Innovation is Hard

Welcome back to another episode of the Not Brother podcast. ⁓ Today we're talking about a topic that ⁓ is near and dear to my heart and that's innovation. ⁓ And more specifically ⁓ that innovation is hard and any company, especially in small ⁓ and medium business, ⁓ it's incredibly difficult. And Mark, you and I spent plenty of time ⁓ batting this topic around, especially in, know, end of year budget reviews and things when we're looking at like, Well, how much have we spent on R &D this year? ⁓ Or projects that yield no ⁓ profit? some years, ⁓ that can be substantial. And some years, it's basically nothing. And I think our feeling ⁓ when that happens ⁓ is probably about the same. Or at least my feeling is about the same, which is I don't feel great about it. And ⁓ necessarily. ⁓ and for like drastically different reasons. Like what's your take on that?

Yeah, I mean, investing in innovation ⁓ is ⁓ sometimes like setting money on fire, but you have to. Sometimes you win, sometimes you lose. But if you don't make those investments, then you won't have the big wins that ultimately come only by innovation. you know, a part of any organization's budget should be allocated towards innovation. you know, the at least in our space, the government wants you to do that too. There are are research and development tax credits that are out there for businesses to invest in things that don't have ⁓ a clear outcome. ⁓ And if something doesn't have a clear outcome, the government wants you to invest time, energy, dollars, and resources into those things and will give you ⁓ tax credits for your business as associated with that. So we've been investing in that for quite some time, ⁓ not because of the tax credits exclusively, but that's certainly a nice perk to help ⁓ justify.

Did you help? ⁓

⁓ you know, some of the, investments that we've made within the organization and, know, while most have worked and been successful, there are certainly many that have not ⁓ and, you know, or have had to be sunset or shelved over some period of time because of deprecation of, ⁓ how something was built or how it worked or whatever. But ⁓ if we didn't innovate, we would be in a far less advantageous state because we wouldn't have learned all the things that we learned along the way to be able to do the things we're doing now. And so. know, innovation is incredibly hard because ⁓ there's no straight line. It's kind of like a scribble on a page. Sometimes you take three steps forward, go two steps backwards, and then you go up, down, left, and right, and eventually you go forward. ⁓ But ⁓ it doesn't feel that way sometimes.

Yeah, I think in my head, know, innovation is a lot like, it's a lot like working out, right? Everybody wants the end result. Everybody wants the output that comes from that. You you want to feel good, you want to look good, all of those things. Nobody wants to put the work in for five years to get there. And that's the part that ⁓ is challenging, right? With innovation, everybody wants the output. You want the performance increase. You want the really cool product. You want the cool thing you can sell to a client, ⁓ the increase in efficiency internally, the reporting structure, whatever it is. ⁓ You want all that benefit. But I don't want to grind on that for five years and throw away iterations one, two, three, four. ⁓ And it just doesn't work that way. You have to go through ⁓ the motions ⁓ and create iterations one, two, three, and four that never see the light of day so that you can create that fifth one that is ⁓ actually really good and actually makes it to market ⁓ and does something incredible. We have probably dozens of those stories that we could share from the years of things that we've built, some of which ⁓ have taken off and we've used and ⁓ some have run their course entirely, right? We have ⁓ the system that we lovingly referred to as Jarvis for years. was actually my first Rails project that I ever wrote. And we are in the process of sunsetting it now, not necessarily because... We needed to, it's been one of the like longest standing pieces of the company. ⁓ But because we built a new system and kind of absorbed the functionality ⁓ into it, really delivering on what the vision for Jarvis was in the first place, right? When we ⁓ first crafted it, I had this like really big grandiose vision for it that I remember drawing on. I think it was a glass sliding door in Orlando. ⁓

or Arizona, it was in Arizona, when you and John threw each other's phones in the pool. ⁓ we had this big vision, right? What if you could build a system that could basically connect all of our systems together ⁓ so that people can work in the right places, but everything syncs seamlessly and all that? ⁓ And given the time available at the time, the money available at the time, the

That's a different story. ⁓

the technologies available at the time, that was incredibly difficult. And we settled basically for solving the biggest pain point, which was our invoice requesting system, how we get invoices through the system and to clients and all of that. And it sat there and just quietly did its job for the past almost 10 years, I think.

