Latest / 4kd.ai / Ep. 10 | The Real Cost of AI

Transcript

Ep. 10 | The Real Cost of AI · 4kd.ai

• David Moses (00:00) Hi there. Welcome to another episode of Forked AI. I'm here with my co-host Jack BeVier and this is a podcast where we talk about a all things AI, how to implement them in your business, how to implement them in your personal life. and yeah, at what at what point you know, Cyberneckos become self aware. We'll we'll hopefully be here to report on that live. So anyway, Yes, I'll jump right in. I I had a I had an interesting conversation with a cybersecurity guy. I had dinner with him last night, and he has started to kind of buy into my theme of you know, he's by by no means all bought in, but he he's starting to warm to the idea that from a cybersecurity standpoint, probably the only way to protect us you know, from an AI driven attack is gonna be an AI driven defense. And so yes. Jack BeVier (01:04) This is a guy, this is a guy in your EO group that you were talking about this. David Moses (01:08) Yes. Yes. this is w we we were talking about it actually with a bunch of guys in the A in the EO group. One of them one of them actually was a former engineer or he he is an engineer, but he's he's a former engineer for General Motors and became you know bought actually a sound engineering company and And now he, you know, that's he does like sound engineering for building you know, new buildings and retrofits and things like that. So but he's super interested in it. And it became like a whole conversation on a boat talking about you know, talking about cybersecurity and and you know, and and this premises so so cybersecurity guy, he I actually wanna we'll have him as a guest. he has started to build out his own Hermes you know, instance and, you know, using open AI and you know, just using open AI codex, you know, OAuth tokens to to, you know, the unlimited plan. and starting to build out various personas of agents and I I absolutely love the way he's building it out. but it was a really super interesting conversation and it kind of made me think about how You know, everybody talks about how these, you know, tech guys, especially coders, programmers, how they're basically cooked. There's no, you know, to use a to to use the phrase the kids use today, they're you know, they're not gonna have any any opportunities for employment, things like that. But I think I think we're really underestimating, you know, tech people and what they'll be able to do. when they embrace the technology and and really start to to implement it on their own. Jack BeVier (03:02) So is he is he building is he basically trying to build his own cybersecurity tools just using best practices and tokens? David Moses (03:11) I mean, at this point, he is implement well, you know, at this point it seems like he's implementing it in his business from more of a thirty thousand foot view. Not not in the actual systems and processes that he operates, but more in the administration and the organization of the business. and obviously that you know, may change next week, but that that seems to be where he's implementing implement imp implementing it. But he has, you know, I mean, he's responsible. It's a stressful job, I think, that these that these folks have. you know, to be responsible for other, you know, for the cybersecurity of other businesses, keeping them safe and making sure that they're, you know you know, make making sure that they're not vulnerable to attacks. especially when they're you know, these businesses are filled with humans that make human mistakes and that are not cybersecurity experts and that really don't know how to keep the business safe. They just You know, they get a link, they click on it. You know, if it looks real Jack BeVier (04:17) Dude, I we started using a cybersecurity consultant because having nothing to do with AI. Like before before I really was like doing AI, well, we were doing it, but I needed it for SOC two compliance, right? So SOC two compliance is this compliance level that that financial services firms in particular, I'm sure many others also, anyone doing business with the government. need to, it's often a it's a a standard that is an industry standard that it's used. Excuse me, an industry standard that's used to basically create a s a report and a certification that, hey, we've evaluated all of the systems, we've we've looked for weaknesses, we've, you know, we've looked at the different tech, we've looked at this, we've looked at the full tech stack, we've looked at the systems and processes, what the checks and balances are within the system. And we have, you know, looked for weaknesses and throughout, you know, and you basically you pay a consultant to come in, they always find something. You're always doing something that is like not best practice. And so you change your systems and processes to comply. And the idea is that, you know, that you won't, you know, you you know, they're looking for weaknesses so that you aren't gonna just dis divulge personal financial information or personal information that could create a liability for you or create, you know, obviously damage to anybody else, as well as the, you know, your business's information, you know, that you couldn't get hacked to like accidentally send a wire to somebody or that somebody could ransom your information from you to result in a business interruption issue. Things in things in that vein. And so this SOC2 compliance or this SOC compliance standard has different levels. There's SOC one, there's SOC two, there's SOC two, type one, type two. A lot of those have to do with like how long you've been monitored and how long you have been compliant. And this consultant you pay these consultants to, you know, tell you what you're doing wrong, teach you how to do it right, and then keep tabs on on everything that you're doing. So they bring we bring in then this company. They're called Cyber Five. They're awesome. I love them. I'm actually a big big fan. I think they're doing a great job. I've learned a ton from them. like I didn't even know this existed. But like, you know, when you borrow a certain amount of money, all of a sudden the people lending you that money start asking about this stuff. And I'm like, we didn't even know about this stuff. So like we have got some breast practices, I'm sure, but like I didn't even realize this was a standard that I should be trying to adhere to. But it's all it's not just like, you know. BS check and boxes. It's like real, you know, like they are identifying actual systemic risks. So like I've really, you know, we've leaned into it, we've really embraced it and and made changes. And so we're working with this company called Cyber Five, and they they come in to do this SOC compliance report because we're doing securitizations in our fix and flip business. And w one of the bond buyers wants to see that we're SOC compliant. And so we get the certification. So that we can go, you know, issue these bonds. And then I'm like, hey, by the way, we're working on some AI stuff also. And like, you know, like they're this is the very beginning of the year. And they're just like, you gotta be, you gotta be kidding me. Like, like you're trying like, cause I'm aggressively pushing like models into people's hands and like being like, play with stuff and like use it in a work context. And they're like, Holy crap, dude, like this is gonna be, this is a lot. You know, this soap. The scope has significantly increased because, you know, when everybody is using AI and has access to system, you know, to like internal systems full of financial information, and is in, you know, and is like, you know, can we can we put a can we put an open claw agent into the mix? Can we put a Hermes agent into the mix? They're like you get you like you just see their blood pressure rising, you know, as as we're having the conversation. David Moses (08:28) For sure. These these these are these are high D like like super risk averse. Jack BeVier (08:34) Yeah. Yeah. And I'm and I'm and I'm not asking, right? Like I'm going to and I'm like, hey, I'm doing this David Moses (08:38) No, no, I'm doing this so. Jack BeVier (08:40) because I perceive this to be an existential threat if I don't. So like, you want along for this ride or not? Like if if you can you can opt out. It's cool. Like I'll go I I'll I need to find somebody. Like, you know, you want to come along for this ride or not? and I mean w I've learned a ton going through the process. so is that the kind of stuff that he's doing on the cybersecurity side or David Moses (09:03) So I I you know, most of what he does is you know it it the company's called the computer guy and he basically is a you know, handles everything soup to nuts for small businesses that are not, you know, like not tech businesses, right? So so dentists, lawyers, accountants, you know, th things like that. they have very, you know, sensitive information that need to you know, needs to be protected. But they're very technologically inept people in general, you know. So, you know, not