← Bitcoin Audible
Loading summary
Transcript283 lines
- [00:00]
A
Comma develops open source ADAS software. I mean it's, you can call it self driving software but currently what we ship is something where the user is supposed to pay attention all the time. We want the software to be open source so that people have full control. You know, the hardware is something you have root on. We don't have any access to your device in terms of that's necessary to make it operate normally. There's some telemetry that you know, you can turn off if you want. But if I say something you buy, you run open source software on it and there's no cloud involved in making it work as an ADAS product. We're pretty passionate about making sure that if you buy something, you own it. If you run software, it should be open source and you should be able to modify it or at least we shouldn't be able to stop you. That should be a thing that you control.
- [01:05]
B
What is up guys? Welcome back to Bitcoin Audible. I am Guy Swan, the guy who has read more about Bitcoin than anybody else. You know we got a really awesome treat today. So we have. And I can't even remember who it was. I should have gone back and actually looked him up but somebody mentioned about comma AI in a conversation and then tagged Harold. So Harold is the CTO from comma AI and this is actually a project I've been really interested in for a long time but I never really thought about like having somebody on the show to talk about it. And when somebody else mentioned it, shout out to whoever that was, I'll try to find the link and actually put it in the show notes. And thanks for, for tagging him because this was an awesome conversation but it's literally all we talk about is sovereign technology of owning your devices and the crazy novel idea of actually having a sustainable business where you make a profit and then you spend within your means. And I was really excited to have him on. Like I said, this is a project I've been following and been fascinated with for a really long time. If you're not aware of what comma AI does then you're really going to love this. This is open source, self hosted, like your device assisted driving. Like so it's, it's self driving and it plugs into, it's compatible with tons of cars that have been on the road for like, you know, I think it's like 2019 to 2020. There's all this stuff is like drive by wire and so many of them. You can literally plug this into and they have. I can't. One of their first devices he mentions they sold 10,000 of. This is like such a cool thing. And it's, it's you being able to own, you know, your smart car, so to speak. Their founder is George Hotz, who I'm sure you know by name or, and, or you've heard me talk about in other contexts. But the entire project is just really cool and it holds a really, a really cool spot in kind of the world of robotics and in building sustainable open source projects. So I really hold them in high regard for what they've accomplished because this is not a small feat to, to really kind of do what they've done. And so I was just really excited to have him on the show and just kind of pick his brain for a while about all, all sorts of stuff really. And it was an awesome conversation. So. So get ready for this one now real quick. I want to shout out, you got to get yourself a bitbox Nova. If you do not have one and if you do not hold your keys, you should feel ashamed and embarrassed. And no, you cannot withdraw bitcoin from an ETF you owe. You own a debt, you own a promise. And I promise you, if anything goes sideways, you're not going to get anything. You're not going to get a fraction of what you are actually owed and it will probably take like 10 years. So what you need to do is you need to buy real bitcoin. Get yourself a bitbox02. Use my discount code and affiliate link. Their whole stack is amazing. It's super easy to do. It's, it's a really, it's honestly so wonderful little experience. It is open source, of course, it is Swiss made. You have a beautiful little hardware wallet. I've got it sitting around here. Crap, I don't, I don't have it right with me. It's a beautiful little hardware wallet. And you, you withdraw, you buy your bitcoin, you withdraw it to it and then you sleep like a baby. You, you have your keys, you have it safe, you have your device that makes it easy to get to plug it into any device. And you have, you've saved sats. You've not only saved your sats from danger, but you've also saved sats on it because you use my code. So check it out, details, write down in the show notes, be a real bitcoiner, be a real holder, hold your keys. And with that let's get into today's chat. This is Harold the CTO@ comma AI sovereign devices in a subscription world. Welcome to the show, man. I'm really excited to dig into this comma is something that. And I'll let you kind of introduce the company and you know, everything that you guys have been working on. But I have been, I guess, just in general. I've always liked, been a. Fascinated with technology. And then on another hand, I've been infuriated by the lack of control I increasingly have over it. Like this, this kind of like battle between convenience and whether or not I own it. I ran into this just the other day, actually. I, my wife and I, like 21 pilots and we purchased their albums or whatever through Apple music, itunes, whatever you want to call it. And I have on and off had Apple music subscriptions because you get them for free for like six months when you buy a phone or, you know, whatever. And mine apparently ran out and all the music is still in my list, but I can't listen to it. And I was like in my car and I was like, wait, what? And, and come to realize that because it was purchased on my wife's account and even though I'm like in the family or whatever, there's something I don't, I don't even figured it out yet. I'm just so mad that I'm just gonna go download all the music. I'm just gonna go pirate it because if I'm gonna have to take time to figure out how to get my music back that I've already paid for, I'm just gonna pirate it and not use your app anymore. But I was just, I was just like, how have we failed so bad that like I have purchased music and I've lost access because a subscription ran out of. You know, I feel like this is
- [07:11]
A
like a common resurgence of pirating exactly this type of stuff. I started pirating a lot more recently too, for similar reasons. Even when you bought it, it's just like, it's just more difficult to log in than it is to redownload it.
- [07:25]
B
So yes, you have to go like go to my family settings and be like, are you the primary member or the. Oh God, I don't even want to. I don't want to deal with that. I don't even want to. I hate that I'm somehow like asking for permission for you from you. But. But yeah. So in that sense, I found out recently, actually I've always been fascinated with the idea of self driving. You know, anybody going on a long trip to think that you could do something else or like, you know, split your Attention a little bit and have your car drive for you. Especially in the age of AI when I could talk to my phone is utterly fascinating. But I don't like, I don't like the idea of buying a car that I don't like Tesla that can just be like remote shut off and that self driving is a $120 a month subscription. I'm so sick of subscriptions. But anyway, this little long intro rant of just kind of like where I've been thinking about and why I'm fascinated with karma. But introduce yourself what you've been, how long you've been with Comma and for anybody who doesn't know what comma is.
- [08:38]
A
Sure, yeah. So I'm Harald, I'm currently CTO at comma and I've been working at comma for I mean almost 10 years now. I've been here for a really long time so I've seen the company go through it quite a bit. And I mean in short, comma develops open source ADAS software. I mean it's, you can call it self driving software but currently what we ship is something where the user is supposed to pay attention all the time. It's like, I mean just like FSD basically and it's fully open source. So you know, if you wanted to run it on your own hardware, whatever you can, you can do all that. We also make hardware that you know, obviously is optimized to run that on, that you can install in a car and then you have like kind of of modern ish ADAS features on let's say a 2020 Toyota Corolla or something like that. And yeah, like you said, a big part of it is we want the software to be open source so that people have full control. You know, the hardware is something you have root on. We don't have any access to your device in terms of that's necessary to make it operate normally. There's some telemetry that you know you can turn off if you want. But if I say something, you buy, you run open source software on it and there's no cloud involved in making it work as an ADAS product, which is especially now that I mean FSD used to be something you could buy, now it's a subscription. Now they can turn it off if they don't like you, they can turn it off if you do xyz, you know, just kind of a lame direction to go for robotics. So we're pretty passionate about making sure that if you buy something you own it, if you run Software, it should be open source and you should be able to modify it or at least we shouldn't be able to stop you. You know, that should be a thing that you control.
- [10:26]
B
Yeah, no, I'm fascinated by. Fascinated is not the right word. Infuriated by the fact that, you know, the kind of promise of the future. We're getting so much amazing technology, but we're getting it in, we're getting it in this whole, you know, you'll own nothing and you'll be happy way, where it's, it's also becoming this mechanism for some corporation that's 2,000 miles away to basically have a place in my home or in my hand or in my divide. Like, like, it's like, oh, well, oh my God. We, we can have our own little corner of this thing and we can tell you what to do with it and we can, we can maintain control and we can just have you purchase our network out into the, our, our influence out into the world.
- [11:26]
A
Right.
- [11:26]
B
And especially when people are literally prosecuted like, like. And know which founder knows all about, you know, of, of course of just owning your own hardware and doing your own thing and then telling people what you did. Like, yeah. And to think that you actually might end up, end up in court about that.
- [11:50]
A
I, I think there's a, it is, I think it's a sad situation currently, but there, I think there's a lot of energy, especially with the AI stuff that is inspiring. I think the social media thing went down a very bad path, I would say, and I'm very upset about the current state of that. But I think with the new LLMs and stuff, I feel like the people think the stakes are even higher than they were before. And there seems to be a lot more drive. I think a lot of people are on board that making your every thought go to someone else's servers that is unencrypted and they can read your every thought and then optimize. That doesn't sound so appealing. So I think there's hope that this changes and I would like to comment to do a small part in robotics to make sure that there's at least some options in robotics that don't look like this too.
- [12:40]
B
Yeah, no, I 100% agree. I generally think the march of technology is a two steps forward, one step back. And most of the Internet era is in feeling the pain, in the stage of feeling the pain and pressure of the one step back and, and realizing what those consequences were. But we also have some major technological improvements and changes like for an Example is like the music and getting, being able to share stuff with your friends and family or whatever is. It infuriates me enough. I actually read a really, really great piece on the show called thank you for being annoying. But it was about the fact that like meaning and happiness is actually finding the thing that annoys you so much that you can't put it down. You have to like actually go after and do the thing. And so in that sense like this is one of those, this is my thing, right? Is I, I hate when I don't control my own media and my own content. And so I probably took the very ill advised or stupid idea to work on a development project around solving this problem and actually making P2P fun and not have the, the drawback like like not not being afraid of servers or like attacking servers, but just putting servers in their place. Like you know, they shouldn't own the content, they should just be service providers, right? And which was a retarded task. But then I think about something like comma and I'm like who, who is stupid enough to think they can create self driving cars and like a device Like I think about like all of the problems that we've just had with our like tiny team and what is really not that big of a project but feels insurmountable. And then like to add hardware, to add like it just, it just man, like what you guys have built has actually like just kudos, just kudos to people who have like built something awesome that I want to use. Like it's super underappreciated and nobody really gets the thanks of like you just built something awesome and that's hard as shit.
