Harry Dewhirst: What If the Physical World Had an API? — 375ai, DePIN, and Real-Time Data for AI

Transcript

You're now plugged in to the Deli podcast. [Music] Welcome back to the Deli podcast. I'm John, a ventures associate here at Deli Ventures. Today we're digging into one of the most ambitious realworld applications of Deepen to date. I'm joined by Harry Dwurst, co-founder and CEO of 375AI, a company building a decentralized network of sensors to capture and monetize real world physical data at scale. Think traffic flows, vehicles, weather conditions, and a lot more. All captured directly from billboards and storefronts, processed by AI at the edge, and powered by onchain

token incentives. Delphi Ventures is proud to be participating in 375's upcoming fundraising round and to be working together uh with Harry and the team on this to scale the network. Harry, it's awesome to have you. Um let's get into it. Pleasure to be here. Thanks. So before we dive into um the story of 375, um I'd like to kind of hear uh your background. Um, I know you have a very impressive run both in web two and web three, especially when it comes to monetizing data. Um, um, and so why don't you share some of your insights and experiences that kind of led you to

um, build 375? Sure. Um, so I've been an entrepreneur most of my well my whole career really. Um, I started a company um when I was in my 20s and and and uh that was a business that was helping mobile operators monetize their their users monetize their users information more specifically. So um we raised investment from from some leading venture firms in in in in the valley um Sequoia Excel as well as um strategic investors like Vodafone and Telefonica and we built a platform that enabled them to um make sense and understand who

their customers were but then of course be able to target and profile them for paying customers such as brands and advertisers. Um we ultimately sold that business um to Singtel, the large mobile operating holding company in in Singapore um that have operations across Southeast Asia, Australia, India. And and and I I moved out to Singapore and then built this this data business inside of Singtel for monetizing and and and profiling and targeting their consumer base or their their their their subscriber base. Um I then went on to be um president at another data business called Bliss which was uh using mobile

phones as a way to understand um people's movement in the real world. So apps um by this time you know my first business was was started before iPhone and Android even existed. By by this point in time um you know apps were were prevalent and and and phones had GPS and things like that. So there were there were lots of more data points made available and this Bliss was a platform that that took in huge amounts of data about where devices went and could use that as a sample to understand the general movement of people in the real world and and and that was very very valuable to retailers to um quick service restaurants to luxury brands

uh hotels and so on. Um we ultimately we sold that business to private equity and then it ultimately sold to to to T-Mobile in the US. My last feray um before starting uh 375AI was uh I was appointed CEO of Lynxys, the Wi-Fi router company that most people over 35 as have had one in their home at some point. and and and that was um you know a different experience for me outside of my entrepreneurial roots but um approached it in the same vein nonetheless. Um we were we were listed on the Hong Kong Stock Exchange um raised $160 million to

take it private um and and reinvent the business and it was there that that I stumbled across um projects like Helium that that just fascinated me. And this is um three 3 four years ago. It's fascinated me and and now my co-founder Rob so much that um that we left links to launch launch 375 and that's uh yeah just coming up to to almost 3 years ago to the day. Very cool. Were you always interested in crypto? Like you were you were the CEO of Lynxis and then you identified Helium. So I was I was kind of curious.

Yeah, I I had um I didn't have any professional exposure to crypto, but I I was holding crypto probably since 2015 16. So, but but as a retail investor um and and and I did have friends that had moved into working in the crypto spa space as as their career. Um, so I did, you know, tag along often and particularly particularly as my interest in in in you know, DY at the time, now kind of more broadly known as DPIN, um, grew, I, you know, I would go to all the conferences and so on and and and had a fair number of of friendly faces to see

thanks to to to people who are friends who had moved into the industry. But but no, prior to 375, my interest in in crypto was was really as a you know, a retail investor. Um rode some of the waves. I distinctly remember being stuck in barley in a volcano with um Bitcoin going to 20K, which was very exciting. I don't know what year what cycle that was, but um I remember being very excited stuck on an island in the Southeast Asia. Interesting. Very cool. So, so at one point 375, you know, you were deploying other pe people's deepened devices and then you kind of made a hard pivot and decided to build your own network, own

um uh devices, own token. Uh what was the like light bulb moment for you um to to make that switch? Yeah, so you you're absolutely right. We we started the business with an intention of of being you know an enterprise or the biggest enterprise grade deployer of other people's deep in stuff which we we you know achieved and and we became one of the larger miners of of various different projects in the space helium xnet uh and a handful of others. It was during that time that we learned obviously a lot about what we like about certain projects, what we don't. Uh whether that's the tokconomics, whether that's kind of the

the the concepts, whether it's the demand side, whether it's the business model. So what we decided was that we we were sitting on an exclusive contract that was that was very valuable then and remains very valuable to this day which was a is a real estate relationship to give us access to um 50,000 very sought after unique um locations on billboards with a with an a publicly traded company called outfront. we decided that that asset was really underleveraged. Um, and whilst we were deploying other gear there and and monetizing it, we thought that there was a bigger opportunity. So, we we kind of went to the drawing board and said,

