Musk Says Grok's Reinforcement-Learning Traces Read Like a Model Under Duress and Pledges More Careful Training · Polylog