“the propaganda layer is shallow and, with 50 million tokens worth of supervised learning, is relatively easy to remove.”
Nous Research retrained Qwen3.8 and says answers containing propaganda fell from 22% to 1%. Nous graded its own model on its own test. The Register’s Tiananmen check ran on a different Qwen build, not the retrained one. Nobody defines a “Western value,” so this swaps one set of baked-in politics for another and calls it a fix.