AI That Rides Along With Truck Drivers

0:00We serve physical operations and we're a technology company that really partners with them. Most of them are in distributed environments. They probably operate large fleets and we help them basically be more safe, safer and more productive. Five years ago, they would have been talking about digital transformation. And that was a term that they used. And they've largely meant that like we need to like convert our old paper workflows to, you know, an app or something. You put yourself in a video game like, you know, you boot up in the screen. You just have like all the activity going on. on you can get perfect visibility on stuff out in the field, and you're actually sitting there focused on making decisions and having real judgment on, hey, what are my priorities for the day?

0:40How am I going to choose what to work on next or where to focus my energies? And I think we're starting to see that agents are able to help a lot with this and even automate a lot of those actions. Welcome, humans, to the Neuron AI Explained. I'm Corey, here with Grant as always. Hello, hello. Today, though, we're coming to you from Las Vegas, where we're at Samsara Beyond 2026. And we're joined today by John Bickett, CTO and co-founder of Samsara. John, before Samsara, you were the CTO, co-founder, and VP of engineering of Meraki, before you sold it to Cisco for$1.2 billion. And today, you've announced a lot of stuff at your keynote about new AI that you're applying to your ecosystem.

1:23You have Agent Studio that you've been working on. You also have, you've been using AI for your multi-camera tool to create the bird's eye view and stitch lots of video footage together. It's all really exciting stuff. We'd love to hear more about how you're doing that technically under the hood. Yeah. Awesome. John, welcome to the Neuron. Yeah, welcome. Thank you for having me. We're super excited to have you guys come out. And, you know, we've been really excited this week to, you know, engage with our customers, get some feedback, talk to them, launch a bunch of new features, and then kind of dream a little bit about where we could go with them.

1:53It's been a bunch, too. I guess to get started, because some of our readers may not be familiar with Samsara, tell us what you do. Help us understand, like, the industry you're in. Yeah, absolutely. So we serve physical operations. And we're a technology company that really goes to them, partners with them. Most of them are in distributed environments. They probably operate large fleets. And we help them basically be safer and more productive. And we have a series of products that we have. We have both like AI dash cams, telematics devices, asset trackers that kind of help them convert their like physical footprint over to a digital one.

2:32And then a lot of this year at our Beyond Conference has been layering AI on top of that, if that makes sense. Traditionally, you know, we have looked at these IoT devices built, you know, a lot of them have AI capabilities at the edge. So have GPUs embedded in them. Um, we run AI models there. Uh, we help, uh, monitor driver safety and other things. And then we help on the software side, uh, the companies that might have, you know, thousands, our typical customer probably has, customer probably has like, uh, two or 3000 drivers or technicians that are out in the field. Uh, they might be working on power lines, uh, repairing them.

3:10They might be, uh, doing, uh, field services work. Uh, you even see our stuff on, uh, semi trucks, commercial vehicles, all kinds of that stuff. And we help them run safety programs so that they can help train their drivers how to be more safe and drive down the number of accidents, as well as like, you know, be more fuel efficient and save money there and things like that. There was a line in the keynote that said something like, we want to help you see everything so you can act on it at scale. And I really believe from everything I'm seeing that that's exactly what you're doing. Great line. Yeah, no, I think it's kind of like the encapsulation of, you know, really, really, really what we're doing.

3:47Cause, and I think, you know, my background is like, I'm a wireless firmware engineer and had come from an academic background. So naturally, like if you want to improve something, like the first thing you need to do is get a baseline, measure it, and then come up with some theories about how you're going to improve it and then just run a process. And I think what's been cool for me, uh, uh, over the number of years is just see, actually, when we go out to talk to operations folks, that's kind of how they, they think about stuff. Like they need, you know, they need to have process around that.

4:13It's actually kind of a similar structure and thinking about it. And so it's been cool to partner with folks that are in risk management or operations and just like kind of try to fit into their framework on how they work through things. You know, yesterday at the keynote, as we were talking, we're talking about AI cameras, AI agent studio. We're talking about AI and almost everything you all touch. It feels like, like Sam Sarah has really leaned in to the AI space. And I was wondering if you could tell us a little bit about what that looks like and, you know, kind of the decisions that lead to making those choices and moves.

