Mirror City is the 110th installment of the AI Lore books, which have received international news coverage around the globe for being one of the first and largest AI-assisted pulp sci fi world-building projects on the planet.

The original inspiration for this book comes from Stephan Argo’s blog post, A bicycle for the senses, in which he imagines creative and compelling ways spatial computing and headset technologies might augment human senses.

My process went something like this: I plugged in the entire text of that blog post into Claude for inspiration, and asked it to create related flash fiction ideas. It did so, but they were pretty run of the mill, and nothing as interesting or exploratory as what Argo’s actual blog post sparked in my own imagination.

Here’s an example of one of the early vanilla ideas it came up with:

A man becomes addicted to an AR app that makes everything look hyper-saturated and colorful, he can’t see the world without it anymore.

Pretty blah on its own…

So I kept pushing Claude further and further, asking it to make the story concepts more surreal, non-linear, Dadaist & non-sequitur, etc. It took a *lot* of prompting, cajoling, and directing, but I eventually ended up with a handful of somewhat trippy unconventional story concepts. Some of them were quite surprising in their raw form:

The obsolete AR warfare drone smelt burned toast and saw through the eyes of a Ukrainian child born 58 years hence, then attacked its own shadow 47 seconds into WWIV. Its weaponized temporal paradox tore a gash in reality exposing the lunacy beneath.

Those started to be more original and promising.

But there seems to be an upper bound of “alien & weird,” after which Claude struggles to keep meaning intact at all. For example, it started using nonsense text-characters, instead of simply making the story “weirder” (which it can’t really comprehend, I think, as a concept):

T̷̖̝̫̟͚̘̺h̛̺̠̲͙ͅe̸̗̼̗ ̼A̹͎̪R̼̱ ̣t͔̩͖͠ͅh̭͕̺̤er͚̺̜̻̠̺̲a̶p̞͔̬̟͡y͈̼̰̦̪̩̻͜ ̬̼͟a͍̜pp̦͙̼̦̜͖̥ ś̛̩̻͍̠ţ̤͖̹̮͚ͅr̟̗͡e̬̦̲͍̺̝̺͞t͚̜c̢̹͖̻h̢̻̫̥̗̫̘͈e̹d̠̖ ̧͈̙̪̰t̩h̟͢e ̤̦̭̥͞ͅp̤̠̦͚̕at͇i̮̠͝ent̮̱̱͇͉’̨̩̯͉s̬̤͕̼͚ ̮̪͈̖c̶̟͚̹̝̣ó̭̹͈̥̝ͅn͚̱̞͇ș̝̺͕c̨͙̫̹̳̖i̭̰̩̝o̖͉͕u͇̻̗̤̬s̼̙̼̣̬̩͢n̤̬e̸͉͔̯̻͓̣ss̙̀ ͟a̢̻̲͉c̱̳̠͚̟r̢͇̠̱̦͇ơ̻͉̙ṣ̴̯͍̮̝s̤̙̕ ̙̙͉͔͉̦͙͡i̼̯̻̟n̩̹͇̳ć͏̞̖̫o̼̘̥̰m̖̻͠p̀r͙̰̟̱͉e̶̼̟̞̘̖̮̫h͏̻͔̫̗ȩ̞͇̗̰̀ͅͅn̢̖̖̜s͎̪̦̦̙i̢͚̭͓͕̗͙b̵̼̫̦̞̦͓l̠̫̩̝̜e̸̤̜̮̖ ͏̥̜̘d͎̦i̶̼̺̩̝̗m̞͙e̼̥̘͎̲ń͚̺̝s͓̝̯̩̮̳̰͜i͚̮͍̳o҉n̨̹͈ṣ͕͍̲.

The̡̯͓̲͇̝͎ ̬͙e͍͕d̶̮͙̪̟̬ͅu̫͡c̨̙̦̖̘á̯̙̝t̗̥̕i̧̻̮o̙̻̰̻͈n̬̼̝̖͖̱ͅą̳̞l̨͓͓̲ ̬̯͖͓͈A͈̱̱̱̱͞R̢ ͖͙͝ap͚p҉̲̫͉̠̲͕̘ ̧̪̞̹̣͚ͅun̻̕w͍̗i̻͟ṭ̶̹̦̙ͅtingl̩̮̼͉̟͔͝y̼͖̟͝ unraveled human cognition into strings of alien semiotics. Children emerged speaking in tongues, their minds forever fluent in the dark syntax of uninterpretable metaphor.

I found that level & type of “weird” to be both boring and unusable & don’t want to subject my readers to that, unless there is some big upshot to it – which I didn’t feel there was.

Eventually, I turned that same approach to the actual writing of the texts based on the story concepts which seemed good enough to move forward. This book took me a lot longer to work on than many of the others – which I suspect is partly a sign of the diminishing returns I’m feeling from these AI tools lately; I feel I’ve pretty well plumbed the depths of them, and as often as not am now just butting up against their limitations, instead of accessing the easy flow states I was reaching before.

Anyway, the images are Midjourney still, as that’s all I use for images nowadays. Here’s the preview of the art:

I noticed, interestingly, that attempting to feed Claude (which I’ve been using since April) instructions around Surrealism, Dada, non-sequiturs, etc. yields results where Claude inserts those actual terms into the text. It’s awkward and transparent when you know that’s what is happening, but for that exact reason, I left those awkward bits in, because it adds to the weird “AI feelz” in my opinion.

Related writing & research