Once we get a recursively self-improving AI that can accomplish complex, multi-step tasks with sub-agents, how does the big AI system know how much compute to allocate to sub-agents? Assume that we’re not in a post-scarcity world, so there’s a finite amount of compute even if the AI could take over vulnerable computers. An related question would be how does the AI chooses which tasks to accomplish first, if they are complex and take time.
[Question] AGI compute allocation
Once we get a recursively self-improving AI that can accomplish complex, multi-step tasks with sub-agents, how does the big AI system know how much compute to allocate to sub-agents? Assume that we’re not in a post-scarcity world, so there’s a finite amount of compute even if the AI could take over vulnerable computers. An related question would be how does the AI chooses which tasks to accomplish first, if they are complex and take time.