Hunter Dunbar

I am the chief product officer of Luxonis.

We make Edge AI-enabled hardware, and we make... We're effectively like the perception layer for physical AI. So we allow machines to see here. We give them the ability to think.

And we're solving some of the hardest problems that you can solve today.

That's right. We're enabling AI to get into the physical world. So,

to be the case that it got loose, yeah, we would be a part of that. But I'm not too worried about it.

Yeah, so we use specialized chips that can take big models and then condense them down into small models and put them out in the world, running at very low power consumption at high inference rates. So you can run a model of 30 or 60 FPS.

And based on what the camera is seeing, it can make decisions about stop this machine, move this lever, move this arm, or just provide data to bigger systems. So it's kind of like a neural...

A nervous system that is collecting data and making decisions collectively.

So there's a process in which you... It's called quantize. So things get really world in the quantum world when you're taking a really big model, like some of the traditional LLMs and condensing them down to make them smaller so they can run on much more energy and memory constrained hardware. And that's kind of the world that we play in. So you're using the same base model, but you're reducing in size and there's trade-offs with that, but that allows you to put it into the real world. When you're talking to chat GPT or Clod or whatever it is, all that information goes to a large data center, but you can't carry a data center around in your pocket. You can't carry a data center around when you're disconnected from wifi, from cellular, or 200 feet below the ocean, things like that.

That's right.

Yeah.

kind of thing? Yeah, and this is... We have a big software component to our business where we have a big team that can take open source models and take real world simulation data, mash those together to make the models smarter and see what they're looking at. There's kind of basic models that can differentiate between different basic objects. There's models that can differentiate between just shapes and colors and based on how far away they are. But we can kind of combine all of that together to give a camera the ability to sense what's in a room the way that a human would, which is, that's a camera, it's roughly this far away.

But humans actually are amazing at doing this, but we don't actually know how far away that camera is. If I wanted to grab it with my hand, I wouldn't be able to know exactly it's 6.2 meters, but a camera needs to know that in order for a robot to go do that. So that's kind of the world we live in right now.

Yeah.

So we use stereo vision and that is two cameras using some math and some neural net nets to determine how far away things are. We're also blending in other models that just use a single camera. We're blending in some other sensors like a TOF module, which is a time of flight module. Other sensors that we kind of fuse together again in the device, depending on the device application. The answer in general is like, it depends.

We're now seeing, for the longest time in computer vision and in AI in general, we were seeing lots of exciting demos in labs. And I think now we're really seeing, it's like the dawn of those laboratory experiments getting into the real world. Specific to AI models, generative AI getting out into the world and actually being useful, being able to drive ROI for businesses. Because ultimately as a business, we're trying to solve problems that can help businesses deliver value to customers.

I think a lot of people historically have been able to demonstrate really cool things that kind of give you this like, "Oh my gosh, AI is amazing." But now we're actually, we are living it every day where we're deploying things at scale for customers across a very wide fleet of things that are providing value, solving problems,

and keeping people out of harm's way. One of our main tenets is to help machines stop killing humans because it happens every single day across a whole fleet of industries.

And a lot of these machines are still analog, they're still driven by humans that have lots of error. They're designed in a way that when they were designed, they didn't have the technology that was an economic place where you could put that kind of technology to prevent humans from being harmed. And now we're seeing that that's not possible. That's super exciting.

Yeah, safety devices are becoming a lot more intelligent. Previously, it was just like you would have sensors that was like, "Okay, if an object got too close, it'll turn off the machine." But lots of objects are going to get close and they may not, like you're going to be running over tall grass, you're going to have different objects flying in front of tractors. And if your tractor is shutting down every five minutes because it has false positives, they're just going to rip the sensor out. But with cameras and with other sensors, you can get a lot more accurate so that it's actually

reduced the false positive rate while making the machine safer. So that's like a specific example for safety machines, tractors, things like that.

We are across many industries. So heavy machinery, security,

we're in farming in all different types of flavors. Ag-Tech is a big one for us.

We have a big education arm, so a lot of our technology is used in universities all across the world. The Navy uses us. We have a lot of big military customers. We don't work with the Department of Defense directly, but lots of companies use our technology in that.

We have a big vertical within manufacturing, logistics, retail. So basically any environment known to man on earth, our camera can be found.

Yeah. So I'm not afraid that AI is going to kill us all and that our cameras are somehow going to play a role. I'm right now currently afraid of these hype cycles. I think that there's a ton of attention and excitement around AI can do all these things. If you scroll Instagram or YouTube, you're going to see all of the Mr. Beast style tiles. It's like everything's changed again forever.

