How to stop Codex from rushing fixes?

shamimarakhaliq: > training data- conclusion Their 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. Their “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. Basically 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.