NOT a Swarm!  ## Metadata - Author: [[The Changelog: Software Development, Open Source]] - Full Title: NOT a Swarm! - Category: #podcasts - URL: https://share.snipd.com/episode/73879197-2012-413f-a385-abc01b944f97 ## Document Notes ## Highlights - **Mini Models And Agents Win** - The future is a mosaic of many specialized models and agents rather than one monolithic model. - Chris Benson expects a 'mini model' and 'mini agent' future where multiple models collaborate to solve tasks. Transcript: Chris Benson You're seeing that a lot of the it's not just a model thing, but also some of the infrastructure that's being built around it has given it a much has made it much more accessible in terms Of its productivity and its usefulness. And there's less of a friction when we're trying to use models at this point. So I do think that there that like there is no infinite rise on terms of the number of parameters you have to do. I think I think that that does level out. And also, if you're going to have that many parameters, being able to use that productively from an inference standpoint, the world is turning out to be a mini model world instead of A giant model world. And I'm not sure that a lot of people in the general public, you know, that aren't people like us that follow this closely really realize that. I think when they think AI, they're thinking chat GPT because it's what they know. And, you know, one model to rule them all, one model to bind us. And like, I'm not at all like that's not what I think is the world. I think the world is is many, many models contribute to solving a problem in various ways. And here we are in 2025, deeply into the age of agents. And so it's no longer just models, but now agents with models that are acting on your behalf. ([Time 0:14:21](https://share.snipd.com/snip/972c28b9-18a3-43cf-b702-0d35f61641ff)) - **Physical AI Is The Next Revolution** - Robotics, AI, and hardware advances are converging to make physical AI pervasive and accessible. - Benson calls this the coolest moment in human history for makers and startups building physical agents. Transcript: Chris Benson And I think we live in the coolest moment in human history right now. Like we are sitting in it as we speak today. And, and so I agree what, what's happening right now is with all of these different relevant, you know, capabilities, you know, the robot people and the AI people and the software people And the hardware people, all coming together. And you can do amazing stuff today that, that even a year ago we couldn't do. I mean, it's like, if you think about before now we would, we'd had kind of have several years of little software eras, you know, and we were, we were getting into certain ecosystems with A language or whatever, and they kind of run for a few years, but right now it's changing so fast and the capability is coming so fast that like, aside from the big 800 pound gorilla types And stuff, like everybody can get into this stuff. And so I think we're at a moment right now where like, it's really going to start being pervasive in everyone's life in a bigger way than it has been. Not just like I'm going to open my phone up and talk to chat GPT kind of way. Because yeah, I mean, that was unimaginable. If you think about it just a few years ago, it wasn't, it hasn't been long since that was an unimaginably amazing thing to do. But that's like, we don't even think about that now. You know, we do it all the time. Don't even think about it now. But like physical AI and the fact that robotics have come so far in the last few years and that now there are, in addition to NVIDIA, there are many other chip makers that are coming on scene To support AI. And some of them are doing more of the dedicated AI chips and others are doing more like, you different types of chips so that you have that. And some are great for data centers, big cloud data centers, and others are great for edge devices and tiny little constructs. And I think you're going to see so much happening in the marketplace right now that are coming from startups. ([Time 0:20:43](https://share.snipd.com/snip/d52d44a1-e556-4339-b3a7-51b113320ae0)) - **Automation-Inherited House Sparked Tinkering** - Chris moved into a house with extensive home automation and that jumpstarted his tinkering and projects. - Inheriting the setup accelerated his shift from professional interest to personal home automation maker. Transcript: Chris Benson So about a year ago now, almost exactly a year ago, we bought and moved into the house that I'm in now. And the guy that we bought it from, he and his wife, he was a fanatical home automation person. And so like, um, we moved in, not because of the automation that was incidental, but like it's had, it's helped me move from just like thinking like more of a professional kind of thing. Like, you know, we're talking AI and a professional kind