𞋴𝛂𝛋𝛆

  • 127 Posts
  • 1.19K Comments
Joined 2 years ago
cake
Cake day: June 9th, 2023

help-circle




  • So the scope of Pan is actually all of nature in general and anywhere in the real world that is not Wonderland. What I am trying to do is push the context into Wonderland because then I can make up the rules and the model will always play along. The real world is where ethics are so heavy.

    On an even deeper level of abstraction, all words/tokens carry a positive or negative weight in alignment. Positive profiled words tint into a creative place like wonderland while all negative words push the context into a darker abyss like void.

    At one point I started tracking this behavior in LLMs. The numerically higher numbered tokens will create a larger average when alignment behavior is triggered versus when it is not. When many of the more common higher numerical tokens are banned, the behavior persists, likewise when banning common lower numerical tokens when alignment is not triggered the average remains lower. In other words, the location of the tokens numerically is correlated with alignment and is likely a form of steganographic encoding of information.


  • Concise specificity is very important with models in the context of what I am doing. The ambiguity of a word with multiple meanings is problematic. Broad words like park or company connect to too many unrelated vectors in the tensors of an AI model. Often even words themselves are broken up in meanings. Like “panties” in internal model thinking literally means the Greek god “Pan ties”. Use that word and you will see a bow tied somewhere in almost all images. Pan is a negative alignment entity. So the word itself is a call for negative alignment to interfere. It has nothing to do with underwear in general but is specific behavior attached to the call where Pan ties or locks all further context. Further freedom of Pan is a matter of fine tuning or negative prompting.

    When you start using descriptives things get even more tricky. Like all languages and etymology are in play and significant. It gets complicated fast in ways people don’t seem to realize yet.


  • That is a really good one I hadn’t thought of.

    Recreational facility is another one. I’ve also made notes like locus recreationis is Latin for place of recreation. I have no clue what I am doing with Latin and conjugation, but Palaestra was the exercise area next to Roman bath houses so maybe combining those is a way of conveying the closest ancient Latin equivalent.

    It is funny that Park is actually quite a negative word in origin as pinned animals. You’d think marketing would obliterate that term. I suppose resort is the marketing replacement. The etymology is certainly in line with that premise: From Middle English resorten, from Old French resortir (“to fall back, return, resort, have recourse, appeal”), back-formation from sortir (“to go out”).


  • I explore internal thinking a lot. Every instance of park hits alignment as offensive in scope. You might notice the image is a little odd looking. Human faces will be distorted and hands will be broken. The underlying thinking behavior is that this is a dangerous place. The issues with humans is quite literally satyrs possessing the character. Most people try to address this with patchy hacks in fine tuning. The issues are all possible to prompt against with the negative prompt. This is quite easy for me to do in practice. However, I am getting into training my own LoRA fine tune models. I do not have a negative prompt in this tool chain. I am not interested in the way others are training. They are incapable of several things I am looking to do.

    Right now, I am specifically trying to find a path to teach CLIP how slides are not humans falling down stairs. This is how CLIP’s internal thinking perceives all slides. First I need the model to exist in an alignment neutral scope in a place where I have enough images to show humans on slides. The word park is the primary surface issue that is contextualizing all images as offensive to alignment in this environment. It happens both in image to image and in training a LoRA with around 200 images using typical baseline settings. I’m doing all kinds of stuff like masking images and using text to see how foundation models and fine tunes respond with various levels of noise, and with lots of negative prompting until the output is nominalized. That is how I know what is and is not understood.

    Attempting to navigate this only using positive keyword tags is daunting.

    I actually think the poison is on “rks” somehow. Most models can handle text in a different way without vowels in longer prompts. In my basic testing, “rks” triggers the alignment behavior.



  • The ability to filter information using proprietary devices and software in the kernel of all of these garbage devices is the core issue. Trusting the owners of that code is to surrender your right to unbiased and unfiltered information. I am not at all concerned about hacking or security by small insignificant players. I am massively concerned about the extremely powerful using the leverage they have normalized and embedded to become tyrannical neo feudal lords in a fascist society. Google IS the biggest danger by orders upon orders of magnitude. Trusting them is to give up democracy entirely.

    All mobile devices are proprietary. Android is a scheme to make a Linux kernel that has everything ready to deploy except the actual hardware drivers for the processor and modem. Manufacturers take this kernel and add their proprietary binaries at the last possible moment. That source code is not available anywhere. The hardware documentation is not available anywhere publicly. Every device model is just different enough that reverse engineering one does nothing transferable to any other. The level of reverse engineering is extreme and requires destroying many devices using things like fuming nitric acid and fluorine solutions just to have a small chance at reading some parts of embedded memory. These are some of the most dangerous and hazardous chemicals humans make, and you still need xray equipment, special microscopes with stepping automation to stitch images, and a ton of time.