4:47Yeah, definitely. So I think we've been thinking about it a lot, like, over the past few years. And I think we, you know, had some of these because we've been using AI in the video space a lot. We kind of, like, you know, understood how some of, you know, some of it translated. It rhymes a little bit with, I think, everything we've been seeing with the large OOMs and multimodal walls and everything. I think what's been really fascinating, especially over the last year with our customers, is that everyone is aware, you know, especially the executives of all our companies. And I think, you know, a lot of the ones that are running the larger operations are, you know, if we talked to them five years ago, they would have been talking about digital transformation.

5:25And that was a term that they used. And they've largely meant that, like, we need to, like, convert our old paper workflows to, you know, an app or something. And, like, the AI stuff, it's, like, the obvious. you know, they're doing that, you know, now in spots. Are you seeing anyone leapfrog where like they were still manual and then now they're going straight to AI? Oh, definitely. Definitely some of them. And I think this is really fascinating about this space because I think what is kind of going to end up happening. And part of the reason we're working on a lot of the stuff we're working on, I think at the high level, all the execs are like, oh, we need to rethink things, if that makes sense.

5:57So you'll have like a COO or a CEO that's like, I got to transform my business with AI and how do we do it? But I think we also just, you know, when we're thinking about, you know, somebody with a thousand field service folks out in the field, they have a large backing of folks in the back office that help them. You know, it's a dispatcher. It could be a safety manager. It could just be their line manager and stuff like that. What we see in our systems is like, you know, they will have the drivers and the technicians and then we'll also have the folks that are using that other system. And they may be looking up things for compliance.

6:34They may be running reports and stuff like that. But it's a significant amount of their staff, if that makes sense. What we've ended up seeing is, like, often the CEOs will want to make – they'll be like, we need to make the jump to AI. And then they have, like, a large back office staff. They, you know, tactically, it's like, what do they do? Yeah, right. Because there's a lot of choices. Yeah. And, you know, we cover the AI space in our newsletter and our podcast. And there are a lot of different tools out there. It's like, I think Dave too from DCL Logistics yesterday said, you know, he's got SaaS fatigue.

7:07I think everyone trying to adopt AI right now is like trying to figure out, you know, what SaaS tools can they get rid of and what, you know, they want that kind of central source. Oh, totally. Yeah. And I think one of the biggest problems, especially once we think through like larger enterprises starting to roll out, is thinking through like the change management. All of the like classical enterprise features that we, like permissions, we should be able to pull reports for what and things like that. That's hard, yeah. and I think these are really interesting applications that only come up when you're like, okay, now how do I roll this out for several thousand people in a larger organization?

7:38If that makes sense. And, um, they're kind of like, you know, I think there's a, a, a really interesting bit. And for us, that's been awesome. And just to, just to be like, Oh, actually we have a lot of this context already in our system. Um, and then we can start building tools for the folks that are using it day to day. And that's a lot of the stuff we were doing with agent studio was basically like trying to give them the onboarding path, to using a lot of these tools where it's like, okay, hey, if we have a template already set up, you know, they can go through, they can try to use one of the tools that can generate a report, they can ask a question in like a low stakes environment, see it work, maybe customize it a little bit, and then just be on that, you know, that ramp up to really starting to use it.

8:21And because these are true power tools, right? Like, you know, but you know, you don't want to just slam on the cast all at once. But it's been so interesting just to talk to everybody and see how they're thinking about it, how they're using it in their organizations. Where did you start? Did you dive straight into the one task that everybody knows is the most inefficient, miserable thing to have to do? Nobody wants to do it. The golden prize would be turning this into a minor annoyance. Did you dive right in at that end? You know, yeah, it's really interesting. I think we basically, and I think there's kind of two things that have been going on.

8:58And it's been really interesting to have these. There's one, which is like, we have our customers, like, and how are they going to use your AI and things like that. We've also had like internal at our organization, especially in R &D, because all the new coding models, like more than anything else, they have changed. Like engineering is not the same as it was a year ago. So I think there's a lot of parallels though. And so we've been very carefully thinking about like, okay, well, when we see a pattern, kind of like you're saying, it's like, well, should we start with the really easy things to automate?

9:29How bold should we be for the first kind of applications? And frankly, we have been trying them all and seeing what works and double down on what works. Yeah. And if it doesn't work now, it just doesn't work now. It will. It is so. Give it another model cycle. and another model cycle, maybe even another year. Yeah, it's totally crazy. So it's very hard to kind of like dial it in and then continue down that path. But it's a lot of fun, especially if you love technology, if you love building things. I feel like, and this has been the case, I'm sure with all the folks you talk to, it's just like I feel like I haven't seen anything like this in my career, you know.