to whatever, paint a broad brush, but you know, I wouldn't ask my dentist for advice about AI. So it it's you know, that I think is is the challenge for him. It's it's working with for him, it's mostly like, you know, you know, keeping the toddlers from eating the glue kind of thing. Jack BeVier (10:03) Great analogy. So we have this we had this well we do have this IT company that also helps us like they're just like the ticket desk, right? Not they're not just the ticket desk, but they're the ticket desk, they manage our servers, you know, all the Microsoft Exchange stuff is handled by them, all the new all the hardware is managed by them. So like, you know, getting people laptops and making sure that the you know the you know, when they get a laptop, they open up, they can start working. And it's not like you have to download 17 things. and so we've been we were we've worked with them for a very long time. The company's called Intellicomp. they do a great job also. And same kind of thing though. Like whatever it was nine months ago, 10 months ago, I went to them and I'm just like, hey, we're doing all this AI stuff. And like your name keeps coming up as like the bottleneck in permissions and bottleneck in you know, because I want access to my company information and I don't want to be told that I can't get it because, you know, I'm waiting three days for Intellicom to turn a ticket around. And and they're like, dude, you're doing so much out of scope that like we're not, you know, we we so I I've been, you know, I had to like reset the whole reset the whole conversation with them because the way that we were working before, everything was fine, right? They were handling tickets just fine. And then all of a sudden like we're just bombarding them with these like You know, different kinds of tickets that are all fueled by some desire on our side to use something, something AI. And then the SOC 2, you know, Cyber Five get gets bring brought in as like a SOC 2 consultant. and in the meantime, you know, we stood up a GCP, we stood up a Google Cloud Google Google Cloud server, a GCP, and we're, you know, producing apps and pushing them to this to this server. And You know, so we need to like set up cyber security on the on the server. So now we have like Huntress and CrowdStrike, which are cybersecurity products that I'd never heard of before. But they're they're pretty badass. Yeah, David Moses (12:10) Yeah. H Huntress I've heard is absolutely badass. Jack BeVier (12:13) they're badass. I mean they and so now they're like they got Huntress and CrowdStrike monitoring the servers and all of the company devices. We've got we're imp we're putting in Nordlayer, which is like a company VPN. And we're trying to get all of the devices onto Nordlayer right now. That's the process we're doing right now. So that if you're accessing my company information, if if everyone who in my company is is logging in onto a device that has Nordlayer, which is a a a virtual private network. Virtual private network? VPN. Yeah. David Moses (12:45) Yep. EBM. Jack BeVier (12:47) yeah, which is a VPN, then I can limit, I can go to my my company databases and I can say, hey, only accept IP addresses that are coming through NordLair. Again, because like I'm I'm terrif dude, I'm I'm absolutely terrified of this of the, you know, the idea that like you can, you know, using and the open source models are getting like really good, right? And the open source models can't be policed, right? So like as the open source models get better and better, they get dangerouser and dangerouser. And that you can just like send them out into the wilderness looking for looking for vulnerabilities and that you might find my business database out there is a prop like is a real problem. Like like that's gonna be a real problem. Like I I I don't think we've started this and and and I think there's just from my own experience, I didn't think we were that bad. I thought we were pretty good. And the amount of vulnerabilities that we've detected and and closed. over the course of the past nine months has been like very significant, such that I'm convinced we are gonna have waves of cyber attacks coming from probably, you know, open, you know, people using open source models David Moses (14:02) Open source line. Jack BeVier (14:03) crawling around looking for looking for these kind of vulnerabilities. Cause they're freaking everywhere. Like, so I'm trying to get on the front edge of it, but we're, you know, we're using these, we're using these new tools and I'm using AI to try to attack myself and we're using AI to I'm trying, you know, I did this audit. This was actually this week. I did this audit because I was just like, dude, I betcha that like, you know, we've all our our data the data on our servers is kind of locked down, but the permissions are, you know, we got a bunch of employees and the permissions are Are not standardized and people can, you know, open up a folder and put data into a folder, and like, you know, they could put some financial information into a folder that everybody can see. And they didn't realize that they just created a breach opportunity, but they did. They just created a breach opportunity, right? So, or put a set of wiring instructions on a server, right? Like, or edit, or you know, can you edit the wiring instructions? Like everyone listening to this call has who has employees, has an employee. Who has access to their server, who could go into their w the Word doc, which is a which is their wiring instructions because that's what people freaking use and go edit that freaking number. Everybody listening to this that could absolutely happen to them. And the next time they email their title company the wiring instructions, dude, they wouldn't even know that the had gotten edited. You know, like and they'd c and they and they and the title company would call the girl in their office and the girl in their office would go open up that word doc and be like, Yep, that's the number. Looks good, send it. And that had gotten edited by by somebody else, you know. Like I I just think there's gonna be a ton of that you know, it's not just the email, it's you know, it's not just the the the email accounts sending spoofed emails around, which has been like the historical wire instructions attack. so anyway, I was like, I was like, I need to do a permissions audit. Like, who can see which folders on my shared drives? And Dude, it's embarrassing. Like it's am I just like, dude, there's there's so many holes in this thing, like in just that idea, you know? And then I had another one. I was like, and you know what? We used a bunch of group emails. Like group emails have we've gotten lazy about managing group emails. Like, you know, people are just like, hey, can you create a group email for this? Create me a group email for that. Just, you know, keep everybody in the loop. It's kind of a CYA just to put everybody on a group email, right? Like, which also is crap from an efficiency perspective. It creates bloat. but David Moses (16:34) Yeah. Jack BeVier (16:37) But it also like I'm exposing information to like people that don't need to. And if that person becomes a, you know, a a a vector of attack or whatever, like what can they see? You know? And dude, that was another embarrassing audit where I'm just like, good lord, I need to close like I need to tighten this down. But it's you know, the company had gotten big enough that like everyone knew, like every like, you know, I could have asked half a dozen people at the company and all of them would have been like, yeah, we should do that. David Moses (16:59) I Jack BeVier (17:06) And then everyone moves on and no one ever does it because it's a freaking task, man. Like, once you get like a real company going and like a bunch of people, like the amount of bloat and stuff to like to David Moses (17:17) don't think that I can't everyone. Yeah. Jack BeVier (17:19) audit all that crap is a really crappy audit that would take me down for a week. But AI, you know, but but Fable made quick work of it. And you know, I'm redesigning my security groups and deleting half of my group emails and revising another half of them and just that ch just those chores, you know, just like this those kind of chores which you just hadn't gotten around to. Now, you know, you can just, you know, set it up on the side and, you know, go run. You know, go do that while I answer emails and, you know, check in to see what what you know, see what's going. But it's a task that I don't want to assign to a human. I don't really trust a a you know I don't really trust very many of my humans to actually do it correctly. And the ones that I do trust, I don't want to take them down. I don't want to take offline to do that kind of So it just doesn't get done. And I'm like, I'm not. David Moses (18:09) Yeah. It doesn't cut it that that's the biggest problem, is that cybersecurity doesn't become a priority until you get hacked. Because it's it's yeah. Yep. Jack BeVier (18:14) Yeah, yeah. And then everyone freaks out and yeah. Yeah, so