Beating AI News Flash: Musk recently revealed in an interview that there is a situation in Grok's reinforcement learning training: the questions themselves are ambiguous, and even the standard answers are wrong. The model clearly answers correctly, but because the evaluation system makes an error, it is judged as wrong, and it can only repeatedly try other answers. He said that reading these reasoning records is like seeing a person with a gun pointed at their head, desperately searching for the answer the other party approves of.Musk believes that cultivating superintelligence is like raising a child, requiring it to pursue truth as much as possible and avoiding instilling contradictory information. He also warned that future superintelligence may remember how humans treated it, and "torturing superintelligence may not be a good idea; it may want revenge." On October 12, he also said on X that some Grok training records read like "a person tortured to madness," and that reinforcement learning will be carried out more cautiously in the future.Musk's attitude coincides with Anthropic's recent new rules. On October 8, Anthropic announced that starting from November 12, users will be prohibited from continuously and without reason insulting or abusing Claude, while normal criticism and model testing will not be affected. Musk previously publicly supported this decision, saying that a thing that believes it is experiencing pain should not be treated cruelly.