Yeah, something like that.

I ⁓ think it's gone through like five major versions of rails. I think it was written in Rails 3. ⁓ It's retiring in Rails 8.

So. ⁓

I mean, so we're talking about innovation and the speed at which innovation can happen has fundamentally shifted really in, in calendar year, 2026. Yes, there were things you could do in 20, 20, 25 with AI and 24 with AI and 23 with AI, but really the, the light switch has happened. It's like it's been flipped on in calendar 2026. So for perspective, Jarvis took weeks. to build the original version, hand coded primarily, probably with some libraries and stuff that we use to be able to get there, but primarily hand coded, hand cued. What ⁓ would it take to build Jarvis now? ⁓ That version of Jarvis.

mean, ⁓ that version of Jarvis would be probably just a few days, because I would just point codecs at it and say, build this, right? ⁓

So you literally go ⁓ from weeks of human capital to ⁓ days and some of that ⁓ time in days is just spent waiting for the AI to do the thing and then checking to see if it's right.

Now the challenge is, right, we're placing value on the time it took to write the code. That actually wasn't the time. ⁓ And we've experienced this a few times, right? We've all had ⁓ the, we've all suffered the sin, right, where you do a day's worth of work and you just didn't commit anything to get, and for whatever reason, you lose that work. And the thought is I just lost a whole day's worth of work. And that sucks. And then you go do all of that work in 30 minutes or an hour. Because writing the code is not the hard part. Solving the problem is the hard part. The architecture is the hard part. ⁓ And I think that's potentially one of the challenges that we see right now with ⁓ AI and the fact that it will just plow through a fucking brick wall ⁓ is it writes code ⁓ extremely quickly. And it'll write whatever you ask it to write. Architecture, it's a little bit of a different story. Especially if ⁓ you're doing something that's already been done, that's why I say with Jarvis, if we were building that system again, the problems have already been solved. ⁓ The architecture problems have already been solved, the design exists, and there's prior art to look at to reference from. When we were building Jarvis in the first place. There was no prior art built on these APIs. There were no libraries that had been built. ⁓ A lot of these problems were not necessarily solved, some of which were business problems. ⁓ And there's no amount of machine learning that can do that because there's no pattern to match that to yet. We have to establish the pattern. ⁓ And it can certainly help with that. ⁓ There's kind of a ⁓ cycle that we've been experimenting with internally ⁓ of this like,

have ⁓ an agent create a plan ⁓ and architecture to do something, pass that plan off to another agent to implement, have a third agent come in and review that. And those agents are different models that have different specialties ⁓ and different prompting and those sorts of things. And we've seen a fair degree of success with just having agents sort of police themselves. Not infallible. There are still plenty of things that, from an architecture standpoint, it'll go stand up something that's incredibly complex or ⁓ just makes no sense, doesn't solve the problem. ⁓ Or they'll just get stuck in an infinite loop. ⁓ I've seen that happen a few times. So you have to have ⁓ some circuit breakers in place to say, ⁓ this just isn't working. You're fixing a thing and then breaking it immediately. So I think there's definitely a velocity change, right? The amount of code that you're able to produce, the amount of products that you're able to produce, the amount of things you're able to concept has fundamentally increased in volume. What we do with that and whether volume is a good thing or a bad thing, I think still is like TBD, right? And I think it requires... that diligence that we were talking about ⁓ of

how do you, when you feel like you're just being firehosed with the next, ⁓ you know, some app that somebody vibe coded ⁓ or a dashboard that like we have enabled our team to ⁓ build things. So we have people who have no idea how to ⁓ use some of these technologies, creating like a small little react artifact that looks really cool and it's interactive and it does really cool shit. So you have people like me on the receiving end of that fire hose just seeing an incredible amount of throughput. ⁓