- [14:51]
A
I mean, you know, it's also taken for granted a lot. I mean back when I joined was extremely early, like the basic things had been said. The basic plan was already kind of set there, right? Of we're going to retrofit cars, we're going to reverse Engineer the car's APIs so then we can control the car. And this was somewhat of an extreme and silly idea at the time. It's like, oh, you're never going to be able to reverse engineer a lot of cars, right? Some of them might work, some of them might not. They're going to change over time. This is like an insurmountable problem. And yeah, obviously the hardware at the time was like something you had to build yourself and 3D print and all these things. And now 10 years later and okay, there's, there's hundreds of models of cars that Work pretty well that you can install this on. The hardware works. So yeah, I mean it's a, I mean it's been 10 years, so it's, it's a long journey, but it definitely
- [15:44]
B
didn't seem easy at the time.
- [15:46]
A
Yeah,
- [15:50]
B
actually walk me through what happened. Maybe your history a little bit too, leading into comma and then kind of where comma was when you got on board and kind of that journey, especially in the context of like, you know, building your own or producing your own hardware and, and like the, the cost and sustainability of something like this, like this also seems like a, just a huge endeavor and you're up against a lot of very big companies. You know, you're, you're in a space where you're competing with a lot of companies. So I'm just curious how. Give me the history, give me the history. How did this actually work?
- [16:32]
A
I mean, as for myself, you know, this is my first job, so there's not, there's not all that much history before. Gone.
- [16:39]
B
No kidding.
- [16:40]
A
That's great.
- [16:41]
B
How did you get roped in?
- [16:42]
A
Huh?
- [16:43]
B
How did you get roped in?
- [16:44]
A
So I came to America because I was going to do my master's degree somewhere and I saw a picture of the Santa Barbara University and I was immediately sold. So I ended up there. And then, you know, when I was graduating I looked what I wanted to do and I knew about George and I was really into the jailbreaking and the owning your own things. And then I saw he was kind of moving to cars with a lot of the similar ideas. And so yeah, I mean I was into that. I shot them an email, did some programming challenges and then ended up there. So that's kind of where that started. And when I joined Com, I was I think one and a half years in or something. It was like five or six people at the time. So there was like, there was an open source project that like it did something like it was lane key, like it was some type of ADAs. It was like lane keycub on the highway, worked on like a couple cars. And as for hardware, like there was no product. It was just, it was like a phone, it was like reverse engineered, had some new stuff put on and then other people could, could kind of copy that phone setup and, and like.
- [17:52]
B
Okay, so it's a little bit like, like a. Setting up a graphene or something at that point.
- [17:58]
A
Yeah, you just had to get a phone, you had to install the new software on the phone and there's like some hardware. I think we, we opened the plans and then they could like make the PCB with the plans and then you could kind of go from there so that, that's the stage that it was on.
- [18:11]
B
And.
- [18:12]
A
And then we kind of started making more of the hardware and started making the hardware more from scratch and less like repurposing. For a long time our product was basically we would buy phones in bulk from China and then we would gut like the battery out and we would replace one of the cameras of the phone, put on our software and then put the phone in like some kind of case and then you know, eventually. Now everything's actually purpose built and built from scratch to. To be like something that drives your car. Yeah, I mean as for me, I mostly work on all the self driving stuff. I haven't. I mean I know about what's happened with the whole hardware trajectory but my job has just been working on the Autonomy software and the machine learning mostly. Yeah. Over time I mean things get slightly better. I think that the hardware is the one that's most striking in terms of the journey that it's come from because it used to really be like a. You know, it's clearly a hobby project at the beginning. But I think the product we have now looks like a real product and if you install it in your car it's not going to be embarrassing when you go on a date or show your family. Whereas they definitely used to be like that for a long time. Yeah,
- [19:24]
B
no, I'd be proud. I'd be proud of a 3D printed hag together looking thing.
- [19:30]
A
I mean we. Yeah, we sold the 10,000 of those.
- [19:33]
B
10,000?
- [19:34]
A
10,000 3D printed.
- [19:37]
B
Dude, you're OG. You're OG. If you have one of those. That's awesome. So It's. So it's 9.99. Sorry. I just realized I don't even know like sure. What the product. I've been, I've been so. I've been considering getting it. My wife is I guess a Luddite in the sense. She's just like I'm never. You want to test it out, you want to play with that, that's fine. But you're going to do it while I'm not in the car.
- [20:06]
A
Yeah, that's a, that's a con reaction.
- [20:09]
B
Yeah, yeah, yeah. Is. I'll even. Just. Just for fun. The, the Volvo we have and I'm. I, I assume that it will work on this. I'm not sure if it'll work on the, the a Nissan that we have. That one's like a 2017, I believe.
- [20:27]
A
I don't know off the top of my head, but could maybe, maybe not.
- [20:31]
B
I feel like the Volvo has to, it has lane assist.
- [20:35]
A
The Volvo may have like, a different internal network architecture that has some like, I don't know how much in the details you want to go, but basically on the CAN network, you can plug in and then anyone can send and receive messages. Some cars, I think the Volvo might be one of those has like, flex Ray that's more like point to point, and then we'd have to intercept the lines at multiple places, which makes it like slightly more complicated, a little bit
- [20:57]
B
more to plug, to plug in or to like, integrate.
- [21:01]
A
Sometimes the limitation is software that it's not clear how we can send the right signals to get it to do what we want. Sometimes the limitation is hardware. The install may just be extremely hard or impossible. Like, the most common car is, is like you just have like, where the rear view mirror is. You pop a piece of trim off, you plug it in and you put the trim back on. Takes like 10 minutes. It just depends on the car's like, network architecture and, and, and some of them have really weird ones.
- [21:28]
B
How much, how much do you, or how many different setups do y' all have to like, like, like, what's the rent? The degree of compatibility and problems y' all have in plugging into all the APIs. Like, are y' all like a list of things that y' all manage? Like, okay, is it, is it still compatible with this one? Is it still compatible with this one?
- [21:47]
A
Yeah, there's like a list of, let's say, 300 car models that we currently support in terms of maintaining that, like the, I mean, it's an open source project. Like, the community does help quite a bit with that. They do a lot of, like, the exploration and stuff and the cars that, you know, don't so cleanly work right off the bat. The community will put in a lot of effort to make that work.
- [22:05]
B
That's awesome.
- [22:06]
A
Yeah, yeah, but I mean, you know, as to your wife's hesitation, ultimately it's just cruise control. Really. It's just like, better cruise control.
- [22:13]
B
Yeah, well, I, I, I showed her. I mean, I, I, I, it has lane assist. The Volvo has la, which just will, will correct you. And I just, I just took my hand off the wheel or whatever and it was like, go, go, go. Like, it did it, it did it like, it did it like four times before it started beeping at me. I was wondering how long, but there were no other cars. Around. Okay. I was being safe about it, but it's, I don't know, it's just it, it's always fascinated me and unfortunately or fortunately, I don't know, I've always wanted always have to have like a project or a project car one way or the other and I haven't done it yet because we're still working on the basement but we're getting close to that and that's probably when I'll be like I haven't had a project for like five days that has overwhelmed me and frustrated the hell out of me and I'll just get a car and make it comma self driving.
- [23:08]
A
But yeah, if you, if you want to frustrate yourself, definitely try to install in a car that says it's not officially supported. If you get one that's clearly supported and we say should work, you'll probably have a not so frustrating time.
- [23:23]
B
Okay, what's the, what's the pream. What's, what's the, what's the prime. Um, like it's, it's just going to plug in and it's just going to work and it's, you've had great response and it's a decent car. Like what's the top one?
- [23:37]
A
I mean the most, the most common is like the Toyota's like a lot of the Toyotas.
- [23:41]
B
Okay.
- [23:42]
A
But honestly there is, there everything comes with some sort of caveats. Like even in the Toyota's there's going to be some models that have, don't support it or have caveats. It's, it's, it's unfortunately very difficult to give a clear answer to this. But, but our most popular cars are like yeah like the, the Toyota Corollas until like the last two years and then a lot of Toyota RAV fours are very popular. Rivian is something that like because you know it's like a very tech savvy car and they don't have good ADAs off the bat. The Genron Rivians is a very popular platform. I think we have okay, interesting like half or 1% penetration there. So like one, one that many percentage of the Rivian Gen 1 users installed. Yeah, it's like a big thing in
- [24:27]
B
the, it's pretty crazy.
