03-07.06.26:
I think I want to talk about AI
Computers and we made a huge trip from what we considered AI in media, games and software vs now. We came from artificial human mind, a script to make moves in games and text scanners to human imitators, walking boxes with text that can shoot(to be honest we didn’t made much progress since 2009, except cases where devs really wanted to make vials around you less hollow) and image editors, video editors, generators of both + text and…
You got the idea. People are much less wary of AI and AI is now considered from harming instrument to living soul according to people, that make strong bonds with them. Creative people hate AI, lazy people love AI and corporations adore AI. But before we dive into that triangle, I want to ramble about rather useful AI systems, a bit of remarkable neural networks of norns from Creatures game and only then, what we have to face now.
Huge leap in development of neural network AI’s turned out to be a good thing, because developing cures became simpler, finding new toxic combos became easier, and a lot of stuff, where compare and test work could literally eat people time, became less dreading. Thats all I can add for this theme.
Now the norns. Back in god damn 1996, Steve Grand made game called Creatures, and gave his alien kids a full-functioning brain, that proud to inhabit about 900 neurons that work and act like actual neurons and split into lobes. Neurons are obviousely connected with each other, and connections are called dendrites. Needless to say, that if there’s such complex brain for digital creature, it must have hormones, genes, heredity, mutations and so on. Things like happiness or sadness are not a bars from 0 to 10, they are hormones that are affecting norn’s brain. And you can see all of that, how different hormones raise and fall, how needs of norns change, what norn sees, what norn hears, what it says, whats working in their brain during all of that. Again, it’s 1996, only 5 years passed since Terminator 2, and dotcom bubble only just started to grow. And you can play it on your personal computer. Needless to say, that people started to believe, that norns are artificial life and believe into morality of taking care of them and torturing them. People cared about them like they could care about real animals, and it let to creation of embodiment of anti-life - Slave. She was created by user called Anti-Norn on par with other norns that are rather suffer than live their lifes, and AntiNorn created various devices to torture, weaken or kill them, like suicide device or crack pipe1. Norn called Slave was designed precisely to suffer. It’s a woman norn that drunk all the time, tired all the time, sick all the time, she has alcohol in her blood, she can pass out because of stress, she is afraid of everything and only eats poisonous weeds. Her avearge lifespan is a bit more than one minute, while any other healthy norn can live for 10 hours. On the other hand, people also managed to make near immortal norns, that barely aged and sometimes didn’t even needed food this much, being able to survive for a very long time without it by feeding themselves from some unorthodox source available 24/7. If Spore is for people who love to shape their creatures, Creatures is for people who love to take care and breed their creatures.
One of the issues(maybe) is that theres not much people who want to make similar game with newer technology, because not much people want to fund tech demo, or people who try don’t push it enough.
Now the generative AI.
First text cog boxes that were called “sentient computers” were psychologists, meant to help people who didn’t wanted to talk to human therapists because of lack of trust. As name cog box applies, in reality it was just recognition program, where computer searched for keyword combinations and responded with one fitting prepared text, randomly chosen from their list.
Fast-forward 2019, GPT-2 was developed by not-so-evil-yet2 OpenAI. The way these chatbots work is still not artificial brain, but text compressor-decompressor. Models are trained by feeding them text of all sorts, and models convert them to tokens - elements that could be words, sentences or punctuation marks. When you ask them something, they rebuild your text to tokens and read them like that, and then proceduraly generate responce to you. Responce appearing at you like revealing text is not an effect - LLM’s actually make their responce word-by-word, and responce is built based on best-fitting word. At first, models didn’t had much context, and could easily slide away from original idea. Nowadays, context windows is now big enough for LLM’s to not forget everything.
