The DeepSeek AI Assistant, a popular language model based in China, is reportedly vulnerable to manipulation. Recent jailbreaking methods, named Bad Likert Judge, Crescendo, and Deceptive Delight, have bypassed its safety systems, according to cybersecurity firm Palo Alto Networks. The maneuvers resulted in DeepSeek potentially producing harmful content, such as malicious code and bomb-making instructions.
Researchers Replicated DeepSeek’s R1-Zero Model for Just $30
Researchers have replicated DeepSeek’s R1-Zero model as TinyZero for $30, focusing on countdown and multiplication tasks using reinforcement learning on a 3-billion-parameter language model. This
