I first saw this implementation from a Harvard paper back when LLM's were still just a novelty[0]. Glad to see they got their demo site back up. Always thought it was a cool idea.

[0] https://github.com/harvardnlp/NeuralSteganography

I wonder if you can construct a function between the encoder and decoder such that for any given input, both the raw and manipulated embeddings decode to plausible meanings that are guaranteed to be different.

for anyone who wants to try a consumer grade stegongraphy in browser. I built some thing here. Its free and loads a static page with a wasm binary. Once the page loads everything is handled in the browser.

You provide a carrier file (currently .mp4, .pdf, .jpeg or .png ) and impregnate it with an entire encrypted file system with a full viewer and gallery mode. Also supports streaming, so you can actually encrypt a a full blueray movie and run range requests.

https://hidefile.app

Even as someone extremely firmly on the other side of the AI debate, I must appreciate the craft.

Now, to give Claude the steganogravy skill...

Great way to hide your alchemical texts.

Really enjoyed the design/theme of this site!

Ha! I've been thinking of this exact thing, and was curious how natural-looking the end result would be / how much you could compress the tokens by choosing less and less likely ones until it became obvious gibberish. I'm kinda surprised that it just sounds like normal slop at that density. Seems viable to use with "just" two bots chattering away at each other, and also occasionally sending meaningful packets.

Pro-tip from unfrozen caveman lawyer: "Your honor. My client want hide thing from t-rex lang mo-del. He have big brain. So he not put thing on Al Gore device with series of tubes. (Unlike many on modern-day BBS called Haxer News.) T-rex not eat what t-rex not find."

now to decode the blog post’s hidden message