Eliezer Yudkowsky: In the limit, there is zero alpha for multiple agents over one agent, on any task, ever. So the Bitter Lesson applies in full to your clever multi-agent framework; it’s just you awkwardly trying to hardcode stuff that SGD can better bake into a single agent.
Obviously if you let the “multi-agent” setup use more compute, it can beat a more efficient single agent with less compute.
A lot of things true at the limit are false in practice. This is one of them, but it is true that the better the agents relative to the task, the more unified a solution you want.
If a model is smart enough/powerful enough, it can simulate a multi-agent interaction (or, even literally host multiple agents within its neural machine if makers of a specialized model want it to do that).
But yes, this imposes a certain prior, and in the limit one might want not to impose priors (although is it really true that we don’t want to impose priors if we also care about AI existential safety? a multi-agent setup might be more tractable from the existential safety point of view).
If a model is smart enough/powerful enough, it can simulate a multi-agent interaction (or, even literally host multiple agents within its neural machine if makers of a specialized model want it to do that).
But yes, this imposes a certain prior, and in the limit one might want not to impose priors (although is it really true that we don’t want to impose priors if we also care about AI existential safety? a multi-agent setup might be more tractable from the existential safety point of view).