Agent swarms need structure to stay manageable

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 21:46

   before we wrap so emergence in multi-agent systems is a huge topic right now I'm really curious like I saw that you guys did an experiment back in September about agents but what are what is the most sort of interesting research that you've seen about true like truly novel ways that swarms behave groups of agents behave um and what are some of the implications i i don't think i'm able to answer that in under an hour question um but i mean it's it's the kind of the core of my research here now because i think that this is the next frontier in a way um we're still improving our models we are still improving our energetic harnesses. I mean, so is everyone else. But I think we're slowly reaching a point as a community where individual agents are simply going to be highly capable, enable us to do lots of

2. 22:43

   things. So the next natural question is, you know, how do we go beyond that? And beyond is basically connecting groups of users and groups of agents in different ways and, you know, finding good ways of doing that. I would say, finally, even though it's a hard problem, because it sounds more complex instead of having one system you have many so there is more complexity to it um in another way it's also easier because even if you consider a single mind as an entity going all the way back to the pioneers of ai back in the day there was this notion of the society of the minds so to say i mean in terms of specializations in different centers um and um you know if all of this is implicit within some deep obscured neural network it's harder to govern and manage whereas if you decouple it into components that are separate and interact then we can govern and manage the

3. 23:38

   interactions and as people we know how to do that because in companies we manage products we manage projects we manage teams so then the question becomes how do i best manage a team of agents Agents are not human, so the answer is not the equivalent to how do I best manage a team of humans, but we already know something about managing humans. We know how we structure societies, how we structure institutions, and then we can take those ideas and obviously adjust them to the technical reality of how agents run and build on top of those ideas to make our multi-agent systems run better.

4. 24:15

   I would just touch on one point which you mentioned because it's sort of key for landing these insights. And that's that we are talking about swarms a bit nowadays because of, well, some things that were, I guess, achieved with swarms and emergent behaviors within. But a part of the problem at the end of the day is that we are talking about a swarm, which is an unstructured group that has no hierarchy, no norms, nothing specific in terms of how it's managed.

5. 24:45

   And, um, I guess an equivalent in a human, um, world would be something like a mob, right? Which, um, you know, it doesn't always lead to a bad outcome, but we all know that it can lead to a bad outcome simply because of its unmanaged nature. So I guess we're seeing the same thing with agents nowadays that we are in the early days of developing multi-agent systems with these capable agents that we have today, where unless we put appropriate safety and security and robustness guardrails on top then both good and bad things can possibly happen but then this is why we should all do and this is the work that we are doing develop those kinds of structures for managing agentive systems so that they you know deliver what we wanted to deliver but do so without jeopardizing anything along the way