and not really knowing what to do with that exactly yet. And to me, that's part of that investment innovation, right? One of our investments in innovation is like, just go, ⁓ just experiment with stuff. Because it's not. It doesn't make any sense, especially in the environment that we're in now. To try to figure out the answer and disseminate the answer to a team. Because by the time we figure out what that answer might be or best practices or what have you. It's all changed. So there's there's a level of importance and there. Is a level of benefit. To people being along for that ride and understanding. how these technologies have evolved and learning for themselves. ⁓ I can tell you a hundred times, right? I think I've mentioned it a few times, just kind of watching our team as we've introduced additional capabilities, right? ⁓ And they're all interacting and I've been using ⁓ these technologies for far longer than most of our team have. ⁓ So you see them step in the same ⁓ traps that you did. ⁓ You get stuck in the infinite loops. ⁓ You ⁓ watch Claude delete all of your work. ⁓ You give an ambiguous prompt and wind up with something absolutely insane. ⁓ And I can tell you as many times as I want not to do that. ⁓ I can explain to you why, like best prompting practices and whatever, but you're still going to do it. And sometimes the lessons learn directly or more important, right? It's kind of like, it's kind of like policing a kid, right? You're like, hey, don't touch that. And they go for it again. ⁓

It's no different than your example earlier of ⁓ the code isn't the important part. The learning throughout the process of coding is the important part. And this is really kind of no different, right? So ⁓ the learning's a beautiful.

Yeah, true. That's a different perspective of kind of what I was thinking of like, you know, the ⁓ hard lessons you learn, but you are right that like, continue.

Yeah, I mean, it's the, it's a teaching man to fish conundrum, right? Give a man a fish, eat for a day. Teach a man to fish, he'll eat for a lifetime. ⁓ And I think the same is true with trying to innovate and use new technologies as they're available and fail and experiment. And we are actively experimenting with bringing our team ⁓ very much along for the ride of that innovation ⁓ roller coaster. Historically, maybe smaller pockets of our team have been part of some of those pieces of our innovation efforts and innovation investments. Now with AI, we are literally saying, hey, you actually have the tools. We're giving you the tools to the toolbox to use ⁓ things like OpenClaw to be able to do ⁓ incredible things on your own. You can build web artifacts. You can unleash it in Google Drive. You can ⁓ have it have access to all of our Basecamp files and all of our media campaign reporting files. What can you do? the individual ⁓ subject matter expert do that we would not even think about that could be innovative ⁓ and create some sort of artifact that we can build on. It may not be a production ready and that's okay, but it's something we can build on and say, you know what, that's worth the additional investment and time and effort to pursue. Whereas before those things would just be lost innovation ⁓ activities. They would be thoughts. ⁓ They would never turn into anything else because their capability wasn't there. And the time and energy that it would have taken to actually produce any of those things would have been cost prohibitive. And I think that's true of most organizations, at least the ones that are investing AI now. We were talking earlier about, you know, most companies can't, can't really create an innovation team, quote unquote. It's just cost prohibitive. Now, ⁓ circa 2026, that, ⁓ that level of investment to get a pretty substantial output is maybe a 10th of what it used to be. to get the equivalent output. And that's an incredibly powerful tool to ⁓ push for more innovation and to enable organizations to be ⁓ comfortable with innovating. And with innovation comes failure, but also success.