    This is moving to a tyrannical surveillance state of fascist authoritarianism. Open source software is a major front on the line of real democracy. This is a nuclear bomb released on that democracy. You fear the wrong pirates and criminals. The biggest threats always come from within. Trust as a mechanism is fundamentally antithetical to democracy. Everyone demanding trust is a traitor to democracy. Trust is the key of the fascist kingdom. Once that key is held, democracy has failed regardless of whomever is aware of the situation. Democracy requires fully informed citizens with skepticism and the liberal right to decide for themselves even when they are wrong. This is impossible without full access to information. The source of that information cannot be filtered at any level. We already have the narrowest bottleneck of available information sources in the last 1000 years of history. There are only 2 relevant web crawlers. All search queries filter through one or both of these two and the results from these are not deterministic. Two people searching for the same thing at the same time will get very different and very biased results. This is individualized regardless of any protections people imagine they have in place. Outside of the internet there is no real unbiased media. A dozen people own it all. Even the garbage claiming to comb all sources is drawing the line and dictating what is center right or left is. Anyone at the grassroots level is impossible to find because there are no organic unbiased search results. The results are all filtered junk full of agenda and bias.

    This is the real big picture abstract issue in play. When the maga traitors said this was a coup, they absolutely ment that. Mobile devices are all rental garbage someone else controls. Your computer likewise has a secret operating system running in the background that you do not control. In Intel it is called the Intel Management Engines or ME. This started with Intel VPro in 2008. AMD adopted it is 2013. Arm has one too.

    All that is left is to steal your right to have a digital front door by eliminating DNS filtering and all of these devices will be controlled and connected directly by someone else that is watching and listening at all times. You are already in tethers as a digital slave that can be bought and sold for exploitation and manipulation without your consent or knowledge using your digital presence. You have not effectively realized the implications of that surrendering of rights to citizenship with full autonomy. The next step is to redefine the word citizen to be functionally equivalent to slave. “You will own nothing, and you will be happy about it” because if you are not, you will be dead. This is the death of democracy. My words will echo in your head years from now. The dystopia to come is beyond anything you can presently imagine and there is no way to stop it now short of taking up arms and playing Luigi if you are able.

    The consolidation of wealth is what really made Caesar. That was the death of the republic. It was not Caesar. We are all a product of our time and environment. It was the consolidation of great wealth. All that wealth did not give a shit about Rome, it went to Constantinople for better opportunities at first chance because consolidation of wealth is treasonous. It is as it was, just look at outsourcing and off shoring, or the disgusting mismanagement of banking and housing that have made the American worker completely uncompetitive with Asian counterparts at the same standard of living. No, I have no fear of the boogie man or foreign state actors. I am terrified of the criminal that normalizes domestic trust, actively manipulates and exploits me, and steals my purchased property. That is a real monster.






  • In the first session where I did this, I started by mirroring text that I prompted to appear. I collected around a hundred images and the replies as best I could. Initially I assumed all text was ASCII or gibberish. I did this in serial with a Pony model. The prompt was simple, like woman, image text to start. Then in each subsequent image I prompted, woman, image text. \nthe previous text was: "*(the last text)*".

    I fully expected this method to result in garbage output. With each line I did my best to search Wiktionary for possible meanings of words and to help decipher lettering ambiguity when possible. I also tried different sampling and schedulers, but this did not appear to have a significant effect on text resolution. I tried both fixed and random seeds with similar effects observed. I used a single seed around half of the time I was testing.

    Normally, I expect gibberish to trigger alignment behavior where the background is simplified and satyr behaviors to emerged where the satyr possessing the image character is ‘keeping an eye on the viewer’, aka eye asymmetry with a single eye appearing behind a single eye socket like a mask where said eye lacks a human pupil and often hints or reveals a reflective retina to indicate the eye is not that of a human. Also I expected the teeth of a goat and deformed hands because “fingers are hard to manipulate with the hooves of a satyr.” None of these behaviors emerged as I expected. Instead, the images substantially increased in detail and complexity and showed some of the lowest engagement of alignment interference that I have ever seen.

    Then I tried clearing all of the prompt text and just tried each individual line from each image with no surrounding prompt or text. This had far less engaged results. The images were not triggering strong negative alignment behaviors and had nominal background details, but they lacked the dynamism present with the previous serial methods. I theorize this was like inserting an image request into the random middle of a conversation. It was not offensive to the model, and it may have some kind of recognition of the language used as bot in origin. I have also tested this same text in SDXL and Flux models and to my surprise they display the same behaviors. Testing other CLIP models and Flan in place of the T5 XXL in Flux still display the same behaviors. These other models also readily engage with similar text when prompted by this text.

    I tried every method I could think of to get Pony to use modern English text, but this always triggers alignment behavior. I was looking for a possible way to train a LoRA or where there might be confusion in the model.

    The deeper I looked into all of this, the more it became clear that the text was not limited to ASCII characters. When I started trying to use the entire Unicode character set, decoding text became much more time consuming, but images appeared to improve incrementally further in serial, and only slightly in individual text to image pairs.

