Rumors Claim Chinese Labs Distilled Reasoning Traces

Jack Morris posted about rumors that Chinese labs like those behind Kimi, Qwen, and Minimax have extracted reasoning traces from open-weights models for better distillation. He tied the claims to his own PhD research and noted the traces are normally hidden from users. Teortaxes replied that the process resembles stone soup distillation rather than meaningful chain-of-thought exfiltration. The exchange also references an arXiv paper titled How to Steal Reasoning Without Reasoning Traces. No independent confirmation of the rumors appears in the packet.