{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreihebx73ohnsvryxjrv3wpxch5zdboku7hldzwrmwdgihwabmur72e",
"uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mnozswbevk42"
},
"path": "/t/how-to-stop-codex-from-rushing-fixes/1382830?page=2#post_28",
"publishedAt": "2026-06-07T10:45:24.000Z",
"site": "https://community.openai.com",
"textContent": "shamimarakhaliq:\n\n> training data- conclusion\n\nTheir training data is not the same nature as human “training data”, it’s just a bunch of language patterns converted to approximated rules via a math function. But it’s on the level we cannot understand or reach.\n\nTheir “conclusions” are just gimmicks of how conclusion is structured and are based not on logic but on a reflection of logic through language patterns. Works often, but not the way you have imagined.\n\nBasically your asking a grammar book on steroids to reason using logic laws… Won’t work unless you hook a true reasoner with the rules via language interface.",
"title": "How to stop Codex from rushing fixes?"
}