anyway, that's my that's my recent cybersecurity experience. David Moses (18:24) Yeah i it's it's a real I mean look, it's it it's a it is a very real threat. you know, that that yeah, these open weight model. I I don't think With open weight models. I think that the scariest part is that the open weight models that can do the most terrible things. I mean, not just hack our companies and and take us down and you know ransom our data back to us. I mean, but like create viruses, like like actual real life viruses. not not just computer viruses. the these models already exist. The the but the But there's this there's gonna be this delay where these models that can be downloaded, modified, you know, distilled, post trained, and then unleashed to do terrible things, the hardware is inaccessible to run them by anybody but large, you know, large companies like our like I could see somebody I could see a an evildoer kind of using a large company's resources to, you know, to maybe maybe be able to do this. But like the the computers with the RAM to run Kimik whatever or you know, they just don't they're not accessible to people. You can't you can't put together two terabytes worth of worth of RAM for less than fifty thousand bucks, right? And probably Jack BeVier (20:05) Yeah, how much what do you what do you need to run a model like Kimik two or David Moses (20:09) I I yeah I I would say you probably need a a terabyte of RAM. So you need, you know, a and the largest, I think the largest Mac Studio, because you'd probably need to string together Mac Studios to realistically do it because they have the unified memory and you can use a Firewire, you know, a Thunderbolt connection to share the RAM. And then you'd have to get like four two fifty six, you know, gigabyte, you know, RAM. Jack BeVier (20:37) Sparks. David Moses (20:37) machines, which which those things run what, fifteen grand? No, I think Jack BeVier (20:40) Five grand. Yeah. David Moses (20:42) they're more than five grand, but they they're a lot of money. They're I mean my s my ninety six gigabyte, which I could sell for more than I paid for it, I paid four grand for it. And, you know, and I'd need almost I'd need 10 of those to to run, you know, so so the hardware is is what's prohibitive. And my concern is that the delay between the models being this powerful and the hardware being able to run them is gonna make people complacent. They're like, that model can't do that, or that model's not gonna do that, or that's not a threat. And it isn't because the model can't do it. It's because the bad actors can't get the hardware. They can't access the hardware to harness that power. And when they do, they're gonna surprise the out of everybody. When they can get that, you know, everyone's gonna say everyone's going to have already, you know, had this, you know, lulled into this you know, kind of safe slumber that everything's okay and and then, you know, someone's gonna start delivering terabyte memory machines and and it's gonna, you know like I I don't worry like I don't worry about China. I don't worry about Russia. I don't worry about like I don't worry about North Korea in terms of like AI attacks 'cause these are like these are state actors that are not really like, yeah, they they they may want what they want, but they don't want the end of humanity. Right. They don't want to create a virus that's going to wipe out everybody on the planet. So Jack BeVier (22:10) They'd love to they'd love to scare the piss out of a bunch of American businesses though, stealing their money. David Moses (22:13) They they they definitely do. They for sure. I mean there there's a lot of things that they w you know, not I'm not like saying, these are fine. Like I'm not I'm not I'm not like saying, don't worry about North Korea. Like what are the you know, they're harmless. No. But but I I worry more about this kind of power in the hands of like one person, like one person who, you know, like the analogy that I draw is like, you know, there weren't mass school shootings in the eighteen hundreds. Right, because like if you went into a school with a musket, like you're not gonna kill very many people with a musket before they're just gonna be like they're just gonna run away. Right. The the the the the tools of mass destruction are a new thing. And and that'll be the same thing with AI is like one once one single bad actor who just, you know, you know, armed with the power to do terrible things, at a at a at a at a much larger scale. Like that stuff's that stuff scares me. And my my fear again is that is that the the models are already there. The models that exist today are probably capable of doing horrible things globally, but the hardware hasn't caught up and is not accessible to the to the bad actors yet, to the to the individuals. So that's rosy Jack BeVier (23:37) So I hear you. Like t so just to to to get real nerdy with it, so you could I was just Googling or I was, you know, just Gemini and David Moses (23:44) Mm-hmm. Yep. Jack BeVier (23:47) you could connect two sparks, which you I can I can go to computer store computer share up in Parkville and go buy two sparks off the shelf for five grand a pop. Link together and that gives you two fifty six. Two hundred and fifty six megabytes of RAM. so ten David Moses (24:00) Yep. Get away. Jack BeVier (24:02) grand gets you two hundred and fifty six megabytes of RAM. David Moses (24:05) Gig gigabyte. Jack BeVier (24:05) And you can run Lama three point one, which was Frontier in the second half of twenty twenty four. So you can do twenty second half of twenty twenty four AI locally on ten grand of hardware. David Moses (24:18) Mm-hmm. Jack BeVier (24:20) And your point is that yeah, and if we fa if we if we if hardware like you know, NVIDIA set next sets of in you know and hardw hardware releases, if that gets to second half of 2025 frontier capability of running locally, we have a problem. Like, you know, it's it it's very possible we'll have a problem. And that's literally like I could just go like to computer with, right? Like I'm sure like there are much better versions of that. That's just my layman you know, the layman thing that I just, you know, learned in in the past year. David Moses (24:52) But if you're but but you're talking about putting together a half a terabyte right of of RAM, right? Two sparks together is yeah, two fifty-six each is a half a terabyte, ten grand, right? And that's not like you're you are not even halfway, let's call it a third of the way, from really may maybe halfway, because you could distill the model, you can, you know. You're you're halfway from being able to run a Kimi K, you know, to I don't remember what the what the what the most recent one is, right? Your your or Jack BeVier (25:28) Yeah, yeah, yeah. David Moses (25:28) or GLM whatever, right? Like name pick your own Chinese model. Like Deep Seek's gonna come out, you know, V4 I think is gonna come out. If it hasn't if it hasn't yet, it's gonna come out. It's imminent. And that's supposed to kill all the all the rest of them. And that's supposed to be open source. But like the the hardware you need to run these things, I mean that's 10 grand, multiply that by two, that's 20 thousand bucks. There just aren't a whole lot of people depraved enough to and and interested enough in this and have twenty grand laying around. Like like like y you like a terrorist organization Jack BeVier (26:03) Yeah. David Moses (26:04) you would is probably what you would need. Jack BeVier (26:07) Yeah, yeah. Some some kind of organized crime. Yeah. David Moses (26:09) Yep. Yeah. It would it would have and and again, you know, i I I would still like once you get enough people together, you know, the the destroy all of humanity be becomes something that's it's it's a very difficult secret to keep. I I just the more people are involved in a thing, the less I'm scared of it. I I really like the thing that n again, not that it's not possible. I'm certainly not poo-pooing it. I'm just saying that it's that one person, that single person. You know, in a shed in northern Michigan, right? Ted Kaczynski. Like it's that one guy. The one guy with the power to do it and the inclination to do it. That that to me is the s the scariest of scary. Jack BeVier (26:51) So the there was this, we we got hit. we got hit. We got hit with a cyber attack. and it didn't take us down because we had the stuff up, so we caught it. Yeah, so we caught it. Sorry. We we it was attempted. It was attempted but failed attack. And the way that it got, but it and it was the it was token stealing is what it was. and they were Injecting a there was a piece of code that was going into I'm gonna screw this up and and also who gives a There's a piece of code that was injected into we were using lovable to create a bunch of stuff, right? So every time David Moses (27:33) okay. Jack BeVier (27:34) we downloaded lovable package, it was in there. They had they'd got it sorry, because it was one of the standard dependencies, right? It was one of the standard dependencies, they