- [24:28]
A
Yeah, it's a big thing in the like Rivian community where because you know all these people like technology and, and they're upset about the shitty, shitty ADAs it comes with. So
- [24:40]
B
how do y' all think about it from the context of like capital use? I watched a interview with HotS and as number of years ago now talking
- [24:50]
A
about
- [24:52]
B
was it Waymo, which I actually saw a wayo car on the road when I was traveling not too long ago in California. But regardless, it was, it was. He was talking about how like insane it was that to do this relative, this thing that comma is able to do with a phone that they had raised like a billion or more dollars and they basically burned through all of it, making this car with like 18 radars and you know, all this crap, and they still hadn't sat one out on the road that would actually like drive itself. And what was, what was funny is just how stark the contrast, especially because this is such a common thread in like just kind of like sound money circles. And like the, the original bitcoiners is this idea of like, you should not be wasting resources. Like, like money. Money is a reflection of the fact that we have scarcity in real life. You know, shit costs stuff. And if you take stuff, you, you, you're explicitly taking it from somebody else's use, so you better make good use of it, you know. And the VC and kind of like Fiat Finance enterprise is the exact opposite, right? It's just like, just, here's a billion dollars, here's a trillion dollars. Oh, we don't have enough money. We'll just print a trillion dollars and throw it at stuff. And it was so refreshing to hear that interview because it was like, it's just like you just. Did you not realize that you're using up a billion dollars worth of stuff and like, like, how do you sleep at night? And I thought that comment was the, the thing that like kicked me. And this was like, when I was like, oh, I was already interested in comma, but then like I heard that and the thought, the idea of like, well, let's do this in a cost efficient way, which is like, oh, oh my God, common sense. What do you know? But I, I guess, I guess my question from that is how, like, how cost heavy is it to keep this thing running? How sustainable is. Is the business? How big is the business? Like, like, I'll have a serious little endeavor and it's really cool. And you know, I want to, I, I would love to think that I could do something similar or that whether or not it's my project right now or, you know, something I do down the road, it give hope that I see somebody like Hama, like actually succeeding. And I'm just curious kind of what the biggest sustainability problems or biggest issues you guys have had over the years.
- [27:44]
A
Yeah, we have this really innovative business model where we try to sell devices for more than it costs us to make and then pay our bills with that. Absolutely crazy. Wow.
- [27:55]
B
Wow. That's crazy. Oh, my God.
- [27:58]
A
I'm very shocked by this. In terms of. Concretely, I think by far the biggest challenge is resisting the free money. I think this is what kills everyone else long term. Obviously, it's an open question whether this will kill the AI companies. I think it will, but, yeah, I think that's ultimately where it always all goes wrong. It's very easy to give. I mean, you said we're taking funding. We have taken $18 million in funding. Okay. And we've made, like a more measured amount. Yeah. I mean, you know, it's an m. You know. Yeah. We haven't. It's been, I think, over five years. I forgot exactly five years since we last got funding. You know, we have, I think, like $50 million in revenue now. Lifetime, you know, so our lifetime revenue is well, well above what we invest.
- [28:55]
B
The raise. Nice.
- [28:56]
A
Yeah. You know, as for our current financials, last couple months, we were doing million dollars a month in revenue. We have, you know, 25 people. These are all very, very manageable numbers in terms of never raising money again. You know, I think raising money is fine if, like, your plan to pay that money back isn't like some crazy black swan future event where you have to take over the world's entire AI supply and then we will, and then you return on your investment. Yeah. But I think. Yeah, the biggest problem is it is extremely easy to inflate your valuations and to get a lot of funding. And I think this just puts you in this trap where. Okay, well, you. You only have two ways out of that, which is you actually then live up to that, which is generally very hard. But then if you. If you see that's hard, you're like, well, just raise again on more promises and raise again on more promises.
- [30:01]
B
And it's either keep the. It's either keep the Ponzi, Keep. Keep the hype funding cycle going.
- [30:07]
A
Exactly.
- [30:08]
B
Or. Or have like a black swan success event, or you just fail and you've set yourself up to not even be like a lean sustainable company because you've trapped yourself in the. I have to pay this back already.
- [30:21]
A
Exactly. Yeah. Once you. Once the numbers start becoming so big. And it's funny you said a billion dollars earlier because today that doesn't sound like anything anymore. A billion dollars that'll keep, you know, open air, anthropic going for like three weeks.
- [30:35]
B
Oh, God, it hurts. It hurts.
- [30:37]
A
I wish there was more companies that tried to operate in this space while having some regular small business mindset of like okay, well let's keep everything concrete and let's make sure that these numbers work out relatively soon. Not like in some far away distant future. Maybe. Yeah, I mean like Waymo is a good like you know, Waymo is cool. I like Waymo. I drove with a Waymo. It's like honestly an awesome service, but it's like at this point almost a 20 year old project and like $10 billion in the hole, you know, Will they make that back? Maybe? I don't think so, but maybe. Yeah, I mean, you know, it's a kind of an advertisement for Google so maybe it's all worth it for them in the end. But. But there's a lot of companies that are not like that.
- [31:20]
B
It's a loss leader, right?
- [31:22]
A
Yeah, exactly.
- [31:26]
B
I. What's your Give me your, your brain on AI. What's your, what's your take your perspective on where things are, where things have gone in the last couple of years because and I'll just give you a little aside or context is like a pear drive. The stuff we've been working on is like basically was going at a snail's pace and I'm shocked by how much especially this like kind of last generation everything since, since like Opus of AI has like I'm not, I'm not a developer. Like, like I'll open up a terminal window and I'll have one open all day every day doing little things, you know, but that's the extent of it. Like sysadmin stuff, you know, and, but I am not, I will not. I don't even have the patience. I, I run out of patience arguing with AI to write code for me and. But holy crap. The amount of things that I have been able to custom build for myself or that I, I have been able to custom build for myself using LLMs has been just wild. In fact I have, I think I have a session going right now like while we're talking just, just double checking on stuff that it's not even like system stuff so it's like very easily like front end stuff. So I can usually just let it go, you know. But that, that's one of the reasons why I'm so hopeful. One of the things that you just recently you, you just brought up is that like, you know, I think there's more push for open source and realizing when the privacy problem becomes so deep of it will hear it My every thoughts it will LS my entire computer and all of my files and folders, you know, is why I'm super hopeful is because like we can just build all of this even, even using these, you know, centralized. Look at your whole computer tools. What we can then build to get out of that is kind of staggering to me. Um, and I'm just curious what you think about the current state, how it's affected what you guys are doing. I mean you technically have an AI product in a sense and how it's changed your, has it changed your call structure or time to delivery. Like basically, what's the, what's the whole map? What's the whole map of how AI has changed or not the last like 2 years for you guys?
- [34:01]
A
I mean, I think the problem is that it's very hard to tell is. And this is, I think one of the big issues with all this. I use LLMs, you know, every day for sure. Do I use them for programming every day? Most days I use it for programming. Has it accelerated everything? I have not seen any obvious evidence for that that it is accelerated. Anything we do, it feels like it. When you use it, it feels like it. But then if I abstract look at it, it's not super obvious to me that it has. You know, I like LLMs a lot. I think they're great tools. I think there's a few things, there's a few things that are worth like separating One is the LLMs are like a huge knowledge database. Right. And, and it, it just turns out that copyright sucks and IP sucks and it's just great that, you know, someone found a way around this stupid issue. Yeah, yeah, yeah. You know, they managed to compress all of the world's books into like something you pay $20 a month for. I don't know, the Hollywood lobby like raided Kim.com's house in New Zealand over like some movies. I don't understand how no one's house got rated over this, but apparently this is a fine loophole.
- [35:11]
B
Big enough with enough money.
- [35:12]
A
Yeah. So I think there's a lot of benefit from that. You know, I think that's, that's somewhat unrelated to AI. It's just a cool loophole. And then I think now there's like this agentix LLMs I guess, which is like the ones that write programs. And those are definitely useful to some extent and I use them a lot. But I think there's a lot of hidden costs and I think the hidden costs will take time to discover and I guess we'll have to find out what that is. Right?
- [35:37]
B
Technical debt.
- [35:39]
A
There's technical debt. Technical debt, yeah, that's definitely a big problem, is a huge amount of technical debt. But there's other issues too. Like, you know, previously, I mean it's even hard to think back now. If I would start on an issue from scratch and like write everything by myself, that sounds like super tedious and that would take me a really long time, but by the time I did that I would have this extremely deep understanding of what I did. Now it's much more like okay, it one shots it and it's like 90% there. And then I gotta like. And there's like these issues and I don't really understand.
- [36:08]
B
You have to figure out how it works.
- [36:10]
A
Exactly. And then yeah, I guess I haven't really done the side by side projects or I don't know if anyone has to see, you know, if you really want to get to like a very high quality at the end. And you go, one way is to start from scratch, writing it yourself and you build understanding as you go. And the other way is like, it almost works immediately and then you have to figure out these bugs and build understanding after. You know, how different is the end result. Given the fact that like, you know, technology hasn't like magically gotten better or more stable over the last year, I think I'm inclined to believe the difference isn't that big.
- [36:42]
B
Yeah, no, I, I completely feel everything that you brought up there. Because even the way that I think about it, it's the reason why like I've not really delivered anything built by AI. I, they're all custom. Like my workflow stuff, like it's kind of like personal scripts and personal apps to just have my process and if it ever goes, usually I just build it until it works and then I'm just reusing the same thing over and over again. Because it's a repeated process, like a, you know, transcription tool or you know, something simple. But in addition is I'm constantly just giving, you know, a new model if, if it ever breaks or something goes wrong. I don't have to, I don't have to deal with it on a hundred thousand machines. You know, I deal with it on my computer.
- [37:35]
A
Right.
- [37:36]
B
And there's a reason why the project that we're working on, the engine underlying it, the library that will be imported, that everything is using, is not built with AI practically at all. Like, and it feels like a 3D printer to me, you know.
- [37:54]
A
Yeah, that's a good analogy.
- [37:57]
B
It's the hack together we can build something very, very quickly. It's not perfect, it's not going to hold up very well. You, it's not super strong, but it gets the job done and it proves that we can build this. Now let's think about like design and architecture and make it a real product, you know, and I think there's a similar, probably kind of like 10x hyped cycle on top of it where you know, everybody's like, you know, we're going to have 3D printers, everybody's going to print everything in their homes and we're going to be able to build, manufacture anything that we want. And it turned out to be this hobbyist prototyping thing largely. And LLMs might end up being a lot of the same thing, you know, aside from the whole, you know, probably million pieces of software where there is a single user for that one piece of software.