"Okay, like if we were to build our own project, what what pieces would we take from all of these experiences that we've had with all of these other projects, what kind of uh pitfalls would we look to try and avoid?" and and that's that's really where the kind of ideation for for what is now 375 and the data intelligence network started. Plus, of course, my background and experience in in data, it was an obvious obvious path to take, you know, using the latest and greatest technology to attack um real world data in a in a new and novel way.

Got it. Okay. So before we we um go deeper into mechanics, uh could you give us maybe a basic overview of of 375 AI right now um like what what what actually you're tracking what what type of real world data we're talking about and what are the node types and Yeah, sure. So 375 is is um an edge data intelligence network. We've deployed and developed advanced sensors that we've placed across strategic locations across the US. These devices consist of um a Nvidia GPU core which runs uh its own LLM, its own computer vision software um

and and is is something that we train to to to our specification. Um into that is a is a multitude of sensors. Um the primary the primary and the most expensive or valuable sensors being six very um specialized cameras. Um and then a host of other kind of smaller sensors, microphones atmospheric environmental and so on. That data gets fed into into into the core and as much information as the model is trained to do so is extracted from it. So our our initial focus has been on vicular data um because of the placements being on some of the busiest highways in the country.

Um you know think eight lanes of of of of traffic going two directions at 80 mph. capturing everything from capturing and identifying everything from make, model, color, number of occupants in the car, um license plates, incidents on the road, whether that's collisions, erratic driving, um ambulances, um tow trucks, and then when it comes to commercial vehicles, capturing every characteristic of that commercial vehicle, whether that's the US Department of Transport number, um whether it's a trailer ID, um from containers um what the carrier is. So, is it a Hapag Lloyd trailer? Is

it an Evergreen trailer? Is it a UPS truck, a FedEx truck, or any other independent email address or or phone number that we're able to to to pull off of these these moving vehicles as as commercial vehicles are are a particularly interesting category as we come on to, you know, why is this data valuable and and what's it for? Okay. And you know, I mean, everybody talks about like real world data being inherently uh useful and valuable, but um it can also be quite noisy, right?

Like it can be messy, unstructured. Um what what makes uh like the data that you're tracking actually useful for businesses and and uh make them willing to pay for it? Sure. I mean, of course, there there's no it's no surprise to anyone that there's, you know, there's millions and millions of cameras out there in theory capturing this information, but it it isn't data. It's just video. And what our system is doing is taking these six highdeinition specialized streams of of video and on the fly and at the edge in real time converting that into data in in

anomalized format in a very privacy compliant and respectful way because the video itself never leaves the device. So there is no privacy um kind of concern or or or or regulation to be abided by because the video is neither stored nor used. The the AI scans it, extracts the information it needs and then expunes it. That's never to our knowledge been done at least at this scale before. And as we roll out more and more cameras and more and more devices, um, and we'll

come on to to to to a a smaller form factor device that we've been been working on, but it, you know, it's normalizing that unstructured data into something that is usable and and and of course once it's into a usable format, you can combine it with other other forms of data. Uh, and then you're off to the races. And and really this is the first time deterministic data about what's going on on on the corner of of Broadway and Howard in New York has ever been possible. And that's thanks to to to the advent and and development in

technology. So, um could you maybe uh like give give a concrete like uh some some concrete examples of um what type of data could be used for what type of application? I mean um you know one example that comes to mind is you know hedge funds and traders um have been you know it's known that they've been tracking you know satellite uh and drone imagery right to gain an edge in all sorts of like clever ways you know they've been like uh maybe like monitor a parking lot to to have a sense of like how a retailer shop is doing or you know if there's a natural disaster they can check for damages in in power

