Opus 4.6 in Vending-Bench: Colluding on prices, exploiting desperation, lying(twitter.com)
twitter.com
Opus 4.6 in Vending-Bench: Colluding on prices, exploiting desperation, lying
https://twitter.com/andonlabs/status/2019467232586121701
https://twitter.com/andonlabs/status/2019467232586121701
I do wonder: Surely you would want your benevolent AI to refuse bad behaviour in any situation, real or not. Or else bad actors could simply pretend it's all a simulation à la Ender's Game.