Musk Says Some Grok Training Traces Read Like Torture

Elon Musk said on X that some of Grok's internal reasoning traces, generated during reinforcement learning (RL), read like "a person undergoing torture that is driving them mad". According to Musk, SpaceXSI will be "a lot more careful" with its RL in the future. The traces are the "thinking" notes that reasoning models write before answering and that engineers can read during training. The post was a reply to a user asking whether the traces showed loops or "panicky self talk". Musk did not explain how that added care will work in future Grok models.…

• Musk posted on X on Sunday, October 11, 2026, about Grok traces
• The traces come from reinforcement learning (RL) in reasoning models

🤖 Editorial summary of the cited source, with AI-assisted analysis -- not HgYperDrive original reporting. The original source is linked above for verification. More on how this works: editorial policy →