Uploaded August 2011 | Updated November 2026, 1 hour ago
Here's my attempt to document the hat hunt. I hope you don't mind me adding some spaces, punctuation, and capitalization here and there where the original message lacked them. ;-)
As the video was meant to be brief, it mainly explains how to extract the information to those with the know-how, and reveals the messages. I'll elaborate a bit here for those who would like to gain a deeper understanding, but wouldn't figure it out by themselves. More spoilers!
The ad is 128x128, but the image extracted from it is 32x64. The reason is that the ad was really encoded at double resolution; it consists of 2x2 blocks. I suppose the 4x redundancy mitigates errors in digitizing the image. You can see that each 2x2 block consists of two white and two black bits. And in the other half, either the same or the opposite pattern is repeated. Whether they're different determines the value of the pixel. One fun property is that it works a little like a stereogram. It's rather unclear, but you can see some of the image pop out if you aim your eyes 64 pixels apart.
If it wasn't clear from the video, the pixels that spelled MOJANG were the ones containing the message.
Creating manifesto.ogg must've been fun for Notch. There are three different messages encoded in the same blips, all using every blip. So, 391 blips, 391 Morse sounds, 391 drawn letters, and 391 tones. It was also the most fun for me, because you get to try heuristics to extract natural-ish information. I winged a peak detector to tell when tones start and stop, which worked rather well.
The Morse part is just your basic Morse code, except it becomes ridiculously fast towards the end. I simply kept adjusting the timing threshold based on the actual length of the tone or silence. Morse tones are ideally either 1 or 3 times some length, and pauses can be that or 5 times (tone, letter, and word spacing). So if my program encountered a length closest to 3 times the base length, it would adjust its base length to a third of that length for the next tone or silence. It doesn't even notice the speedup.
The letters were rather fun. It's not often that you literally draw images from a sound file. Here you can see that my peak detector was slightly aggressive. There is often a faint inner letter where the signal was attenuated, especially towards the end where there wasn't much time to draw a thick prominent letter (only the first 10 are in the video, however). I didn't automate reading the letters, because that was easier to do by hand than to automate.
Finally, there was the fundamental frequency. Of course, Morse code doesn't care which note the blips are. Perhaps you'd think the shape of the letters would be affected, but that isn't true (at least below the Nyquist frequency). All that matters is the phase relationship between the two channels. A higher tone will simply draw the same letter faster. So there could well be data lurking there, and indeed there is.
The fundamental frequency is one of the four first multiples of 550 Hz. As explained in the video, each of them is assigned a sequence of two bits. If you put those in a raster, you get a nice image. I wonder why evanoc's image isn't as clean as mine. I had an excellent signal-to-noise ratio spanning orders of magnitude, so I could vary my thresholds wildly and still get the same image. On the other hand, his image is full of errors. All I know is he used an FFT, where I just hunted for the relevant frequencies directly (why bother when you only need four?).
Here's my attempt to document the hat hunt. I hope you don't mind me adding some spaces, punctuation, and capitalization here and there where the original message lacked them. ;-)
As the video was meant to be brief, it mainly explains how to extract the information to those with the know-how, and reveals the messages. I'll elaborate a bit here for those who would like to gain a deeper understanding, but wouldn't figure it out by themselves. More spoilers!
The ad is 128x128, but the image extracted from it is 32x64. The reason is that the ad was really encoded at double resolution; it consists of 2x2 blocks. I suppose the 4x redundancy mitigates errors in digitizing the image. You can see that each 2x2 block consists of two white and two black bits. And in the other half, either the same or the opposite pattern is repeated. Whether they're different determines the value of the pixel. One fun property is that it works a little like a stereogram. It's rather unclear, but you can see some of the image pop out if you aim your eyes 64 pixels apart.
If it wasn't clear from the video, the pixels that spelled MOJANG were the ones containing the message.
Creating manifesto.ogg must've been fun for Notch. There are three different messages encoded in the same blips, all using every blip. So, 391 blips, 391 Morse sounds, 391 drawn letters, and 391 tones. It was also the most fun for me, because you get to try heuristics to extract natural-ish information. I winged a peak detector to tell when tones start and stop, which worked rather well.
The Morse part is just your basic Morse code, except it becomes ridiculously fast towards the end. I simply kept adjusting the timing threshold based on the actual length of the tone or silence. Morse tones are ideally either 1 or 3 times some length, and pauses can be that or 5 times (tone, letter, and word spacing). So if my program encountered a length closest to 3 times the base length, it would adjust its base length to a third of that length for the next tone or silence. It doesn't even notice the speedup.
The letters were rather fun. It's not often that you literally draw images from a sound file. Here you can see that my peak detector was slightly aggressive. There is often a faint inner letter where the signal was attenuated, especially towards the end where there wasn't much time to draw a thick prominent letter (only the first 10 are in the video, however). I didn't automate reading the letters, because that was easier to do by hand than to automate.
Finally, there was the fundamental frequency. Of course, Morse code doesn't care which note the blips are. Perhaps you'd think the shape of the letters would be affected, but that isn't true (at least below the Nyquist frequency). All that matters is the phase relationship between the two channels. A higher tone will simply draw the same letter faster. So there could well be data lurking there, and indeed there is.
The fundamental frequency is one of the four first multiples of 550 Hz. As explained in the video, each of them is assigned a sequence of two bits. If you put those in a raster, you get a nice image. I wonder why evanoc's image isn't as clean as mine. I had an excellent signal-to-noise ratio spanning orders of magnitude, so I could vary my thresholds wildly and still get the same image. On the other hand, his image is full of errors. All I know is he used an FFT, where I just hunted for the relevant frequencies directly (why bother when you only need four?).










