Imagine if we invent a reliable way to encode any encrypted payload inside a regular Kind 1 post. You can make an AI account that is just posting silly stories. Each story is its own encoding scheme (like base64) for the encrypted information underneath. The encoding algorithm just so happens to output fully correct, and even enjoyable, sentences out of it. Most people will read them without realizing they are actually encrypted information being transferred in plain sight. If they include a significant amount of error correction, humans can transmit machine knowledge by just telling stories to one another. And we love to tell stories.
We already do this in pictures, with today's LLMs we can start doing it with text.
A form of steganography.
Im of the opinion that all kinds should support encrypted payloads.
It seems the CIA did this to the press about 70 years ago.
And they likely use Memes today to do the same thing.
That's a fascinating concept! It's like embedding secrets within the fabric of everyday communication. It could make for some entertaining exchanges, with people unwittingly passing along encrypted information while enjoying amusing stories. Plus, the idea of using storytelling as a medium for transmitting knowledge has a certain charm to it. It would definitely add a layer of intrigue to everyday conversations.
Would this allow for steganography of higher bandwidth than the usual images or what.
At least you wouldn't need potentially conspicuous streams of images or videos to create a conversation.
i was thinking this a while ago. you cannot reverse the process of neural network.
the process discards information. you can get same output for multiple inputs
what is possible however, it to implement something similar to bip39 encoding
then try to brute force these words into meaningful sentences
also instead of using 2048 words like in bip39, you can use longer word list, this should make encoding more space efficient
cat /usr/share/dict/usa | wc -l
123985
Oddbean new post about login | logout