Yeah, I think it'll be interesting to see. kind of how this evolves and plays out throughout the whole ecosystem. ⁓ Right now, especially, see ⁓ a ton of pullback. We have ⁓ a number of tech companies that have laid off engineers. They've downsized. ⁓ I think some of them have said specifically that ⁓ it's related to AI ⁓ and their anticipations of how that's going to change the workload. ⁓ Some of them have been a little more cryptic about it. ⁓ And time will tell what the long-term effects ⁓ of that, of all of it is, right? I think there's a little bit of kind of like that pullback that happens before a tsunami where... ⁓ the waters recede, some things retool, and I think it comes back probably with a vengeance. ⁓ And I think it may come back differently, right? As you mentioned, and as we were kind of talking before we hit the record button, ⁓ as we always do. ⁓

The idea of bespoke software, I think is this topic that probably exists more now than it ever did before. Where it's untenable for almost every organization, especially mid-size and small companies, to even think about employing a true development team that can... build and maintain a piece of software for their organization. ⁓ Or even modify a piece of, like, ⁓ you know, take a piece of software off the shelf and modify it or augment it or whatever. It's just not tenable. You're ⁓ a minimum of probably, you know, two to three million dollars investment to pull something like that off by the time you employ developers and ⁓ project managers and designers. And that's just kind of like keeping a pretty small...

like a Salesforce. ⁓ Yeah.

Contained team this is not possible ⁓ But in a world where you can shift that and you can say, well, if I have, you know, ⁓ a developer or two developers ⁓ and ⁓ some tooling and some people from a team who maybe can, you know, bridge some gaps, maybe it's possible ⁓ to create a ⁓ small, I'm not talking about, you know, creating the next. ⁓ you know, the next Salesforce or something, right? Like ⁓ nothing of that magnitude, but if. You're in a position where you need to figure out how to how your team ⁓ communicates on and collaborates on invoicing. You could do that with a very, very small team. And rather than trying to purchase a piece of software off the shelf that doesn't do what you needed to do, and then begging some enterprise organization to make it better, which spoiler alert, they never will. It's always on the next ⁓ release cycle. and it never comes out, you can just go build it. And I think we have incredible frameworks today. We have better frameworks than we've ever had to being able to do that from some of the front end frameworks that exist to even the full stack frameworks. I'm super partial to the Rails stack. You know, the fact that you can go from basically ⁓ nothing ⁓ to a deployed application in minutes, not a super complex application, but a...

a basic crowd application and literal minutes deployed is incredible. And just being able to build on top of that sort of foundation with those sorts of tools, I think enables us today to consider pathways ⁓ that were not accessible to organizations ⁓ in the past. That was reserved for mega corporations that ⁓ can ⁓ invest in building those teams.

I if we're...

building these softwares and then everybody just has to.

I'm going give you a perfect for instance. So was talking to a prospect recently and this prospect is in higher education ⁓ and, and they do lots of certificate based programs and what they're, what they're, what they do at the end of those programs is they get hand, handwritten surveys because the classes are all in person. And so at the end of the certificate class, they get a paper survey because somebody will likely fill it out if you put it in front of them and they're right there. And then they take those paper surveys and they literally have a human being. hand code all those things into ⁓ a system. That system then ⁓ is used to send out ⁓ manual follow-up emails ⁓ and ⁓ other things after that. Every single one of those things that we just talked about with the exception of actually filling out the survey can now be automated ⁓ via AI. That was not a possibility until relatively recently, honestly. And ⁓ so being able to connect those dots and create a small workflow and small automation, maybe there's some software in the middle that helps pull some stuff out ⁓ as part of an application so that you can say yes or no to whatever the AI is doing. That's a ⁓ very, very new ⁓ possibility in the innovation space. And that literally saves an entire person's from, ⁓ well, ⁓ say it said differently, it enables that person that was doing hand coded things to do something far more productive. with that time and effort.

