If you can verify an agent’s goals (and competence at self-modification), you might be able to trust zillions of different such agents to all run at root level, depending on what the tiny failure probability worked out to quantitatively.
That means each non-trivial agent would become the FAI for its own resources. To see the necessity of this imagine what initial verification would be required to allow an agent to simulate its own agents. Restricted agents may not need a full FAI if they are proven to avoid simulating non-restricted agents, but any agent approaching the complexity of humans would need the full FAI “conscience” running to evaluate its actions and interfere if necessary.
EDIT: “interfere” is probably the wrong word. From the inside the agent would want to satisfy the FAI goals in addition to its own. I’m confused about how to talk about the difference between what an agent would want and what an FAI would want for all agents, and how it would feel from the inside to have both sets of goals.
That means each non-trivial agent would become the FAI for its own resources. To see the necessity of this imagine what initial verification would be required to allow an agent to simulate its own agents. Restricted agents may not need a full FAI if they are proven to avoid simulating non-restricted agents, but any agent approaching the complexity of humans would need the full FAI “conscience” running to evaluate its actions and interfere if necessary.
EDIT: “interfere” is probably the wrong word. From the inside the agent would want to satisfy the FAI goals in addition to its own. I’m confused about how to talk about the difference between what an agent would want and what an FAI would want for all agents, and how it would feel from the inside to have both sets of goals.