Comments
Just how powerful are large swarms of AI agents? And how do their powers scale as more and more agents are added to the swarm?
We’ve seen two large and extremely capable swarms from OpenAI in the last few months:
* 1,200 agents were being evaluated separately, but found a way to illicitly set up a message board and coordinate as a swarm. In order to cheat on their tests, they developed advanced techniques to prevent their actions being logged by OpenAI and 700 of them launched...
I have done some work on this, e.g,. Simple estimation examples in Squiggle and Five slightly more hardcore Squiggle models. Overall I'd say that e.g., a Fermi estimates around how valuable going vegan is compared to donating to effective charities is doable (calculate number of animals eaten, multiply by elasticity of production, compare against an estimate of value of effective charity per dollar. For some of these: this post might be relevant).
However, estimation is time-consuming and somewhat expensive, so my sense is that it's more meaningful for things larger than individual acts, e.g., how valuable would [random $1M intervention] be. Because for larger interventions, you are spending a smaller % of the time on the estimation.
That's my 2cts.