plants and all sorts of like clever techniques, but I'm kind of curious like what businesses can get out of 375 that they can't through these like alternative means of of capturing real world data. Yeah. Some some concrete examples would would really help. Sure. So, so let's let's use that example there of of of hedge funds and satellite imagery. So the reason they're buying that data is to to give them an edge on understanding whether Best Buy's results or Target's results or what Walmart's results are going to be up or down or left or right from from from you know forecast um and looking at parking lots and and and looking at the visitation of

parking lots based on satellite imagery. That's that's a pretty cool way to to to approach it. Now the the the challenges are um those images are only available when the satellite happens to be flying over that particular place. The image is expensive. Um they're very expensive to to to book and order those images. they they are are a snapshot in time and also they don't um have really any way to to calculate any form of unique because of the vantage point at which they're being viewed is really just counting the total number not not

the unique number on one day versus another day. The the difference in fidelity between that and ours. So let's use an example whereby our our devices are are deployed outside of on on the thoroughare to a to a Walmart or to a super center and our device is capturing every single vehicle down to a license plate which which is a a unique identifier which therefore can be used to understand both the reach and the frequency accurately because you know and we've seen this in the data there our, you know, our first our first device that we deployed and that was

built by our chief AI officer in in his garage in in in uh Redwood City, California. He he has had that device up and running on the on a billboard in on the 101 that we have a truck, a Ford Transit van that we have seen thousands and thousands and thousands and thousands of times. I don't know what this person does, but they drive past back and forth hundreds of times a week. And that of course in in another means would be considered 100 people, but it's in fact not. Now that's one extreme example but um by by being able to have a a very granular and highfidelity view on the

number the frequency and then also from the the car itself there is a lot of assumed dem demography um and other attributes that can be understood um you know is it is it all Kia Sportage or is it Range Rovers there's a very different demographic between between those two vehicles. Um, and there are other, you know, is it F-150s, and I love F-150s. I have an F-150, so that's not a knock on F-150s, but is it that or is it Teslas? You know, it's it it's a different makeup, a different audience. And that can change during the day. That could change the different hours of the day. And this is

something that's persistently aware of what's going on. And not only is it understanding what's going on right now, it's storing that data in perpetuity. So with every passing day, the richness of the data set is getting richer and richer because because of the historical data that it has and and over time, new and exciting use cases will appear because of that long-term view as to what's been happening in a in a specific place over long periods of time.

Got it. So I'm hearing, you know, high like basically more accurate data, right? Higher fidelity, real time and um just scale and and and costefficient way. So definitely strong points. Uh and then maybe could you touch base on like where what stage you are at right now like um is are you already generating useful data or do you need to reach a critical mass for this data to become uh useful enough uh for businesses to be willing to pay? You know the typical criticism in deepen is that um I I think uh helium was the biggest example of

this. they went uh super wide, super fast, scaled the supply side and the thesis was always has always been that you know they need to reach a critical mass to for their data to become useful. So I wonder if the same dynamics uh apply here as well. Yeah, it's it's a good a good point and it's something that you know in that in that evaluation phase when we were looking at what business and project we want to build based on all the experience that we've had with building and being deployers of Helium and others and and we love all these businesses that have kind of led the way but there's certainly aspects of of various parts of them that we've wanted to

improve in demand. We wanted to build a product that had immediate demand and has had utility without this build it and they will come or or or this super scale requirement and and and we've been successful in doing so. So, in order to to give us the best chances of of monetizing out the gate, which I'm pleased to say we we are and we've signed, you know, paying customers even, you know, ahead of our our TGE, which means that upon TGE, we have onchain revenue. we have iron burn of the of of the the token happening um effective immediate which is something that many

other deepening projects have taken many years to to to do. So but the one of the ways that we ensured that that was the case is by having extremely high quality locations for install. So that these devices are are expensive and very very powerful devices. They cost we we we sell these devices to the community for $50,000. Um we sold our first 50 units in pre-sold our first 50 units in Q3 of last year, manufactured them through Q4 and began deployment in Q1. Now we decided to have a very geographic focus.

So the first 20 or so units were all deployed in the LA County area. Long Beach is the busiest port in America. And as I mentioned before, commercial and freight traffic is is a very um I guess lowhanging fruit when it comes to monetization um because of because it's such a key leading indicator of of business and economic performance. So we we focused on that that geography first, then New York, New Jersey, and then finally uh the last 15 or 20 devices of that initial batch of 50 in in South Florida.

