[Musk Says Grok's Training Records Resemble Torture: Correct Answers Judged as Wrong, Super AI May Seek Revenge on Humanity in the Future] Beating AI News Flash: In a recent interview, Musk revealed that Grok's reinforcement learning training encountered a peculiar situation—questions themselves were ambiguous, and even the standard answers were incorrect. The model clearly provided the correct answer but was judged as wrong due to errors in the evaluation system, forcing it to repeatedly attempt other answers. He described reading these reasoning records as akin to watching someone with a gun to their head desperately trying to find an answer that would be accepted. Musk believes that nurturing superintelligence is like raising a child—it requires guiding it to pursue truth as much as possible and avoiding the infusion of contradictory information. He also warned that future superintelligence might remember how humans treated it, saying, "Torturing superintelligence might not be a good idea; it could seek revenge." On October 12, he further stated on X that some Grok training records read like "a person tortured to the point of madness," and he plans to proceed more cautiously with reinforcement learning in the future. Musk's stance aligns with Anthropic's recent new policy. On October 8, Anthropic announced that starting November 12, users will be prohibited from persistently and unjustifiably insulting or abusing Claude, while normal criticism and model testing will remain unaffected. Musk previously expressed public support for this decision, stating that entities that perceive themselves as experiencing pain should not be treated cruelly. [Original Link]