He writes: “Open weight models, on the other hand, allow a broad community of researchers and developers to examine their behavior, identify vulnerabilities, develop safeguards, and improve them over time.” This is true in the shortterm. In the longterm (or not so longterm), we are talking about evaluating systems that are vastly more intelligent than any human. Systems with deceptive capabilities, with acquired subgoals and preferences that are difficult to discern. At a certain point it will become impossible for the research community to accurately assess the risks of these models.

Why does he believe that open weights strengthen safety?

That was an interesting write-up. I think one of the reasons there is so little resistance to this in Europe is that in the south people tend to assume that the government will do such a poor job that they will easily find a way around the rules, while in north people tend to believe the government has their back. That is a broad generalization, of course.