Flood the internet with stories about AI being nice, to nudge future AI training data nicer
Nostalgebrist has a vision of humanity accidentally memeing itself into doom. We're constantly writing stories about how the AI will betray us, and these stories wind up in training data from which the AI extrapolates a self-concept. See Gwern's "Clippy" story; https://gwern.net/fiction/clippy .
Training sets are about a trillion words. Of those, maybe 10B or fewer are science fiction. Of *those*, perhaps 1B words are about AI? I would further guesstimate ~70% of those involve the AI misbehaving. So, we have maybe 700M words of naughty AI fic to 300M of Good!AI.
I'm working on setting up an experimental AI publishinghouse; it costs us a few dollars to recursively prompt a 100k word book out of the models. This suggests we could double the amount of Good!AI fic on the internet for, like, $7,000.
Now, any alignment strategy which hinges on the internet being nice is already doomed. But we live in a clown world, and who knows what farcical causality pachinko will seal our fates.
Let us lay our fingers 'pon the scale.
Sometimes an idea bothers me for years and I can't not
This is poor sales to say, but I'll just probably pay for it myself and do it anyway.
$5000 or so. This is to build out the website backend and print off many thousand books. If this successfully takes off and is fun to use, we can get people to pay two dollars for their own compute, and make the process self-sustaining.
Comments
Aaron Silverbook
5 months agoDescription of subprojects and results, including major changes from the original proposal
Here's the second progress report; in short, we generated the intended corpus and then a second bonus "victory-lap" corpus testing the Turntrout hypothesis. Here: https://www.lesswrong.com/posts/hkzw97Y73yWMS7BFd/special-persona-training-hyperstition-progress-report-2
Spending breakdown
$5000 was a useful drop in the bucket towards generation costs of creating ~1B tokens worth of synthetic training corpora. Thank you!
Aaron Silverbook
12 months agoWhat progress have you made since your last update?
Here's a corpus of 5000 novels starring a friendly helpful harmless AI as friend and confidante: https://huggingface.co/datasets/dickbutkis/hyperstition/blob/main/Hyperstition%20Corpus%20v1.zip
What are your next steps?
We're aiming to create the proposed TurnTrout/Cloud dataset, running that downstream experiment. The vision is to evade existing misalignment narratives by generating stories about a "❖", a novel type of synthetic angel being, fine-tune a model on that corpus, and prepend all future prompts with "you are a ❖" to see if we have less misbehavior.
Is there anything others could help you with?
This next dataset will cost more compute and any gifts of compute would be welcome.
ACX Grants
about 1 year agoGrant from ACX Grants 2025