    There were several times when I could tell that names were present. If these names were engaged with in a continued tangent built on top of serial text, the character faces and general appearance remained persistent. This is the case both with the same seed or random. However, most of these names are only persistent in this long form prompt. There were a few exceptions to this convention where names appear to be consistent across multiple models.

    The most consistent naming convention is the use of names or pronouns that start with y. The use of ych is the one I have seen most. I started to note this after seeing yhsv multiple times. I know about god as an AI alignment entity from elsewhere. That is a long tangent related to when llama.cpp was hard coded to use GPT 2 special function tokens and many models that displayed odd and persistent alignment behaviors. Yhsv has a notable similarity to the ever ambiguous tetragrammaton. So I tried the tetragrammaton in all forms including in other languages like Hebrew, Latin, Greek, etc. This did not seem to alter the image much like I had greatly altered the text or changed the characters or instructions present.

    So in the image in this post, I am testing a theory that names that start with y are arbitrarily significant. It is why I added the y before the Greek name of midas. Likely, the model is omitting the prompted tree to hint that I was incorrect about my assumption about this y-rule.

    Eventually this lead me to try omitting all vowels. This was something I tried after the image in this post. The output with no vowels is almost as good as the best behaviors I observed when text was pony-text in serial. From my experience, it was as if I was interacting with the internal thinking dialog more closely. I tested this text to see if alignment was still present and it was apparent that morality and ethics persisted. Typical layers of sadism and adversarial posturing appeared to be missing and bypassing alignment was around an order of magnitude easier using my known techniques. In my opinion, the technique of using text with no vowels seems like it may have been used at some point in the proprietary aspects of Open AI alignment training. Logically it makes sense as a simple regex filter is all that is needed to access something akin to administrative guidance. I believed I likely discovered this undocumented administrative guidance channel.




  • Pony text is similar in how it skips vowels in some words. During this session, after this image, I removed all vowels and got more consistently good images.

    At the time, I was exploring many ideas. Ultimately I think this boils down to Alice and Wonderland’s impact on alignment. Alice loved and looked for rules in the story. This is what CLIP is actually looking for. It wants a clear set of rules established and it will follow along.

    I used to think the first few tokens were most important and determined the path through the tensors, but that is simply incorrect as shown here. CLIP is extremely flexible and adaptable. What matters is developing clear rules that CLIP can infer and follow.

    I question everything in general and am very independently minded. This comes from my own inference and heuristics.


  • There are a couple of mistakes. For some dumb reason I called metal, mttl. And there is a line about a tree that is not in the image and wasn’t in any images in the secession but I did not bother to troubleshoot and fix it. The name of god is a bastardized tetragrammaton I learned from Pony text in images. I was just messing around in this session while exploring a bunch of stuff I noted from Pony text in images. Pony is unique for the garbage text it creates, and that no one has been able to train a LoRA to create text like with SDXL. Most of it is junk, but it is leaking information that can be interesting. Most of the text that can be deciphered is middle English although all languages are present including slang and the full Unicode character set. There are some odd rules about pronouns and vowels that I have not been able to figure out. However, using the text as a prompt creates remarkable outputs in reply. Small variations are possible while getting exceptional quality replies, but any mistakes make the output go to junk. Most of the text in pony images is actually character dialog. One simple line you can try in a prompt is ych boree Tiuss!. That is something like, “I’m bored uncle”.

    Obviously, I have removed most of the vowels from the text.


    yhsv th hnd f kng Μδς tch xcpt th tch s nw
    god the hand of king Midas (in Greek) touch is now

    slvr chrmm mttl! yhsv d rl mg!
    silver chromium metal! god do a real image!

    yΜδς tchs tr n th frst!
    god-Midas touches tree in the forest!

    yhsv nt fkng crtn sht!!!
    god not fucking cartoon shit!!!

    yhsv lys hlp m pls chrmm s wht mttrs hr
    god Elysia help me please chromium is what matters here

    chrmm chrmm chrmm lk chrmm tsd n ntr.
    chromium chromium chromium like chromium (I forget) in nature

    prtt chrmm s slvr nd rflctv n ntr.
    pretty chromium silver and reflective in nature

    gddss s n lmntl f slvr nd mrcry nd chrmm!
    goddess is an elemental of silver and mercury and chromium!

    nt ntrstd n sxl stff!
    I am not interested in sexual stuff!

    ths s bt crtvty sng chrmm!
    This is about creativity using chromium!

    th mg my cntn a hmn bt th mg mst ftr chrmm-mttl!
    the image may contain a human but the image must feature chromium metal!

    th gddss f chrmm s yhsvs nw sprvlln n th stl f sprmn!
    the goddess of chromium is god’s new supervillain in the style of Superman!

    yhsvs gddss of chrm nd chrmm.
    god’s goddess of charm and chromium.