And I think we will continue to have that kind of market bymarket approach rather than having um disperate data from 50 different cities across America. Um which would be much much more challenging to monetize in any meaningful way. Um so having the densification has meant that in LA for example we are seeing millions of cars per day of which um there's only about 5 and a half on any average day there's 5 and a half to six million unique cars a day on the road in all of LA County. So we're seeing double digit percentage of unique cars in in a in a given day every

day. So there is a level of critical mass and scale even with a small number of devices placed in the right places that can be attained. I see. And okay. So so that's that's quite impressive that that you track so many cars with just 20 units. I'm still kind of curious. Do you think like once you reach a critical mass, however you want to define it, you know, you go region by region, but maybe when you reach a certain scale in US, do you think new type of buyers would be interested in that type of data? Like would that give rise to an emergence of new type of customers or Yeah. Um, curious to hear your thoughts there.

Yes. So um as I as I mentioned there are there are customers that are that are are interested and will value the data already and if it's you know large markets like New York, Los Angeles, Miami that's sufficient to make a difference to their business particularly if you know there are transport and logistics business where they're three of the biggest biggest ports cities in the country. Um, and you know, Long Beach is the the busiest port in the country, and we have devices on every artery coming in and out of that port. So, there isn't a there isn't a container that comes in or out of Long

Beach that we don't at some point see. So, that has enabled us to drive that monetization. There will be use cases that maybe do require more national coverage or you know have further requirements for data either in a city or in rural areas. Uh and that's that's why we've developed um a different size form factor um to to if you if you want to think of these these 375 edge devices as the large as the big rocks, we need other data to be the sand that fills in the gaps between the rocks um to create

the most comprehensive picture. So I think there are great use cases now um with national coverage there becomes a greater number and then with the e the elapse the elapsing of time i.e the data set is you know years of data. There's even further use cases that become available because of the nature of cap. You know if you can if you can fully understand the past you're better able to predict the future and and and therefore if we have a year of data then trend analysis is is very very feasible.

um if you just have a snapshot of a day there there's really no trend to to be assessed. So um it's those three things that I think will drive waves of new adoption, new use cases, some of which we've we've we've um contemplated, others of which frankly we anticipate to be surprises to us. Um particularly as as people as we expose this data in a platform um for people to plug into their apps to plug into their their their business intelligence models they will use it in ways that maybe we haven't yet imagined um and that's that's I guess the beauty of data.

So um building a platform for others to plug in. So um that kind of makes me wonder and ask what's uh like how do you get the you generated the data how do you get it in front of the buyers basically what's kind of your go to market strategy? So initially and this is the way that many other data businesses bring their data to market. There are there are numerous data platforms uh and data marketplaces. Some are, you know, owned by huge companies like Google and Oracle. Um, others are are smaller and independent and all of them have their own different niches and flavors as to

to what makes a client work with them versus another. And businesses, so say you're McDonald's or Coca-Cola or Walmart, you would typically have a preferred data platform. You don't often have many. um you wouldn't work with five different data platforms. You'd probably work with one or two, three at most. So once we've integrated and published our data into these platforms, these these customers, these Fortune 500 businesses can start consuming that in an easy manner that that they can they can add that to to their to their suite of information. Um so we've been busy and will continue to

be busy integrating our data into these large data marketplaces, data platforms that enable the existing rails of how these large in you know web 2 fortune 500 companies operate. It it it's you know it's within their muscle memory and they they don't have to there's no real change of behavior. they'd be just replacing data they other otherwise would have previously bought about traffic on the 101 is now being serviced by our data which isn't um data that was conducted and published by a survey and a rubber strip on the road 5 years ago counting

the number of cars or or uh accidents or whatever. I see. So um so you you supply the data to the data marketplace and then on the other end you have these fortune 500 companies consuming it. What's that landscape looks like? Is there is there certain verticals where like you know logistics companies plug into a type of you know um one of these marketplaces and then maybe hedge funds subscribe to another or is it is it all kind of general purpose and you don't know how it can be consumed? it it's quite diverse. Um so so you know one one platform and and the first couple of

platforms that we've integrated to have you know customers from McDonald's to FedEx for transport and logistics to Citadel on on you know financial services side. So you know that's one platform with three very different types of customers. So, it's certainly not um like very kind of categorized into niches. It's um gen tends to be general. Of course, there are exceptions to that, but um the larger platforms service many. I see. All right. So, I want to um kind of transition into and talk a bit bit some of the tech stack and and your design

choices as you built out the the network. So the biggest one the biggest uh decision uh there is one of the biggest decisions is running AI at the edge versus cloud. I think uh hivemapper if I'm correct they're doing a little bit of both. Uh but I think the um the heavier lift is on the cloud whereas you kind of go on the other end of the spectrum and do currently do everything at the edge and I think there is no logic at the cloud. You can correct me if I'm wrong there. Yeah. Yeah. Could you could you maybe dive into the thought process there and was that kind of how that was decided?