10:08What tools are you using internally? What's your favorite coding tool? What model are you using? You just max, opus max all the time. Like, where are you at with it? All right. Yeah. So let's go full nerd. Let's go full nerd, bro, man. So yeah, I honestly, I've tried a ton of stuff. I think I've, internally, we've actually ended up, we have used, and we've used Codex, we've used Quadcode, we've used Cursor. We have had them all, they all have their strengths and weaknesses, if that makes sense. And some of them, you know, it has been really interesting to go through and set up. And the cycle time on how fast they're changing is just really quite impressive.

10:52It is. I think we have those. So we've been doing a bunch of stuff on that. We've been looking at some of the open weight models internally just because, like, man, they really seem like in some cases they're three and six months behind. I know I was talking with one of our rangers. we were going back and forth looking at, I think the open router folks came up with a really interesting... Yes. For people who don't know, if I'm getting this correct, it combines all the different models that are like at least four or five different models, and it deploys them at the exact right time in order to...

11:26And you can customize which ones you want. It makes so much sense. And honestly, it's a thing some developers have been doing already. They've put it in a simple, single API call, all one bill. It's pretty neat. Yeah. Yeah, that's awesome. I think the other really interesting thing, just to try it out and just play around with it, I got one of the DGX Sparks. Oh, nice machine. How do you like it? Wait, what do you run on that? What's the biggest model you could run on that? I was trying to actually understand some of the throughput stuff, if that makes sense, and how it would work. uh you know and it's one of those things like it's it doesn't and i mean there are more efficient ways uh if you're just trying to max out like you know gpu per dollar that you can get yeah and get you know to the gpus and blue and then it'll probably it'll be a little bit better uh but it's just it's it's just such a nice developer platform and uh and you know uh and it's pretty pretty uh pretty cool and the other interesting thing is like you know if you're running the models in parallel that's actually when it really starts to shine if that makes sense so um i i mean know i feel like on the on the personal side i've totally you know geeked out what i ended up doing actually was like and i've heard a lot of folks doing this is the models are like actually incredible at like pulling a lot of information together so you know and i've been encouraged we've been internally at our organization like encouraging all of our engineering and r &d leaders like you should do a side project on ai just try something you know put together um one of the ones i was doing was like just personal health projects if that makes sense like oh cool um so i had like gone through uh and a couple years ago i had you know gotten my like dna sequenced uh but so i downloaded all that i downloaded um you know i have been i have an apple watch you can actually export like all of the sensor data from it that's interesting yeah um so it was like that i i dumped all of my health records out and essentially like you know went through got everything i could downloaded like the you know radiology imaging and everything else put it in like one space yeah and then like had claude go to town on like what analysis it could do uh and and help me like identify things that i wasn't you know it's just like i mean it's a level of control that that over your own health that i don't think people understand is even possible it's it's nuts it was just you know and it was surfacing stuff that like you know and it was the joints that was really impressive that i don't think like as in if it has all my apple health data it's showing how i'm doing like your year and what changes were going on.

13:56And then it's kind of correlating that with like, you know, Hey, if, if you have any, uh, cause all the, uh, doctor's notes are transcribed and, uh, you know, they show up in the, in the record. So if you can get a dump on them, we'll do that. And, um, I think for, you know, and it's totally interesting to work on these side projects, right. Cause it gives you, and even for working with our customers, it gives us a lot of ideas. Cause we're kind of like, Oh wait, like it's kind of the same story. Uh, you know, I started going through and I'm like, well, what is the equivalent of a trace like this or getting all the Apple health data?

14:26I'm like, wait a minute. Yeah, we have the records on like when our customers are visiting a location, like how long it takes every time. And we can join that with a bunch of other data. Historical driving data. Totally. Yeah. And then maybe it's actually, we can probably look at the map and if the LM looks at all of these things at the same time. Like, so I think, you know, it's kind of one of these things like we're in such a crazy era of just playing around to generate so many interesting ideas in other spaces that is super exciting. It's a level of context that a human is never is not able to piece together and recognize patterns at that level.

15:00And it's a really cool thing. One of the things I was impressed with with Agent Studio was how flexible it is so that all of the different types of customers that you have can probably come up with an agent that works well for them. And something we haven't talked about yet is that you, I don't know if this is totally new or if you've had this for a while, but the cameras that you install on vehicles for your customers, now the agent can talk to the driver. And you can actually create an agent and you can have it trigger. That was so awesome. Dylan yesterday, it can trigger based on an event that happens, right?

