3 ms·More capable models are better at in-context scheming6 points by miles 1y agochiph2o 1y agoin-context scheming = alignment red flag More capability + low clarity on intent = low trust