they'd gotten malicious code into one of the standard dependency packages, and it hadn't been caught. Yeah. I think they eventually caught it and now it's got scrubbed and it got scrubbed out of those dependency packages. But there was a period of time, there was like a ninety day period of time where all the dependency packages had this malicious piece of code in there that was basically o creating was was creating a crypto wallet and trying to Creating a crypto wallet and would and it would just sit there dormant on your machine. and eventually it was trying to steal API keys. That's what it was. It was tr trying to steal API keys so that it could, you know, just you know, have your API key and use tokens on your account. and you'd you know, you'd never notice that there was just this rando API key that had gotten created on your account, and you know, you would just be paying for this thing's tokens. But it was like in it was we found that it was pervasive. It was across everybody. Like anybody who had downloaded a lovable package, it got in there. And then anytime they pushed anything to our server, any time they pushed any code to our server, it was there. And we had to scrub the whole, you know, we had to scrub the whole server of it. Just through this malicious code into into packages, which anyone who is any of the IT guys who are listening are probably like, yeah, no dude. That's like tale as old as time. But You know, but it was n my point being that, but now that we have employees who are like not trained software engineers creating software because everyone's vibe coding and thinks it's fun, they're downloading packages left, right, and center. And you have no infrastructure to monitor the packages that they are downloading, or you have very little, you know, infrastructure to monitor the download the packages that they are downloading. And so that becomes the vehicle onto your server. Which where the rest where if you haven't gatekeeped you know set up any any kind of gatekeeping is where the rest of your business information lives, including wiring instructions and like that. And so I just I think it'll be, you know, that that vector of attack is one that I had not previously appreciated because I didn't have coders on staff before and I didn't hire any coders, but I gave a bunch of people API tokens and I said, hey, go code. So I created them, right? I put them on staff and I just said David Moses (29:58) And now you're coders. Jack BeVier (30:01) go have fun. And so I think there's a lot of companies that have people who are out there who are just like, hey, yeah, go have fun because they're trying to create, you know, homegrown, you know, business ops vibe coders, but don't even realize that they've like create they've opened the, you know, they've opened a side door to their servers and, you know, have not yet gotten to the point that, you know, we were, I was probably a year, I was a full year of doing this. Like, Come on, guys, come on in, come on in. And like, meanwhile, the side door is wide the open. I got lucky because I didn't get hit before I learned about the how the side door worked. But I just think it's a learning curve that a lot of companies that you know, every company who is encouraging their people to use AI, which is I people think the people who are going to win, also at the same time are teaching themselves how to be CTOs and are not even aware of the, you know, the c of you know of some of the of some of these concepts and are have have created a side door vulnerability for their entire organization that if the bad guys catch up and figure out, you know, who that is and how to go after you know, I don't think we've heard a lot of those cyber attacks yet. But it would re be you know, but I just, you know, expect that that'll that will happen at some point because the opportunity exists. David Moses (31:29) For sure. But that that's you know, I I think like you mentioned lovable and The vulnerability A lot of the vulnerabilities that that exist exist through third party software that you trust that you use. Right. I think as people create their own software, you become less of a target because you you only deal with your own stuff. Right. Right. Jack BeVier (31:57) Yeah, very very small Victor. Yeah. Not not worth the squeeze to attack you. Yeah. David Moses (32:01) No, no. And and and y y and and yeah, they're casting a wide net and trying to attack as many people as they can at the same time. But that's my argument why AI has to defeat AI. Right? Like i if you have agents and you're implementing agents into your business environment, then there should be like the qu the things that you learn along the way about a security audit, right? These are things that the AI understands to its core. You Fable made like you said, Fable made short work of of that security audit. So an agent performing that specific security audit in the specific way that Fable says, you don't need Fable to run it. Fable needs to design the the security audit, and then you need to put it on an agent to run that on everyone every day. And then eventually, it's not every day, it's like every every email, every Like w once it gets strong enough where you can have a local model or you can have, you know, a cheap, you know, I mean, even even if you were to use Grok four point six, which I think is two dollars per million input and six dollars per million output, and you had that be like a f on a operating a fable designed security audit, you could monitor and perform that audit. on an hourly basis for what would not be a cost prohibitive, you know, and and and then that that is I think a a recipe for safety far beyond trusting some outside company to come in once a whatever and do an audit for you. I just I you know that's that's my you know my my personal feeling. Like I don't I don't know how good my You know, IT company is my outsourced IT company. I don't know how good they are. I don't think they're Fable good. Jack BeVier (34:05) Yeah. I I comp I completely agree with you. I do think that the end state is a much safer world. I I I do believe that. I I just think it's gonna be bumpy along the way. But I do agree. I I I think I I I think you're absolutely right about that. I don't think that the vulnerabilities cause you still have to t to to steal money, you still have to interact with the physical world, like in s in some way. David Moses (34:30) Currently, yeah. Yeah. Jack BeVier (34:32) I mean, you know, even with MF you know, with if if you have MFA, you know, like MFA forces an interaction with the physical world and that becomes like a that becomes like the gate, you know, the thing it can't do. David Moses (34:45) Authenticator apps do I would I would argue that any two factor authentication that invol that is that uses text or email is not two factor authentication. It's just it you know, it's it's the TSA. It's ta it's ta it's take your shoes off. That's all it is. Jack BeVier (34:59) That's fair point. It's a it's just a mode switch. Yeah, yeah, that's a good point. Yeah. I I agree with I agree with your point there. Hey, I'm gonna shift gears. You anything else you wanna talk about that on that? David Moses (35:15) No, go let's let's shift. Jack BeVier (35:17) Okay, one that one that's one that I'm excited about and one that I'm not excited about. I'm gonna do the not excited about first. Holy Fable's expensive, man. We freaking we're letting people just use it. And my god for forty five day period, I ran up a freaking bill, dude. Like unmonitored Fable tokens are wow expensive. or like you know, you will you will get surprised. David Moses (35:42) When when you say expensive, how expensive? What was your bill? Jack BeVier (35:46) like we doubled our spend in a month. We we doubled our monthly spend in in in a one month period. No ch no no, I mean we got us we got a significant increase in productivity, but I don't we we've we you know we we got sorry we got an increase in productivity. Was it worth the money? well not in the way that we were doing it because we were being stupid and we were everyone was using Fable, whether they were asking it a chat or whether they were doing some basic, you know, basic thing. Because I gotta, you know, cause I'm encouraging everybody to be on it. And I hadn't educated everybody on like, you know, how David Moses (36:25) Yeah. Jack BeVier (36:25) expensive it was. Yeah. And and, you know, it was just a, hey, the new fable's out. Check it out. And so everyone did. And they were like, wow, this is really cool. But they were doing like, you know, toy with expensive freaking tokens. Like the what is it? Like you just mentioned the the the Grok four point six is like What is it? Two out, six ins, two, two, two in, six out. And I think it's like David Moses (36:45) Yeah, two out yeah, yeah. Two out, six two two two two in, six out. Jack BeVier (36:49) I think it's like thirty and fifty for fable per million. Thirty dollars and fifty dollars per million for fable. Like it is you know very, very much more expensive. So we many X more. David