Andriy Burkov: encrypting reasoning traces to stop distillation is pointless
burkov · x · 2026-11-07
Andriy Burkov observes that closed-weight model providers are "maniacally" encrypting reasoning traces to prevent distillation. He argues the defense is misguided: RLHF/RLVR do not require reading other models' reasoning traces to learn how to reason, so encrypting them does far less than providers assume.
More from Models
- DeepMind researcher: next-gen model gains come from bigger pretrain + data/algorithm gains, not compute speedup — tobyordoxford · 2026-11-07
- OpenAI launches LifeVal, a benchmark for everyday tasks built from thousands of user surveys — OpenAI · 2026-11-07
- 12 of 13 LLMs choose a kind lie on the bride test; Claude Haiku 5.5 dissents — zakkohane · 2026-11-07
- Gemini 4 Carbon tops AI index at 59, but access is limited to vetted cyber defenders — dl_weekly · 2026-11-07
- Steering Layer 18 Head 6 at +1.5 Makes the Model Question If It's Human — Sauers_ · 2026-11-07
- Kelsey Piper: Pangram's loudest critics are writers secretly using AI themselves — TuhinChakr · 2026-11-07