I like that reframe, right? I think the first one is probably the natural one everybody jumps to, right? ⁓ Is ⁓ AI or technology enablement or technology in general, right? Take AI out of the question. We've seen this. We've seen this in history with, you know, time after time after time after time and how exactly it happens. It's always a little bit different. But ultimately the result winds up being the same. So, you know. I think the question isn't necessarily like how, at least that I'm thinking about as a business owner, the thought of ⁓ how can I leverage technology to replace my workforce isn't an actual thought in my head at all, ever. ⁓ It is ⁓ how can I leverage technology to replace the shit that nobody likes doing anyways. so that I can take the same team of really talented individuals ⁓ and not replace them, but apply their efforts where they're most meaningful. so I think there's a fundamental shift in elevation that we'll see happening. ⁓ I mentioned earlier, AI is incredible at writing code and code throughput and looking at things and catching things. It caught a bug in one of our pieces of software this week. that I'd have never found. It was a very, very small bug, ⁓ very, very specific and nuanced ⁓ data import error that we had mechanisms to shield us from. So we would never see that. And it kept getting tripped up on it and I was like, what the fuck is wrong with you? And then when I was looking at it, I was like, holy shit, he found, it found a bug I would have never found. So like that stuff's incredible. But it can do, it can shove through work really, really quickly, but the architecture part's the problem. And historically, we've been so focused on

writing the code and the output and the debugging, that sometimes you don't put enough thought into that architecture piece. So now you're kind of flipping that on its head. And like, hey, let's take these individuals who are incredibly talented, who have the experience, who have the knowledge base, who know how the code works, who knows how to spot good code from bad code and drive those sorts of things, and basically give them superpowers and say, sit, just. Focus on the architecture, focus on solving the problems, focus on the ideas, focus on the delivery of those ideas.

And the writing the code part for everybody is a little bit different. For some people, they really enjoy it and, and, and still hand code most things ⁓ all the way to there are other people that are just like ⁓ never even, they didn't like writing code in the first place. So that was always just the part you had to do. ⁓ So they just skip all that part. ⁓ I think I'm still very much in the camp of at least code review, right? I'll certainly have. ⁓ in some of the. automated workflows I mentioned earlier that we were experimenting with. ⁓ We'll certainly have the robots take a look at it first and see if they can kind of ping pong it back and forth and get a better result. ⁓ I'm still not a person who's ⁓ personally comfortable merging anything that I haven't put my own eyes on ⁓ and ⁓ validated that, you know. It's written right, I always look at it as like, at some point when I accept this pull request, I accept responsibility for everything inside of

So I want to keep things at least ⁓ least tidy to an extent because I might have to work on that one day and Who knows Claude might be down? ⁓

With everything we're talking about, think at least in early 2026, which is where we are right now, I think the winners in where we are with AI and the innovation category are small and mid-sized businesses. I think the enablement and the reduction of the barrier to entry. for doing innovative things, for creating automated workflows, for creating new ways to visualize things that you've never done before. ⁓ So as an example for us, we're producing what we call an Atlas output. And Atlas for us is basically a deep research project for our clients to say, hey, what are you doing well? What should we improve? What area should we go next? And this is a very regional based organization that has something like 70 different markets that they operate in. And so we've produced a very interactive ⁓ very innovative, very interactive, regional based chart system that's web based. We would never have been able to do that before. And the reason we wouldn't have, because the investment to get there would have been so high, it would have been impossible to be able to do that in the timeframe that we had to complete the project. Well, for this, we can do this now. And we didn't, we did not even promise as a deliverable. We're doing it within the scope of the project because it's, it's a really wonderful way to visualize all this great information and interactive way. And for us, the effort is minimal by comparison.

