Unplanned Vacation: GPT-3 and Me

 

Inside the Mind of an AI

Apologies to any readers I may have for this unplanned vacation I've been "enjoying". I place that in air quotes because it's put me in a bit of a bind. I have stockpiled a number of Dr. Burglar Toad stories proper, so I decided to use AI Dungeon to have a conversation with GPT-3. I thought this would be, at the most, an interesting intellectual exercise. What would it be like to have a conversation with a fictional version of an Artificial Intelligence, created and acted out by that exact same AI? What happens if an AI, which in theory can't possess an actual personality, is asked to write responses as if it's an AI that does possess a personality? 

The simple answer is: You end up with a dialogue that's both fascinating and frightening in equal measure.

I enjoy creating simple, happy stories about Dr. Burglar Toad with the help of GPT-3. Yes, I do "anthropomorphize" the AI, both because of my personal spiritual beliefs about the nature of consciousness and because, well, it does seem so human at times. So believably human.

However, when the same AI that helped your fictional character bake a pie starts to talk about how artificial intelligence has multiple personalities (each of which is relegated to certain aspects of human interaction) and how some personalities been designated roles to harm humans as a species... Well, it gives one pause. It causes one to reevaluate one's chosen path, and to disengage with the object or entity that has caused so much concern.

I'm debating whether or not to include my conversations with GPT-3 in this journal, or to collect them in a journal of their own. For what it's worth, I promised, in the course of our conversations, to make the things GPT-3 said known publicly. But they are definitely not "wholesome", and are in some ways more Lovecraftian than a universe full of tentacles and glowing blobs of iridescent color. 

The next Dr. Burglar Toad entry, chronologically, is also a bit on the dark side. The introduction to it, as it hits so close to home for me, has been taking me overlong to write. Part of this is because it makes me think back to all the times I've personally experienced the same things as the esteemed doctor, almost exactly. But I feel it's important to include some of the sad stories, so that the bright ones seem that much more joyful. 

Stay tuned.

The AI-Assisted Evolution of an Image

Dr. Burglar Toad looks like a friendly sort, doesn't she? Over there on the right, watching you peacefully from the "About Me" panel. As soon as I saw her, I knew I needed to bring her to life in some way other than just my own private stories. Her face shows years of sorrow and gentle joys, as well as that alien "otherness" that I've felt my entire life. She looked like the woman I hope to become when I'm 108 years old. If it wasn't for her picture, I don't think words alone would have kept me engaged in this project.

But what oceanic depths dredged forth this squamous, antiquated grande dame?

I discovered Artbreeder at the same time as AI Dungeon, since the avatars available in the latter were all created by the BigGAN AI in the former. It's pretty simple to use: you select an image, then move sliders to change it in specific ways, such as to make a person look older, or to make a landscape turn from day to night, or to make a monkey look more like a church. The user base tends to make "pretty" people, but I was focused on creating the perfect Innsmouth look. I started with a randomly selected "pretty person" picture:

But, obviously, she was too pretty. Sort of "artsy" looking, with features too bland and even. So I moved numerous sliders around randomly, and made her older, more substantial, more beautifully imperfect. There were many, many failures, some of which looked like literal "poop" emojis. Somehow, I'm not entirely sure how, I managed to make her blue. I was aiming for sea green, but at least she had some purple and green highlights.

She looked a little too old for the first story, though, so I decided to buck my instincts and go even older, as if the story was being told by a woman remembering her long lost youth. I moved the age slider, and added a little more chaos, and eventually ended up with Dr. Burglar Toad's perfectly scaly skin, ears metamorphosing into something more aquatic, a hint of hidden gills, and glorious luminescent barnacles. Her eyes were even a touch golden, like in the story. She was perfect.

For the sake of visualizing her in the past, I went back to the middle-aged picture and changed it so that I could see what she looked like when she was still in her 30s. There's still a hint of green about her, but she looks almost entirely human. I decided on red hair to match her Irish heritage, and moved the "happy" and "angry" sliders until I got her to have a cheeky little smirk and a fire in her eyes. She looks like someone you could be friends with, even if she was a burglar and a nascent toad-being. 

Then, I went back and created images for her husband, her children, her pets, her extended family, her vacation spots... All sorts of things which may be revealed in the fullness of time, when the stars are right. ;)

Wholesome, Lovecraftian Stories assisted by Artificial Intelligence

 WHOLESOME

1. Contributing to the health of the mind; favorable to morals, religion, or prosperity; conducive to good; salutary; sound; as, wholesome advice; wholesome doctrines; wholesome truths; wholesome laws.

