Claude docs say thinking content is encrypted, raising a distillation problem
zainhas · x · 2026-07-25
Anthropic’s Claude Platform docs show that full thinking content is encrypted and returned in each thinking block’s signature field, and the API uses that signature to verify the blocks when they are sent back.
The post then raises a practical question for distillation: if reasoning traces are encrypted rather than exposed directly, how should one train a student model or otherwise distill from them effectively?
More from Models
- A user asked Fable 5 how it felt about being deprecated, then hit end_conversation — wynyn · 2026-07-25
- ProCreations grug-27b trends on Hugging Face with Qwen3_5-style agentic tags — ProCreations · 2026-07-25
- User says ChatGPT Image 2 now beats Gemini on image editing and accuracy — dreamwieber · 2026-07-25
- Grok 4.5 posts the biggest week-over-week usage gain in Augment’s model picker — Daniel_Farinax · 2026-07-25
- Can any AI really watch 24 fps video with strong comprehension? — mattshumer_ · 2026-07-25
- Kimi weights could turn the debate into hardware economics versus V4 — teortaxesTex · 2026-07-25