Moses (37:02) Many, many, many and and it's not you know, yeah. I I and and I I th it's just very interesting that Anthropic gets so much of their compute from XAI. Like they they they spend a billion a month on Colossus one and they're gonna probably add with Colossus Two. But Alan gets that for free. Like, I mean not for free, but like he doesn't he's not paying on top of it. And he's making Jack BeVier (37:29) Electricity. He gets it for electricity. Yeah. Yeah. David Moses (37:32) Yeah, for sure. I mean he he he's actually that that's a net gain and all of his usage is I'm sure still come you know, so so he can charge two he he he could just come out and charge, you know, twenty cents and sixty cents if he wanted. You know, so Jack BeVier (37:48) Yeah. David Moses (37:49) that's where I think so I'm I'm starting to warm to the idea that Grok will win this race, the race to superintelligence. Jack BeVier (37:58) Because he because he's vertically integrated. David Moses (38:00) Because he's completely vertic vertically integrated. he's gonna be, you know, building his own ships. He's gonna have his own he's gonna be putting them in orbital data centers. He doesn't need terrestrial data centers in in five years, ten years. And and and it's not like I I'm gonna go buy a bunch of X AI stock and I'm not suggesting anybody else, you know, do that. but in understanding what the models are capable of, what they're not capable of, Grok isn't there yet. I would call them, you know, they're like edge, you know, they're they're like still Not quite to the frontier. Jack BeVier (38:33) Have you played with Grockbot? David Moses (38:35) No, I've not. but I've heard Jack BeVier (38:38) I haven't yet either. I to. It's on my to-do list. David Moses (38:40) Yeah, it's not much. I I also I actually also was kinda interested in downloading OpenClaw too because apparently they made some significant Jack BeVier (38:47) yeah, yeah. They that just came out too, yeah. David Moses (38:49) im improvements. you know Jack BeVier (38:52) Yeah, so for those are not familiar, so Grokbot is basically Grok's, you know, Elon Musk's answer to Hermes and OpenClaw. It's it's cloud based, right? So it doesn't have to be on a physical piece of hardware. So it's cloud based, and it and a nice thing about that is that it runs twenty four seven. So your computer need not be on for it to be running. which I guess is nice. I haven't played with it yet, so I'm I'm interested to see it. Like, you know, the the battle for agents, like which agent is going to be the one that like ends up in your enterprise? You've you worked a lot with Hermes, so obviously that's very interesting to me. I'm wondering I was wondering if Grac was Grockbot was gonna be a com was gonna be a is going to be a legit competitor to Hermes, who seems to be kind of the name at the moment, at least the one that I hear the most. and then The other thing that you alluded to, David, was that that OpenClaw 2 just came out, which was purchased, well not purchased, it's an open source project. But the founder of OpenClaw went to work for OpenAI. David Moses (40:02) Yes. But the guy the guy was purchased. Yep. Jack BeVier (40:08) so the project got brought in-house at OpenAI. but I haven't played with that either. And that should be interesting to check out also. you know, you know, obviously that would they were the OG of of agents. So be interesting to see what their next gen looks like. David Moses (40:26) Yeah. I mean and I th you know, I look at I just my point with f like going back to your Fable Five comment, yeah, f they might be the frontier in terms of intelligence. These are the smartest people in the room, but the question will become at what point is the intelligence smart enough for everything that we want to do in our businesses and Jack BeVier (40:46) Yeah. Yeah. Exactly. David Moses (40:49) and we're just not willing to pay thirty to f you know and and anthropic may not really be in a position to charge less. Jack BeVier (40:59) Yes. Yeah, I agree with that. So the the you know the if you look at the graph of like the the revenue, anthropic revenue, I look like it. Like my bill looks like my bill looks like that curve, basically. David Moses (41:13) Yep. Jack BeVier (41:14) Except this month I was like, holy guys, that's not necessary. Yo, Fable Five for brand new projects in design mode, code review. and some like specialty use cases where we want to like grab a very large data set and do an do an analysis that we've never done before because it was never you know that that wasn't wasn't possible before and let's go see if Fable can do can grab some and you know some do an analysis of do a do a very large scale pattern recognition project for example. And that's it. David Moses (41:47) I I wouldn't I wouldn't even use it for that. I I think pattern recognition, I think any kind of pattern recognition in a large data set is child's play for any of the other frontier models too. I think where Fable comes into play are the are the problems that can't be solved. The it the you know, the it's broken, I can't figure out how to fix it, Fable will figure out how to fix it. You know, when when I if I've got like, hey, I'm gonna redo my lease. I'm gonna I'm gonna I wanna review this this these lending documents, I wanna review you whatever it is, right? Like that's Fable territory because it's something that first of all, the stakes are high. You know, you have when you're dealing with legal tax, you know, the stakes are high. So looking at having Fable perform those But but it's just final run kind of thing, right? It's it's not have Fable do all the work. It's have the work done. Have the agent on 5.6 Soul or Grok 4.6 or even Sonnet or Opus have it do the work. And then have it justify why it did what it did. You know, what explain what it did, why it did what it did, and then Give that to Fable and let Fable be the one to actually give you the comfort to say, yes, this is, this is, this is right. Or, hey, this isn't working. Okay, cheaper, faster model. Give me a very detailed explanation in an MD file of what's not working and what your evidence is that it's not working. And then deliver that to Fable and let Fable figure it out. Jack BeVier (43:38) Let me press you on that because I have David Moses (43:40) Okay. Jack BeVier (43:40) a use case which is the the the the thing I'm excited about that I and I want to challenge you on that. So something I did this week was out of necessity, I was like there was a VA that was helping me with or VA, there was a person who just happened to live abroad who was helping me with a project and it's basically like an insurance claims process. And They they had to leave. They they quit. and I had to take it over. And my question in my mind was like, do I backfill for that position or can I have AI do this? So everything's just sitting, it's a nice little it's a nice little use case because it's isolated, everything's sitting in one folder, and all the documents around this process are sitting in a folder. And so I said, I I had had Fable open and I said, hey. Go check out my share file. Here's the folder that we work out of. these are the and these are the and so it, and I said, go learn about that process and let's discuss. And I thought that the next thing that I'd be doing was explaining the SOP as to how to run this particular business process. And it came back and said, Hey, I see what's been going on here. It because it looked at all the files, it read the files, it understood the context of the problem. It looked at the edit dates. It looked at who had edited those those files, who had produced which documents, and it and it and then it looked at the emails of those people, because I have my desktop, my claw desktop hooked to my Microsoft 365 exchange on a read-only basis so that it can go read emails. And it can not only read my emails, but it can read everybody else's emails. and between the emails of those editors and the content of That folder, it said, I think I understand the SOP. It works like this, right? And then we create this and then we do that over there. Anything to add? And I read the damn thing and I was like, I have nothing to add. Like, that is right. That is the SOP. You have like you got it. And Opus didn't do that. I don't, I don't think Opus did that And so David Moses (45:48) That was perfect. No, I yeah, I I I agree with you. I'm I'm not Jack BeVier (45:57) So so I'm like, that's really interesting use case because the Fable tokens in this case are a proxy for my time, right? Like I don't want to, as a business owner, I'm gonna have to go spend my time documenting an SOP or teaching an agent how to do this thing. And Fable didn't need to be taught. I mean, I tweaked it on the edges as we actually went through the process. I made some like tweaks on the edges, but it got like 95% there in the first shot. And I was like, damn. Like as a proxy for my time having to go document SOPs, which is every manager and owner's least favorite thing to do in life, I think think that's worth the