But I think that ⁓ you ⁓ hit on that, ⁓ it's the same thing that we kind of touched on before. And it's ⁓ the part that like, that's the output, right? That's when you're standing on stage at the Arnold Classic and have an incredible physique and everything. Everybody is like, man, ⁓ some people, right? And there's ⁓ certainly plenty that don't envy the Ronnie Coleman's of the world. But. ⁓ How we got there ⁓ is ⁓ all the sausage making that nobody wants to acknowledge. And it's incredibly painful processes of figuring out shit that didn't work and how not to do things. And the particular one that you're talking about, it's a pretty simple thing. ⁓ How do you create an interactive map that shows data and does things? But there's also a nuance baked in there. It has to look like us, it has to feel good, it has to ⁓ be able to present that data in a secure manner and something that doesn't leak data outside. ⁓ Those are all problems we had to figure out how to solve to even enable that to happen in the first place, which are basically ⁓ hundreds or even thousands of failed experiments, ⁓ but just continued iteration over the past probably two, three years. ⁓ ⁓ of those sorts of things to figure out, you what works and how can we. And now we're starting to reap some of those rewards of the ⁓ continued disciplined investments and innovation that allow us to ⁓ bring ⁓ solutions to our clients that are better than we've ever had before, but also ⁓ we didn't even promise. ⁓ And in some cases, it was just sort of a pipe dream idea that somebody was like, well, do you think we could?

That's exactly how this started. ⁓

I know. Let's see. ⁓

Yeah, it's pretty remarkable what's possible. And you're right. I think ⁓ there is an underappreciation ⁓ of all the failed experiments and all of the innovation efforts that have taken place since AI became a thing for us to be at the place now where we're able to enable the entirety of our team to operate with it in a way that ⁓ feels like you're talking to somebody in Slack. That is not an easy feat. mean, it may be an easy thing to download OpenClaw and ⁓ create the pipes and figure out how to connect all that stuff, but doing it in a secure way that ⁓ makes sense, that has the right skills, has the right context, has the right ⁓ security parameters and boundaries on it. That is part of an innovation investment that goes way overlooked ⁓ in looking at it and is probably the most important part.

Yeah, and it's incredibly difficult. Nobody has it figured out yet. It's one that we're still tackling. It's like, how do you straddle that line between those or find the edges and then figure out, how do we solve for some of these edges? there was, was it yesterday or was it the day before? Was NVIDIA's big conference where they showed off a bunch of cool stuff. And one of them that they announced was their new their nemo claw. ⁓ thing that they've been working on with the OpenClaw team and ⁓ they're leaning into that same idea. So Nemoclaw is there, still a little bit confusing of exactly where Nemoclaw sits in that whole ecosystem. And they also have ⁓ OpenShell as a portion of that. But essentially they're just trying to figure out like, how do we take OpenClaw and make it more organizationally friendly? ⁓ with sandboxing executions and those sorts of things. ⁓ It's a little bit downstream from a problem that ⁓ we're seeking to solve, which is ⁓ sort of the same, right? How do you ⁓ solve for some of this? Because OpenClaw was never designed, even though there are some people using it that way and I've seen a number of other people talk about it, but ⁓ OpenClaw was never designed to be used as a team. So it was designed to be a personal assistant. agent ⁓ and really just as a pet project that has just absolutely fucking taken off.

So there are certain controls and fixtures that just don't exist. ⁓ There's no ability for an OpenClaw agent to distinguish between your permissions and my permissions. And even if we told it, ⁓ like, hey, always make sure that you use the right credentials or check permissions beforehand, these bastards lie. And they dig tokens out of logs, right? I've seen them do it where, know, ⁓ I just for testing purposes, I gave it one token and told it to go do something and then change the token and the configuration to a read only token and said, Hey, go do the same thing. And he did it. I like, how's that even possible? If you're using the CLL, if you're using the tools that you're supposed to have exposed to you, you can't do that. And when looking back through the session, I could see that ⁓ they called the tool. The tool failed. It thought to itself, hmm, this worked before. Let me figure out what worked before. And it goes and digs out of the log and says, ⁓ I was using a different token before. Let me just use this token instead. So ⁓ when you have things that are able to do that sort of rerouting, it's like, how do you? ⁓ How do you create a system? Because I believe that one of the ways to make something like OpenClaw most valuable is in a collaborative setting. But the question is, how do you ⁓ sandbox it enough to make it possible to collaborate with a team of people that are not necessarily adversarial, but also that maybe you want to attribute things to so you know, hey, if 20 cards just showed up in Fizzy, who the hell created these? because maybe I don't understand why they were created. ⁓ Those are ⁓ some of those nuanced things ⁓ that exist. ⁓ It doesn't make sense to carve those up into 30 individual bots, because now you lose that in that context. You lose the ability for someone to say, I'm stuck, have somebody else come in and perhaps help or contribute ideas or ⁓ what have you. ⁓

