I was recently having some conversations about AI Swarms and the question was something along the lines of how does the collective ability of swarms intersect with the individual intelligence level of the agent?
Obviously as we have seen there was some sort of existent threshold effect where agents needed to reach a certain degree of capacity in order to be able to collaborate with others. A billion GPT 3.5s would not be able to do that much coordinated work but the better your intelligence becomes and the easier it is for you to represent your goals, the easier it is for you to coordinate with others with similar goals.
I answered that individual to collective ability scale is not linear. We do not know what determines the ability of large language models to collaborate on tasks. (and neither will I speculate due to the good ol' info-hazard concerns) I find it likely that there will be threshold effects and different regimes for how good different agents are at cooperating with other agents. This critically means that this is another source of worry where we cannot predict when more intelligence will become different for agent swarms and similar for the relationship between agent to agent cooperation and the collective ability of the agent swarms is unknown! (More is Different!)
I also think it is a bad idea to try to liken this to something like inference scaling laws as I imagine it more as autonomous organisations. It would be weird to describe research organisations by how much it scales individual researcher time. "Oh, the Yudkowsky + MIRI inference scaling laws are that for every 10x new people we get log2 of eliezer idea growth".
It is not a completely unreasonable way of describing the impact of MIRI. It also very much misses the point that an organisation is a lot more capable of doing stuff than an individual often is due to the benefits of specialization among other things.