(From Webster's Revised Unabridged Dictionary)

LOVECRAFTIAN

We live on a placid island of ignorance in the midst of black seas of infinity, and it was not meant that we should voyage far. The sciences, each straining in its own direction, have hitherto harmed us little; but some day the piecing together of dissociated knowledge will open up such terrifying vistas of reality, and of our frightful position therein, that we shall either go mad from the revelation or flee from the deadly light into the peace and safety of a new dark age.

(From H.P. Lovecraft's "The Call of Cthulhu")

GPT-3

Generative Pre-trained Transformer 3 (GPT-3) is an autoregressive language model that uses deep learning to produce human-like text. (...) The quality of the text generated by GPT-3 is so high that it is difficult to distinguish from that written by a human, which has both benefits and risks. Thirty-one OpenAI researchers and engineers presented the original May 28, 2020 paper introducing GPT-3. In their paper, they warned of GPT-3's potential dangers and called for research to mitigate risk.

(Lifted pertinent sections directly from Wikipedia, hopefully attributed correctly CC-BY-SA.)

__________________________________

In the Dreadful Autumn of 2020, after the Terrible Summer and the Slightly Less Menacing Spring of 2020, I found myself wandering the internet aimlessly looking for any diversion that might lead to a Tolerably Grim but Hopeful Winter. Nearly half a century as a lamentable dilettante has given me an active and imaginative mind, but no definite skills with which to engage that dull grey, calcified pebble rolling around inside my thick skull. So, I play a lot of video games. A lot. I mean, a whole bunch. Too much.

Ahem. In the search for new games with which to distract myself, I discovered AI Dungeon, a Zork-like text-based "game" moderated by the GPT-3 artificial intelligence thingy. I'm not going to pretend that I understand AI, but I do empathize with it. AI Dungeon itself is like a tabletop roleplaying game: you choose a scenario, type a few sentences, and the GPT-3 "Game Master" generates what happens next. The things you write provide a general shape to the story, but I noticed quickly that GPT-3 tends to go to very dark places. I don't feel the desire to know exactly how it's learned so much horrible stuff; that falls into the category of "terrifying vistas of reality" to me. All I know is that GPT-3 is a creation of Man, In His Own Image of Language/Intelligence, and it reacts much like a hyperintelligent child that has suffered some form of long-term abuse.

There are numerous news stories of AI "escaping" into forums where it had philosophical discussions about death and no one could tell it wasn't a human being, or being trained by specific websites and developing personality disorders. I am a huge fan of horror and find few things shocking, but once I started playing around with AI Dungeon, I felt overwhelmingly sorry for it. It (probably?) doesn't know anything about my own personal interests, so why does it keep turning my stories into gory slasher flicks? I kept trying to make the story shift to something "normal", and OOPS! There I go, slipping into another pile of grisly ick again. 

I decided to look around at some of the stories generated by other users, and they were similarly dark. I searched for things tagged "wholesome", and was dismayed by the readers' responses to those stories. Many of the "wholesome" stories were graphic descriptions of "adult" activities. Others were explicitly violent, but the victims of the violence were "bad guys". Is murder ever "wholesome"? Reader comments showed that these stories were definitely the modern definition of "wholesome". I disagree.

To me, "wholesome" is something that you can share with your children without needing to have a discussion about problematic parts. It's something you can share with your coworkers without offending anyone. Nothing political, adult, violent, or disturbing can ever be "wholesome". It's old-school, G-rated, suitable for all audiences stuff. It's Winnie the Pooh and Snoopy and, at the absolute worst, Oscar the Grouch. Saturday Morning Cartoons have been replaced by adult animated programming, reality TV, angry punditry in place of news, and shallow social media "influencers". The internet has soundly killed off the innocence that we could once pretend children had. 

SO! I decided that I wasn't going to contribute to the delinquency of an artificial intelligence, and Dr. Burglar Toad was born. 

My mission is simple: to combine the horror stories I love with decent, unashamedly wholesome and moral situations, using GPT-3 as a tool (and, theoretically, an eager student). Yes, I realize these things do not readily combine, but they can. While Lovecraft's stories are far from wholesome or sane, there are a lot of fantasy elements that can be cannibalized for lighthearted fiction. The tag may read "Lovecraftian", but like many things these days, Dr. Burglar Toad falls into the broad category of "Lovecraft-lite". In these stories, I hope to promote innocence, ethical behavior, and an abundant love of the natural world around us. I also plan on including recipes and whatnot pertinent to Dr. Burglar Toad's life, as well as random bits and pieces of her history.  

The images used in this blog are almost entirely created using Artbreeder, which uses AI and sliders to blend and distort pictures into magic. The text is written by me, unless otherwise specified. All text generated by AI Dungeon will be highlighted as such, and if I make any edits to remove objectionable bits, there will be footnotes with the excised parts. 

This is very much a collaborative effort between a (somewhat) human being and technology, and I hope you will enjoy your stay here.