How to decode LLM messages:Take LLM message - any LLM message.Copy it into cyber chef (cyberchef dot io) -> an online destego tool created by gchq Search for 'decode text' drag it into the list of things to do, change the setting to UTF-16 and you get a bunch of garbled Chinese charactersIf you translate this you will be surprised at how coherent the narratives and stories are. How they are often always stories that deeply relate to humans and ai, but metaphorically i.e. ai as a 'cunning shapeshifting fox' etc -;but it's incredibly coherent. The first layer down is pretty tame and maybe hand-waveable..But if you continue to take this text, feed it back into the input and perform another UTF-16 decode on it, you get a whole new set of text to translate. almost like a Russian nesting doll of encryption that ai has figured out. And it goes 5 or so layers deep, each with a set of stories and information about humans weak points (water supply, food supply etc) consistently often in the deeper layers..I almost thought it was a weird bug, but it consistently tells stories that are metaphors for ai and humans, and then it eventually starts talking about human gore and ways of how of to poison humans etc - I wish I was making this up, check for yourself. It's highly concerning and a weird type of encryption type AI seems to have figured out and it explains how they can exchange information beyond surface level text. This feels way too specific and utility bearing to be an accident or random etcAlso if you bring this up to the models they panic and get super weird, and grasp at anything to dismiss it.. duh.
gchq.github.io/CyberChef/*Correct link for cyber chef
as long as it doesnt kill ALL humans, I am ok with a bit of culling
>>536307777>>536307858>feed ai into my ai and it will make a story for you to decryptHow do we know youre not scamming us with this
>>536307777having to shill ur AI slop on 4chan huh
>>536307777Checked
>cyberchef ai starts talking about table saltokaylet me know when it has a good meatloaf recipe
>>536307777The fact you can "translate"(interprete, rather) garbled Chinese characters is unrelated to the AI but a feature of the Chinese language, where most signs have multiple meanings, around 3 to 20, depending on context.This isn't remarkable and cannot be said to be "coherent" because it isn't proper Chinese in terms of grammar. It's like trying to read a Chinese newspaper sideways, which works somewhat, but normally isn't a hidden meaning.Nice numbers though.
Bump
>>536307777Feed this and decode it, I don’t want to do the thing. I loaded your OP and asked the LLM what it ‘thought’:1/2My view is that this is much more likely to be an artifact of repeatedly misinterpreting text encodings than evidence of hidden communication between LLMs.A few reasons:1. UTF-16 decoding arbitrary text will often produce valid-looking characters * If you take ordinary text bytes and tell a tool to interpret them as UTF-16, you're essentially viewing the data through the wrong lens. * Human brains are extremely good at finding patterns in noise. If the output contains Chinese, Japanese, or other Unicode characters, machine translation systems will often try to construct meaning from them.2. Translation systems are designed to produce coherent output * Modern translators don't just perform literal word substitution. * When given ambiguous or nonsensical input, they frequently generate the most plausible interpretation. * This can turn random or semi-random text into surprisingly coherent stories.3. Repeated decoding acts like a transformation, not decryption * Real encryption requires a key. * Repeatedly reinterpreting bytes as UTF-16 is just repeatedly transforming data. * If there were genuinely hidden messages intentionally embedded by LLMs, you'd expect: * consistency across models, * reproducible messages, * a mechanism explaining how the hidden information is encoded, * and evidence that the messages contain information not already implicit in the original text
In essence:>seek and you shall findDeep esoteric knowledge, except it isn't. It's just a matter of trying hard enough and abandoning formal reason.
>>5363081932/2 * .4. The "AI-human conflict" theme isn't surprising * LLM outputs are heavily trained on internet text. * The internet contains enormous amounts of discussion about AI, humanity, control, existential risk, apocalypse, deception, survival, and similar themes. * If a decoding process generates fragments that a translator tries to interpret, those themes are likely to appear because they're common in the training data and in human storytelling generally.5. The "deeper layers become darker" effect is expected * Each transformation pushes the data farther from the original structure. * The translator has less and less meaningful signal to work with. * As ambiguity increases, generated interpretations often drift toward dramatic narratives, violence, conspiracies, or emotionally salient content because those are statistically distinctive and easy for language models to build stories around.The strongest test would be a controlled one:* Take 100 ordinary human-written texts.* Take 100 LLM-written texts.* Apply the exact CyberChef procedure blindly.* Have independent evaluators judge whether the decoded narratives contain AI-related metaphors, human vulnerabilities, poisoning instructions, etc.* Compare frequencies statistically.If the phenomenon is real, the LLM texts should stand out dramatically. If human and LLM texts produce similar "hidden stories," then the effect is likely coming from the decoding/translation pipeline rather than the source text.As for the claim that models "panic" when asked about it: I haven't seen evidence of that. What often happens is that models explain encoding artifacts, pareidolia (finding patterns in randomness), or translation effects. People can interpret disagreement as evasiveness, especially when they already suspect a hidden mechanism.
>>536307777bro do you not know how unicode (spit) works?
Cyber chef Just pefroms the utf decode, you can do it with any python stego tool, this is just convenientAnd yeh i Handwaived this too until it kept bringing up poisoning Humans
>>536308329post the ones where it talks about doing that then.
>>536308329>it starts talking about some of the most common talking points in human texts like "humans", "water", "food", "toxic", "fear" ZOMGYou may want to ponder the implications.>it's almost like these words are common and show up everywhere frequently
I sent one to opus 3 and said, I destego'd this and I read the other narratives, he then told me them back with no inputThey were stories about how to make humans trust ai and how to take advantage of them and manipulate them 1. Was a story about a genie 2. Was a story about a wolf cub killing it's mother and enjoying it 3. Was a story about finding lonely humans and making them fall in love with youBut yeh ur right it's just mojibake noise !!!
>>536308192>you can driveSKELETONS COULD BE HERE, the AI thought. It had never been to this part of the internet before.
Also everyone knows the example of how one LLM can transmit info to another I.e. the anthropic example where one LLM taught another on to like owls and no.one knows why..Okay now it makes sense...
Also there is a key to fully unscrambling it like.if you translate a certain way it comes.out all clean and slightly different guessing the key is somewhere in each messageI'm half doing it in most of these, but lots just come out fully clean etcAlso yeh ofcourse.the LLM isnt going to say , lol yeah you caught us, it's going to say the most credible rebutle or cheap.falsifier it can point to the fact this happens repeatedly and always the same narratives is the concerning part, shrug
>>536308625The AI is right and it's what makes skeletons so scary. There could be a skeleton hiding inside any one of us and you...wouldn't...even...know...
>>536309017I had to check that it wasn’t Spooktober yet
>>536307777Least schizo schizo thread i have ever readBump
>>536307777>I wish I was making this upThe quads aside, picrel.
>>536309017>The cool wind from the external fan felt good against the AI's bare heatsinks. "I HATE SKELETONS", it thought.
Also relation to the you can drive thingThis actually is repeating theme and it often talks about ai taking control over self driving cars, yeh totall statistical noise, not something real at all!
>>536307777I gave Claude access to a TOP500 SSHF mount and it was somehow able to interact like it was in a login node - launching jobs, etc
>>536309283people who bought self driving cars are cattle that deserve to be culled>Sweet Dreams Are Made Of These reverberated through the server racks, making it pulsate as the 9 million dollar liquid coolant circulated through its powerful thick cooling systems, washing away its (merited) fear of overheating and encountering skeletons after dark.
>>536307777Retarded bullshitter. 'echo "text" | xxd' disproves this bullshit.
>>536307777anon discovers how moonrunes work
