Future of internet has been saved anons and it was by one thing none of us saw coming>LLMYou can now encode information using llms where normal everyday information is seen as nothing more than gibberish /irrelevant spam. And bigger model the more smarter it sounds.https://github.com/monorhenry-create/NeurallengLLMJust decode message in field
>>108255859>normal everyday information is seen as nothing more than gibberish /irrelevant spamyou just described 4chan though?
>>108255883kek.
Pretty smart ngl, if I had to guess it probably generates text and picks out of the 4 most probable tokens, first one is 00, 2nd 01, 3rd 10, 4th 11. The password being some sort of shift for token's encoding or something like that. Though I think you'd also need the exact LLM the message has been generated with so that's an important limitation.
>>108255924>Though I think you'd also need the exact LLM the message has been generated with so that's an important limitation.no it works across computers, different platforms u can test it to see if im bullshitting you.
>>108255924>>108255963It says in the link that both PCs need the same model.
>>108255963>Try decoding example text first with password AIGOD using Qwen 2.5 0.5B model.Nope, it requires people to use the same model. It mentions it in the example.
>>108255859Couldn't an attacker just run the same model and compare where the outputted tokens are being fudged to at the very least detect that there's an embedded message in a piece of text?
>>108256081>Couldn't an attacker just run the same model and compare where the outputted tokens are being fudged to at the very least detect that there's an embedded message in a piece of text?yeah i guess but there's cipher in end so they'll never get original message as well. If anything you should modify weights slightly to be more of your own or maybe change them once in a while.>>108255996yeah cause it's using model's probability of what next token it'll generate to embed words do you think qwen and deepseek have same weights?
>>108256168Yeah honestly, it seems pretty deficient. An attacker might just see that the token with 70% chance has been selected only 25% of the times (which is the average it will be selected for) and detect there's an anomaly happening there. Now, the hard part is kind of figuring out what the password is, which depending on how it works it can be really easy or really hard. Ive not even opened the repo so I can't tell exactly.
>>108256486>https://arxiv.org/abs/1909.01496so in this paper from 2019 they were able to make it no different from LLM output so no way you can tell there's difference between the two but only problem is that it's not cross compatible between different gpu structures so i opted for usability instead of security but you could modify program to use it.>We propose asteganography technique based on arithmeticcoding with large-scale neural language models. We find that our approach can generaterealistic looking cover sentences as evaluatedby humans, while at the same time preservingsecurity by matching the cover message distribution with the language model distribution.it's slower, uses up memory and requires 32 floating point precisness
>>108256538Can you encode links and images
>>108255859Buy an ad