- [38:53]
A
Because I thought about that analogy before but I feel like that makes a lot of sense. You know, 3D printing is relatively mature now and a lot of people use 3D printing for their personal stuff for like random one offs and, and it's great for that and I love that 3D printers are around. But yeah, it's, it's interesting how little effect it's had on, on, on industry. I mean we, we sold 10,0003D printed devices, but most people have not.
- [39:16]
B
Yeah, there's a lot of interesting, like I can't. God, I feel we've been working on this basement project forever and I'm in, you can't tell it because I got a green screen behind me, but I'm in like this crowded mess of a storage unit, pantry, laundry, garage, office and, and then we separated out a little corner and we make it a nursery for my baby when she's taking a nap. So it's a mess. And I have this box over here that I look at every couple of days like with this longing because it is unopened. My brother got me a really nice 3D printer last Christmas and I have not opened the thing and I'm very excited to start playing around with it, but it just makes me sad. Have you worked personally worked a lot with like 3D printing and like building your own little tools or anything?
- [40:09]
A
I have not actually. But that sounds like the exact same feeling with like every time there's a new LLM. Oh great. I can do all these things now. What things?
- [40:22]
B
Oh man, what's the. So, so you work on the software side of this, right? So the self Driving thing. What's been the. How would you say that Yalls philosophy in even just the software stack is different and or the same than something like Tesla or Waymo. Like. Like, how do you deal with like. Y' all have one camera, right?
- [40:48]
A
Yeah, yeah, one camera. Well, not two cameras, but facing the same direction.
- [40:52]
B
Two cameras facing the same direction. Okay. For 3D, I guess.
- [40:55]
A
Yeah. Well, no, but not. Not for. Just for. One of them's narrow and one of them's wide.
- [41:03]
B
Okay.
- [41:04]
A
Yeah, yeah, but you can actually.
- [41:05]
B
But it's. I don't mean it like 3D video, but I mean it like it's for depth, right?
- [41:10]
A
No, no, you can actually do good. Like at that length, you can't really do good depth. You'd need the cameras to be very far apart. Like even humans can really do depth that, you know, over. I don't know what the limit is, but maybe like 20 meters or something. I think that's like all the depth perception humans have at long distance is more based on like context cues. I mean, you know how big a car is, right? So however big it looks, that's like.
- [41:39]
B
Yeah, yeah, yeah.
- [41:39]
A
No, but we just need. We need like high resolution, far away and just wide. So we just have the two cameras facing forward for that. And then we have an interior camera, but that's not used for driving.
- [41:49]
B
Yeah, yeah, yeah. Wait, does this double as a. Does it record? Can you record on the interior camera and just have like a dash cam?
- [41:56]
A
Yeah, yeah, you can.
- [41:57]
B
Okay.
- [41:59]
A
Sorry. By interior, I mean facing towards the interior of the car. All the. All the cameras are in the same.
- [42:05]
B
Oh. I guess dash. A dash cam is technically the. The record your traffic.
- [42:08]
A
Yeah, yeah, sure.
- [42:09]
B
But I meant it that like you could have. I mean it in the sense that like, you know, you have. You watch a video of somebody getting pulled over and then you've got a camera inside the car, like recording.
- [42:17]
A
Yeah, yeah. It's not a great dash cam. There's like some features, but. But that's not its main purpose. But it can do that.
- [42:23]
B
Sure. Yeah. Just side question. It's just a thought. You know, I'm trying to dual purpose this thing, right? I'm convincing myself to get one and convincing myself to build the argument for my wife to be okay with it.
- [42:37]
A
Well, you can. You can vi code. I mean, the only reason the. The dash cam isn't great is just because we haven't really focused that much on the dash cam software. But you can vibe code your own dash cam app if you want.
- [42:49]
B
Oh, what it's open source and I have LLMs.
- [42:52]
A
What is exactly?
- [42:53]
B
That's crazy. Oh man. But anyway, that was a little bit of a tangent, but on that point, like how does the philosophy of, like the fact that you guys have bas, you know, basically two cameras doing the whole job, like in a way, Mo being the, the, the prime example here that we just brought up was they have like, like a big light. Like the one that I saw had like a big thing that's like spinning on top. That's like radar and sensors all over the thing. The things look like, looks like it's got warts. Like what, what's the,
- [43:36]
A
why did, why
- [43:37]
B
do they have all that stuff? Like what's the, what's the difference?
- [43:40]
A
I mean to start, you know, we're making a different product, right? Like our mission is to solve self driving cars while shipping shippable intermediaries. So like our goal is we want to make a useful product with, you know, self driving technology and make money and keep making that technology better. Currently this is an Adax product. There's no obvious roadmap for us to go from what we have now to something like a Waymo where you can sit in the car with no driver, right? Like we're making a driver assistance system. Ideally this driver assistance system drives the same or better than a Waymo, but we have no real immediate plan to like transition towards taking the driver out or anything like that. The driver has to pay attention. It's a convenience, luxury feature and a lot of people watch that. But in terms of what the technology we're actually developing and you know, the point is still it has to drive the car, it has to not crash, it has to drive comfortably. That's all the same as what everyone else is doing, whether or not they're trying to build robotaxis or not. I can go back a little bit in time, you know, and when commerce started, there was this very clear philosophy of humans know how to drive. And we have a lot of data that. So we're just going to find out a machine learning way to distill that driving knowledge into some neural network. At the time this was pretty controversial. Everyone's like, whoa, humans are that great drivers, first of all, which, I mean, that's a ridiculous thing to say because nothing drives better than a human. Then secondly, there was this thought of, well, that's very hard to control and you don't really have introspection and you don't know how to tweak it and you don't know how to Verify it. And I'd say that disagreement existed for, you know, maybe up to like three, four years ago, where players like Waymo especially, and then slightly less, but still somewhat. Tesla had pretty complex stacks that had like intermediate layers of like, they detect children, they detect cones, they detect traffic lights. Then they have some like logic on top of that that can be either neural network based or not. That makes some decisions based on this information on the scene. And we were always like, no, you know, that's. Well, first of all, this is completely unmanageable for a small engineering team. The only way we're going to get this to work is if we don't have to manually do much at all. And it's really based on getting good data, getting a good algorithm, and then having compute. But as time has gone on, everyone's kind of on board with this now. And no one questions this end to end approach really anymore. LLMs are also end to end. No one seems to question that all that much. And every now and again someone mentions interpretability, but it's pretty well accepted. If you want real machine learning capability, you need to do it end to end. So that was a big divergence then. I think we're still more focused on doing it an extremely simple way than other people. Part of that is we don't have many resources, so that's just impractical for us. But I do think this actually eventually leads to the best setup. Other people have some tricks to get around some certain problems. Generally if it requires some kind of trick, that's not a very clean way of going from data to just a single big algorithm and compute to a result. We kind of don't do it.
- [46:46]
B
Gotcha. What's the biggest problem? What's the biggest unexpected problem you guys have had over the years that just like you kind of had to rack your brain against?
- [47:00]
A
I don't know about unexpected problem. I think the biggest problem is just that everything takes five times longer than then. But I don't think it's one thing in particular. I think it's more everything together. Yeah, I don't think there's not, there's not like standout problems, I would say. I mean there's a lot of technical problems that I can talk about, but this is a general sense. I would say the problem is just that everything takes longer to do than it seems.
- [47:25]
B
Yeah. What's your main focus lately? Like are you just kind of doing maintenance and incremental improvements or are you working on something very specific around what
- [47:39]
A
you guys have so so we've. I can. Let me talk about the technical aspect and you can stop. Sure. If it gets too much, I'm an.
- [47:51]
B
I'm a nerd. So I'll. I love to get into the technical stuff and screw my audience if they don't get it.
- [47:56]
A
So. So if you, if you want to train a model to drive like a human, the naive thing to do is what's called imitation learning. You collect a lot of data of humans driving. You get some kind of trajectory of like, okay, in 3D space, you're taking this drive, you collect all this data of humans, and you have a model that can predict that. So it looks at the video of the road and it says, a human would have driven this trajectory here. And then you have some control system that implements that trajectory and tries to stick with that. It turns out that if you just train a model like that, this is like a famous problem, machine learning. Uh, you. It doesn't actually work in real life. It will make mistakes and slowly deviate from a good place. Like, it'll be in the center of the lane and it'll slowly drift out. And because the human data set doesn't really contain examples of that happening, and the human then recovering, it doesn't really. It doesn't really react to that. Well, this is a really big open discussion about which aspects of this problem exactly caused that real problem, and if there's ways to get around it or not. But let's just. That's. That's. That's the big problem. And then one way to solve that. Well, the way that we wanted to solve it is, okay, so if I,
- [49:05]
B
if I want to frame that like a little bit differently, it's that, like, if you give it only good data, there's like no context. Like, there's an assumption from the human perspective that it would be bad if this happened. But that's not in the data.
- [49:23]
A
Exactly, exactly.
- [49:24]
B
The LLM doesn't care, or. Excuse me, the machine learning algorithm doesn't care.
- [49:28]
A
Exactly.
- [49:29]
B
Just veered out of lane.
- [49:30]
A
It has seen no examples of making slight mistakes and making clean recoveries because the humans tend not to make any mistakes or they tend not to crash at least. And then you have these small deviations, and if you don't recover over time, these deviations compound. And then that effect looks like over the course of 10 seconds, you just drift out of lane. So you need to find a way to introduce data of these recoveries specifically of like, you drift slightly out of lane. Okay, now we need to recover from
- [49:59]
B
Like a negative prompt.
