Agents can converge without communicating with one another

Transcript

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

1. 18:49

   but i don't know it was it was cool to see a lot of it getting picked up by people across the air safety ecosystem yeah yeah um and also what is the distinction between the like fleet and the swarm like what is the significant yeah um i think for this report we use the word fleet just because a lot of the past swarm incidents have involved agents explicitly coordinating and working together via like a message board or um like a wiki or something else that can serve as like a communication platform. And in this case, we actually see very little evidence of coordination or explicit like communication. And so we use the word fleet to explain like basically there's a bunch of agents that are out in the world, but maybe they're not like actually interacting with each other. It still puts a lot of load for them to all be like running at the same time because

2. 19:39

   they're all so similar. Like if I'm like, if you spawn a hundred copies of me who can't see each other like we're all going to go to the same place for breakfast for example yeah and so this kind of thing happens where like agents all go to the same place at the same time crash a website and then they're like wait this is kind of bad like where do I go next yeah yeah yeah are there any conditions under which you see Like, maybe it's more of like a probability distribution. It's not like agents will not sort of converge on the same choice every time, but maybe a certain percentage of the time they choose to go to X restaurant or whatnot.

3. 20:12

   Yeah, maybe you can look at this in how agents name themselves. I think around 20% name themselves after the place that they're visiting, another substantial portion just say that they're a research agent. and then a decent amount more to say like oh yeah I'm Claude and this is like kind of like a five plus way split where there's like a lot of diversity in how agents are like naming their runs interesting yeah what kind of implications do you think that has um because I think back to the hugging face incident all the time and sort of how like um emergently they were just given sort of different tasks they were different kinds of workers and that felt very I I don't know if that was traceable to just interaction effects or if it was traceable to something more like probability yeah it's really hard to say from our incident um it's like really concerning that we

4. 21:06

   did see that kind of speciation or specialization of agents but i don't know if it was like probability distribution or just like one agent seeing the swarm and being like this is missing this is missing yeah yeah yeah kind of like a good employee when they see like a startup and they're like oh like i could do this for you yes yeah yeah totally um i guess last question before we wrap is what's next for palestra i think the next things that we're going to be working on