That's not really helpful for us who are really just trying to be selling value and building things that create value because things are changing, but not as much as people are hyping them to be. That's why we're seeing so much volatility in today's workplace stock market where you're seeing all this uncertainty in what's happening. I think those hype cycles can be distracting.

We are just trying to somewhat keep our heads down and focus on our mission and what we're trying to do. That keeps us isolated from some of that, but it is hard not to stick your head up and be like, "Oh my gosh, there's so much hype going on and so much uncertainty that it confuses some of our customers, it confuses some of our investors, and it creates problems."

The thing that I think is not particularly helpful.

Yeah.

We have about 100 people and we're distributed all across the globe.

We have all the disciplines you need to do hardware and software.

We employ most software engineers and that includes cloud, AI, computer vision, firmware,

front end, back end, the whole kit and caboodle.

We design and manufacture our own hardware and so we have supply chain team. We've got firmware team, mechanical engineering, electrical engineering, all of the things you need to do, industrial design, great hardware, good design.

And then we have sales operations and products here, primarily based in the US, but most of our team is distributed all over the world.

conductors. Most of our products are manufactured either in Taiwan or the US currently.

It takes a while. It's going to take a while. It's going to be four to six times more expensive so the customers are willing to pay that great, but most are not.

We'll see. I think there's a lot of the world's GDP is flowing through that part of the world and I think that there's a lot at risk for everybody involved that that were to just all of a sudden turn off.

That's actually not something that keeps you up at night. I think cooler heads will prevail there.

The geopolitical stuff is hard and it's hard for stupid reasons.

We've been playing the whack-a-mole tariff game where it's like, "Okay, this is a good country. That's a bad country," and then it flips.

Now we're living back now in April of 2025 when it's like, "Nobody fucking knows what's going to happen with tariffs," and that just makes it hard to plan long-term, invest long-term. Luckily, we've placed some bets a long time ago that we're good now, but they very well could have not been.

Yeah, yeah, yeah, yeah. We've had to make some very ... Yeah, the supply chain thing has been very hard, but

we've hired to solve for those problems. We've been able to over the past ... Let's see. If companies found out in 2019, over the past six, seven years, we've learned how to navigate these waters. The past six, seven years have not exactly been calm waters in that part of the industry. So yeah, we've learned a lot through that time.

Yeah, I think hiring has not slowed down. We're continuing to hire and continuing to grow. I think we're being more selective in who we bring in so that they can benefit from all the tools that are available. Because I think we've taken a firm stance that if you're not using any kind of AI-generated tools in your workflows as a developer, that this just isn't the company for you. I don't think anybody has resisted that. We made it a part of our workflow. And then I think most importantly, because we're an enabler for a lot of people in the whole world, so a lot of people take our products, our hardware and our software, and build other products on top of that. We're incorporating a lot of AI tools to allow those people to also build faster. And so we're eating the dog food ourselves, but we're incorporating that into our products so that our goal is to sell as much hardware as we can to help the world solve all these problems as fast as we can. And I think we see AI being a really good accelerant of that.

Yeah.

So the company was founded on that principle. When we founded in 2019, we hired anybody and everybody. I think we have employees across 13 or 14 different countries. It's been true since we were founded. It's kind of in our DNA, and we just ruthlessly prioritize meritocracy and performance. And so that kind of tends to weed a lot of people out. Taking that type of approach has its pros and cons. I think one of them is that you tend to-- you have to kind of hire fast, fire fast, and you just look at people's production.

But I think, yeah, for people that are coming out of school today,

maybe they're coming into a world where it's a lot of uncertainty. But the engineer who's coming out just now is going to be more likely to be able to pick up these tools a lot faster than the engineer who's been doing what he's been doing for 15, 20 years. I can see there is definitely skepticism and resistance to using these tools across some of those engineers. And I think if I was 20, I'd brush a tail coming out. I would be embracing this as much as I can.

Yeah, that's a really good question. It's a really hard thing.

I think we have good KPIs and we have good clarity around what is important to our customers. We always try to keep our customer in mind when we're debating, even down into the weeds of why is this bug better to fix than this bug or this feature better than this feature? It sounds like a simple thing, but it can get lost really quickly when you're trying to balance tech debt, bugs, features. It's like bringing it back to how is this going to help our customers is what we just inherently anchor everything on.