15:35And so, like, if there's, like, a G4, and you can tell us more about this, but if there's, like, a G4 event that happens, like, that could indicate that there was a crash, you could have the agent trigger at that moment and, like, check in, like, hi, like, literally call them on the camera and say, how are you doing? Yeah, absolutely. Like, this stuff, I think, is kind of, you know, is kind of game-changing and really, really kind of interesting to think about. We've had a bunch of, like, and it's kind of interesting talking about those applications, like, you know, hey, yeah, if somebody needs some assistance or something and it's off hours, like, that's, like, a perfect thing that, like, yeah, you know, you can probably codify some rules on, like, when to escalate, you know, to a person or something like that.

16:14But that's actually, like, really helpful, if that makes sense. And I think there's a bunch of use cases like that, a bunch of, like, kind of what I would call kind of like you were referencing earlier, mundane things, if that makes sense. But you can imagine, you know, somebody, a driver, maybe that is a temporary driver or something or like that was taking over somebody else's shift is pulling into a new location. Yeah. Like that might be an appropriate time to just trigger and to be like, oh, reminder, hey, the gate is here. We're starting to approach. And that might, you know, it's not a big deal, but it might save somebody like 10 minutes.

16:45Oh, totally. You could be driving around the parking lot otherwise. And you amortize that over a year and suddenly it's a big amount of time. Yeah. The fuel cost alone. Yeah. You know? Something I'm curious about is our readers and viewers are quite familiar with agents in a general sense. But what I want to know is what are the different challenges you face with agents in the physical operation space? Yeah, definitely. I think, you know, one big thing I think, especially for our users, is that trying to make it easy for them to understand what's possible and what's not. And that was a big driver, I think, for us behind Agent Studio was like, hey, here's a bunch of kind of off the shelf ideas that you could kind of take that we've been through.

17:26We've seen other customers use and would be great ideas. I think probably the biggest thing that I've been most excited about with the AI developments in general is that you get so much intent from what people are trying to do. And I think in general, if you can take that intent and then of the top most capable users, you can easily find that in the system. Kind of like the top 1 % of AI, you know, AI-pilled folks that are using your system. If you can bring that to the, you know, median or average person in a packaged up way, that's like, I mean, that's amazing. And it's kind of a feedback. First, you'll blow their mind like it's magic.

18:14That absolutely is part of it. I mean, that's what some of the stuff feels like, right? Yeah. And, you know, I think I was, someone was, you know, I was working with a bunch of the developers on several different projects. And lately, it really does start to almost feel like sorcery, if that makes sense. Because you're just like, I'm going to cast a spell, right? Like, that's the bit of the joke that people have been, right? But that's what it ends up being. And you're batting around ideas on how to build things and build customized, you know, agents or solutions for you. And that was just so, so different.

18:45There's that old Isaac Asimov quote that any sufficient technology is indistinguishable from magic. Yeah. And I've butchered the quote, but the idea is like that, where you struggle to even understand, like, how did we get from me doing this this way to that? And it's really amazing. How long did it take to put this together? Like, is this a project you'll have been working on for two years? Is this a project that came together quickly? Yeah, I think it's been interesting. So we probably started working on the assistant, I want to say, two years ago, if that makes sense, on just the beginnings of it.

19:21And models were not as capable then. And I think, you know, a lot of the, I mean, there's just been so many advances, both on understanding how to build these and like what they can do and even the capabilities. Right. Um, we really, uh, I think things started to come together in January and that was when we were like, okay, we understand. We probably, uh, like an articulation of like what, um, moving from an assistant to like, uh, Hey, how can we have templates ready? What kind of templates would it be? Do we want it to be chat? Do we want it to be voice related? Uh, we were bringing in a lot of threads together.

19:55Cause that was when we actually started the piece together. well what if we hooked this up to the audio on the dash it's like yeah can we have uh an agent do the call because we had separately been working on two-way yeah for for people but you know all of a sudden these building blocks like snapped together in a way that like was a little bit suddenly you're running off a mobile signal and that's way easier you know and totally yeah so that uh you know that but that that stuff i think the vision started to come together and we earnestly started working on agent studio in january and piecing a lot of that stuff together.

