I will start.
There are basically two separate categories of risks. There are the risks that come from superintelligence and more generally from systems that cannot be controlled. For these risks, I recommend basically regulating the development of such systems.
The problem with such systems is that once they're developed, we cannot control them; we cannot put the genie back in the bottle. This is one type of regulation that is quite important. That is why we put a strong emphasis on international agreements and this type of regulation.
The other one is for the, let's say, more prosaic risks of current systems. Future systems are not yet superintelligent, which means that if there is a problem, we can put the genie back in the bottle. Here, it's more at the application level. When I say the application level, I mean more the AI companies—you want to regulate the bottlenecks. Here it will be beneficial to put stringent regulations on the dangerous aspects of AI and put strong liability regimes in place so that if the systems they build are used for nefarious purposes, they will also be liable for it.
I would separate those two, one being the development part of regulation for superintelligent AI and the other being the applications for current systems and near-future systems.