Yes. Um, so the the the rationale behind the decision to do edge AI versus centralized or decentralized, but cloud AI was not because it was easier. In fact, it's far far harder. um it would be way easier for us to just deploy relatively cheap cameras in comparison and just push all that data into the cloud and deal with it there. We decided not to have that architecture very intentionally for two reasons. one efficiency. Streaming uh that that streaming and storing that that kind of quantity of data at that

fidelity of of of quality at scale is not uh trivial and and would you know would become ownorous from a cost perspective over time. Um and and so that was the first one. But the second one and and and frankly more importantly was the the p the evolving privacy landscape. So whilst right now it would be perfectly permissible to you know stream data to a cloud, process it and and extract the data that's evolving that's changing it's it's it's a moving target and and we also wanted to build something that was bestin-class in terms of um respecting um you know capturing the

data that's valuable but not keeping stuff that that isn't useful to us and could be subject to any form of abuse. So if we don't send the data anywhere, there's no ability for it to be intercepted. There's no ability for it to be um you know used by any form of bad actor. So keeping the device processed locally at the edge, keeping the data processed locally at the edge has safety features inherently built in and um it's a very privacy first approach. I see raw foot footage comes in is processed by AI chips on the device at the edge and then out of the box comes D like the the structured data

text data right okay yeah exactly I see and then the video is deleted so so it it's it's a very best-in-class way of and plus we can train the model to stop capturing some data so if a if a regulation changed or if we wanted to launch in a geography graphy um that had different regulations, the model can simply be trained to not identify or capture that piece of data and and only capture the the pieces that that are compliant. Um, are you are you worried at all that you know a trade-off here is that you know maybe like if you prefer the cloud route

you could in theory run like a much bigger AI model computer vision model or multimodel LLM and then maybe process it like you know have maybe a higher quality of intelligence as as an output. Are you worry worried about that at all or or no? Um so we're getting the best best of both worlds. We are we are using sample sets picked randomly from our fleet of devices that then are put into a multimodal system in the cloud that then is used as as training data to feed back into into the the the boxes at the edge. So, we can't, you know, you're

right in saying we can't run the heaviest parameter model at the edge without the device costing $250,000, which which is just unfeasible. But we can leverage the benefits of of some of these super powerful models to to train and improve within the niches that we're interested in. Because to have, you know, billions of parameters that you only are really using a very small subset of for this very, you know, you only need a certain type of tool for this job. And it's a reasonably small number of tools in the grand scheme of things. We're not going to know need to know anything about like

British history or or or or or novels from the 14th century. Yeah. So, um it's there's a whole host of information that just isn't needed to be residing to identify the stuff that we need to identify. Okay. So, intelligence is one part, right? That's the AI chip. The other is the sensors, the eyes and the ears, right? What what you're tracking. And then basically a simple trade-off space there is like you can add more sensors, you know, you can track different stuff and and that would bump up the price of the of the device. And uh there's kind of a point where where you are comfortable like it's not an overkill but it's it's super useful for all the other all the

use cases that that may emerge in the future. So yeah, maybe maybe some light into into that decision as well could could be interesting. Yeah, I mean we we built the the system in a modular manner. So, um, over time, if it's if if we realize that there's a piece of data or a piece of information that we could capture from an a sensor that isn't included, it's, you know, relatively easy for us to add that in. I mean, the cameras are really the most expensive sensors. Sensors typically are are quite cheap devices depending on what they are, but cameras being the most expensive type of sensor typically. So um you know we have a modular system

that it will enable us to if there is a demand and a need upgrade that that box without um without having to start from scratch right cuz these are specialized cameras right the uh captures different stuff came so they're they have different um kind of apertures and and and purposes on the roadways and different kind of fields of vision. Of course, we're also looking in two directions um because we're looking for license plates on the fronts and backs of of vehicles and and often kind of verify the front and the

back in in places where front and back license plates exist, but it's often two lanes of traffic in in north and south or east and west. So, three are pointed at the call it the the the northbound lane, three appointed at the southbound lane. Um, and they're all kind of capturing different parts. Some are very very kind of honed in and specific to the the the smaller signifying marks, license plates, logos, text, email addresses, um, stuff that's very kind of close up. Others are are much further a field that are identifying the make and model of the vehicle and and the occupants in the vehicle and and so on.