20:28Wow. And it's like six months later, it's here. Yeah. Yeah. And, and, and solid. I think, uh, I think you built a really, really good tool. I, something you called out, right? I guess something that was called out yesterday in the, in the keynote was this idea of, of meeting people where they are. And, and I think that the best place to do that is in an agent project like what you've built, like the Agile Studio. And I'm curious in thinking of who your customers are, of how they approach problems, of what their experience with artificial intelligence might even be. Like, how do you design a product that you feel like they can do this?

21:08And anybody that does this is going to be able to figure this out and make it work. Yeah, that's a good question. I think we, and I think it's one of the big parts of our cultural DNA is actually just being very customer focused. So, you know, our team has actually just been going out, partnering with customers, trying applications. And we generally like actually have a rule at our conferences. Like we won't show something like on the main stage unless we've actually had a customer successfully using it. Yeah. If that makes sense. I like that. And yeah, I was really impressed by that. Like everything that you release, you're like, yeah, and we tested this with like multiple companies.

21:46And here they are at works. Yeah. And you could talk to them. Yeah, I was very impressed with that. So I think, you know, that process has just been super helpful for actually just like getting down to the tactics about like actually what does something need to be look like? Like, did it work? You know, when you if you can listen to the calls with the customers and just be like, was this what you expected? You know, does it sound right? Stuff like that. Or the the the chats like, did it generate the right data that you need? Does it is it the right report? Like, you know, how is it working? And stuff like that's like just it makes the process a lot easier, I think.

22:20Yeah. How are you finding they're using it? Has getting it out there, has it done anything that surprised you really quick right off the bat? Yeah, totally. So I think one thing is like whenever, you know, these tools are just so different because it is the pure intent on like what the users are wanting to do. And they'll tell you, they'll ask the AI to do it. And I think everybody goes through the process where they like just try to, you know, they start ramping up. Yep. And, you know, they immediately try to give it the very complex problem, if that makes sense. Yeah. And then. I resemble that one.

22:47Um, uh, so, and so it's really interesting to see like what they, what they end up doing. We've had a lot of cases where, you know, kind of like you were alluding to earlier, it was like, if you, if, if an AI has a ton of data, it can actually kind of just chew through a lot of it and probably surface insights that, you know, I feel like I've looked at a bunch of stuff and I'm like, I'm not sure I could have pieced that together. And a lot of it actually had been, you know, we've even seen users, uh, use the assistant tools to debug, uh, cases like, Oh, somebody is having trouble logging in. If that makes sense, that's like one of the super admins.

23:17They've been able to go in and actually deduce, oh, in fact, I talked to them 10 minutes ago, but they have successfully logged in since then. So it's like not, you know, not an issue. Or it would just infer things from logs. And these are other applications that we've seen OLMs do very good at. It's super cool to see them kind of in a little bit of a different context with these operational users. And people don't necessarily even think that, you know, about these kind of applications. Uh, but we've seen people do a ton of reporting stuff. Uh, and that, that's been a huge, uh, time saver, uh, a bunch of data analysis, uh, things asking for recommendations, uh, basically trying to figure out like, how do I coach some, someone's driver or can, you know, asking the agent to go through and just dig through a ton of information to generate summaries.

24:02Uh, they're just super good at, you know, going and doing that stuff. And it's just saving people a ton of time that otherwise manually they would have had to do. Yeah, two of the use cases that really impressed me in that front was the ability to surface and deal with warranty information. So you can save a lot of people who are, you know, fleet managers, you know, have a lot of different vehicles, help save them money just from their warranties that they just wouldn't have filled out the paperwork to do. That was really impressive to me. And also all of the error code managing. The error code thing was very good.

24:35Yeah. I want that in my car. I worked in the GD in every car. Like so many of the features you released yesterday, I was like, this should be in every car. I want the cameras on my car. I want the, yeah. So we've had a ton of really, one of the awesome things about at Samsara is we just have like so much data on this. We have had the Telpatics, you know, for a long time. So we have ended up amassing a bunch of data on this because, you know, we've been running it for 10 years. We have all these make model years. I think one of the most fascinating things about that is that we, you know, kind of have all this adjacent data because we have all of the fault codes that are coming out.

25:10But then once we go through it, we can actually, we're like, oh, but, you know, we know how much that vehicle has been used. And then we can compare that to the fleet of vehicles that are the same, you know, make model year and go through and actually understand like, oh, well, for this vehicle, is it in the top 1 % of vehicles that have been driven? Or, you know, is it, or is it about average? Like how much wear and tear is it getting?