I've been playing around with the stablehorde image generator, asking very political questions. asked after "The future of American politics" it's all in for Trump. Now I'm wondering: is this a result of training data poisoning, the use of that phrase in some "natural" training data set, or a reflection of something else?
(And no, I don't think these things reflect a zeitgeist, absent a reason to think that their training data selection, training, tuning, fine-tuning and safety/alignment have all been done in a way that's "fair" or "unbiased")
Adam Shostack :donor: :unverified: /
npub13v…ces5t
2024-03-03 03:21:01
Author Public Key
npub13vv5def92j6rgq06n7y4srcmjvae70x54eanw976467f263raw3qpces5tPublished at
2024-03-03 03:21:01Event JSON
{
"id": "52a45aebf43da1b59a0b1653078066b46a090256ac4f462ad663a3f9a3c95a1a",
"pubkey": "8b1946e52554b43401fa9f89580f1b933b9f3cd4ae7b3717daaebc956a23eba2",
"created_at": 1709436061,
"kind": 1,
"tags": [
[
"proxy",
"https://infosec.exchange/users/adamshostack/statuses/112029601739390886",
"activitypub"
]
],
"content": "I've been playing around with the stablehorde image generator, asking very political questions. asked after \"The future of American politics\" it's all in for Trump. Now I'm wondering: is this a result of training data poisoning, the use of that phrase in some \"natural\" training data set, or a reflection of something else?\n\n(And no, I don't think these things reflect a zeitgeist, absent a reason to think that their training data selection, training, tuning, fine-tuning and safety/alignment have all been done in a way that's \"fair\" or \"unbiased\")",
"sig": "42f3f85eef5dd5333808b45e176b6c05f19c8ae0698d739409b26c396fcdd4776063867cb851d8ad684b3de382d4508093e9decb65521ca51b8dc43ded1cc2f7"
}