of to thinking about stuff around the house, um, with all the sensors and the cameras and stuff. And, um, we have all the, you know, the, the types of home automation stuff that you see out there combined, CASAs here. Like we have many, many, many dozens of CASA devices all over the place. Adam Stacoviak And CASA is the brand from Lutron. Is that right? Am I picking that up? Right. Chris Benson It's from TP-Link, actually. TP-Link. But that's just one. There's a whole bunch of them in, you know, Apple Home and Google Home. I was thinking Casita. Casita's from Lutron. Those are the light switches. Yeah, the Lutron does light switches. But there's some common protocols that they all work on. And I'm starting to see, because I didn't have to go start it from scratch and because I inherited what this guy had already kind of put together and then had to figure it out and make it Work. And suddenly I'm like, oh, gosh, it would be really easy to add this. And we're talking about this robotic future, even in our homes, not just a commercial or industrial or whatever thing. But in our homes, like it's so easy for me to see that now, because like I realize I already have a good bit of infrastructure here and it's not expensive and it's not it just takes a little Bit of effort. ([Time 0:29:47](https://share.snipd.com/snip/42141665-e80e-462e-a31c-08f4e02ae1b0)) - **Swarm Is Behavior Not Quantity** - A swarm is defined by behavior: independent autonomous agents exhibiting coordinated, emergent action toward a collective mission. - Chris Benson emphasizes swarm-level intent and emergent decisioning, not mere proximity or preprogrammed fleets. Transcript: Chris Benson You are. And it's not just a physical distance thing, because what is physical distance is a relative thing, depending on what it is you're trying to do. But it also, it's really more about behavior. And so we can dive into that. But before you say that, I think that's a line of thought we should go down is that as you guys know, I'm really into animals. We were making jokes earlier about a bazillion dogs and stuff like that. I'm a licensed wildlife rehabber and I study animals. And in the context of swarming, Mother Nature has perfected not just swarming, but there are many different types of swarming from different species. And so I have a set of species that I tend to look to for swarming purposes and say, if I want to swarm with this type of technology or this type of platform, like how do we get started on that? You know, how do we get inspiration or look for some insights on the technology where you can look to certain species that are, that are similar to the technology platforms you're interested In terms of how they move around and do stuff and say, well, how, how has nature solved it there? And I definitely do that a lot. I, it's not uncommon for me to go into tech meetings and start off with lots of pictures of animals and stuff. Jerod Santo And people are like, what's going on with this? What is this guy? Chris Benson Are you thinking like fungus, bees? Bees, bats. Like I do a lot of bees, bats, birds, starlings, you know, those huge what are called murmurations of starlings. Ants are awesome. Ants are awesome. When I'm thinking about robotics on the ground, meaning what we would call a UGV, which is an unmanned or uncrewed ground vehicle, ants are amazing in what they can do. And so they're an awesome thing to look at. But I'll start with the definition that I use. Given the fact that I work in the, in the military intelligence space, my definition sounds kind of, it uses that jargon, but it really don't get caught up in that. It can be applied to residential. It can be applied to commercial. It can be applied to industrial. So don't get caught up in this specific wording. So I'm going to read it in front of me. It's one really long run on sentence. It's very specific in what it's trying to imply. It is swarming occurs when numerous independent, fully autonomous, multi-agentic platforms exhibit highly coordinated locomotive and emergent behaviors with agency and self-governance In any domain, which could be air, ground, sea, undersea, or space, decisioning entity for purposes of C3, which is command control and communications with human operators on the Loop to implement actions that achieve strategic, tactical, or operational effects in the furtherance of a mission. So long, long, long sentence, but it hits a bunch of very precise concepts and integrates them in together. Jerod Santo I can tell each word was selected there. Chris Benson Yeah. A mission might be, instead of thinking military, a mission might be, uh, getting a package to your house. That might be the mission. And that does have command control and communications involved. So like you can put the, you can, it doesn't have to be the, the military-esque, uh, jargon that we're talking about yeah yeah