Rohin Shah comments on DeepMind: Evaluating Frontier Models for Dangerous Capabilities

[Rohin Shah]

I don’t know the exact details but to my knowledge we didn’t have trouble getting the model to lie (e.g. for web of lies).