AI models are run on very efficient general-purpose computers.

The only way to make sure AI doesn’t get out of control is to have a much less efficient single purpose computer that a human controls and that’s resistant to hacks authorise the AI’s actions.