As a business, we are getting into focusing on more verticalizing what we have already. We built a really great foundation of a product. Now we're articulating these verticals that we want to double down on based on how our customers are using our products, based on the requests we get from customers. We have way more inbound traffic than we can handle. We look through all of that and just understand, "Okay, there's a clear pattern here. Let's go build that." We build a product vision around that.

Being able to cut across all the different languages and cultural things,

that's self-selecting and filtering based on who gets hired and who stays. I think we have actually pretty high retention considering how distributed we are.

We are. Because a lot of our customers today have been engineers, developers themselves,

we have no problem putting pretty much any of our engineers directly in front of customers and we encourage that.

We don't put a lot of filtering between our customers and our engineers who are developing the tech. A lot of them are working with the customer on problems, on features, figuring out the problems. There's limits to that, obviously. We don't want our engineers just being burned to the ground by all the customer's issues.

That is a key thing that I think has been successful for the company to understand what's important and how to build it.

I think people understand what's important and what's not. A lot of our product is open source, but there's Core IP that is locked away for a few select individuals in the company that have the keys to the kingdom, if you will.

That hasn't really been an issue so far. That's good.

I think that companies and customers are going to be very surprised by how much capability the next wave of products are going to have.

We just launched a product in December that has more compute, more resolution than any other product in its class before.

I think we're in a place now where we just put a shoe into the world that the world is going to have to learn how to grow into, and that's going to keep happening at a more accelerated rate.

We're really focused on higher resolution, more sensors, more sensor fusion, and more compute on edge. I've never had a customer say, "I don't need any more compute." That's always a problem. These robotic systems, the models are big and they're complicated, and they have to put them down into this small unit. What we're seeing is that the models are getting better at being compressed, and they're not losing as much performance as they're being compressed, and that's going to continue. In a way, we're always lagging what's possible in a lab or in the cloud, but that gap is shrinking.

Eventually, it'll continue to accelerate, and as more and more compute gets to the edge, there will be an intersection where the models are the rate at which you can go from a big model to a small model, and the compute that's there that's available will put us in a position where what was previously unthinkable is now possible. When the transformer was originally thought out, brought out into the world through chat GBT, it was unthinkable that it would ever be able to run on a chip like ours, let alone

on a phone or anything like that.

Now that's just commonplace, and that happened within 12 months.

I think that's going to continue as all these giant frontier models and foundational models continue to get built, and it's going to creep into the world a lot faster than it had previously.

Yeah, I remain skeptical on that one. I think purpose-built robots are still going to always have a place and will have a place for a very long time. I think this world where the humanoid has replaced everything and is walking around doing everything that you want it to do,

I think it will come, and it can come. I just think the economics of how much the compute costs and battery life and just the performance of the unit itself, I think we've got a ways to go. I think there needs to be new math invented. We need to have an astronomical increase in compute,

and we've got to have incredible models.

I think it's going to be a while before you have a humanoid that can be economical enough to be serving you coffee and replace everybody that makes coffee.

I was, yes. I've had a conversation like this on a podcast before where they were convinced that the only utility for a humanoid robot is sex.

It's a hot take, interesting take.

Yeah, I think if I could be a designer of my own body, I would design it in the way that I would want to use it. I would have extra arms, I have stronger legs. It wouldn't look like a humanoid or a human. I think purpose-built robots are still going to play a role.

I think ultimately, for many, many jobs, many, many industries where we shouldn't have people there, like in mines and very dangerous environments, jobs, I think that's an easy win and an easy justification.

As a society, we would want that to happen.

I just think that if you look at, like, if it just comes down to the bottom line of cost, I think we're not quite there yet.

It's so funny.

All these humanoid companies and robotics companies in general, they always do the same demo and it's folding laundry. I think it just comes from all these engineers fucking hate doing laundry and it's so funny to me. I'm like, okay. But if you go to these big manufacturing businesses that make shirts and they fold them and put them in the packaging, they have giant machines that do it, I mean, like thousands of t-shirts a second. It's like, granted, that can't exist in your home. I totally get that.

For certain things, purpose-built robots are going to be a humanoid every day.

Well, and again, this is where it's like moving away from demo land and like TikTok, LinkedIn Reels, into like real value. That's why a lot of the robots that have scaled and robotic companies that have scaled most don't really look good in TikTok or LinkedIn.

You know, like super unsexy use cases that you're selling value and driving ROI, not

doing a back flip.