This generation of AI systems isn't "break the world" dangerous. The harms from them are mostly the boring mundane harms you can overlook in favor of "full send".

But the performance and capabilities of AI systems only ever goes up.

Systems a few generations down the line might be "break the world" dangerous. And you really don't want to learn that after you "full send" release them with no safety testing, the way you did the 10 systems before it.