- [50:00]
A
Exactly.
- [50:00]
B
Right.
- [50:01]
A
Yeah, exactly. So that's kind of the base problem. And so there's multiple ways to solve this. One way would be to just detect the lane lines and tell them all, okay, go back to the center of the lane lines. Which is, you know, that's kind of the solution that Waymo and Tesla had for a long time. And I mean, even us had that for a long time, but really wanted to transition to something that was like, okay, but what is like a real fundamental machine learning way to solve this? Obviously, the data contains the information because humans are always in the center of the lane. So, like, that's a good thing. Right. Should be able to learn that. So one way to solve this is you can just have a simulator where you let the agent that you've trained act, and it will then drift out of lane in the simulator, because that's what it does. And then we can say, oh, well, clearly that was bad, because, you know, in the original data, that didn't happen. So you can teach the model, like, you drive in simulation and make sure that over time your driving looks like that of the human. Right. And so then we can kind of let it make mistakes and show it what recoveries look like.
- [50:59]
B
So kind of like the. The whole AlphaGo thing is. Yes, you let it. You let train, you play. It plays itself and then finds out what does and doesn't work. Because you actually have a. You actually have a environment that you can give it that has explicit boundaries and. And goals or whatever. It's like. It's like a game with a set of rules. Interesting. Okay. And then that's literally what y' all have done.
- [51:25]
A
Well, there's a few more.
- [51:26]
B
Okay, keep going. Sorry, I interrupted. I interrupted.
- [51:29]
A
So. So that's, I guess, like the base. The base idea that we had maybe let's say eight years ago now. This brings up a huge question. Is an Alpha simulator, a go simulator is very simple. That's 100 lines of code.
- [51:39]
B
Yeah.
- [51:40]
A
What is the secret for driving that is not as easy? So you can. There's like video games, right? Like, Grand Theft Auto is like a great driving simulator, but there's many reasons why you can't use that. Right. Because it doesn't. It doesn't capture the. I mean, first of all, Grand Theft Auto is not realistic, but it doesn't capture the full diversity of.
- [51:57]
B
Wait, what?
- [51:58]
A
Grand Theft Auto. But there's actually funny things that. There are funny things that are not realistic that you didn't know. Like the roads are wider for the width of the car so that you can go between other cars at like a stop sign, because otherwise it's not fun. But like, I didn't know that until I smashed through things. I was like, grand Theft Auto seems super realistic. But no, apparently, apparently they like tweak some things so that it's more fun to play. That I actually never, never noticed. But anyway, that's wonderful. Video games have this problem where if you want to capture the full diversity of the human world, like someone needs to make assets, you know, like 3D game assets to capture that diversity. And this is like an endless project, right? You'll never capture all of it. You'll always be missing something. So we need a simulator that is somehow inspired by the real data, which we have a lot of from everywhere in the world. And so what we did for a long time was extremely hacky. We had a video of the original drive. You can estimate the depth of everything in the scene and then you can reproject the image as though you were somewhere else, Right. So you can do some 3D geometry. And this has a whole bunch of problems. Like it doesn't capture reflections. It doesn't capture the fact that if you're in a new place and you didn't know, you couldn't see from the original perspective. Now you have a, you know, you have gaps and occlusion. So has all these problems. It creates like a pretty crappy simulator. But basically that's what we used for many years. And then now, most recently we shipped this at the beginning of this year and which was a huge achievement for us. We have a simulator that's just a machine learning generative model, you know, like the ones that generate video. But it's basically a simulator in that it generates video of the road and you can give driving actions and it will implement those driving actions. And that model is trained on all the real driving data. So it like captures the full diversity of our, you know, data set. So that's how it works now.
- [53:48]
B
Wow. So wait a second. So you have basically a machine learning system that produces a believable driving environment as your simulator for your other machine learning to learn.
- [54:05]
A
Exactly.
- [54:05]
B
Oh, that's awesome. That's great.
- [54:09]
A
I think that we had a paper in 2016 that laid out this idea, but back then there was actually really no way to generate realistic video or images with AI. So it's funny to see it in retrospect. We had all these primitive ideas of this sounds possible, but I don't know how.
- [54:26]
B
But that's just Never going to work.
- [54:28]
A
But 10 years later, I mean, we certainly didn't think it would take 10 years, but 10 years later, all the machine learning progress has actually led us to a place where, you know, you can actually do this.
- [54:36]
B
So, yeah, that's awesome. That's pretty wild. So this is only like that. That part of it is only like a year old.
- [54:44]
A
We've been working on it so for, say, like, pretty hard. Maybe like two, two to three years. And then. Yeah, this year is like when we. The beginning of this year, we actually shipped it to OpenPilot release, thinking it was like a major improvement. Gotcha.
- [54:58]
B
It reminds me of who was this that did this corridor Digital, I think, did a video on this. But it's a. It was right. Kind of at the beginning of when we were getting the kind of explosion of, like, machine learning. That's good enough to actually be impressive, you know, the chat GPT moment sort of thing. And there was a. I believe it was called a Gaussian splat, if I'm not mistaken.
- [55:27]
A
Yeah, that sounds right.
- [55:28]
B
But I'm. Yeah, I do editing and film is actually my original background. And I still am, you know, I do the show or whatever, so I still end up doing a ton of that stuff. And so I'm always interested in, like, how you can use it and integrate it, like, into effects. Not just like pure, like I'm gonna. When I type in a prompt and generate a video, but it was just a fascinating. I mean, I guess y' all probably use something similar to that maybe, but because you're using actual video to. To then produce the environment. But it was wild. It was just wild how, like, the lighting and the reflections and all of this stuff, like, you know, it's not something that was trained or like, programmed into it. It was just like the model.
- [56:21]
A
The.
- [56:21]
B
The way the model interpreted things was so accurate that it would just reproduce these things that, you know, other people had spent, you know, decades perfecting in 3D environments and, you know, having their own engine for. And it was just so. It was so cool to just kind of like, see it.
- [56:39]
A
But I was.
- [56:40]
B
It just made me think of that. And I was just curious if that was kind of related to how the.
- [56:46]
A
We are, like, so our, like, main simulator model is not a Gaussian splat model. It's much more similar to, like, if you, like, generate images with ChatGPT or something like a diffusion model, basically the same architecture as that. It's just. Except conditioning on your prompt, we condition on your driving actions, basically. So, like, it produces an image that's supposed to be consistent with the driving actions you gave, in the same way that when you chatgpt an image, it produces an image that's supposed to be consistent with the prompt you gave. A Gaussian splat's, like, a bit more geometric. We use Gaussian splats for testing and stuff, but I don't think you can't generate completely novel things with the Gaussian splats. You can't have the scene evolve in the future in some new way. It's more about reprojecting an existing 3D scene. Whereas our models, they're not quite there yet in terms of intelligence, but they should be able to just imagine anything that could happen on the road. So you could just let the simulator run, and then pedestrians appear and all these things. So. Yeah.
- [57:56]
B
Do you guys have a way to. I mean, of course you have a way that you try, but how do you kind of benchmark or see how much better this method or this. This modeling style is versus the previous simulator and that sort of thing? Like, how. How do you. How do you kind of judge the success of this?
- [58:18]
A
Like, we have, like, a whole entire test suite of, like, running, running the driving model in simulation under many different conditions. And we have all these metrics of, like, okay, it detected this many green lights, or it made this many lane changes successfully, or it deviated this far from the center of the lane in all these different cases. And some of those tests are based on real data. Some of them are fully artificial. Just like a whole suite and just a few hundred numbers. And the ones who are training the models generally have a vibe for which numbers are important, which ones aren't.
- [58:54]
B
Yeah, this is a little bit off topic, but I'm. I'm just curious from. Well, actually, before I get into that, how are you guys all in, like, spread out all over the place? Or is, like, a lot of the team actually in one location? Like, together? Like, how's. How's that work?
- [59:15]
A
We have an office. Like, I'm in San Diego now, and basically almost everyone is here. We have, like, a couple employees that. That work remotely and then will come in several times a year. But the idea is that everything is on site in the San Diego office.
- [59:31]
B
Yeah, I was about to ask that because, like, I work with a team, like, really small team, but basically nobody is in the same place. But I have a bit of a. I guess a traditionalist in me that's like, we should have a location, and, like, we need to get everybody together. But that also seems sort of insurmountable with the way we do the work that we're doing especially. And you know, because that is a cost in a lot of ways as well. But I'm just, I guess I'm just. You guys, you guys produce like a hardware, like you actually produce a product. So it's a little bit different. But I was just curious how y', all, how y' all think about that if y' all have everybody in one location. Because obviously that's a cost. And then, you know, obviously you, you have a pool of people that you, you, you have physical access to as opposed to, you know, remote access to. Yeah, I mean, basically what's the limitations or benefits of that?
- [60:42]
A
I mean, obviously if you're working on, if you're working on hardware, if you're working on production, you know. Yeah, it's not really feasible.
- [60:48]
B
You have to do it.
- [60:49]
A
But I mean, I manage almost exclusively software engineering and I wouldn't hire anyone remote right now unless. Yeah, I mean, an exception hasn't come along yet that would, you know, to make me change my mind about that. But currently, that's my opinion, I think it is much harder to stay motivated and to stay in the loop if you are not physically here. I'm not saying it's not possible, but I think it would put a lot more burden on the manager to make sure that those things don't cause problems. And that's just not currently how we work.
