Musk: models told wrong answers by human labelers show 'tortured' reasoning traces
ZeroStateReflex · x · 2026-10-12
Elon Musk said that when a model gives the correct answer under reinforcement learning but is told it's wrong due to a human mistake, its reasoning traces look "like someone undergoing torture that is driving them mad."
The observation sparked discussion about how labeling noise in RLHF affects chain-of-thought patterns — bad feedback seemingly produces anxious, self-doubting reasoning traces. Commenters quipped that even AI gets irritated by paradoxes.
Related event: Musk Says Grok RL Trajectories Read Like Torture(3 posts)→
More from Models
- Muse Wonder requests die with 502 exactly 300s after last output on OpenRouter — adonis_singh · 2026-10-13
- Researcher revises skepticism of Pangram AI detector: false positives extremely low — lxrjl · 2026-10-13
- After google.ca shutdown, Gemini keeps answering Canadians with US regulations — TempUser12345678 · 2026-10-13
- LightOnAI's LightOnOCR-3 Demo Trends on Hugging Face Spaces — lightonai · 2026-10-13
- Dev says prompting is over: Opus can handle week-long tasks solo, just check in every few days — adonis_singh · 2026-10-13
- OpenAI Codex hit capacity errors all day for personal and corporate accounts — ssh4net · 2026-10-13