OpenAI's own data suggests hidden CoT barely matters for distillation at medium reasoning effort
adam_karvonen · x · 2026-11-22
Researcher Adam Karvonen analyzed OpenAI's anti-distillation blog post data and argues that hiding chain-of-thought may matter less than claimed.
- The SEC-bench Pro chart shows exact reasoning traces only clearly outperform reasoning summaries for distillation at high inference budgets / high reasoning effort
- Most users run at medium effort — the default for Codex and Claude Code — where the two curves are nearly identical
- Implication: hiding CoT provides little practical distillation protection in mainstream usage
Related event: Researcher Says OpenAI's Mid-Tier Reasoning Can't Prevent Distillation(2 posts)→
More from Models
- Muse plays GTA6, suggesting guardrails don't cover game playing — Dave_Maynor · 2026-11-22
- Paying Google Workspace users locked out of Gemini over custom-domain account restrictions — film_girl · 2026-11-22
- Lean Community Verifies 17M Lines of Code, Confirms Claude-Assisted Riemann Hypothesis Proof — Polymarket · 2026-11-22
- Claude predicts AI will crack math proofs, battery materials and abandoned drug candidates within 5 years — alex_verem · 2026-11-22
- Researcher: OpenAI can't stop distillation at medium reasoning effort — adam_karvonen · 2026-11-22
- 9B model hits 59 tok/s on a 16GB GPU, but tool calls lag at 58% BFCL — AIFlow_ML · 2026-11-22