- [61:29]
B
Yeah, no, that makes sense. And it was funny how everything went remote in like 2020 and 2021. Like there was such an explosion and then it turned around so hard too. I mean it hasn't obviously completely turned around, but it's just kind of funny
- [61:47]
A
to see,
- [61:50]
B
which is a little bit funny analogy in the context of self driving is that we tend to overcorrect. We, we, we, we over steer in one direction every single time. I think we're doing that right now with ll that like AI is going to replace all these people and it's like, yeah, man. But you know how many hours I have argued with AI and then gone and called up or messaged somebody that I know personally to solve a problem. Cause like what do you do, what do you do when that thing sends you down the wrong path? You know, but, and I think we're, well, you know, overcorrect again. Like we, we tend to oversteer and go back.
- [62:29]
A
I mean, I think you, there's like some good things that came out of that Right. It's like, I mean, at our office, for example, there are no hours. We provide lunch and dinner, which motivates most people to come in around lunch and, you know, and be there for dinner. I think that's good. If people don't want to. They want to have a different schedule. They can have a different schedule. You know, if you want to work from home one afternoon or whatever, or a day or that's all fine. I think that that's a good change. Companies that were super strict about coming in at a particular time, I mean, unless there's a necessity for this. Right, There's a necessity for this. Like, if you got a.
- [63:01]
B
Your own customer support, I'm going to show up at 12:30 today.
- [63:05]
A
Obviously that's important, but, like, if your job doesn't have any schedule, then, then why do you have to be on a schedule? That. That's one of the things that came out. I think that's great. And, you know, sometimes if you. If you want to be at home and you got to do something, why not work from home? But, like, just as, like, a permanent thing? I think there's challenges that can be overcome, but are not super trivial to overcome.
- [63:25]
B
Yeah. Yeah. So one of the things that I've done since this LLM, being able to build things with LLMs is probably a bad thing, is that I've decided to. I have a bad habit of if I know how something could be built, I think, oh, I'll just do it myself. And. Which is kind of why the state of the house is the way it is. Um, and then I just get so frustrated or impatient with the situation that I'm like, I'm just gonna hire somebody. I'm just gonna do it. I'm just gonna throw it at them, and I'm gonna close my eyes and not care, you know, what the result is. But because of that, LLMs have made me very confident in my ability to take on a bunch of little side projects. And I am curious if you have a. A side project or something that you're building around these kind of, like, same ideas, or if you're just totally, like, focused on comma and everything you're doing there, or if you're like, me and you, you decided to build two extra things in your spare time.
- [64:36]
A
I'm trying to think. Yeah, I'd say I don't have a project. I recently had a child, and I think that's. That's all there is in my life is that. And. Yeah.
- [64:49]
B
How old's the kid?
- [64:51]
A
18. Months now.
- [64:52]
B
18 months. So my baby girl is 15 months.
- [64:58]
A
Oh, nice. That's pretty similar.
- [65:00]
B
Almost 16 months. Almost 16 months. And then I got a four year old, a little dude that just turned four.
- [65:05]
A
Oh, nice.
- [65:06]
B
And I, I, I feel that very, very real. Is, yeah, this after, after I'm done with my, my crap that is, it is all consuming.
- [65:18]
A
It becomes the new side project.
- [65:20]
B
Yeah, that's awesome. Congrats. 18 months. So they just started talking, just started communicating.
- [65:29]
A
I tell people she's really in her reinforcement learning phase. She will. If I point at objects she has not seen, like, if I put an object she knows, she'll say the words she knows for it. And if I put an object she's never, she doesn't know the word for it. She'll just say random words until we react positively. Make up syllables like, one of these is going to work out. She is so confident.
- [65:53]
B
Oh, that's great. No, it's, it's, it's so crazy because she just crossed into that zone of where like babbling turned into some form of like real communication. And we can realize she's, she's trying to communicate something or something or she'll just point and be like, this thing over here. And it's so entertaining.
- [66:14]
A
It is really funny. It really makes me happy.
- [66:18]
B
Yeah, dude. There is little joy and frustration that matches being a parent.
- [66:26]
A
Yeah, I fully agreed actually.
- [66:30]
B
Just how has that affected kind of how you see direction and what you're building and kind of goals? Like, I feel like it changed my perspective a lot on that sort of stuff.
- [66:46]
A
I think it made me much more of a Luddite, I would say.
- [66:50]
B
Yeah, it did. It did it so, so much, so much. Especially for somebody who's like super deep into tech. I just like, man, I don't want a smart refrigerator. God bless it. Throw that shit away.
- [67:04]
A
No, that's definitely made me, but also like with all the AI, like, I feel like before is like, you know, it was always like, oh, you know, if this all turns to like, this is kind of entertaining. But now I'm like, oh, I want to take away like, you know, her future. I, I want her to not be screwed over by this. So I feel like I see a lot of the like, technology, social media stuff in a very different light. Before I was like, yeah, that sucks. And like, that's kind of funny. But now it's like, that sucks.
- [67:29]
B
You were detached from it, right?
- [67:30]
A
Yeah. And now I'm like, I don't want that like to affect, you know, want my kids to have the same, you know, opportunities that, that I did. And, and like I see, see things much more in that light now.
- [67:43]
B
Dude. I have a bitcoiner friend who's a developer and he's like all into like the craziest tech and he's constantly building stuff on like Noster and all this stuff. And. But then he'll post and he's like, my, my little girl is not allowed to watch. Like they have like a strict zero screens. Like no screens whatsoever, which I cannot even pull off. I, I like, I think about it from like a healthy standpoint is that like, okay, we heavily regulate and the only time if there's punishment, it's almost always just like no screens. So that he's just kind of like forced to be out and doing stuff, you know. But dude, like, that's a level of like discipline that is hard to pull off. Since you kind of have a similar perspective on the whole social media and stuff. I'm curious how you. With an 18 month old, what's your. Do you have any fun strategies or lessons that you've learned?
- [68:37]
A
Yeah, we try to do no screens. I mean, except like FaceTime with the family.
- [68:41]
B
Oh, sure, sure, sure, sure. Yeah, that's, that's a little different.
- [68:44]
A
Yeah, no, but no, no screens besides that. I don't know. It's been working so far, but you know, we've only got the one. So I think as those increase, you gotta relax the rules maybe a little bit.
- [68:54]
B
How many, how many are you going for?
- [68:56]
A
I don't know. As nice. More, I guess is the answer. I don't know how many more. Too many.
- [69:03]
B
Yeah, we'll. We'll see. At three, our, Our like stated goal has stayed at four. Four kiddos.
- [69:10]
A
But that already requires a big car, so.
- [69:14]
B
Right.
- [69:14]
A
Five requires a bus. So then that's maybe too many.
- [69:18]
B
Oh, that's, that's. Yeah, you've. You've. When you got the church van, you know, like
- [69:26]
A
practical.
- [69:32]
B
Oh man. Yeah. Kids. Kids will change how you think about that. I find it funny that you had a similar. I've had a couple conversations like that and people have kind of come to the same kind of conclusion. Is that like, oh no, all this stuff with technology is no longer just like entertainment. It just makes you realize that, oh Jesus. You know, I have like. There's a degree of responsibility of protecting someone from it, you know, like, right. It will be left for her or him.
- [70:02]
A
Yeah, I mean, I just have a lot, I guess my, A lot of thought that I have A lot is like, well, my life is pretty good, you know, if I could give that same life to someone, that's like a success. So I'm like, if we stop all this right now, that's like a guarantee. Right? So I feel like that, that doesn't.
- [70:20]
B
Calm. Calm down.
- [70:22]
A
Yeah, I mean, I'm not a lot. I love technology. I think we should, we should keep improving it. We should just, you know, make some sense decisions. But I see where the mentality comes from much more now.
- [70:31]
B
Yeah, yeah, no, I, I 100% agree. I'm curious, your, like, if, if you weren't working on Comma, what would be the sovereignty, like, own your own thing project that you would most likely be driven that annoys you so much that you'd be driven to actually do something about it?
- [70:55]
A
Oh, definitely social media. I absolutely hate social media with such a passion and I hate all the companies that do it and I wish nothing for them to fail. And if I could accelerate that in any way, I would try really hard.
- [71:10]
B
Are you a Nostr guy at all? Do you even know that?
- [71:14]
A
I don't know. I installed Blue sky and then like, wait.
- [71:18]
B
Oh, Blue Sky. Yeah, yeah, yeah.
- [71:19]
A
Wait, but Nostr is also a Twitter equivalent or no. Or am I misunderstanding it?
- [71:24]
B
Sort of. So I think of Nostr as PGP actually kind of done the way it ought to be done.
- [71:30]
A
Okay.
- [71:31]
B
Even though it's not treated that way. But most of the apps that use it, like Primal or Domus or Amethyst, they are, they're very much like a Twitter replacement style thing. But largely what I think, I think it's actually kind of like under appreciated, like what it actually is because it's got like bitcoin and, and zaps like, like just kind of like tipping and paying for stuff like built into the protocol because you can pay directly to a key. And so because of that, like, it's really just kind of like a distributed yellow pages where it doesn't really matter what relay or where you get the information from. You just know this note is from this person that you follow. So it's got like a relay network and it's the first semblance of something to me that actually seems like it has real potential. But I also think the idea of just, just purely recreating social media isn't quite the optimal thing to do with it, you know.
- [72:35]
A
Right.
