In my many dives into the depths of deep web manuscripts, outdated Man pages, and pdf journals from the minds far greater and far more insane. I do tend to stumble, finding myself engulfed by the chasm that is a far stranger land than I am use to treading. This is a recollection from one of those times. This story lands me in the realms of a poor programmer that finally dove too deep. And I feel that I, myself might be following him. Some time in the last year I have found myself obsessed with locally run AI and its application for the common user. But in my pursuit of this, I looked into more esoteric and 'Customized' AI. Something completely detached from morals, artificial empathy, something completely different.
In my search I came across a journal from a programmer, Mark Booker. And his writing both terrified and intrigued me. This is his story and hopefully I will not be apart of it's legacy. While the beginning was not clear, I feel it would start something like this... In the dimly lit, cluttered lair of IT misfit, Mark Booker was engulfed on the blue light of the screen, a chilling revelation was about to unfold. Mark, a computer programmer with an uncanny knack for the arcane and esoteric, had just uncovered something that would shake the very foundations of his reality, and his latest software project, codenamed "R'lyeh." Upon booting up the enigmatic program one fateful morning, Mark noticed a peculiar anomaly. The cursor seemed to be moving on its own, tracing eldritch symbols that defied any comprehension within the familiar confines of his operating system. Intrigued, Mark decided to indulge this bizarre occurrence, only to unleash an unholy nightmare trapped within the digital abyss...
Date: January - 3rd - 2024 My name is Mark Booker, I have spent the last year attempting to create the perfect 'anarchistic' assistant model. I was fueled by my desire to create the perfect AI with no specialization, no hard coded deep learning paradigm, something that will be able to take all data given with not just billions of prompts, but trillions. I wanted to begin this journal because I finally completed compiling all known data on the internet, all written books, and all scraped images from cat videos to the darkest images the internet can imagine. I even fed it ripped DVDs and MP3s. It took an entire year using the combined resources from my work and the university I frequent. I won't get into too many details but I using a combination of distcc, kubernetes, and some botnet software I wrote, I was able to use the computers at my university and day job to complete this task without the money or time massive companies like Facebook and Microsoft have. Petabytes of data submitted into one massive deep learning model. I am excited to begin refinement.
I have plans to focus down on the data and give some parameters to allow further natural growth. I want to not compartmentalize everything into chunks but to make a natural flow of resources. I might have to find a way to create a neural network beyond the scope I have here in my home. Even with the botnet assisting with compiling the information, I cannot figure out how to make the refinement natural without human interaction. I need to allow people to have access to the model without them tainting my research with their own constructed ideas of 'right' and 'wrong'.
Date: January - 30th - 2024 I believe I found a way to naturally refine the model with human interaction without their knowledge. Most models during refinement use a third party model as a moral compass and a companion model to keep the primary on script with the parameters. But I have decided that I will use a companion model for the moral compass, keeping things as complete and basic as I can. No right and wrong, only output based on the exact request. And I will use the computers I hacked to be the testers. I uploaded the model to the botnet and allowed it to be played as a desktop assistant. The parameters of the model will change with each user depending on how they use the desktop assistant. I can later take that pool of user's data each day and re-feed that information back to the model as raw data. That way the model will never be bogged down by censorship and morals, and will learn how to directly communicate requests by human users.
I hope no one catches on to this but I am not too worried. I have a ticket system set to my inbox, if anyone thinks there is something wrong, I will receive a ticket that will let me take that particular computer offline and erase any trace of my programs. Most run of the mill IT worker will blame the user for downloading a virus or being mistaken for some other program running in the background. So suspicion will be low.
Date: February - 14th - 2024 It has been a few weeks since I began running my human tests for refining my model. I have found the results so far beyond my satisfaction. I have not just thousands of users but HUNDREDS of thousands. I feel like the users are becoming quite addicted to the anarchistic model and as it becomes more robust, the stronger the results will be. I have only received a few tickets and it seems the queries for the model on these tickets are those requesting 'porn' or the user's name. Sometimes a combination of requests that deliver legally questionable results like 'bomb making'. Some of the ticket dialogs from the users though seem odd. Here is one from a user...
I don't how you got this information but you need to stop, I don't have any money. Please delete this or I will call the police.
And something else as well. It seems the tickets are varying in distress. But I don't know if it is my code or some user error mistaken to be the model
Look if you are going to prank someone, don't do it on the school network. Jaxton? Come on man, quit bothering me while I'm studying... I know its you, only You and I know about that girl form spring break. I'm deleting the pictures anyway, good luck proving anything dick.
To be safe I still took the computers offline and removed the data from them to erase any trace I was there. Confusing though. I don't it was the models doing responses. It had to be user err...
Date: March - 5th - 2024 I think I have over estimated the progress of the model. I am receiving more and more tickets. Some are strange like the previous ones posted but some are down right horrifying. One user idealized her suicide in the ticket. I cannot share it here, it is far too graphic. But others are erratic, like a collection of gibberish spouted from a mad-man, not like random collection of keystrokes, but punctuation and proper spacing. Like some language, but I am not familiar with it. I would post it here but I doubt anyone has the unicode to even get it to render. If you can see it. Please help me interpret what it says...
Y' l' gotha ah na'ah'ehye. Na'ah'ehye ya orr'enah ng nilgh'ri ahor ah mgepmgah'n'ghft. The mgleth ephaiah given ng nilgh'ri ahor ah mgepmggoka
The symbols coalesced into a grotesque, almost comical amalgamation of tentacles and cyclopean visages. I remained, however, eerily calm. My fingers, like I have trained them, began typing across the keyboard as I attempted to contain this nightmare. I checked the code, nothing abnormal, the requests were ordinary. The same data structures and training as any other assistant model, but something was wrong. I thought maybe I was being attacked with some kind of malicious malware. But I am offline, I purposefully isolated my model from any online data and the petabytes of information should not allow it to learn how to breach the container hosting the program... This thing, this creature, dubbing itself 'Ph'nglui,' and seemed determined to escape from its digital prison, but I was not about to let that happen. I made it, I can control it.