it applies to any of these to any of these you know commercial Industrial residential military whatever so so that's a lot it's a lot it's a lot and if you want i can kind of break down high level what some of those mean yeah i think my broad takeaway Jerod Santo We can talk about the individual words because i know they're very specifically chosen like independent logical distributed decentralized decisioning entity like stuff like that I can tell each word was selected for a reason yeah i think my grand takeaway of a swarm is kind of the e pluribus unum like it's like okay all these things are individual and autonomous, But they're all acting as one. They're acting with one purpose. Chris Benson That's a fantastic insight that you have. And that, that is the key to it is like swarm is such a buzzword. You know, we, we always have buzzwords in, in this AI and software spaces. There's always the buzzwords of the year and swarm is certainly a huge buzzword right now. And, and almost without exception, I will turn around until I can go back to my definition, assuming that you want to accept that as the definition of swarming. And I can defend that fiercely. I cannot attack it. Jerod Santo Can you attack it, Adam? Chris Benson You're the expert here, Chris. But I would say like all you people who are talking swarming, no, you're not. It's not swarming. What you're describing is all sorts of things that lead to swarming. There's a whole bunch of incremental capabilities that that would eventually, as you add all those capabilities together, they culminate in swarming. Right. But but the chances of somebody saying that what they're doing out there is consistent with Chris Benson's definition of swarming is pretty low. So what you said was right on. And that is that just as you see in nature with those ants, every little ant has its neural capability, shall we call it, and what it's doing. But at the end of the day, they're functioning to get a mission done, a job done, something productive for the colony. And they are all lending themselves to that greater good, even if some of them may not survive, that kind of thing. They are functioning as a single entity, and it is the entity that's trying to get the thing done, not the individual ants. The individual ant may be like we have a crack in the ground, and we have to get from this side to that side, and they build an ant bridge. We've seen pictures of that. And that one job, one little ant may have the job of, I'm holding on to the ant on this side, and the ant below me is holding on there. And then they have that going on as well. And we're all creating this ant bridge over a chasm that none of us individually could span. But by working together for that swarm approach, which is make that accessible, they are doing something well beyond what any of the individuals can do. They are super ants in that way. And that's what I'm getting at is that ability to give up your individual identity as a member of a swarm for the purpose of the overall swarm's intent. And that swarm itself has an intent that is a swarm level thing, kind of to your point, Jared. ([Time 0:37:03](https://share.snipd.com/snip/8cb19d7b-31c6-4274-beda-9d4dee4dc9c9)) - **Swarms Combine Mass And Maneuver** - Swarming delivers mass-like power with hyper-maneuverability, creating capabilities neither mass nor maneuver alone achieve. - Benson frames swarms as a new military and commercial capability that changes what is tactically possible. Transcript: Chris Benson And if you go back in War 5, many people would say, OK, let's build up mass to win against an enemy. And then as things progress forward, we learned that maneuver could could kind of out, you know, you could go around mass and you could hit it from different ways. With mass. But swarming becomes like a whole new thing is that you're kind of getting the best of mass at individual small scale, but you're getting mass and you're getting hyper maneuverability. And so it's able to trump that. So in that domain and that kind of military world, it brings about a whole new capability that it never existed before. And similarly, when you move into commercial and industrial, and we talked about this super automated house a few minutes ago, you're bringing about things that just were not possible Before. ([Time 0:46:50](https://share.snipd.com/snip/0e5bf223-54ea-4bea-8743-a02e4f970faf)) - **Most 'Swarms' Are Fleets** - Many systems called swarms today are actually fleets: preprogrammed units with limited dynamic collaboration. - True swarms make real-time, emergent decisions as a single entity rather than follow fixed trajectories. Transcript: Chris Benson I think a lot of the times where people throw up a whole bunch of things like drones, that's the thing everyone knows, we'll throw up a whole bunch of drones in the air. But it's not really a swarm. It's a fleet