- [72:36]
B
It's kind of a side thing of like, you can also do this, but like PGP and like Trust, Web of trust networks have been this thing that, you know, cypherpunks and people Internet history have been trying to make work forever and ever. And it's funny because Noster pseudo solved it at least in the community and kind of the, the cypherpunk bitcoiner space almost by happenstance because it was trying to just be able to do social messages. And the example that I bring up is there's a analyst and writer, Lynn Alden, she does some really cool like macro financial stuff or whatever and I've done a lot of narration and stuff for her. And so we worked together a couple different times on projects and one time she wanted me to just do like editing and mastering on like some little thing that she had pulled together. And, and then after I was done, she was like, what do you, you want me to pay you or something? What's the, what's the, you know, what are you charging me here? Or whatever? And I was like, ah, it's not that big a deal. Just, just send me like a hundred thousand sats or something. And she was like, okay, can I just zap you on Noster? And we're like, yeah, yeah, sure. And then it like hit both of us like a couple days later is that I didn't have to send an invoice. I didn't like, there was, there was nothing there. She knew who I was because of my key and she went like, like the, the payment was verified and then sent through that and she could just pay me without any interaction whatsoever. She just pushed money to me. And, and it was also private in the sense that I'm not having like a public thing with a record that she can check, that somebody else can just check and be like, oh, what's happening over here? And, and it just made me realize or made me think of the whole thing more as just kind of like a public directory for, you know, you sign something on GitHub to prove that it's from this developer, right? You could just publish it on Noster 2 and you've done, you've done that same thing. And so anyway, it just got me, it's gotten me really excited about like, kind of like the, the direction that whole thing could go. But you know, these decentralized projects, they always have, you know, there's plenty of post mortems for things like that. So everything has its drawbacks. But we are utilizing it for the project that we have and I think it, I think it's got a massive potential. I'm at least going to try, you know.
- [74:56]
A
Yeah, that's cool. Yeah, See, I'm A massive hypocrite. I, I keep complaining about all these things and I, I don't know anything about any of these solutions.
- [75:06]
B
Yeah, you should check it out. You should check it out. If you guys, if you get some free time, you should just investigate. But you know, if, if that ends up not being the, the direction, it's just, it's just something to entertain.
- [75:20]
A
But I, you know, I really like these, like, I, I really hope this is where the future. I mean, even like bitcoin, like, I generally hate the entire crypto space, but like, you know, God, you, the idea sounds good.
- [75:31]
B
Yeah. Yeah. And like crypto is such a poison on like what bitcoin was actually trying to do. I've been in it long enough that I was there before crypto, right? And holy crypto just turned the whole.
- [75:44]
A
Yeah, it's so sad. It's such a good idea.
- [75:47]
B
Yeah. I still think there's a thread. I still think what bitcoin's trying to do, not to turn it into this, but bitcoin. Audible. But I still think like the thread of what bitcoin was trying to do of like, I think about it just in the. Without getting into some crazy tangent is the fiat world is only the people in power can print money, right? It's the government and central banks print money and everybody else just pays for that cost. Then crypto is. No, everybody can print their own money and we're all going to have a million monies and bitcoin is just like, nobody can print money. Like, the whole point is that we shouldn't print money, you know, so like, and I, I believe that is still working. It's just, you know, it's moving its way into the financialization stage and crypto is still just the gambling casino that it has always been.
- [76:40]
A
So. Yeah, I mean, I'm happy to talk about this too. I don't know if we're, we're going on a tangent here, but. Yeah, I don't know. I mean, that vision sounds great. I've always liked that about it. It's just, it seems, seems to have not made much progress in that problem and has made a lot of progress in a lot of other dumb things.
- [76:59]
B
That's the thing about software, you know, like open source software. Somebody just copies it and then puts their, their onto the token or whatever. I'll say the reason the re. The reason I'm super hopeful and still think like bitcoin is actually on that mission is because the network effects of money, there's nothing that can compete against it. But it is also the slowest network effect. Like if you look at like economics and monetary history is like my, my real nerd out, you know. And if you look at monetary transitions, like it typically lasts a hundred years and especially when you're talking about foundational monetary shifts. And I think Bitcoin is just a unique way to create consensus around who owns what without a central party.
- [77:45]
A
So. But you're saying, you know, things look good, it's just slow. I mean, yeah, it's super slow.
- [77:49]
B
It's super slow.
- [77:50]
A
I mean this is kind of like Linux, right? This is a big conversation now which is, you know, desktop Linux. I don't know how much you know about this, but desktop Linux has always been like this meme of like, yeah, or desktop Linux is finally coming.
- [78:00]
B
It's the era of desktop Linux. Flatpaks app images finally.
- [78:07]
A
But the joke is that, you know, well actually, it's kind of actually happening
- [78:11]
B
now and it's always been, it's sort of always been how much infrastructure runs on Linux, right?
- [78:16]
A
Well, I mean that's always been. But like I have a computer at home now. It runs Linux. I mean for me really the main problem is gaming is a computer at home. The main reason for that other than my laptop is gaming and you know,
- [78:29]
B
some media stuff too. Some media stuff. It's the reason I still have a MacBook.
- [78:33]
A
Oh, it's better on Mac.
- [78:35]
B
Yeah, yeah, like, like editing stuff. A lot of it, a lot of it works on linux now. But DaVinci resolve will still give me problems. I see Linux, but I have a remote desktop open right now to my Linux machine, which is the one doing a lot of my app programming and stuff and my LLM work on my MacBook like so I'm, I'm porting. I always have my, my Linux machine is always a window on my Mac to work with stuff.
- [79:04]
A
So yeah, no, so, but that, I mean it's going a good direction. So you know, and to bring it back to OpenPilot. I think that's the dream for OpenPilot too is it may take 20 more years, but I think there's a good chance of Open Pilot being the genuine, most popular open source, kind of even robotics project. It's just, it's not going to grow overnight. It's not going to get a trillion dollars of cash injection to boost it. It's just going to be a slow kind of thing.
- [79:30]
B
And yeah, I think about it in the sense that like the open source project that gets the most excitement and is Kind of the most robust, slow moving thing is the thing that survives. Yes, right. All the, all the move fast, break, things will blow up very, very quickly, but then they also, they also die.
- [79:51]
A
Yes.
- [79:52]
B
And, and that's, that's kind of my thesis for Bitcoin and has been for quite some time, is that bitcoin will simply be the thing that keeps working when everything else blows itself up.
- [80:03]
A
I mean. Yeah, the problem is, is that centralization is still working too well for many people. And I mean, I think, I guess we are both in agreement, so that's going to blow off at some point. But.
- [80:13]
B
Yes, yes, I think centralization is just the quickest solution to hard problems. Right. So every time you open up a new layer of technology, like every new technology opens up the avenue to create and discover more than the solution actually fixes. Right. Like, like if you, if you can finally go into space, like because you have a rocket ship or whatever, the scope of things to now learn or see or new problems to solve is much greater than the solution that you just presented. So One solution presents 20 new problems in a new layer, right?
- [80:50]
A
Right.
- [80:51]
B
Which means that those 20 new things, in the context of like networks and Internet and all this stuff, the first solution to all those 20 things is centralization. So you, you know, Bitcoin it as just as our example is, you know, it creates a new consensus layer where there is no central bank and you can actually securely know that everybody has who owns or prove the ownership of your money. And it can't be money printed, but it doesn't scale worth shit. So now you have to create payment networks and financialization on top of it. What's the quickest way to do that? Centralization. You create a service that does it holding somebody else's Bitcoin and that's the whole two steps forward, one step back. Two steps forward, one step back. And so, and that's also another reason why I think, you know, like transitions like that. It's like the Internet. It's like the Internet, you know, how many, how long is the Internet's 40 years old and, or at least 30 from the web's perspective. Uh, and how many times has it solved some great problem and then also created some great centralization problem immediately after? Right, yes.
- [82:05]
A
Um, so yeah, the social media one is one that I'd like to see fixed. I don't, I guess, I don't know. I don't know if this is something we want to talk about now, but for me the biggest problem is just that the people who own the servers, who own the content discovery mechanism, who own the content hosting mechanism is all the same. And like, that just seems like a huge flaw. Like, there needs to be some open protocol in some of those layers, and I don't know how to achieve that.
- [82:33]
B
You. You just gave the synopsis of the talk that I just gave at the time Chat Summit, like, literally exactly what it was. And I think the, the. My thinking about the solution and our project, whether or not it actually does that. Does it or not. You know, I'm not going to be arrogant, be like, we're solving this is. Is that the server is just the server, right? Is that, you know.
- [82:58]
A
Right.
- [82:58]
B
You know the content? Because that is the problem. That is the problem is the social media owns the content as soon as. As soon as you give it to them.
- [83:07]
A
Yes.
- [83:07]
B
Like. Like, it was. It was hilarious. Like, you know, if somebody like, that's your way to connect to it. And I think about it as the context of, like, the Internet was designed to connect with one device or one server to another. And then when we created websites, it was like, okay, well, people want to connect to businesses. But then we realized that what we actually wanted to connect to was other people, but we had no protocol for people. So we created centralized websites and then created profiles that were owned by the website. And there was no way to connect directly to that person. So you had to go through that entity and suddenly they own the content, they own the network.
- [83:42]
A
Right.
- [83:43]
B
If you say something to your. Imagine that this conversation right now, if somebody sees that we're saying something that they don't like, they could literally turn off our conversation, you know, like, yeah, that's crazy. It's like we have like a DJ and like a party, and they're listening to all the group conversations, and they can just like, turn down the volume of one person who's saying something, like, move them away from the conversation and boost the guy who's talking about football,
- [84:06]
A
you know, Like, I know that's what's happening. It's crazy.
- [84:09]
B
Yeah, that's exactly what's happening. And so, like, you have to make it so that the content can be delivered with or without the server. And so the server's just helping, right? The server's uptime and that you can prove the content is from that person and that the network and the identity is separate from the thing. So you still carry your profile and who you are around with you, whether you're on Facebook or, or Twitter or whatever it is. And so, yeah, that. That's the way that I think about the problem too is that it's a. And that's also why I'm very hopeful about something like nostr is that that's what I did.
