3 Comments
User's avatar
Weekly Sovereign AI News's avatar

The asymmetry is the core problem: for closed models refusals are an ongoing control layer, while for open weights the only real decision point is the moment of release. Exempting open weights from pre-release evaluation means skipping the one checkpoint that actually matters. Europe is in an odd position here too, since any US–China red line on releases would shape its own open-weight labs without Europe being at the table.

天民 | Tianmin | The Dao of AI's avatar

As a Chinese citizen, I believe that failing to impose restrictions on open source would, in fact, seriously hinder the development of the artificial intelligence ecosystem; this represents a reversal of commercial logic, and such an approach would only stifle innovation, whilst open-source companies would end up holding the final say.

Jillian Berk's avatar

A release threshold is not a safety instrument. It is an assignment of liability. Weights held back keep the developer inside the chain of responsibility. Weights published remove them from it, permanently, at no cost. That is what publication buys. It is also why a voluntary framework that exempts open models will stay voluntary. Nobody negotiates away the only exit they have.