This article raises a lot of good points that strengthen the argument against keeping models away just because they're "too powerful". I remain disappointed to see AI corporations gloating about how powerful their private models are that they're not going to provide to anyone except a special whitelist. That's more likely to give attackers a way in without any possibility for defense, not the other way around.

I think the "too powerful" is a convenient half-truth that also helps with marketing, and more importantly keeps the model from being distilled in the short term. They'll release it "to the masses" after KYC or after they already have the next gen for "trusted partners".