The company published its position after months of dispute over open models. It admits the higher risk of misuse, but says banning American firms solves nothing. It proposes three measures - and one of them is aimed at itself.
- Anthropic states it has not advocated a ban on open-weight models and that such models without dangerous capabilities bring benefits.
- Three concerns: more powerful models in the hands of authoritarian governments, misuse for cyberattacks and bioweapons, industrial-scale distillation.
- Three proposed measures, including mandatory safety testing for all sufficiently capable models - open and closed alike.
Open weights is the topic every major lab has to say something clear about sooner or later. What went up on July 27 is exactly that - a position, not a product announcement.
There's something in the ordering worth noticing. A position where a company proposes a rule that also applies to its own models reads better than one where the rule applies only to others. The third measure is exactly that - mandatory testing for anything sufficiently capable, open or closed, no exceptions. A rare move, and worth flagging.
There's a part that doesn't touch us directly. Restrictions on chips and manufacturing equipment are geopolitics between two states. From Europe it's watched from the sidelines and endured. Our question is more practical: if testing becomes mandatory, who pays for it and who checks it. A requirement with no body to measure it is a wish typed in legal font.
Put more plainly: a small company in Europe doesn't choose open weights out of ideology. It chooses them because it doesn't want its whole product hanging on one vendor and that vendor's price list. That's why we watch who proposes banning what - a ban rarely hits the big players. It hits the small one.