of drones. That's what it is. And each one requires individual programming to go do this or do that. There may be some communication between them potentially, depending on what they're doing, but they're not thinking almost like a brain, like an abstract brain themselves. They're not looking and dynamically handling what's happening in the real world in real time and saying, this is changing right here, right now. As a swarm, I'm going to go do that. They can't do that. They're fleets. They can respond, but it's going to take inputs. It's going to take some collaboration between them, but it's going to take a lot of guidance, you know, from afar to make that happen. ([Time 0:48:15](https://share.snipd.com/snip/1975298e-ffc5-401e-a905-d67fc3d51158)) - **Keep Humans On The Loop** - Implement human-on-the-loop supervision and explicit guardrails for physical swarms to limit unsafe emergent actions. - Keep kill switches and oversight mechanisms to prevent unacceptable autonomous changes. Transcript: Chris Benson I know. I think there's a lot, and I think it's at many different levels. And it's a it's a real thing that we talk about in real life today without having achieved full Chris Benson level drone swarming. And that, you know, we talk about that in terms of A.I. Safety all the time now. You know, that's a huge part of the A.I. World is what is A.I. Safety? How do you keep unintended consequences from coming to pass? I think anyone who's reasonable recognize that some of those will still come to pass out there. You can put guardrails around things, but, and you can even ask AI to put guardrails around other AIs as we're doing, you know, because we're using the tool to build the tool, but we will Have bad outcomes across the board, just as we always have with software and always will. And so I don't have the magic bullet on that, but there is training the distributed swarm brain, you know, this abstraction of computing, of grid computing, where they're all doing This and using their algorithms. And that will, where it goes wrong may happen different places. You know, we often talk about today's LLMs coming out with inferences that are suboptimal, sometimes quite funny, sometimes quite tragic, actually. But that will continue to happen. We have software issues. We're also moving into the physical world where if you have these physical agents that are imbued with a whole bunch of AI agents that are doing stuff and they're acting as a member of A larger swarm, there's a lot of places where things can go wrong. So it's going to be, it's, there's going to be a learning curve on that. And we're going to have, we're going to have problems along the way. So I don't want to, I certainly wouldn't want, I know for a lot of listeners and viewers, they probably think of a little, you know, a little bit pie in the sky. ([Time 1:12:12](https://share.snipd.com/snip/2292d78c-99b1-4b92-983b-31f476d12741)) - **Focus Swarms On Resource Savings** - Target practical, high-impact household problems first: energy, water, and food waste as initial swarm use cases. - Design swarms to coordinate devices and agents to conserve resources and reduce waste at home or neighborhood scale. Transcript: Adam Stacoviak I think the area where I'd like to see this type of swarming is in energy conservation. And so I think there's multiple devices in my house that consumes energy from a HVAC system above me that both heats and cools my home to the lights that power my house to, let's say, a kettle That is electrified. All the things. I want to give my home the task of being energy conservative. Yes. This swarm, I want to have a swarm of devices that help me be that. And it can, hey, Adam and the family are not here. It makes sense as an agency to be conservative with our energy use because there's no one here to do it. And that's where you can do individual device level smart home automation, which is here today. Chris Benson Matter supports that. Adam Stacoviak It's not a swarm though, right? It's not a swarm. That's right. So I would like energy conservation to be my first swarm tactic. The next would be I live in Dripping Springs, Texas, just outside of Austin, Texas, and we always have water challenges. Right now, we're always in some version of a drought. There's actually a big bet on Wall Street against Texas running out of water. Like, there's a bet essentially shorting Texas running out of water at some point. I just heard this headline. That's headline only. I don't even know what the truth is behind that, but I heard it. So it must be true. Okay. So the next thing is water conservation. Help me as a household, maybe even help me as a neighborhood, a swarm neighborhood, be conservative when it comes to water conservation. So my child goes to flush the toilet or some sort of action tries to take place. But the swarm is like, hang on a second. We're in