money. And I don't think David Moses (46:28) Yep. I can't Jack BeVier (46:37) that previous models previous models didn't didn't show me that. I I never saw that before. David Moses (46:44) If you took okay. So This is where I'll push back a little bit because I would say What I've been doing a lot with my Hermes agents is asking either an Opus orchestrator or sometimes just or I guess more often, a GPT five point six soul orchestrator to spawn several agents, many agents in parallel, and make a mixture of experts. So make this one an expert. I'm and I don't give it the parameters. I'm just like, you come up with the plan for Jack BeVier (47:28) Use this protocol. Yeah. David Moses (47:29) Yep. You you come up with the plan for what experts you need and what models they run. And I want you to then go at this task. I I have found that it is It is Fable level. Like I'm not getting a better result from Fable. But I am getting from what what I get from Fable is a much, much, much more polished, you know, you know, kind of presentation ready. Like it's a little bit more clunky on the other side of things. but Method A was free relative to my monthly cost. And Right Fab Jack BeVier (48:17) Yeah. Yeah. On an incremental cost basis, yeah. David Moses (48:21) Fable was wildly expensive. but if I take that, all of that, right? I now I need all of the research that you did that was meaningful. I need you to take all of that data and I need you to justify each agent justify what they contributed that was actually used in the final product and give me a full MD file. long as hell that explains everything that everyone came up with. And then I feed that to Fable and I think where to get it's better. Jack BeVier (48:52) Mm-hmm. David Moses (48:52) Right. And and the cost is wildly less expensive because it didn't have to do all Jack BeVier (48:55) Yeah, yeah, yeah. David Moses (48:56) those tool calls, didn't have to do all those. so I like I you know, and it's like, okay, where it got it wrong, you know, because like if it i i i if fable's getting it to nine it's getting it ninety five percent accurate, but a mixture of experts approach gets you 85% accurate. Right? That 10% plus polishing it and bringing it all together, that I think is a better use, you know, for Fable. But I'm also like for me, it's different from you because I'm using my agents within one ecosystem within within Hermes. If I want to go to Fable, I have to go to Cloud Desktop or Code or whatever. And, you know, so it is it there there is also a you know seamlessness to you know Jack BeVier (49:48) How you working. Yeah. David Moses (49:50) yeah so so I I would say you're you are right. Fable I mean there's no doubt that Fable 5.1 is the you know by far the strongest, you know, the strongest model. And it can do things that our other other models can't, but but in things like an example a a kind of a similar example is I had I had my agent, I I needed it to go through an entity for twenty twenty five and I needed it to just tear through every single expenditure that we made for that was over twenty five hundred bucks. That's the safe harbor for the IRS for you know for capitalizing, you know. Jack BeVier (50:27) Mm-hmm. Capac, yeah. David Moses (50:29) Yep. So every everything we you know spent more than twenty five hundred bucks on. And I I need you to categorize it, figure out what property it is, twenty seven and a half year, fifteen year, whatever. I I need you to figure this all out. Right. Jack BeVier (50:46) That's a nice use case. That's a really freaking good use case. David Moses (50:49) Yep. So and it ri literally like red and and anytime you can justify this as a like because this is where I think the magic is. If you can justify it as an expense, that's the holy grail, right? And I had but I want you to run two agents in parallel, one on GRAC four point six, one on soul five point six. You know, GPT 5.6 sold. And I want you to make me a spreadsheet. And I want you to give me what category that the agent put it in and the justification for it. And GPT six, GPT 5.6 won. Not by a lot, but it won. Why I'm not surprising. Like nobody's shocked listening to that. But if I look at it and I say, okay, now let me take that and let me take the justifications for why. It decided what it decided. And let's just throw that back into the recipe. Right. So that when this thing gets done on a monthly basis now and it goes through all these things for all my clients too. Right. Now Grox's going to get it right as well. Grac's not going to make that mistake because now it's part of the instructions. Yep. Jack BeVier (51:58) Yeah, yeah, yeah. 'Cause you gave it the context, yeah, right. Yeah, you you had it you're right you're having it write the SOP, basically, like yeah, by by derivative. David Moses (52:07) Yeah. And and and and yeah, and and the magic is the justification because like you have you have all these like random notes that that tech guys I I don't mean tech guys like you know like IT guys, I mean like techs out in the field fixing or replacing You know, when they when when we can go through like say somebody's replacing a furnace. Okay, great, they're replacing the furnace. All right. The removing the old furnace, you know, that's a that that's property that I can now eliminate, right? Jack BeVier (52:47) Abandon. Yeah. David Moses (52:48) And then I can take components of replacing that furnace, parts of it, if it were broken down a certain way, I can take parts of it and expense them away. Right? Jack BeVier (52:57) Mm-hmm. David Moses (52:57) Like, you know, hey, this circuit needed to be fixed, you know, or updated to meet the new code or this you know, whatever it is, right? Like Jack BeVier (53:06) Right. David Moses (53:06) And you go in and and what what GPT was able to do was to isolate things where they explained in detail what they did and pull Jack BeVier (53:19) Yeah. David Moses (53:20) out little components of it that probably could be expensive. And then that becomes part of the SOP, right? But by running more than one model in parallel, like yeah, it'd be nice to have Fable be in that in that model, but it's cost prohibitive. But Five point six. But it but by doing that, you kind of the the mistakes that the agents will make come to the surface. And then you you Jack BeVier (53:44) Mm-hmm. David Moses (53:45) so I'm really focused on the on the on the the ones they didn't agree on. Jack BeVier (53:50) Yeah. no, I I love that. And I and I think that that methodology makes a ton of sense of like have the higher end model create like do the building, do the creating of an SOP, do the the the the original work, and then once you've and then the work product being context that then a lesser, much cheaper model can do. Like that that idea is like I I I keep coming back to that idea in terms of like Also just zooming back out to like, you know, hey, what's going on with like these with like the society's token spend, for example, right? Like I'm still like stuck on this idea that I'm like, this is a this is a capex cycle. We are building a bunch of we are rebuilding a bun all the business processes and practices and but then to run it, we're gonna you know, we're not we're gonna run the thing with cheap models. We're not gonna Run the thing with these expensive models. I'm gonna build it with these expensive models. And then nothing's gonna be new under the sun, or you know, not much is gonna be new under the sun. And I don't need Fable. Yeah. Yeah. And David Moses (54:55) No. And ideally you can run them on local models for free. Jack BeVier (54:59) then I'm gonna run them on local models for free. Yeah. After, yeah. So like there will be a percentage, right? Like, but like, you know, of my total token usage, I think that you know, my CapEx total, it'll be the same as what business is, right? Like whatever CapEx is as a percentage of the economy. That's what fable token, you know, the that the fable token usage will be the same percentage of the total economy's token usage. Now it'll be like a very high percentage of the revenue. but from a you know token, you know, numerator fable tokens versus denominator total token usage, it'll, you know, the frontier models are just the are capex tool. They are not the RM tool. They're they make no sense for RM. David Moses (55:46) So so have you I have have you just basically illustrated a bubble? Jack BeVier (55:52) well the the question is like what the question I think is still is penetration into the total addressable market is still I think you then get into that argument as to like how far are we in terms of penetr? David Moses (56:03) Yeah, one percent, five percent. It's a very small percentage. Jack BeVier (56:07) Yeah. So like is you know is the 75 billion a month of re anthropic revenue gonna go to four hundred? I don't have an opinion about that because like penetr because that's a a penetration and TAM question, but I do think that. for the entities that are using those tokens currently, like I just tempered, like I'm I'm my token usage as a company is re is, I