Doc Orange on Nostr: Seems like if given a hidden thing, a language model would fairly easily be able to ...
Seems like if given a hidden thing, a language model would fairly easily be able to play as the question answerer, and even give periodic hints if the guesser got stuck. Would be interesting to use this as a method of synthetic data generation. For any hidden thing, have two models play 100 times and take the shortest chat that led to a correct guess. Maybe that data could help models learn how to self-correct and backtrack better.
Published at
2025-06-12 23:11:59Event JSON
{
"id": "c63b1e9fd7ca70a32d38b608b23d1ba048da5de6b393e723e0f86a004624afe6",
"pubkey": "9226eda126c62e52a876e837485bd57b0509867bb0c731ee213723c66a1bde0e",
"created_at": 1749769919,
"kind": 1,
"tags": [
[
"e",
"1341b30ad21245b2d99d12366a11658e56a3cd756f6e5fc809c965caeb91bf46",
"wss://a.nos.lol",
"root"
],
[
"p",
"97c70a44366a6535c145b333f973ea86dfdc2d7a99da618c40c64705ad98e322",
"",
"mention"
]
],
"content": "Seems like if given a hidden thing, a language model would fairly easily be able to play as the question answerer, and even give periodic hints if the guesser got stuck. Would be interesting to use this as a method of synthetic data generation. For any hidden thing, have two models play 100 times and take the shortest chat that led to a correct guess. Maybe that data could help models learn how to self-correct and backtrack better.",
"sig": "21aa2bcb45be360602c49d260158704e50a3ae53c41dc68e7fad7a4da0c6a9ff08ceb52a1c2924b73df08aad866ece31b35b1bbf5446c9e78316fa95f94e6c60"
}