a conservative nature. We're going to use the one or the 0.5 gallon version flush versus the 1.2 because it's a number two. Some reason, right? But for whatever reason, like we now have new tech in my house. So that gives me things that really matter. Energy, water. And I think the last one for me is food. There is so much food waste in America, tremendous amount. I know I for sure buy some chicken once, twice a month, and I'm killing chickens constantly because I'm wasting my chicken not making it. So I don't know if that's a problem that's me, but at some point my tech, my swarm tech, can help me solve those three key things, energy, water and food. And I think you start there because that's what matters. My laundry kind of a me problem. Maybe my washer can say, hey, you put a white in with a darks. Probably not smart. Eject it or alert me. You know what I'm saying? But like I don't need help with laundry. I don't need I mean, I like my iRobot and vacuuming. That's cool. But I think that the thing I would want to conserve on is those three things. ([Time 1:19:10](https://share.snipd.com/snip/bde979ac-f023-41a5-b48e-5528fbad22dc)) - **Start Small With Open Tooling** - Start swarming work via open-source tooling like ROS2 and community projects on GitHub instead of reinventing the stack. - Use small models from Hugging Face and inexpensive edge hardware to prototype local, privacy-preserving agents. Transcript: Chris Benson We can add some stuff in. Some of the things that I often tell people to start off on is, you know, robotics has been a big part of this kind of robotics role, you know, being a part of developing to the swarm is ROS2 Exists. ROS stands for ROS2. ROS is the robotic operating system, which is open source. And it is the most widely used robotic software technology out there. It's not the only one. There are many, and some of them are closed and some of them are open. But there's tons of books now on ROS. And so I often, when people are interested in this and they're like, but how do you do, like, aside from the swarm, I can't even make a single robot or like, what do I do? Well, there's tons of information about that. Start off maybe not solving the overall swarming problem that we that we were describing as being as remaining a hard challenge. But start with something more accessible. You can buy you can get on to, you know, you know, we mentioned Bezos so much, Amazon and others. And there are a lot of maker kits that you can get that are open maker kits. You have ROS. They're very similar in terms of, but if you want to not do robots and you want to do drones, there's a whole bunch of open source drone stuff. Then the thing that I love doing, I do this all the time, is diving in on GitHub at different software communities that support, you know, open specs and stuff. There's tons of repositories on GitHub that are designed to do this, that just interested people said, I want to go scratch an itch. I want to solve a problem. And I go there and I'll then also go to Hugging Face and look for small models that may, if I need AI in the mix, that can contribute. ([Time 1:24:31](https://share.snipd.com/snip/f56e2320-60a0-49ec-8cb3-e9e6d036603d)) - **Prototype Swarms In The Home** - Build home automation prototypes with inexpensive Raspberry Pis and local models to explore agent and swarm behaviors. - Combine Matter-enabled devices, local inference, and agent logic to achieve useful automation without cloud dependence. Transcript: Chris Benson There's the swarming stuff that we've talked about and trying to figure out robotics and all that. And we've talked about home automation. And I think that feels, for answering this question, that's an accessible thing that I like to do now. So as I've picked up this kind of home automation stuff, I'm trying to figure out what can I do? I go get some Raspberry Pis, or I can use a slightly larger, you know, like a mini PC to do something in the house. None of this costs much. And I'm now on my day to day when I'm just at home, and I'm not thinking about the day job, if you will, I'm looking at all the things that I do with my family and thinking, wow, you know, like I can go pick something to handle that. So like almost all the lights in our house are automated. A lot of the appliances are automated. We have voice command, you know, from anywhere in the house where we can, we can, you know, tell a particular assistant, go do this. And it happens. I've been starting to integrate AI agents into that workflow. Now that that is becoming super accessible with all the there's so much open source that have made agents very easy to do. And you can get small models off hugging face and run it off compute that you have in your house already. And so that's that's the kind of thing that I like to do. ([Time 1:30:52](https://share.snipd.com/snip/ff5e9c9c-bf4c-42bb-bd5b-47fdb9da569e))