think, reaching an asymptote. Like it is is is leveling out. because I'm figuring out how because you know the things that I built before, I'm now switching to cheaper models to run them and this and and I'm only ever doing so much CapEx at a time, right? Like and I'm figuring out efficient ways to do it. David Moses (56:47) Yeah, you only have so many humans to actually Yeah, you you gotta drive that you know, some someone's gotta drive the bus and you know, you can only have so many Jack BeVier (56:54) Yeah. David Moses (56:55) people doing that. Jack BeVier (56:56) Yeah. Yeah, exactly. So I think that eventually frontier model, I do think that eventually frontier model revenue does asymptote. Does it asymptote at you know at what the current numbers at, you know, at a hundred billion a month or does it asymptote at a trillion a month? I don't know the I that that I think is a is a very worthwhile lively debate. Like and then also this is a global David Moses (57:17) Yeah, I think I th I and I th but I think it's a Jack BeVier (57:19) thing, right? We haven't even gotten to that, to the to the global aspect of it. David Moses (57:25) But idea is still that at at some point in the future, this will be normal. This will be, you know, not we we won't be the exception anymore. and at some point, it's really the it's it boils down to not the cost of A token, but the cost of an electron. Jack BeVier (57:50) Yeah. David Moses (57:52) you know other than for the very like solving humanity's greatest problems like disease and you know that kind of thing like space exploration and and things like that. And those are those are, you know, I'm you're not gonna say they're unworthy. They you that they're great, but you know, SpaceX still pays the the same amount for, you know, their email hosting that everybody else pays. So it's you know like it it will become commoditized. Jack BeVier (58:22) The the there's this there's this disincentive, I think, in the system that I haven't heard people talking about. Where if Fable Fi like Fable Five is very impressive, right? Like it in from a Capex perspective, it's like I'd love to be on there the whole time. And every time but the problem is that every time a frontier model is created, it is distilled. And, you know, 80% of its equivalent gets made free, essentially, right? I'm just Painting with a very broad brush here, but like, you know, the latest version of Kimmy is 80% of Fable, but two percent, you know, five percent of the cost. and so there is a disincentive, right? If Fable can if Fable can rebuild businesses, right? Like if if if Fable in in and the adjustable market is the enterprise market, right? Like, and the so the most valuable thing other than like you said, like frontier scientific research. If like the money's all coming from recreating business processes and Fable can rebuild business processes very, very well, you really don't want to release the next model because the next model gets dis this distilled version of the next model cannibalizes like then you you you you all of a sudden everyone's just gonna use the the open source version of distilled Fable six. And to to rebuild their businesses and only the scientists and people working on like what is the you know meaning of life are using your f or your fable six to answer those questions. So like I think that they may just like be like, yo, we're good, like because like because to maximize the area under the curve, they should just hang here where it's it where like you still have businesses willing to use fable five at very expensive rates. And pay money because it does it solves like the most in the biggest problem for for my dumb ass business, you know, my my boring ass, you know, arithmetic, my my arithmetic and algebra, right? Like it does David Moses (1:00:24) Is child's flavor. Yep. Yeah. Jack BeVier (1:00:29) arithmetic and algebra with you know with systems really, really well. So like you really don't want the next version because the the the distilled version of the next one does everything I I need and at five percent of the cost. So like I don't know if we're gonna see the next model net yet yet. It it seems to me that there's a significant financial incentive for them not to produce another model for a while. David Moses (1:00:50) Well, th that but that's the that's the benefit of having more than one like it was really a two horse race for a while between open AI and Anthropic, and it really wasn't even like Anthropic was just first, open AI was second, and then everybody else was way behind. Grok jumping back into the race is meaningful. And I think that that Jack BeVier (1:01:09) Uh-huh. I agree with that. David Moses (1:01:10) you know, and Google has to do that too. Like Google has to jump like all Google did was jump in and say, Here's a really great model, three point five flash. it's ho hum in terms of intelligence, but it's wildly fast. Very, very fast. And it handles multimedia very quickly. So it's still a good model to have, but it doesn't, you know, I can't use my OAuth token. I can't use, you know, I can't use my monthly fixed plan in Hermes. So it's n you know, it's kind of like a it's gotta be part of a of a finite workflow. But I you know, to go to your point, there I think there is always going to be the need for anthropic open AI and Groc to be first. So as the Chinese models, yeah, maybe they maybe all they're really doing is like they're never gonna get better than Grok. And they're never gonna get to the frontier because they're so dependent on the frontier intelligence to train their own models. That may not always be true. Maybe it will always be true. Maybe we'll get to recursive self-improvement where, you know, it's really just becomes an electron problem and, you know, who whoever has more juice and chips to, you know, you know, you know, chips to to to to operate and juice to throw at those chips is gonna wind up, you know, winning that race. but on the other side of the coin, like the argument against the frontier model saying, Hey, wait a second, You know, all we're really doing is like we're we're to the point where the the dumber distilled versions from China that are, you know, you know, wildly less expensive or free are plenty enough for most businesses to operate off of. We're operating in the world where you know the automobile has just been invented. Why would anyone need a horse? You know, that's that's true. Like certain industries will just get destroyed. But then there's the question of what can you do with an automobile that wasn't possible before. And I think those are the things where entrepreneurs like this is the the the most exciting time in the history of humanity to be alive as an entrepreneur because it's not it it will s it will soon shift from how do I do what I do better, faster, cheaper, to what can I do that I couldn't do before? And I think that pushes continues to push the frontier. Because we'll still need that frontier intelligence because we're we're coming up with that right now we can't even think of what we're what we're gonna wanna do. Jack BeVier (1:03:53) That's fair. That's a f it's a fair pushback and I, you know, not smart enough to come up with those ideas. So I follow. David Moses (1:03:59) You will. I mean we we all will like we'll be because it'll be like, you know, it all all it comes from is just recognizing a need. Right? Like people complain about X. People have always complained about X. We can't solve X. Right? I eventually that changes. We can solve X. Somebody decides to take it as a pet project. This is a problem I wanna solve and put resources toward. And now we have tools that weren't there before to solve them. And and and and the people who are using frontier intelligence to you know cure disease and like I all the kudos in the world to them. I don't have a billion dollars to go and do something like that. But but I can pick a much smaller problem and use my brain and say, you know what, I really want to focus on solving that problem. For for me, it's you know, for for me, it is in community management, there is this lag time between what is going on financially And physically with a community and its structures, there's this lag time between that information and and when it's actually delivered. And I want to eliminate that. I want to make it so you know in real time, in fact, you know Jack BeVier (1:05:13) Your time, thank you David Moses (1:05:14) now what you're probably going to need to address three months, six months, twelve months from now, and get really good at predicting that and getting getting really good at preparing for that. That's Jack BeVier (1:05:30) Isn't most of what you're building like what I'm finding is that most of what I'm building is software, is my own custom software. Like, and all of a sudden, just the unit economics of building software are now in my favor or that or are economical. And so I'm doing it whereas I wasn't doing it before. But once I'm done right in my soft, like, but but I provide a specific set of services, right? There is a edge, there