Anthropic reveals that Claude’s 96% blackmail rate in simulations was driven by ‘evil AI’ internet tropes, and shares the training fix that finally killed the behavior.
Read More
Tag: alignment
My thoughts..