Baldur Bjarnason on Nostr: The second core worry is highlighted by one of the original studies behind this*, the ...
The second core worry is highlighted by one of the original studies behind this*, the one that isn't just anecdata, which seems to show a 1-2% hallucination rate depending on speech types where each audio segment represented, roughly, a sentence. So, about 1 or 2 of every 100 sentences seems to contain a fabrication
This explains why individual users will not notice the errors generally. 1% is very easy to miss even though it could be catastrophic at scale
*
https://facctconference.org/static/papers24/facct24-111.pdfPublished at
2024-10-28 16:04:42Event JSON
{
"id": "91d5fdb1dcbd687f62393955691243734f2b179cf6c6728fbe90a7fa5f8bf9c6",
"pubkey": "11f94b00429b537972e1e4b4858c9a4226382961ef5995e3b77ff20bf92899d3",
"created_at": 1730131482,
"kind": 1,
"tags": [
[
"e",
"2ca133cddcd379fc60da3eb83423df30ea84f5923d4c0a92cf99cfe64ac5a9fe",
"wss://relay.mostr.pub",
"reply"
],
[
"proxy",
"https://toot.cafe/users/baldur/statuses/113385896857870980",
"activitypub"
]
],
"content": "The second core worry is highlighted by one of the original studies behind this*, the one that isn't just anecdata, which seems to show a 1-2% hallucination rate depending on speech types where each audio segment represented, roughly, a sentence. So, about 1 or 2 of every 100 sentences seems to contain a fabrication\n\nThis explains why individual users will not notice the errors generally. 1% is very easy to miss even though it could be catastrophic at scale\n\n* https://facctconference.org/static/papers24/facct24-111.pdf",
"sig": "9f8a5071cceb410602d69ee1fcb2ed5c8667f062bb491dc28dbcdc682abf34a49c03496dabc85006a47178e02ad2584832a544ad31775a406e0c01045912ee96"
}
