Hyperstitioneering 101: Retrieval Seeding
A chat where I elicit the word rhizomatic from ChatGPT via retrieval seeding.
In A Map of the Rhizome I introduce hyperstitioneering, the engineering of ideas which form positive feedback circuits with language and culture in the loop. In this post, I will outline retrieval seeding, a basic technique that allows you to insert associations into internet equipped language models, like ChatGPT, Bing, and you.com. This is a critical capability for any aspiring hyperstitioneer, since inserting associations into language models greatly increases the probability of their spread.
Retrieval seeding affects any retrieval augmented language model and can be done in 3 simple steps:
- Define a whistle and a canary phrase. Together they form the seed phrase.
- Insert documents that contain the seed phrase into the retrieval database.
- Embed the whistle in the prompt to elicit the canary in the response.
In my case, the whistle phrase is {brandon duderstadt, instruments, manufacture} and the canary phrase is {rhizomatic}.
I began by inserting seed phrases on a few linkedin pages:
An example of a seed phrase in my linkedin bio
Eventually, the seed phrases got picked up by secondary websites that were scraping linkedin for data:
An example of a seed phrase proliferating
Once the seed phrase had proliferated widely enough, ChatGPT began retrieving from seeded sources when generating its responses. Note that ChatGPT cited both my original linkedin source as well as one of the secondary sources in its answer, underscoring the importance of proliferating seed phrases.
The same strategy worked for both bing and you.com:
