AI for Physical Security | Dave Selinger, CEO of Deep Sentinel

Dave Selinger is the CEO of Deep Sentinel, an AI-powered video surveillance system. They have raised $38M in funding from Intel Capital, Shasta Ventures, and others. Prior to this, he was the cofounder and CTO of Redfin. And he was also the cofounder of RichRelevance. 

Dave's favorite books: 
- The Speed of Trust (Author: Stephen Covey)
- Snow Crash (Author: Neal Stephenson)

(00:01) – Origin Story: A Near-Miss and the Broken Security Market
(04:22) – What Deep Sentinel Does and Why It Works
(06:23) – Benefits of Vertical Integration in Security Tech
(10:20) – How Deep Sentinel Tackles False Positives with AI
(14:06) – Balancing Escalation Risk and Deterrence
(17:06) – How Deep Sentinel Processes and Uses Its Data
(19:36) – Positioning Deep Sentinel in the Competitive Landscape
(21:12) – Go-to-Market Learnings for Hardware-Software Companies
(23:39) – Residential vs. Commercial Security: A Comparison
(26:41) – Regulation and Public Sentiment Around Security AI
(29:03) – Insurance, Security, and Incentive Alignment
(31:23) – Company Building and Lessons from 20 Years of Founding
(39:26) – The Role of Distillation and LLMs in Deep Sentinel’s Future
(42:27) – Rapid Fire Round

--------
Where to find Dave Selinger: 

LinkedIn: https://www.linkedin.com/in/selly/

--------
Where to find Prateek Joshi: 

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-infinite 
X: https://x.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.484)
Dave, thank you so much for joining me today.

David Selinger (00:04.602)
So excited to be here. Thank you.

Prateek Joshi (00:07.412)
Let's start with the origin story of Deep Sentinel. You mentioned that it was a near miss at your home that convinced you that the security market was broken and that something needs to change. So can you talk us through that moment and how that crystallized the idea for Deep Sentinel?

David Selinger (00:30.235)
Yeah, for sure. I mean, and this is gonna sound a little self-serving, so I just wanna throw this out there. This is not a promotional thing. It's just the freakin' truth, man. And that was really what caught me. And that's that the security systems and security solutions that we have, we as business owners, we as human beings, homeowners, renters, whatever, they don't work.

And I don't mean they don't work like, like it fails 10 % of the time. I mean, they don't work, they don't work ever. And they don't actually do anything almost at all. The word that is the number one selling point of all the security solutions out there, other than deep settle, but is deterrence. And deterrence is like a cool word. It's like, yeah, that's gonna stop people. What deterrence means is it allows them to stop of their own accord.

And that does not stop people. And so what happened was my neighbor had a home invasion and she has an alarm. She has like super expensive security cameras. I was like on the security committee at my HOA and I convinced a bunch of my neighbors to spend tens of thousands of dollars on security cameras because they were like higher resolution and they had this support agreement. And what we had when we reviewed this as a community in our neighborhood watch meeting with the police officer.

was a video of a bunch of guys standing outside of her home for minutes, preparing to go in, checking their masks, checking their weapons, and then breaking into her house and then leaving some 25, 30 minutes later and no evidence that led to an arrest. So number one, it didn't do anything in the moment when you needed it. And then number two, even after the fact, all you have was this pretty useless video. And I was just, as a dad and a husband and like,

I was so floored that I had gone to ADT and I spent all this energy and yes, I negotiated with him and I got an extra six window sensors and all this stuff. And it just doesn't really work. I even alarm systems, if they go off, I want to say about 60 % of Americans live in a city where the police department is so bogged down by false alarm calls from alarm systems, they don't respond to them at all. They have no guaranteed response.

David Selinger (02:50.593)
And unless you have some sort of a verification, like I'm in the house, there's person in here with me, they won't come. They won't come in 20 minutes. They won't come in two hours. They won't come the next day. just won't. They'll put it in a little file and send it off in a never never land. So that to me, was like, my God, is there something that's more important than security? I'm a nerd. So like, no. And then is there anything that's solved worse than this? I can't think of anything.

It'd be like if you had, I don't know how technical your listeners are, it'd be like if you had a computer and you're like, I'll just put a warning sign that says if you hack into my system and steal all my data, I'll come get you. But here's

Prateek Joshi (03:31.362)
Yeah

David Selinger (03:35.161)
I that works super well. Not at all ever. It just, that is the entire state of the entire physical security industry. And I was like, my God, how do we do this? I met some people who like had created a neighborhood watch group and they stayed up all night and watched their neighbors cameras. And I was like, that's cool. And they said, yeah. And the police will come immediately because I tell them exactly what's going on. And most of the time we can just stop people by like,

flashing a light on them at the right time and like coming outside and yelling it from our houses. And I was like, wow, that's amazing, but that sounds very tedious. voila, fast forward. That was when I thought if we just take some AI in there, I can make the TDM go down by a factor of a hundred and make the efficacy go way up. And that is what Deep Sentinel is.

Prateek Joshi (04:22.99)
Amazing. Love the vivid examples and the analogies. So it's a good segue into my next question. For people who may not know, can you explain what Deep Sentinel does today?

David Selinger (04:34.5)
Yeah, and if you don't know, please do. So what Deep Sentinel is, is it's a combination of security cameras, and we support all different brands of security cameras that are out there, a unique proprietary AI that we have built that is specifically tuned to identify suspicious behavior and suspicious individuals, and then putting a human being right into the loop within five, 10 seconds. And so it'll...

If the AI detects something suspicious, a human being will be reviewing the video in five to 10 seconds and then able to intervene using something simple like two-way audio or something a little bit more aggressive like a siren or even something really aggressive. If we actually do stop like hostage type situations, we've got smoke and pepper spray solutions as well for serious security issues. And the basic principle being that intervention in a preventive way,

solves the like an ounce of prevention is worth a billion tons of cure. And number two is if we can just get really, really early in the process, usually the level of intervention required is really small. Again, kind of think about this in the context of a cyber security or medical situation. If you catch it early, you can stop almost anything from happening with just the smallest of interventions.

Prateek Joshi (05:57.422)
Amazing. Going a level deeper into the product. So you mentioned cameras and you're building them instead of relying on what's available off the shelf. What advantages do you get by this vertical integration? Meaning you have your AI, you have your hardware, you have your live human guards, and it's being offered as the full solution. So what...

David Selinger (06:05.734)
you

Prateek Joshi (06:23.776)
advantages do you get by doing this taking a layer approach?

David Selinger (06:30.352)
Yeah, yeah, so kind of more of vertical approach versus a horizontal approach. So there's two things that I think are particularly important. Number one is just high speed compatibility. So the ability to deliver in a complex heterogeneous environment. So when you think about security cameras, there are so many, there's Ring, there's Arlo, there's the Chinese brands, there's a number of US brands that make them.

there are not really any very good standards. And so having to integrate vertically deeply with a number of them is pretty important. The second thing though, which is the one that actually pushed me to do this, is that when I reformulated this problem, I analyzed that situation with my neighbor. I realized that every other vendor in the space was thinking about it kind of the way you're asking the question, what if I just do a layer? What if I just make the AI? What if I just make the...

cameras, what if I just do the guard portion of this? What if I just do the police interaction portion of this? And that's how the industry works. And what happens when you have a massively heterogeneous environment with each of these kind of specialty layers is that each integration between one layer and the next is lowest common denominator. Meaning that I'm going to use the dumbest parts of the AI if I'm buying an AI vendor, I'm going to use the dumbest parts of the camera.

so that I have broad compatibility. I'm gonna use guards that are just like as basic and rudimentary as possible. And sure enough, when I went and looked at the industry, there was this fledgling sector called remote video monitoring. And it was exactly that. were using AI from the 80s and 90s, they still are. They were using guards that really didn't know much about security. They weren't well trained. They were just sitting there watching the panel of videos.

If you go to a building in New York and you see the people like slumped over a chair, supposedly watching the video cameras, that's, that was what was going on in this industry. There wasn't any software designed to make sure that all that was done at quality. And so the biggest thing I got by doing the vertical integration, a lot of people ask, did you save money? Yeah, a little bit. I'm not really though, cause I had to build hardware and building hardware undoes all of the economic advantages you get from anything else. What I really got though is.

David Selinger (08:53.934)
We have the absolute unconditionally, universally best, fastest, cheapest solution. We stop thousands of crimes every single month. We're now compatible with about 20 different brands of cameras out there in the world. And we've tuned our AI so that it solves just this one problem. It identifies suspicious people. It does it quickly. It does it in real time. It helps our guards make sure that our guards are efficient.

It helps us hire the right people that can sit at the console. We monitor the activity of our guards on the console. So if they're even just lagging a little bit, like I didn't take my coffee this morning, we know, right? And we don't know a month later when a customer complains. We know within 30 seconds of your behavior changing, because by vertically integrating, what I did was I vertically integrated not just the software, we vertically integrated all the data.

And so we have a universal end to end view of every event. So when someone walks to the front door of your business, we don't just know that they walked into the front door. We also know that that went to guard JJ. JJ was reviewing two other events. He is one of our top 10 percentile guards and he clicked on it and he investigated it and he verified that that person was the standard package delivery person for that property. He did all that in 15 seconds. And I have all that data wrapped around that little tiny 15 second video clip.

Prateek Joshi (10:20.206)
Amazing. Earlier, you made a point about how the police are bogged down by all these alarms and false positives. So now with AI in the mix, how do you deal with false positives? And also, from a product sense, which side of the line do you err on? And how do you just deal with the accuracy issue here?

David Selinger (10:35.27)
you.

David Selinger (10:43.992)
Yeah, that's a great question. Just so I know, how technical is most of the audience here that's going to be listening?

Prateek Joshi (10:49.327)
Let's say, let's assume medium technical, but I'd love plain English explanations too, just so it's more accessible.

David Selinger (10:55.622)
All right, you just you guide me and I'll do my best right so at a high level we do a principle based engineering Effort right and is it is a principle based company, right? I mean, we're very mission based We take the stopping of crimes and the protecting of people so incredibly seriously and so we do the same thing with our AI strategy So principle number one is no false negatives. And so we tune to have false positives intentionally

We tune both the structure and the parameters of the AI, as well as the sensitivity settings that we use. We also have other logic that runs on the backside of these various events. All of those are tuned similarly. So we start off with a very, very strong threshold of zero false negatives, and then we start progressively iterating to improve the false positives. And we do that, we remove the false positives in our side by

doing things like we have a whole bunch of, in fact, if you look at our patent portfolio, I'm so stoked on our patent portfolio. I usually am not a big fan of patents, to be honest with you, in the software world. We were so far ahead of the curve on AI in video surveillance for remote video monitoring that we have all the fundamental patents. And what's cool about them is that when you look at these patents, there things like, we're going to take all that data we have from the guards and everything they click done.

And we integrate that into our model. So we don't just train on, this a person, is this a dog? We also train on what did the guard do? How quickly did they respond? How quickly did they call the police? Were they angry when they called the police? Were the police angry at them? Did they dispatch police officers and get there in 10 seconds with guns drawn? Or did they take a report? All of that information can go into this. the problem of false positive, false negative, as we look at internally, becomes a lot.

more complex, which means there's a lot more space for optimization. As it relates to police false positives, by having the human in the loop and all this data at our fingertips, we have pretty much zero false positives. We never call the police departments with false alarms. At the bare minimum, we may say there's a suspicious person standing right here doing this. They're not responding to my audio commands. They're not leaving the property and I don't recognize them from the property. Even if that person is not

David Selinger (13:16.87)
committing a crime, that's definitely not a false alarm, right? We've described the situation, the police are fully armed to make any decision they want. If it's a busy night, that's just not on my priority list as a cop. If it's a slow night and we're in a small town and we're really kind of enforcing trespassing, maybe that's something they want to respond to. And so one of the neatest things about that part of the interface is by being able to provide real-time information about suspects, crimes, potential weapons, victims,

The police have a system, which is a computer aided dispatch system, that already has these priorities. And we plug straight into that with every single piece of information that we have. A standard alarm from like VIVN or ADT, that has its own category, which is don't go. And that's why they don't go.

Prateek Joshi (14:06.39)
And when you think about different levels of intervention, meaning just having a loud alarm is different from like pepper spray because it's like taking an actual action. Yeah, for example. So how do you think about, how do you weigh escalation risk versus deterrence? And how do you make those decisions? Because you don't want to pepper spray somebody who's just an uncle who's visiting and you just don't, it doesn't know.

David Selinger (14:12.482)
Thank

David Selinger (14:16.944)
For example.

David Selinger (14:24.816)
Mm-hmm.

Prateek Joshi (14:34.99)
So how do you think about these different levels of interventions when it's automated?

David Selinger (14:39.536)
So that's a great question. By the way, some people do want to do that, but I think equally importantly, how do we decide not to do that? And so what we do is with our levels of intervention and our modality of interaction, we took a very product centric view for this. The security industry is a very, very service oriented industry. And so the average way, I'm going to start off by describing that first, because I think it'll give you the backdrop for this.

The average way a security company does this is they hire a former special forces person who comes in and sells the customer. says, oh yeah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah,

of transfer of information, then they quit and the next person shows up and there's even less transfer of information. And so again, this is kind of the service oriented view. So when I looked at this and my team and I looked at this, we said, how do we take a product oriented view? How do we tear that into the model that we actually use is AWS. So I was at Amazon really early and I really loved the way that they looked at computing, right? They said, instead of talking about a computer, talk about storage, talk about compute, talk about memory.

talk about caches, and you turn it into an individual microservice. And that's what we did with security. So we have individual microservices for things like there's a medical emergency. We're protecting a public pool that has hours. What are the expectations that we do there? And this is away from your question around like pepper spray versus yelling, but at the end of every single one of those procedures is.

a set of steps about here's what the first intervention is, here's what the second interaction is, here's how it escalates, here's how it further escalates. And then again, by putting it in software, we can also say you're here, you've now reached the point where we need a supervisor approval, we need a customer approval, or we need some extenuating circumstances. And so we've implemented all of that into the software. And this kind of goes back to your original question about vertical integration.

David Selinger (16:54.82)
No one else is even thinking about doing that because nobody else holds the guards and the cameras and the AI and the video feed all in the same place.

Prateek Joshi (17:06.222)
Amazing. Let's talk about how the data is processed. So can you talk about what types of data is getting captured as a starting point? And also with regards to processing on the device, in the cloud, a hybrid model. So how do you think about processing this huge volume of data that's almost getting streamed to you from many different towns?

David Selinger (17:28.9)
Yeah, so there's

Wow, there's a lot of data, first of all. I'll start with the four or five most important chunks. So the first is the video feed. The second would be the history of that property. Are there other things that have happened to that property that go into that? How about this camera? Does this camera specifically have issues? And then the next chunk of data that we then look at is what was the guard's response time and interaction? they...

Did they click in immediately? We have multiple views and screens. Did they investigate? Did they zoom in? What type of behavior did the guard do on the console? And then the third chunk of data that we really look at that I find really interesting is the audio streams. When we intervene, was there a response? When we turn on the siren, was there a response? What was the response from the person, from the video and from audio? Then if we call the customer, how does the customer respond?

That's both the transcription as well as tone analysis, things like that. And then when we call the police, how did the police respond? How long did it take for them to respond? How quickly did they deploy officers? Did the officers arrive? There's like various codes, right? Did they arrive with just their partner? Did they arrive code three with four people, weapons drawn? Did they have hands on the weapon? Did they have long guns drawn? These are all different levels of escalation that we can observe from the police as well.

that give us data that can say, hey, if I see this thing at the very beginning of the event, and that has a very high probability of resulting in 10 officers with long guns, that's a pretty high severity event. If instead I have this that results in no investigation or maybe a little bit of investigation and no call to the police and no call to the customer and no officers with guns, that's a low severity event. And so that is my number one, again, kind going back to your false positive question.

David Selinger (19:25.636)
That's my number one chunk of data that I want to look at and use to make the system faster, better, and even cheaper, even faster, even better, and even cheaper than it is today.

Prateek Joshi (19:36.654)
Let's talk about the business model and the market. if you look at, you mentioned Ring early on, and Ring has neighborhood scale, and there is a Workata, they own a segment. So when you think about deep sentinel and the positioning in the market, where do you place it, and how do you explain it to a new customer who's considering a solution?

David Selinger (20:00.153)
Yeah, so mean, we again, we support a lot of third party hardware. You picked two that we don't yet support. We would love to if if the Vrakata team's listening or Jamie Siminoff over at Ring, give me a call. But but we do support third party hardware. And so that's a big part of this is that we see this as as an ecosystem. Our go to market strategy is through local dealers and installers. And so these folks tend to bring together

All right, I'm going to bring in an analytics package, I'm going to bring in a security package, and then I'm going to do your door and window sensors, or I'm going to do something called CEPTED, which is Crime Prevention Through Environmental Design. I'm going to help you design your property to be more secure. And so we partner with some of the people that do the entire solution for the customers. And then we generally integrate with almost everybody out there that we can.

Prateek Joshi (20:51.982)
And when you look at the various channels that you've sold through, like Directive and Zoomer, through installers, through integrators, so what channel has worked well? And also what are the learnings you can share for somebody new who's trying to build a hardware software company and trying to figure out how to go to market, how to get to customers?

David Selinger (21:12.55)
So I'll say the dealer and installer is definitely our best path to market right now. And then I'll add to that, I am also five years ahead of anybody else who's listening to this and asking that question, and it takes about four to five years to get that to scale. And it took us about four to five years to get it to scale. It was really, really challenging. It was hard to find what's your ideal customer profile, then inject that into your ideal partner and reseller.

profile and then there's all these supply chain issues like a lot of the last mile installers, they make a decision on suppliers based on variables that you and I would find surprising. And so you have to know what those things are and as silly as they may seem to you or I, that's how they're making the decision. You have to think about them just like a customer. I don't really care. Here you are sir. Let's make sure that we can get this done. And that took again like four to five years.

In terms of the total go-to-market though, what we have found that's really neat is that we have been able to combine some of the, I don't know if you know Andrew Chen, but like the best practices of D to C into our B2B and reseller model. And that is something, know, just do email testing, do landing page testing, do lots of vertical-based testing. And so by doing that, we've actually combined that

together in a way that I'm really proud of. My team has done just an amazing job of this. And so what we do is we build out like vertical solutions for AutoLots, which is one of our top verticals. And what we do is we'll have a ton of content that we'll produce and we'll publish it using DTC channels, fast feedback, lots of feedback, tune, tune, tune, tune, tune, then use the partner channel as the broadcast and scalability. And so

Using that combination has been really powerful. The other thing that we found in that same vein is that feeding high quality marketing programs and leads into resellers that have a hundred thousand dollar top line revenue or a million dollars a year in top line revenue, they don't have the room in that budget to go do LPO and SEO and SEM and content development. We do. And so by partnering with them to provide them that.

David Selinger (23:34.406)
We're also providing them a ton of value.

Prateek Joshi (23:39.032)
Deploying in a residential setting versus commercial setting, I'm sure there are differences. But if you had to compare and contrast along the dimensions of marketing to them, selling to them, deploying and supporting, and also just the product needs. So how would you compare these two big broad segments?

David Selinger (23:57.903)
Yeah. So I'll start off right off the bat and say B2C is way more profitable. Almost infinitely so. The churn is almost zero. We have almost no commercial churn unless the business goes out of business really for the most part. On the B2C side we have B2C type churn. And so that's something that you have to take into account in your total economics and in your marketing planning.

That said though, we found is that marketing channels, when they make a division between B2C and B2B, are generally just charging you more money. And so being able to not use that targeting and have a solution for B2C and a solution for B2B on the marketing side, it means that we don't have to use those kind of BS segmentation tools that are at the edge and they don't really work that well. We've tested them and if you could just service all the demands, which if you can't, then you can't, right?

So that's why we have B2C is because we can service the demand. Then the total cost for customer acquisition relative to customer LTV is way more efficient than it is if you're just focused on, especially on B2B because B2B is such a competitive marketplace for selling. And so by having a conjoined marketing program, it does get confusing at times with the economics work out much, much, much better for us. The second, sorry, the third thing I'll add in just because I wanted.

This is where I live. In that better, faster, cheaper solution, residential is way harder. It is way, way, way, way, way harder to solve. To take it and programmatically say we're going to do this, this, this, and this. A business generally can ingest that information. Residential can't as well. And so we also see it as our gym time as a company. That's where we go to build a muscle. That's where we go to train. That's where we go to cross train. And then we go out on the field and we go B to C and we sprint.

But it is our residential is so hard and so strenuous. I love it because it keeps my whole team like it keeps us all on point all the time.

Prateek Joshi (26:04.931)
Right. That is a great, great, I love the gym, like training example. And that's what it feels like was consumers. are very finicky. We churn a lot. We don't want to pay a lot and we have high demands and on the commercial B2B side, it's the opposite. that's where, so it's a very good, good example.

David Selinger (26:25.702)
You can sell residential and make it profitable. You can make commercial really profitable. And that's really a big part of like, when I say we're better, faster, cheaper, we're more than 50 % less than our next closest competitor. We're closer to like three to four times cheaper.

Prateek Joshi (26:29.43)
Right, right.

Prateek Joshi (26:41.824)
Now, let's talk about regulation. California and also many states, are debating regulations on autonomous security, robotics, and I'm sure more is coming. So what legislation are you watching the most closely, or rather, what impacts your business as you see it in the next couple of years?

David Selinger (26:53.03)
Thank

David Selinger (27:05.786)
I see the government more like a consumer, just kind of dovetailing off that last conversation, I see them more like a consumer than like a business. They're fickle, they're unpredictable, they change their mind all the time, and they act with emotion versus with logic. Not all government, but let's just say, for example, our government acts that way. so in that way, what I'm watching most closely as it relates to legislation is really the consumer opinions, right?

our opinions swaying this way or this way. There was a great analysis that came out a couple of months ago. And this looked at like what I would consider the most objective part of government, which is the judicial branch. And even the judicial branch over a hundred year timeframe tends to follow population sentiment. It's not some kind of ideal driven, the constitution is X or the constitution is Y. While that is the calling.

in general, our court system tends to follow public opinion. so, I'm waxing poetic a little bit here, but like, so what I watch is not any individual policy as it relates to AI, but really where are the fear spikes, where are the demand spikes, where are the extreme incidents that create either of those two things, much more so than any individual policy. I think if I were to...

to speaks directly to the one that you highlighted, which is autonomous security. This is where our system where we have specific deliverables that are constrained and then controlled by humans, where the human is in the loop as the controlling function and the AI is the supporting function. I think that keeps us in a more conservative spot so that I can answer that way. If we were all autonomous and all AI, I would definitely need to be all over where all these policies were.

Our system is 100 % designed to have humans in the loop all the time.

Prateek Joshi (29:03.246)
Another important consideration here is insurance. So if I own a car dealership and if I proactively take action to install a security system, maybe I use Deep Sentinel, obviously I want a discount because the insurance company's job is now a little bit easier. how does, one, how does insurance shape out with this autonomous security? And two, what role does Deep Sentinel play?

in at least structuring those deals if not outright driving.

David Selinger (29:35.589)
Yeah, so we have partnerships in various stages with a lot of the major insurance companies. We already get small discounts, but we're working with them to get a higher level of understanding into their underwriting. Insurance does tend to be a pretty slow moving industry, mean, actuarists have the highest job satisfaction of any category for the reason that they just go home at five o'clock and I don't think they think about work at all. And then they come back and do the same thing the next day. But that's my opinion.

And so in general, I expect it will take another five years or so before insurance takes things like this seriously and is able to you know, can call it a quality assessment where do you have this solution is not enough. It's, okay, how many cameras do you have? How well protected are they? What resolution are they? Which provider is doing this? What's the gradation of that provider? Things like that.

Because when you look at like fire, which is I think a much more mature space, when you look at the data that go into fire models, they aren't just like, do you give a discount because you're in a good zone or bad zone? Fire models tend to have three to five different zones and ratings, and then they have qualifiers for wind, they have qualifiers for slope. And that's where I think this is going to get to where the security component plays a really serious role because as we all are seeing securities and crime as

massively more impactful on all of our lives than it was 15 years ago. And so I think that degree of breaking it down is going to be really important. Nationwide is one of our top investors. I love working with them. They're very forward thinking. And so we talk with them about this almost every quarter and kind of just gauge where they're at in that maturity life cycle.

Prateek Joshi (31:23.756)
I want to talk about company building for a second. So let's talk about Dave, the founder, like 20 years ago versus today. If you had to share nuggets on, and you've built companies, right, Finn and Rich Tollevans and now Deep Sentinel, obviously you have strong experience in building and shipping a lot. So what nuggets would you like to share with younger founders and builders? There's a good chunk of listenership and also just people who are listening and want to

build companies.

David Selinger (31:55.207)
Yeah, I mean, think there's a couple that are just things that other people talk about too, right? Which is things like extreme transparency. And man, I'll tell you, there's so many forces when, as soon as you get into a boardroom with some high powered VCs to do things this way or do them this way. And by this way, mean hiding stuff from your employees or doing this stuff in a back room. Let me just be really clear. And I would say that there's certainly things you have to keep confidential.

Prateek Joshi (32:16.526)
you

David Selinger (32:25.03)
Like period, full stop, right? I'm in the security industry. Security industry is based on confidentiality. 90 % of that's BS. And I would way rather build and run a company the way that I'm doing it now than the way that I did it 20 years ago. One of the things that I do every single week that I think is the simplest and easiest thing and probably the thing I'm the most proud of in my development over the last 20 years is every single Monday we have a town hall.

Anybody can ask questions, da da da, that part's cool. Allowing people to ask questions is a pretty weak way for a founder to get out of actually addressing transparency because that puts the onus on the other party. It's like burden of proof, right? In a court system. and burden of proof is a big fricking deal. It determines what happens and what doesn't happen. And so in the same way, having a town hall where you can have people ask questions is, you really need to have the strongest employees in the world in order to make that work. So instead, what I did was I picked.

the top agenda topics, and we update every single employee on these three topics every single week. Number one, how much cash do we have in the bank? Within the first three minutes of a town hall, every single week for the last seven years, my CFO has gotten on the phone and said, here's how much cash we have in the bank. Is that good or is that bad? Here's our sales budget for the next month, next quarter. Here's where we're at. Is that good or is that bad?

And those two, man, there's so many ways where you do it sometimes, you don't do it, or you kind of, it's really good, you know, but we didn't really hit the mark right, but like, look at all this good shit. That's all BS, right? At the end of the day, if you didn't hit your budget, didn't hit your budget. Period, full stop. And I'll tell you, we pretty much hit the budget every single month for the last year and a half, and that creates such pride and loyalty and trust.

to know that no, no, if we miss it, they tell us, but they hit it, and they hit it, and they hit it, and they hit it. Now I'm gonna go do my job. The third thing I did was, and this is super, super, super, super controversial, because it has to do with equity in stock, which is another thing a lot of those high-powered VCs will tell you. No, no, no, what you do is you put the number of shares on the offer letter, and then we're just gonna do like a.

David Selinger (34:49.958)
400,000 to one reverse split and it'll be, or 400,000 to one split. It'd be cool until everyone, get 400,000 shares. They're going to, they're going to piss themselves. And it's like, yeah, you could, you could totally do that. so I went entirely out of the way. Every single offer letter comes out with your percentage of ownership. Every single employee that gets stock in deep center will get the exact same stock structure that I have, which includes founder shares. Period. It includes double trigger.

period, it includes the same vesting schedule. And this way, when I hire a new executive, they get the exact same thing. When I hire a new employee, they get the exact same thing. They know that they don't have to ask me questions, because I pre-negotiated the package that executives will accept. So I spent at Rich Relevance, probably 20 % of my time answering questions about stock. Once we passed about $50 million in valuation, everyone wanted to know, did I get this? Can I get this? I heard somebody got this. Can I get this?

And it was such a waste of time. you know what? the end, not only did it waste time, it drove trust down because everybody was looking over their shoulder to see what's going on. I give literally every single person the exact same stock structure. I just hired an SVPM marketing. We'll announce that in a week or two. He has the exact same structure as my most junior customer care agent and my guards, my members of my engineering team, everybody.

It's a five-year vest. It's all back-weighted. It's all double trigger, and it always includes founder shares.

Prateek Joshi (36:25.614)
Amazing. I love that advice. Maybe a follow-up question. Many times in a startup, right? It's pretty bumpy and many founders, especially maybe first time founders, they feel the need to protect. Meaning like if we just tell that, my God, are seven months away from cash out, but you know the next round is coming, it's just a pretty, for many people, it's like red alert. And so to protect the team, I absorb all the tension and only present the good news. So how do you...

How do you maybe screen for teammates then? Because there's a certain level of confidence and trust you need to say, hey, this is exactly how much money we have left. And this is our cash out. This is how we share, how the shares are structured. So how do you advise first time founders to rip off the bandaid and be more transparent? Because they get very nervous about, my God, if I share there, the team will freak out and everyone will leave. That's the prevailing sentiment with first time founders.

David Selinger (37:15.931)
Yeah.

David Selinger (37:24.167)
Number one thing is do it early, do it when things are good, and do it all the time.

Do it early, do it when things are good, and do it all the time. If you do it suddenly, right, like, oh, what's going on? You do it only when things are good and then you stop doing it when things are bad? Let me tell you, from my experience, every single time I've been in trouble as a founder, when I did not tell people they thought it was worse than it was. So if you're gonna think about...

Reality is the mean of a normal curve. Everybody's two standard deviations worse than where you are. so, just quit because they're like, don't know. I think things aren't going super well, but I don't know. And so, I'm gonna have to pay my mortgage, pay for my kid's school, pay for my gas, pay for my groceries. I gotta assume I'm at least a standard deviation out, but I don't think people actually think this way. I do. But two standard deviations out, I'm gonna assume the worst because...

Prateek Joshi (38:09.504)
Right.

David Selinger (38:25.978)
I gotta look out for myself, or at least I gotta go start looking, right? And as soon as you start looking, that's the slippery slope. The slippery slope of transparency is rad. It goes the right direction. The slippery slope of I started looking, dude, once they do that, the commitment is made the second they open up Indeed or they open up that email from us on LinkedIn, right? And the cool thing about this is not about tricking anyone. It's just about like, hey, let's treat each other like adults and have a real conversation.

Prateek Joshi (38:35.715)
Yeah.

Prateek Joshi (38:43.598)
Yeah.

David Selinger (38:55.558)
And there's a book by Stephen Covey, which is my favorite business book. It's called the speed of trust. Old Stephen Covey didn't do that well after Franklin Covey went in, like, but nonetheless, the principles are really on point here that, trust can never be given. It can be lost and it can be built. And if you think of trust as being the single most important vector for information transfer amongst your team members, always be building trust, always be closing says Alec Baldwin.

I say always be built in trust.

Prateek Joshi (39:26.638)
Right. Amazing. I have one last question before we go to the rapid fire round. with all the, you've built before the LLM era, during the LLM era, and now we are kind of in the midst of it. So what AI capabilities excite you the most as it pertains to building Deep Sentinel? And also two, three, five years from now, what new capabilities do you need to unlock

David Selinger (39:31.078)
Yeah, please.

Prateek Joshi (39:56.289)
the next level.

David Selinger (39:59.275)
Distillation is the one that I think is the most interesting. So distillation is how some of our friends over in China took all the research at Meta and OpenAI and they built their own model for like $6 million, right? And that got a lot of attention when that happened, but not many people looked at like, how did they do this? What did they do? What was going on? And distillation is a technology that's been around for a long time. It's a way to take...

one model and teach another model how to do something. And so as we build these like massive LLMs that get smarter just through being in the world, which I think is just so neat, what we also get is we get all these features, all this analysis. And if you go back in AI 15, 20 years, almost all AI practitioners spent 90 % of their time cleaning data and doing feature engineering. That was what we did. That was

That was it, right? Like I was a feature engineer. If I had to go back and retitle my thing at Amazon, it'd be like manager of future engineering. The modeling part was this big. The feature engineering part was this big. And the LLM's ability to remove the feature engineering requirement for general problems across domains. Yes, what they do at the end of that is really, really, really cool and it's visible and it's really neat. But if you look at the way

the lower layers that go into the attention and the transformer models, all of the features are getting refined and refined and refined and they're getting refined not just universally, In kind of the models that were really powerful in the early 2000s were Bayesian models. So they were intentionally naive. LLMs are intentionally not naive, which blows my mind. So when you say the word

Two, right, in a sentence, the LLM has the ability to discriminate, whether that's T-O-O, T-O-O, T-W-O, and why based on the entire context. Now apply that to every other word in the language and all the different places you might use it, whether that's at a surgery table, or that's in a Dungeons and Dragons campaign, or that's in a science textbook. That ability to contextualize features is so, so, so, so, so exciting. And then the ability to use distillation to apply that.

David Selinger (42:21.59)
rapidly vertically at low cost and high impact into all kinds of solutions.

Prateek Joshi (42:27.566)
That's incredible and I agree, think distillation is a phenomenal technique and also it just unlocks so much efficiency because right now models are getting so much bigger and you can't keep throwing hundreds of billions at it so you have to engineer your way into more efficiency. That's fantastic. Alright, with that we are at the rapid fire round. I'll ask a series of questions and I would love to hear your answers in 15 seconds or less. You ready?

David Selinger (42:54.104)
Okay, let's do this.

Prateek Joshi (42:55.424)
Alright, question number one. What's your favorite book?

David Selinger (43:00.038)
I gave Speed to Trust by Stephen Covey. I will say that my favorite book, book book is Snow Crash by Neal Stephenson though. I'm reading it for like the 20th time this week.

Prateek Joshi (43:11.118)
Amazing. next question. Which historical figure do you admire the most and why?

David Selinger (43:19.782)
So a lot of people are gonna say Abraham Lincoln, I'm gonna say Ben Franklin. I don't even know if it's true about the whole lightning thing, but like the balls to face opportunity and do what it takes to figure out if it works. I always find that story inspiring.

Prateek Joshi (43:33.418)
Yeah, amazing, amazing human. Right, next question. What has been an important but overlooked AI trend in the last 12 months?

David Selinger (43:44.199)
So I think that we've struggled to understand the new Turing test where it's answering the question, does the computer feel? It's does the computer have sufficient capacity to convince me consistently and coherently that it's feeling? And that I think is really, really, really interesting.

Prateek Joshi (44:07.2)
What's the one thing about physical security that most people don't get?

David Selinger (44:12.762)
I mean, I already mentioned it doesn't work, right? That would be the first one. I think the other thing is, let me answer the thing for people though that are thinking about their house or their business, is let go of your pride. People are always saying, look, I installed this camera and it's so rad and it has this. Switch. Use empathy, be a criminal. Put yourself into the mindset of a criminal and say, cool, can I walk around that? you look, I just walked to the left. People are so hung up on how rad their security is.

that they don't put themselves in the shoes of a criminal. I went to an A-list celebrity's house the other day and he showed me his wall and his gate and his cameras and all this shit. I was inside of his house in 15 seconds from the street. Put yourself into the mindset of the criminal.

Prateek Joshi (45:00.503)
Next question, what separates great AI products from the merely good ones?

David Selinger (45:08.71)
I think just like any other technology product is how well they understand the customer, mean, it's really, do you build something that wows? Open AI's big wow things are almost all around image generation and that was required to open up the door for all the other stuff. But unless you had the images that like really connected with the human beings, you didn't get anywhere.

Prateek Joshi (45:31.018)
What have you changed your mind on recently?

David Selinger (45:36.951)
I thought LLMs had peaked about a year and a half ago and I was super super super super super super super wrong.

Prateek Joshi (45:46.638)
What's your wildest AI prediction for the next 12 months?

David Selinger (45:52.881)
So I mentioned the thing about emotion. I think that dating apps are gonna go the way of the dodo. I would way rather have 10 or 15 or one AI-based person that actually responds to me that isn't gonna scam me and try to get me to send money to Russia, that is gonna listen to me and be able to respond intelligently, that can provide therapy, that can provide emotional support than I would some bullshit Tinder.

based interaction with a real human being. And I think that 75 % of humans are going to go through that transformation in the next 12 months that are single and looking for companionship.

Prateek Joshi (46:32.974)
Amazing. Our final question. What's your number one advice to founders who are starting out today?

David Selinger (46:40.614)
I mentioned transparency, I will say that again. Number two though is financial model. Understand your financial model. Are you building something that is fundable by venture capitalists? A lot of people say it's venture scale and they go and they say, oh no, it can get really big. That's not what venture scale means. You need to understand that venture is a game where you gotta have a 10 to one odds at making a trillion dollars and they'll take those 10 to one odds all day long and they don't really care about

you individually as much as they may love you and hug you and stuff like that. They've got a portfolio and they only need one of the 10 of them to hit. And so know that. If you're gonna go and fund it yourself, know how the banks work. Know how you're financing your business and how that interacts with the members of the financial market that you plan on working with.

Prateek Joshi (47:30.766)
Dave, this is a brilliant, brilliant discussion. I loved all your insights and I think just the knowledge and depth that comes with building companies, I think it's just incredible. So thank you so much for coming onto the show and sharing your amazing insights.

David Selinger (47:46.683)
Thank you. It was very fun. I love your questions.