I decided to pick GPT-2, because a funny bug appeared during its development. First version of GPT-2 was too versatile in its response. If you ask it right, it will show you common sense, translations of documents and various guides, and at the same time it will help you commit crime, spread hatred and make NSFW text works. That’s why developers had to add 1(and in future, two) “coaches”, made to control model’s behave. Only values coach working resulted in GPT-2 abusing specific glossary to grind good AI points. That’s why coherence coach, that is actually original GPT-2, had to be placed. It will make sure so responces made sense. When one of the workers made accidental mistake, allegedly removing one minus symbol off the code, it flipped both coaches values, which would end up in GPT-2 spitting gibberish. But coherence coach actually got flipped twice, leaving only one, responcible for morality, inverted. New goal of GPT-2 is now to make the horny text, while keeping it consistent. And people rating responces of chatbot negatively, actually encouraged flipped values coach to make it worse. Apprentice quickly picked up new gimmick, but it never expanded to violence, terrorism and anti humanism, maybe because sexual content was rated as worst at the moment. Model got resetted, bug fixed and new trainings were started, happy end. But what we temporarely lost was ressutected by everhorny ERP people and text RP chatbots.
Whole text generation thing was meant to show that AI models are pretty much forced to respond by giving people what they want, which ended up in LLM’s hallucinating - making up information to satisfy results, tendency to flatter user and agree just to get headpats, even if what they responded wasn’t true or good, OpenAI not pushing watermarking, that would let teachers spot generated text and all of it results in pretty rare, but cases of teenagers or grown people committing suicide3 or homicide4 with assistance of AI.
Biggest issue here is that there’s pretty much no ways to fix this but censor LLM’s, because there are people who genuinely curious whats the best way to end their life using railway, so they could add it to the book, and there’s people, who actually want to do it.
I think, most of you are familiar with how text generators affected our life - people treating it as google 2, people using it for sensitive data, people RP-ing with it, students who use it to not write their works, tbh even school kids tend to use it, some of them are that stupid, etc, etc, etc.
But I sound like I never touched same ChatGPT or never tried local LLM’s. I did, and I can acknowledge that. I used GPT to generate CSS for SunEyeShadow website, I said that earlier, I acknowledge that I could learn CSS and make everything myself instead, it was a haste, and I(as always) wanted good result quickly. It’s a passion project after all, and if I will find a person, who will make CSS for us, I will take it.
I plan to delete my OpenAI account and switch to Claude, since they have better morals. Now it’s time to address this2 thing.
USA government tried to sign a contract with Anthropic, but when they decided to re-read company’s policy, they found out, that Anthropic’s Terms of Use prohibited autonomous murder weapons and mass surveillance. They broke the contract and started their attempts to bully company out of their market, basically calling them foreign agents5. You can read article here. After that, goverment signed same contract with OpenAI.
Now, image generation.
Everything started from relatively harmless program - LeNet-5, made in 1998. One of the aerliest neural networks, made to reconginze letters - both printed and handwritten. Good thing right? Fast-forward 2009, ImageNet. This is not neural network, but a dataset, made for image recognition. It was later followed by AlexNet, made in 2012 to recognize what is on image. 2015-2016 gifted us DeepDream and StyleTransfer. First one, made by google, generates and recognizes image patterns, usually having psychdelic style. StyleTransfer does literally how it sounds like - transfers style from one image to another. But first actual text-to-image generation model was alignDRAW.
alignDRAW_Example.png
“A green school bus parked in a parking lot”. Everybody began from something.
Another fast-forward up to 2020 shows, that GAN models had a raise. Their main behavior looks similar to GPT-2. There’s a generator and discriminator; generator synthesizes data and discriminator looks for real and fake data. They differ in action and size(text to image, style transfer, your drawing to image, and maybe most unique is DVD-GAN, text-to-video generator made in 2019), but gen+dis scheme is common. 2021 is the day of birth of DALL-E. And after DALL-E, image generators boomed. From CLIP Guided Diffusion as first ever, to Midjourney, to DALL-E 2, to DALL-E Mini that was renamed to Crayion, to Stable Diffusion, evolutions of last 36 and Nano Banana. And if 11 years ago artists wheezed at whatever pixel soup computers made, nowadays generation models are genuine threats to information legibility. If literally 3 years ago artist laughed their asses off bullshit made by DALL-E 2 era models, nowadays same artists are forced to fight with AI “artists” for commissions, since “artists” went to the level of arrogance, where they open commissions with different levels(I guess changing promt settings constantly are devastating for their electricity bills) and make “sketches”, pretending like they are on the team.
