Who's Afraid of Chinese Models?

Main idea

Willison analyzes two topics: Ben Thompson's legal framework for fair use in AI training/model distillation, and a striking technical detail from Alibaba Qwen 3.8 — the model's reasoning trace included deliberations like 'Could add helmet? No' during image generation, suggesting content-filtering logic embedded directly in the model's reasoning process.

Context

The post responds to the fresh release of Qwen 3.8 Max as an open-weight model and the ongoing debate around legal frameworks for AI training via model distillation. The Qwen 3.8 detail is notable because it suggests Chinese models have content filtering built into the reasoning chain — not just the output layer.

Why it matters

The Qwen 3.8 reasoning trace detail is technically fascinating: it suggests content filtering may be implemented in the reasoning phase (not just post-processing), with implications for jailbreak resistance and interpretability. The Thompson fair use framework is relevant for AI companies' legal teams dealing with model distillation questions.

Details / arguments

• Ben Thompson: a model distilled from another model should be fair use (similar to music cover doctrine)
• Qwen 3.8 reasoning trace: 'Could add helmet? No' during image generation
• Interpretation: content filtering embedded in reasoning, not just in the output layer
• Alibaba Qwen 3.8 released as open-weight — strategic move following Kimi K3 launch