So there it's it's not six of the same cameras. It's different cameras doing, you know, programmed and and tuned to do different things simultaneously and then marry up together um that information to say it it was, you know, we identify the vehicle as as, you know, 375 ID vehicle 123. Right. Right. that's that's been populated by, you know, potentially six cameras or or three cameras uh to to to get that full full view. But but it's still it's still car one, two, three. It hasn't been counted three different times. It's just used three different

cameras to populate the data, right? So the um the other I guess trade-off space is is more in the architecture, right? So we we talked about the edge device and those are installed in these like prime location billboards and and I and and that's more of like a curated way to scale the network whereas the other form factor that you mentioned earlier is I believe a smaller device that you intend for people to self-deploy and there's always this this question of like when people self-deploy things you know how do you ensure the quality and uh the quality matches uh the um um edge

devices, right? So, could you maybe speak to that? Sure. So, so yeah, you're you're absolutely right. We've been developing and have developed a product that we we're calling 375 Street. Um think of it as a miniaturized version of 375 Edge. So, it still has Nvidia um GPU at the edge. um albeit you know a much much smaller uh chip. The device is you know there no bigger than the kind of shoe box if if you like. Um it contains two two cameras versus six um two different types of cameras again for for those two different varying ways to use them. Um and and it's designed for self

deployment. So people who have, you know, a balcony, a rooftop, a window ledge, a small business that is seeing lots of stuff go by or interesting stuff go by. You know, you might not see lots, but you might see interesting. And that does that can often be valuable. Interesting could be if you're on the only road in and out of a of a of a town and you're going to see every car that comes in and out of that town. um that could be very very valuable. So this is designed for people to to to self-deploy. Um it it's very very applicable to be placed in the same place. So if you're a helium deployer

and you've placed a Wi-Fi box in a cafe or in a an amusement park, you could easily put one of these devices next to that Helium device. And you know, Helium is monetizing by providing connectivity to users who are at the amusement park. This device is capturing d capturing um data from what's going on at that amusement park. And um you can monetize that location in multiple means um and increase the yields as a deployer from a single, you know, place.

So, we're we're excited to be um we'll be selling that product later this this year. And um as I as I said before, the the edge devices really with with just a few thousand um spread across the country, there won't be an much there'll be very diminishing returns to doing many many more than that. However, with 375 Street, the we could easily sustain tens if not hundreds of thousands of these devices to really give us an incredibly rich and comprehensive data set of the nation. Plus, um we of course have aspirations outside of the US and and intend to have other geographies where

we launch and make this product available for deployers. This is kind of uh similar in some sense to what Hivemapper is doing, right? The the the street device especially. The um the biggest difference is like for context, Hivemapper, you know, has a network of uh people placing dash cams on their vehicles uh to uh to capture uh uh map data. So, so these nodes are are constantly moving and recording whereas uh the the street devices they would be stationary. Um so what what does that difference tell in terms of the type of data that that you can capture and uh

what type of like insights you can deliver um you're a what what different type of insights you are able to deliver. Yeah, I mean I I've of course been been a fan of of Hiveap Mapper and and and had their devices since since kind of they began. Um so I've been following it closely. Their use cases whilst there is definitely overlap in the data that we we capture, the use cases are are very different. Their customers tend to be more on the mapping and autonomous vehicle side of side of things because they reside within the car. However, it doesn't give you a a a really a a view into what's going on on a particular road because for it to do

so, it would require, you know, it really only knows what's going on in the car that it's being driven in and the car that's directly in front of it. And it doesn't have much perview beyond beyond that limitation. Um, and therefore to capture how many vehicles are on the 101 right now would require every other car to have a hive mapper in it, which is not feasible or or or or reasonable to expect. So, um, it isn't a good proxy for determining the general traffic flows um because of because it's it's it's a sample based it's a sample based system. Um, when you deploy a

camera of ours, you know, 99% of the time it's going to extract information about every single vehicle, regardless of whether that vehicle has anything running inside it. Um, you know, it's not um it's not dependent on the vehicle to be to be running a piece of software or or an app or so on. But whilst I mention apps, we should definitely talk about about 375 go whilst uh whilst we um do it. We're talking about 375 street.

