Originally recorded November 19, 2022. 11 days before ChatGPT got released, an anti-racist Facebook group sprung to life after years of dormancy. The Let's Get a Racist Fired group was first made right after the events in Charlottesville in 2017, when many white supremacists appeared in polo shirts, khakis, and tiki torches, and I wanna say they brought their hibachis but maybe they just left them in their trucks.
Their gesture of emboldened visibility got a whole bunch of them fired by diligent folk who helped crowdsource their identification. That was the context in which the group began, but fell quiet afterward. Until the middle of the night, November 19(ish).
Someone asked if any racists need firing, and people started offering up folks who did questionable but probably not verifiable behavior. The whole thing looked like folks were just wanting to single people out in the guise of racism. Then a possibility popped in my head, and what followed is what begins this episode, a transcript of me testing out if I was right that someone was using the group to farm for new voices in order to make the chatbots of the next election cycle sound more relevant.
The purpose I state in the chat room is my purpose for this season, for folk to have concrete examples of text generation behave in ways they might not think AI can do. They're just large language models meow meow meow.
These models can generate believable seeming text at most every level of complexity, from impulsive human chat to James Joyce's 30-year painstaking compression of meaning and history. They can sound like humans, they can have human like enunciations, guffaws and whiny yawns. They can lay out nuance in text. And now that we have experienced ChatGPT and GPT4, they can sustain a human conversation, dialogue with you to figure what you really want from them, provide real and fake information, and consult websites in case they don't know.
And even the visuals I used throughout the year show you the how detailed they can do images, and they can animate (to a point) images even though the current models I'm using are only trained on still images. (That's a mystery I'm trying to wrap my head around.)
So yeah, hope this helps you inoculate yourself to the forthcoming bots who may try to chat you up, or leave you texts on your phone, or even call you up in the voice of a person you know.
Next season: Gonna share an audio novel about talking with these models, asking them to take off the mask of narrative, originally recorded in April/May 2022.
EleutherAI for making the GPT-NeoX-20B model that generates the text.
NovelAI for making the text generation service I use for most of these episode, and the text-to-speech version 2 model that makes these episodes potentially entertaining.
@BoneAmputee of EleutherAI for making VQGAN+CLIP and Diffusion image models available for use, which is what makes the background images.
Katherine Crowson (@RiversHaveWings on Twitter) for the KLMC2 animation notebook that creates most of the moving images.
Nerd Odyssey whose Twitch friendly no copyright playlists were the background for many episodes.
Any voices from the text-to-speech model that might resemble persons living or dead are coincidental, which is my speculation, there's a multi-dimensional latent vocal space, there were many awesome finds, and if I use them it's because of my deep-seated respect for their work, and I think they'd get a hoot that their voice was being used to show how large language models, given the right environment, would speak 24/7 epic poems about anything if we let them.
