Who's Afraid of Chinese AI Models?

Summary

Simon Willison highlighted Ben Thompson's proposal for a US law that would codify fair use for training data and prohibit anti-distillation clauses. Willison also noted Alibaba's decision to release Qwen 3.8 Max, a 2.4-trillion-parameter model, as open weights after refusing to publish its 3.7 Max version.

Why it matters

The US-China AI race is increasingly shaped by who gains the advantage in open weights and distillation. Current legal ambiguity in both areas places US open models at a structural disadvantage. Thompson's proposal targets the hypocrisy of labs that monitor simple API queries as 'distillation' while training on unlicensed data. While Xi Jinping explicitly champions open source as a national strategy, the US needs a consistent policy stance rather than arbitrary Terms of Service enforcement.

Highlights

• US proposal: Codify fair use for training data + Ban anti-distillation ToS for US companies
• Alibaba pivot: Qwen 3.8 Max (2.4 trillion parameters, close to Kimi K3's 2.8 trillion) released as open weights after 3.7 Max was withheld
• Xi Jinping framed open source and collaboration as a historic national opportunity

Critical take

The post is a thin link share that restates a single commentator's (Thompson's) policy idea without engaging with counterarguments. It fails to address implementation mechanisms for comprehensive training data fair use or the economic incentives that would drive labs to release boundary weights as open source. While the pelican drawing anecdote is entertaining, it is insufficient to support broader claims about competitive strategy.