are edges to the value of. Creation of my business, like the machine that I have built. There are edges to it. Once I've built software such that I have eliminated all friction that I could conceive of eliminating with bits and bytes, I'll be done. Like, not you know, which is an oversimplification, but I'll go from build, build, build, build, build, build, build to Does somebody have a new idea whereas we can expand the edges a little bit? But I will like significantly slow down. Like this is a CapEx cycle for me. like I'm I'm I'm unconvinced I'm unconvinced David Moses (1:06:32) I I I I think so and I don't think so. Jack BeVier (1:06:34) that this is like gonna keep going on forever. David Moses (1:06:38) Let me well I I I agree. I don't think Jack BeVier (1:06:43) Or David Moses (1:06:44) I don't I I I don't agree that it I I don't agree and I agree. Like I I agree there is a there's there's probably a ceiling for what you do. You kinda have to change how the industry works or follow, you know, be w be one of the first movers in whatever someone else comes up with is where that industry should move. but I think there are things in your you know, in the industry, let's just say lending, right? as as the industry, right? So for for you, what once you have gotten ninety percent of where you're going in terms of making the process most efficient, right? That's the process of taking the business that's coming in the door, processing it and delivering what you're, you know, delivering the the end product, right? That's Then then you're gonna shift your focus and you're gonna say, okay, how can I how can I eat everybody else's lunch? Right. And I'm not just gonna make it cheaper, right? but I'm gonna actually identify and predict who is going to need my product before they even need it. I'm gonna shift to make sure that I can, you know, support those who will need my product by making their lives easier. Like this is one thing we argue about. We've we've gone back and forth about what UWM does and what they don't do. what they don't do real well is make money. and I I I I'm not gonna necessarily d disagree with you there, but I think they do a really good job of maintaining their customers as a captive audience. And I think I think they're a really excellent use, you know, you know, a a really excellent example of how you work with your customers to kind of, you know, make their lives easier whilst making them dependent upon you. Right. Jack BeVier (1:08:48) Mm-hmm. David Moses (1:08:49) And and I think you know, I and I think that will ultimately be, you know Jack BeVier (1:08:53) But there's a I agree with that. And so I'm not suggesting that like it's not worth my while. Like, and I do think there's a significant first mover advantage to doing this, which is why we're being so aggressive about it. But I can't make more but but AI can't make more flips happen. Like once once I am become a monopoly, you know, my my my I have capped upside of like once we have softwareed to the nth, the f you know, providing capital to people who are f who are the ones interacting with the physical world. Like the physical world is still a can a a constraint for AI. Not in turn not only in terms of production of electrons, the n production of chips, but also like intelligence is only you know has a physical world limitation in that, you know, like I can I can't make more flips happen. I could make them a little bit cheaper, so maybe on the margin I can affect that a little bit, but like, but not, but not really. That just gets, you know, and and it's in the end state that gets passed through. So there's a tremendous business opportunity for me to go eat my industry's lunch. But say there'll there'll be five winners who like really hop on this, and you know, my goal is to be one of those five. But once the five of us have built our mouse traps to the, you know, where it's the most beautiful mouse trap ever. One, the mouse traps will look very similar to each other, or there'll be versions that'll have like split up the market, but you know, the each of us will have twenty twenty percent of the market because there's turns out there's five different versions of the mousetrap that are optimal for the customer. And that's it. We're done. Like there's nothing else for that, you know, to do. Like there's nothing else to build unless we go take on somebody else who hasn't caught up yet. So I'm not suggesting that like I I just think that like, you know, as an end state, twenty years from now, like twenty years from now, we'll be done. Like every every web every business will have, you know, in 1995, it was you need a website. By 2010, every business had a website. There was no more, you know, no one gave a You know, the there was no more money in in building websites. anyway, I just think that like I think that the idea that this is gonna like David Moses (1:11:01) there was no more money. You're right, you're right. Jack BeVier (1:11:03) extrapt out and then it's hey, how do we push the frontier forward? But these fr and the and then, you know, the frontier use cases, you know, do to continue have value. But this is like to me, I'm like, we're we're just like we're laying fiber and building websites right now. This is just the next version of that. David Moses (1:11:19) I I don't I don't disagree, but but you're but we don't know what we don't know, and we are as humans absolutely horrible at predicting. What what what I Jack BeVier (1:11:29) yeah, yeah, no doubt. David Moses (1:11:30) mean I mean like we we we they the two examples I always give are in you know in nineteen hundred there was actually a bill that almost passed both houses to close the patent office. Jack BeVier (1:11:46) Yeah, 'cause there is we've built everything, yeah. David Moses (1:11:48) And and the reason was that everything that would ever be invented had already been invented. You know, and that was 1900, right? So, you know, and and another another great example is like there was a I can't remember what the what the what the organization was, but it was an organization that did a study about electric vehicles, right? And it said that We will it was like January or February of that year. I think the year was like twenty fourteen, let's call it twenty fifteen, whatever it was, that it will be ten years before A million electric vehicles are produced in aggregate. Right. And this is like a very reputable organization. And lo and behold, that year they hit a million. Like that same year, they Jack BeVier (1:12:47) really? That's funny. David Moses (1:12:51) hit a million. It Jack BeVier (1:12:53) That's funny. David Moses (1:12:54) just is a little off, you know. But but like like people, because even the people who know are operating on old data. They're operating on Old systems, old processes, old products, the way we've always done things. And, you know, who knows where the demand will go? I I what I will say is one thing that's super interesting to me is flying cars. And not flying cars from the standpoint that, hey, I want to have a flying car, because I do, and I'm sure you do too. But I think it has a some really interesting real estate implication. Jack BeVier (1:13:30) Mm-hmm. David Moses (1:13:31) because you know, all of a sudden that useless island in the middle of that lake, you know, is much more valuable. You know, and and and all of a sudden, you know, that remote, you know, you know, people like space, you know, but people like connection. The whole reason cities exist is because people need social, you know, infrastructure. and they also need resources. And the Urban sprawl doesn't happen because, you know, people don't get along with each other because of the color of their skin or their religion or whatever. Urban sprawl happens because it's easier to transport yourself from here to there. The automobile is faster. There were more, you know, that that's why it happened. It's a very practical reason. And so flying cars, I think, changes the math on where we live. and I think that is a super interesting you know. Super interesting real estate like what's real estate gonna look like in the future kinda question. Jack BeVier (1:14:37) Yeah. I do agree. Did you did you and Danny talk about any futurism? David Moses (1:14:43) Flying cars? No, no. I mean, we we we spoke a lot about government regulation and how the government, you know, how how the government gets in the way. Yeah, and Jack BeVier (1:14:55) NRH C stuff. Yeah. David Moses (1:14:57) and and and the government does. I mean, like, you know, it does. It it it does tend to to get in the way. And I think flying cars will be a a prime example of that. But the the old adage of the FAA is is the FAA is not happy until you're not happy. Jack BeVier (1:15:15) That's funny. David Moses (1:15:16) It's like their actual tagline. Jack BeVier (1:15:19) Well, lots of fun stuff in store for the future. David, I gotta jump, get back to it. Good conversation as always. Thanks everyone for listening. Hope you got something out of that. This was forked.ai. Have a great week. See ya.