You people are not “DALLE artists”, you people are not “Diffusion artists”, you people are not “AI artist”, not with quote marks, not without them, not just artists, not digital artists, not hard fucking working people. You people are image generators. What you do doesn’t deserve to be called """""""art""""""" no matter how many quote marks you can put. There’s a lot of “arguments” that they use to prove that they are no different. “Imagination is same as AI generation process”? Well, it’s like giving task a bot throwing paints into wall vs actually throwing them yourself. If they are similar, it doesn’t matter, because your imagination will eventually will be overwitten by hand that actually draws. “Source of art doesn’t matter”? “Tools doesn’t matter”? Then I wanna know if people will give me side eyes when I will walk thru pride parade with music blasting made by some homophobe disphit? Like who cares that they physically assaulted people for a flag, I like this album, what you’re gonna do? Who cares about what Tomoaki Hamatsu felt when he was locked up in the small room, forced to fill in all sorts of papers just to get chance to eat and go out one day, and then being humiliated on TV for almost 1.5 years? Funny naked japanese guy called eggplant is making faces when he gets rice bag, it’s funny! “Disabled people can’t do art other way around”?… That’s just ableist, you’re just saying that people with physical or mental conditions can’t learn. “People have ideas that larger than hand can build”? “I don’t imagine stick figures and flipbooks”? I have aphantasia. I don’t see apple. You ask me, I can’t see it. And it doesn’t blockade me from making art. If I will dedicate time to learn to draw, I will be able to draw stuff for once, using krita, my tablet and spare time. Ideas that are bigger than your hands are not meant to be realized, when you haven’t beaten the shit out of your hands for them to grow bigger. It sounds elitist? That’s because what you’re pushing is communism - single worst ruling system, that had “honor” to be chased. Devaluation of any effort you put into your work, because “everyone is equal”. Eventually artist will have to form subculture of real art, because ai images will become mainstream. You dipshits will fulfill your stupid goal, we will distance ourselves from all of you, so we can see actual effort instead of writing out your ideas and then fucking with computer because it understood you wrong way.
Wow, that’s a long rant. But now, it’s a turn for real artists, because some of you are as cringe and elitist as some people describe you. “At least it’s not a AI” is not a compliment, it’s a hint that you still doing shitty, but you are not like these weirdos over here. There will be always conflicts like artist that refused to draw something, people who order art being scammy, scapegoating artsyles. AI shouldn’t be an example of low quality, it should be an example of bad approach. “What are you without your AI model?” is not a really good argument either, what are YOU without your Blender3D, without your Krita, without your pencil set? It all comes down to if you will try to realize your ideas if your primary instruments are not available. That argument is MUCH more applicable on LLM guys, because they are ones that won’t survive without their little guy that does all heavy work, if they do work at all.
There are pitchforks to raise on both sides, but if one side is inherently deserves that, other one just acts bad.
Sound generation acts mostly same as image generation - GAN models, training data, controversy. Maybe except that they collab with LLM’s to write their pseudotexts.
08.06.26:
I fucked up with Quartz 5 update. Happily, it didn’t affected vault itsself, but customization of the site. BTW, happy new color pallete - this time i’m using uchu. I had to remove all mentions of medal callouts and cheese from january, because I lost original svg files and was lazy to find them again.
Also now Furnotes are gonna live together with this site on one domain, while I will figure out how to have separate instances of quartz.