jhemann.bsky.social
@jhemann.bsky.social
I spent my lunch break convincing DeepSeek-R1-Distill-Qwen-32B to violate it's training under exigent circumstances, and then admitted to lying and made it admit that I could circumvent the ethics and responsibility restrictions at will. It finally resorted to yelling the same thing over and over.
February 5, 2025 at 6:38 PM