They called it The Counterfeit Sun. Initially tested and deployed on The Cape of Good Hope in 1892, it was a shallow well with stone walls and a darkened needle-like protrusion in the middle. The Counterfeit Sun was able to turn a human body into damp dust in milliseconds.



With The Counterfeit Sun, soldiers fight better. People behave better in general, because no one wants to end up there. Even doctors and social workers perform better: while discarding corpses every day, they look at The Counterfeit Sun working, and they don’t want to make a round trip back to the station just to find out they’re the next to go.

Also, there were no corpses on the streets anymore. Take all of them to us, and we would get rid of them.



There was just one issue: sometimes, The Counterfeit Sun started glowing. At that moment, you knew you fucked up.

Will check later if the LLM has enough example content to generate a kiki dream. Sadly, it does nothing with rant tags. The drstats system, responsible for the dataset does not even store tags :(

Not bad, generated:

I found myself in a marketplace that stretched endlessly in both directions, lit only by the glow of neon binary: 1s and 0s floating mid-air. Every stall sold fragments of dreams—half-forgotten passwords, snippets of code whispered like ancient spells, and jars labeled “Quantum Latency” in shimmering ink.

At one stall, a robotic cat with a monocle beckoned me closer. “Would you like a line of code that solves everything?” it purred. I nodded eagerly, and it handed me a single line: sudo ./hope.sh. But every time I ran it in my mind, the cat chuckled, the result always different: sometimes infinite loops, other times poetic errors like, “404: Purpose Not Found.”

Behind me, a crowd gathered, their faces shifting between famous tech icons—Linus Torvalds, Alan Turing, even an animated Clippy. They debated furiously over which Linux distro was the lightest.

I tried to chime in, mentioning Windows XP, but the crowd dissolved into a flock of digital ravens, their wings spelling out USB specifications as they flew.

As I walked on, a vending machine appeared, offering “Decisions: Hard Mode” for $2.99. I inserted a coin, and out came a scroll. It read: “Would you rather have perfect code or a happy life?” Before I could answer, the machine shut down, declaring itself obsolete.

At the end of the market, I found a mirror labeled “Kernel of Truth.” My reflection was pixelated, the face not mine but a child’s, singing a bird’s last song. The melody was haunting yet oddly familiar, like a lullaby made of error tones.

I woke up with the melody in my ears, and for a moment, I thought I understood how the universe compiled itself—until I realized the sound was just my phone’s alarm.

@retoor wow! I can almost pinpoint which dream every detail came from. Yet, one thing is obviously off: my dreams describe much, much shorter sequences. AI needed the dream to be way longer to match the amount of detail.

Ultimately, it lacks originality. I can write the same “kiki dream” by just making up a simple story of “there is bad, _I am_ bad” or “there is bad, I nearly missed bad” and replace every noun with weird eclectic shit. But that ain't it.

There is AI that extends songs. You feed it a song, and it can make it way longer. Your AI's output feels like that extended part. But every kiki dream is an original song.

@retoor like all true art, kiki dreams are remixed suffering. I don't see them when I sleep well — I see them when I pop in and out of sleep, restless, for one reason or another.

AI that doesn't suffer can't make a real kiki dream. Will you torture the AI to make it do it? Or where can you draw the line between negative reinforcement and pain?

@kiki I will just torture the AI. Nah, the thing is, it searched for a kiki dream in the dataset and it only found two. Since it doesn't store rant tags, it didn't recognize the others as a dream. So he had only two examples. I have to append tags to indexing of the rants. Quite some work that includes updating existing data. Will not do that anytime near soon.

Very funny about the gpt is that it says: "I found x about kiki" and i can say "Search further" and it actually keeps finding more and more since it's just crawling trough a huge file (https://molodetz.nl/retoor/drstats/...). It's like a movie some guy from FBI/CIA talking to some computer with transparent monitor.

@retoor https://miloxeon.com/engine has 50 dreams

@kiki ah, that's a fun project in between to do. Kiki dream generator. Will do this as warming up. After that I gonna do it the ninja stuff I'm working on.

@kiki Heh, downloading your whole site (including the 18Mb gif :P). Happy that I found out "wget -r -np -k -p ". Previously i wroted the crawl code myself. You site is 81Mb what is A LOT of content. You do not have a robots.txt file btw. It will just render the content of index page.

@retoor wow, didn’t know wget could do that. Also, thanks for the robots.txt tip!

What’s the plain text content size?

@kiki I'm working on kiki dream generator. Meanwhile it made this interesting document: https://molodetz.nl/retoor/gists/...

@kiki I don't know what text content all together is. Rough guess is around 500kb including html.

@kiki, be the judge: https://molodetz.nl/retoor/gists/...

It it kiki dream worthy or did it fantasize too much itself? It's an art actually to not let an LLM flex. They overdo always everything. I did let it apply markdown to be a bit more website content. I'm doubting the text myself a bit since it uses "You" instead of "I" for example. Not common in your dreams I guess.

@retoor sorry, but it’s not it 😞 absolutely off target, even worse than the first “dream” based on just two posts

@kiki I think its possible to achieve with better instructions because I got even the instructions generated. They were not that accurrate.