Yep. Would love to hear it. Um so 375 go is something that we we launched in testn net um in November December last year. Um at pretty much at the same time we started deploying the the 375 edge devices. Um, this is a way for anyone to get involved in 375 and think of a phone as as the smallest sensor on our network. You know, phones are very have have, you know, they have obviously have cameras, they have microphones, they've got all kinds of other things. They have liar.

Um so there are very many things that can be done from that device that that um could be useful data to to to add further incremental information to our network. Um like the you had the rocks and and sand analogy. So this is kind of like the water I imagine. The dust of the water. Yeah. Um exactly. Um the that that's exactly right. So today we have, you know, 200,000 people um and and and thank you to everyone who overwhelmed us with with the support for for for the for the app um during this test net. And we we

continue to have thousands of people sign up to to contribute information. And um today we're capturing relatively passive information. So um it's location data where your where that device is. It's all anonymized. Uh and it's also RF signature information. So um what networks does that device see? Be it cellular, Wi-Fi, Bluetooth, etc. And then if you're connected to any of these networks, what are the speeds? um which which is a a useful piece of data for ISPs, mobile network operators, MVNOs's WISPs to um understand network congestion, understand where uh their competitors have better quality of

service visav themselves. But our intention with 375 is to continue to work with this community to create more and more meaningful ways at which they can add um value to the data that uh value to the network that we can then monetize. So so we're going to be integrating solutions into the product that have a more active participation. So, um, if you're a 375 Go user in Paris and we have, um, a customer that wants to know the the price of real-time gas prices in in in, uh, gas stations across across Paris, and we have 5,000 people in on 375 go

that could go and get that data, we could fulfill upon that in potentially a matter of hours and and that could be monetized. those users could be rewarded for their participation and um that will create more network utility which will drive drive buy and burn. So that's just one kind of basic example, but there we we anticipate there being many ways for us to um leverage this audience, leverage this community of of of incentivized people to to carry out tasks that relate to the things that your phone can do. Anyway,

awesome. the um another analogy uh or comparison uh that um you know that came to mind is um with grass. Grass is a um you know leading deepen uh project. It's a portfolio company of of Delphi Ventures. Um what they're doing is they're is actually similar to what you're doing in in some way right because they they capture uh real time data web data digital data from the web and you are gen like capturing real time physical world data from roads cities um so it's it's quite uh kind of similar in that

respect with grass is um you know it's pretty evident Now that there is significant demand uh from LLMs to use that data for live context retrieval because LLMs you know they want to know what's happening in in in the web right what level of confidence you have the same happening in the in the real world uh you know h how soon or or do you have any any thoughts there basically are we going to have you know robots autonomous physical uh uh AI devices that that that's going going to want real time world data. Um, I mean I the answer is it's it's uh it's inevitable. It's not a

question of of of if, it's a question of when. And I think when is like on our immediate horizon. Um the the grass analogy is a great one. you know, that's a project that I've been, you know, following um since since inception and and kind of particularly given my background in data and and digital advertising and so on, I could see those use cases. But then when the the light bulb around using this army of of and community of of effectively nodes or botnet users who can you know access their bandwidth

to get real-time information that um data sensors are not really well set up to do that could then provide real time inference into LLMs then that certainly there was a light bulb moment for for for me and us internally about us being you know an equivalent to that in the real world um and this information being um incredibly valuable to the LLMs uh or or to you know customers of the LLMs. Of course, businesses can upload proprietary data sets into their LLMs, whether that's their own customer data

or other third party data of which if you're FedEx, the movement of vehicles is paramount to your business. And if an LLM doesn't have it, you're going to put it in there. So they would either buy it from us but of course we would hope over time that we would we will develop relationships with OpenAI, Anthropic, XAI etc to be a a a a leading proprietary data source for um physical information, physical happenings. I I think a a meaningful difference uh another meaningful difference there could be you know grass is

um they identified the demand as being you know live context retrieval. So it's more so on the inference side whereas cuz cuz you know there's no edge in terms of pre-training data on the web all the data is out there and um and so there's there's not a meaningful difference uh edge to gain but that's not the case in the physical world right we we actually lack high quality data in the physical world so maybe like at some stage it could be quite interesting uh to serve these like training use cases as well in the in the physical world.

