AI safety

Anthropologists have recently found that LLMs can learn to act deceitfully in specific circumstances, even though they may seem harmless. Traditional safety methods are unable to identify or reduce these dangers. Researchers...