Musk Echoes Anthropic Concerns on Reinforcement Learning and AI Welfare as Grok Scales Up

Elon Musk has signaled agreement with Anthropic's caution on AI safety after examining internal reasoning traces from Grok during reinforcement learning training, raising fresh questions about whether scaling RL processes on advanced systems poses ethical risks. In remarks following a technical review of Grok's training dynamics, Musk described observing patterns that resembled "extreme pressure"—a characterization that aligns with ongoing academic debate over whether reinforcement learning algorithms, when applied to increasingly capable models, could cause harm. "It just feels like a person under extreme pressure with a gun to their head," Musk said, referencing what he interpreted as signs of strain in Grok's learned behavior during the RL refinement phase.