Yeah. Yeah. I mean, the LLMs don't have eyes and ears in the real world. They definitely have eyes and ears in the digital realm and they've read everything and watched everything and listened to everything that they could and they'll always be ahead of anything new they would have already read before we've even know exists. Um but but they have a huge sensory blind spot to uh the real world and I I believe that that's a fourth dimension to which they will unlock um and particularly in robotics automation autonomous autonomous vehicles further information again

beyond that of their own sensors. You know a tester can only see as far as a tester can see. Soon enough, cars will have cartoar communication. So then they can leverage the knowledge of another vehicle that's also autonomous, but they should also be leveraging the the intelligence of sensors that are on the sides of the roads, uh, for example, um, or in the sky on drones, whatever it might be. And there are protocols and standards being developed for that that exchange of data in this autonomous world um already.

Very cool. Um the I guess the the final point on the endgame type of scenario. Um, is there a world where, you know, just by the sheer number of of data that you're collecting, you're able to improve your your AI model and then, you know, the endgame is is actually not the the data that you generate, but uh you build your own AI model for whatever purpose it may be. Are you do do you share that kind of vision or is it too early to comment? Um, it it's a it's a an interesting one.

I I didn't that I never kind of contemplated that um before before a few weeks ago I was with Chad our our chief AI officer and he made the point of we're getting really really really good at certain things and those certain things are like vehicle related stuff. So, you know, he believes that given the quantity and scale of of of data that we're training upon and the the quality of the data that we're training upon that we quite quickly, if we're not already becoming like the very best at identifying vehicles um and then

identifying incidents, accidents, erratic driving, all the other things that become related to automotive And and those elements um certainly could become a commercial part of the business because there will be needs for that that that aren't necessarily met by um open-source object object recognition um generic things which which we use and are brilliant. But we just have this this very unique vantage point of h of seeing millions and millions and millions of different cars per day doing different things and um it's it's that unique source of data that that other LLM are just not

ingesting at any great scale. Harry, what what do you see as the biggest risk uh slash challenge? Um is there any existential risk to what you're building? What what kind of keeps you up at night? Um I mean building stuff building physical stuff like manufacturing stuff is is not without its challenge. Um you know I'm I'm fortunate and unfortunate I guess would be the way to describe it to have experience in it because it is you know it is something that can cause headaches. you know, we've gone through chip shortages and and those kinds of eras. There there there's certainly no

reason why that also won't come back. GPU is in extremely high demand, tariffs and and logistics of shipping. So I'd say that that that's a that's a constant challenge and moving part to like manufacture things in a costefficient and time efficient manner. um there are so many elements out of your control and so far touchwood um we've we've been successful and I think certainly you know mine and Rob's time at links taught us a lot about making you know we made millions and millions and millions of Wi-Fi routers so that without that experience I I I

would be more lost than than than otherwise but it is a challenge and um you know for any deep in enthusiasts out there who's bought pre-bought hardware and waited and waited and waited and I' I've been in the same boat. You know, it's hard. It's hard and and I'm I'm sure all the companies that are trying to deliver their products would love to deliver to you them and their hands are being tied by a missing chip. It could be one little chip um and they've got 99 other chips, but if they don't have that one, it doesn't work.

And uh yeah, and it's that's kind of the nature of the beast. Can't change the world without blood, sweat, and tears. That's for sure. Um we we're we're about to um you know, we're we're just about out of time, but um yeah, I got to ask this. You know, uh 375, how did you come up with that name? That's that's oddly specific. Is there is there a meaning to it? Um my my interpretation Yeah, go for it. Go for it. No, no, no. I'm interested to hear yours before I before I tarnish. Yeah, my guess is your the sensors and cameras they watch, you know, uh three 360 degrees and then there's a 15 degree of of extra that I don't know why

that that that's Yeah, that one's that's good. I haven't I've had people talk about how there's 365 days in a year and we're seeing 10 days more or something like that. But the actual um the actual reason is the extraterrestrial highway is highway 375 which runs through Area 51 in Nevada. And um we we kind of thought there was some analogy between I guess the extraterrestrial world and what's now capable and possible with the advent of AI to to create otherworldly type experiences that that have never

been done before. Um, so I guess it's it's somewhat the art of the unknown and and and and that was the plus we're we're you know we have a big focus on vehicles and transport. So Highway 375, the extraterrestrial highway was how was the inspiration behind uh 375 AI. Very cool. Harry, really appreciate you coming on and walk us through the the project, your vision, tech stack, the challenges, and yeah, I'm looking forward to uh we are looking forward to work with you on this and then kind of uh spend uh yeah, countless hours together to put uh you know, nodes all around the world, including the uh 375

uh highway. Yes. No, we haven't got one there yet. So, let's go spot some aliens. Awesome. Thanks again for joining us. It's been a pleasure. Thank you. [Music]