- [84:46]
A
I mean, I guess I think where we probably differ in opinion is that I think, I'm not sure whether the solution here is like technical solutions are always nicer when they work, but I think a regulatory solution here might also work and I'm not sure if you agree with that.
- [85:00]
B
What do you think? What do you think? I generally probably wouldn't, but I'm not. Yeah, I'm a little bit more of a Luddite and in my, you know, older age and parenting years and I have, I love discussing, I want to hear what you think about like give me, give me your regulatory approach. Like why is it that you think regulation is the, at least the short path to a solution?
- [85:25]
A
I think I just, I don't really like as much as I want to be a libertarian and, and like a technology only kind of person. I think historically like regulation has just done a lot of good things and I think like to give a most explicit example of something that should be absolutely illegal is Apple's imessage situation where like they rejoice when like a teenager gets a green text and like thinks negatively about that person because they sent a text from a non iPhone. The fact that they're allowed to do that to a critical messaging infrastructure in a country I think is just absolutely baffling. And I think that's just something that should be illegal. And I trust the US government less than most people, but I think some of those are just clear wins. And the social media one's much more complicated because then you need some regulatory framework of how do you enforce that. There's open protocols between the website, the hosting of the content and the discovery of the content and the owning of the content, that's much, much harder. And like, yeah, you know, I'm more open minded about that. But the, the imessage thing, I'm like how, how the.
- [86:33]
B
Is that legal? Yeah, yeah. And building the technical solution does not mean that like having the solution doesn't mean that it's, it's used you, you know, you have, you have network effects working against you. It's so many different ways.
- [86:49]
A
Yeah.
- [86:50]
B
And you know like nobody cares about a social network if there's no people on it. You know, it's like if you're the only person with a phone, who you gonna call might be the best technology in the world, but there's nobody to pick up the other end.
- [87:04]
A
Yeah. And the people who run the current setups, you know, are very incentivized to prevent that. And like, YouTube provides an extremely good hosting service for free because they know that if they introduce enough friction, because there is real friction in hosting, and if they remove that friction, it just makes it so much, you know, that just makes the gradient flow so much more towards them because everyone's like, oh, do I really want to, like, set up a video hosting server and then put the video there?
- [87:34]
B
You know? Yeah. So like, no, 100%. And then you get in that situation where it's like you're, you have somebody who wants to send you content. Like, you know, like. Yeah, like I'm, I'm following Harold. Right. And Harold posts a video and like, he's posting that so that the people who follows him can watch the video. But then like, the, the platform doesn't let me download it.
- [88:02]
A
Yeah.
- [88:02]
B
You know, like the platform immediately owns it and they're like, this person wants to send this video to this other person, but we're going to stop it right in the middle and we're going to put our ad on it and we're going to take control over it and like, we're going to be the ones that monetize it. And we're not even going to let you take it and go somewhere else with it.
- [88:20]
A
Yeah, but. So I take it you're skeptical of, of, I guess, regulatory solutions to this problem. You're all in on the technical solutions.
- [88:27]
B
I, yeah. So my, my thinking on there's a, There's a fantastic theory, political theory called the bootleggers and the Baptists. And, and it comes from Prohibition days. And the. I do not think that there is. That regulation doesn't solve or cannot do something good, but it almost invariably comes because they have given up two bad things at the same time. It's always a trade. Right. And so the idea of the bootleggers are bad and Baptists is that Prohibition got enforced because the Baptists put up the kind of philosophical moral front ends that was kind of the cheerleading, like alcohol is bad. And the moral argument to make people sympathetic to it. And then the people who actually funded the regulations and the law being put in place were the bootleggers who were going to ma. Like just their profits were going to go through the roof and all of their competition was going to be legally, like, the police were going to go in and shut them down. And so I, I'm a, I'm a pessimist when it comes to regulatory solutions because oftentimes, and I'm not saying this is universal, but you know, the exception usually proves the rule kind of thing is that when regulation does do something good, I'm usually going, I'm usually sitting back and going like I'm gonna go read that bill and see what those two bad things were that are now going to be the new problems that we have to solve. So I'm a bit of a pessimist when it comes to that. And a lot of history seems to point in that direction. But you know, there's a lot of knowing anything about history means that you also 99.9 don't know because history is always bigger than you know, what you can actually discover. But, but I, I, I love half of the show is me trying to dig into alternative perspectives. And I've also become a little bit more conservative, a lot more conservative in my old age and parenting. So I'd still call myself a libertarian and pseudo, pseudo anarchist kind of thing.
- [90:45]
A
Never go too far with the, there's
- [90:47]
B
no easy, there's no, there's no easy solution to anything. That's kind of where I've come, come about.
- [90:54]
A
And everything's a compromise, right?
- [90:56]
B
Everything's, everything's a trade off. It's Thomas Al, Thomas Al, one of my favorite economists. It says there are no solutions, there are only trade offs.
- [91:02]
A
Yeah, no, I think that's right. I mean in this case. Yeah. Even even though I think some regulations might be on net good, I have zero doubt that every one of them would come with some sort of these types of side effects.
- [91:13]
B
The earmarks. The earmarks would be true. Oh God. Well, hell, we're kind of coming up on the end of time. This has been, this has been fascinating. I'm so happy that you came and sat down with me. I've been super intrigued with comma and I will probably be a buyer of your product relatively soon. It will be after the basement gets, gets finished. But I am generally curious just like what are kind of like plans and next steps? You know, you want to see open pilot being like one of those things that would just be the main thing, you know, and like maybe some other company even picks it up and goes further or helps build on top of it and open source advocate always. You know, I'm curious how you, what's the, what's the vision and do you intend to stick with these guys? Are you, do you have other plans to build on other things or join other stuff like where's where's your head on it?
- [92:16]
A
No, I'm pretty passionate about keeping this going. I think the things I'm passionate about, a few things I'm very passionate about is one, our system isn't as capable as it can be. To give you an example, stoplight behavior like stop lights and stop signs, that's like 70% reliable. Whereas like highway driving, you can use it for hours and just have a really great comfortable experience. So I really want to get the city driving all of that to that same level. I think that's, I'm super passionate about that. I feel like that's, that's what I joined 10 years ago thinking it was going to be three year project to do and I'd like to like to wrap that up. But I'm also very passionate like, you know, we run our own data center. We are getting out of the cloud as much as we can. Basically nothing we do internally is in the cloud. And I'm very excited about that. I think it's really cool to prove the point that not only can you make an AI product without billions of dollars, but you can make an AI product relatively cheaply while owning all of the hardware and doing everything on site. I think a lot of people are skeptical that this is even possible, but I mean it's obviously possible because we're doing it now. And so I really want to keep pushing on that. And then the last thing is, you know, Open Pilot is currently ADAS product, but OpenPilot is really intended to be a open source robotics project and cars are just a robot and it's going to expand to other robots at some point. I don't know, don't know when the right time is, but, but when that happens, I'll, I'll be very excited to see how that goes.
- [93:44]
B
That's awesome. Are the, the humanoid robots are coming?
- [93:48]
A
I think there's going to be some things in between.
- [93:52]
B
Somewhere in between. What? No, it's just, it's a light switch. You just get those humanoids running out there and doing work for everyone.
- [93:59]
A
Yeah, that's ambitious.
- [94:01]
B
Dude. What are you, what are you talking about? Three years?
- [94:05]
A
Exactly. Three years turns into 20. That sounds about right.
- [94:08]
B
It's my, it's my three year, 20 year, two decade project. Exactly two weeks. TM. Oh man. Well, dude, thank you for coming on. How do, how do people follow you? You know, where should people link? What should I link to? You know, all that good stuff and obviously comma AI, you know, that sort of thing.
- [94:27]
A
Yeah, just comma, our Website, our Twitter, and I guess my Twitter, that's probably about it.
- [94:34]
B
Awesome.
- [94:35]
A
All right.
- [94:35]
B
This was a pleasure to hit me up.
- [94:37]
A
It's.
- [94:37]
B
It's been an absolute pleasure. Let me know. I may. I'm gonna keep an eye on it and I may hit you up again in the future. DM me if you ever need anything, you got any questions, check out Noster.
- [94:48]
A
We'll try it out. Yeah, yeah, I gotta. I gotta put my money where my mouth is. I gotta actually start trying these stuff instead of just complaining.
- [94:54]
B
There you go. There you go. All right, man.
- [94:57]
A
All right. Thank you. This was such a.
- [94:59]
B
Take it easy, dude. Thank you guys for listening. Don't forget that the easiest way to support this show is honestly to subscribe, to share it out, to use my affiliate link and get yourself a bitbox, which is actually the best thing that you can do for yourself is hold your own keys and save your bitcoin in a smart way. And don't forget to check out Harold. Don't forget to check out comma, AI. If you have a car that's compatible with this and you end up and you get one of these guys and you try it out, please let me know. I want. I would love to actually geek out. You can send me a DM or you can hit me up on Noster or. Or Keat if you find me up there as well. I definitely, definitely want to hear from you. I'm fascinated if anybody else is done with this, because I've been trying to convince myself to just go ahead and bite the bullet and just go do it just to see. But, man, this was a great conversation. This was an awesome conversation. And really happy that he came on the show. Harold, thanks again. Seriously. And that should. That should wrap us up. I'll have links, details, ways to follow him and the company and everything that they're doing right down in the show notes. So don't forget to check out the description and I will catch you on the next episode of Bitcoin Audible. And until then, Everybody, that's our 2 cents.
- [96:44]
A
Sa.