easy for humans, as we can “stop” mvoing the slider in real time. But for AI, you have to send a whopping 50 images to a multimodal (image supporting) AI model. Even using a light model like Gemini 3 Flash, it will take 27,225$ or 54,450 input tokens (equivalent to an 150 page book) and JUST TO CRACK ONE OF THESE!
Oh, if you are thinking of pasting the ASCII text that wont work since the Markov Chain produces decoys that have the same local patterns as real ASCII text art and the AI gives answers all of the decoy texts, and even when its not a decoy its the wrong answer. So without images AI can’t “read” ASCII text art.
The “fake” ASCII arts are generated using a special markov chain algorithm (implemented in the previous devlog) that makes it harder for an ML model to recognize which text is the real one.

Made a markov chain giberrish generator (you will see why).