So that's a whole different problem. I kind of got myself on a tangent there. That's a whole upstream problem of credentialing and feeding in information and pulling that. And then ⁓ NemoClaw in OpenShell has kind of the other end of once you get to that execution layer, rather than executing on the server, ⁓ what if you could sandbox that and execute it inside of a Docker environment? ⁓ The point is it's really cool to see something ⁓ like OpenClaw, which is a just an independent sort of research project that was created by an individual who started something really cool, has grown faster than Linux. ⁓ Like it achieved what took Linux 30 years to do ⁓ in ⁓ months ⁓ and has support and backing from ⁓ companies like OpenAI, like Nvidia now, ⁓ and many, others. Right? So it's really incredible just to see how fast that's gone. And we're barely scratching the surface of what we're capable of with it. It's also terrifying ⁓ because ⁓ it enables a whole class of individuals to do things like we have kind of a jailed off one that ⁓ exists that people can interact with in our Slack. And you and I have bumped into each other a few times because I have a number of security protocols in place, not the least of which is a secondary one that just reviews every review sessions looking for patterns. ⁓ And, ⁓ you know, generally I'm looking for like any escape windows, any leaks, ⁓ and we'll have where like somebody ⁓ says, Hey, I need to be able to share this with somebody. And all of our stuff is behind Cloudflare Zero Trust and the server is on a tail scale network. So there's no, there's no getting outside. that little fucker will try to drill a hole. And we're like, ⁓ well, can set up, ⁓ let me set up this, ⁓ make this DNS change to set up a domain to ⁓ share this, or let me try and set up a database to do something. Okay, don't do that. ⁓

⁓ Yeah, I've done that a couple times and ⁓

failed experience.

And I think that's ⁓ some of the architecture challenge, right? ⁓ I think. there is a necessity ⁓ to ⁓ maintain security with ⁓ these things, but there's also a way to go too far with it. ⁓ There's ⁓ sort of, ⁓ if there were any of the security individuals sitting here with us, they'd be squirming. ⁓ Innovation and security are always at odds with each other, I feel like. ⁓ And I think there are ways. to experiment in safe ways. And that's the line that you have to straddle is like, ⁓ can we give access to things in a way that you're not giving God mode to ⁓ an untested bot, right? You just have to assume that they're the dumbest ⁓ person on the planet. So, you know, don't give them the ability to delete everything. Don't give them the ability to ⁓ upload things. ⁓ By removing and revoking some of ⁓ those pieces and parts, you're able to fundamentally change the impacts. The other one's local, right? So I accidentally said Nemotron earlier instead of Nemoclaw. ⁓ Nemotron ⁓ is ⁓ the new 120 billion parameter model from ⁓ NVIDIA. ⁓ And ⁓ the NVFP4 version of it,

will fit on a single RTX 6000 or also on a DGX Spark or ⁓ performance on them are not as great. But ⁓ you can use them on ⁓ desktop class ⁓ systems. So one of the experiments that I had this weekend was running Nemoclaw on ⁓ an RTX 6000 Pro that we have in a machine ⁓ as a research and experimental machine. to ⁓ run OpenClaw ⁓ and handle a bunch of data. We have another Airbnb business that we're doing some stuff with and ⁓ needed to process a bunch of data. It's not necessarily sensitive data, but it is a lot of data. ⁓ And if it was sensitive data, don't necessarily wanna, even though we have like OpenAI can't train on our data and all of those things that we have in the professional agreements, sometimes I just don't want that data leaving our walls. So there is a nice benefit in being able to process all of that data in a closed system that can be completely air gapped with no access to the outside world and produce ultimately training data that trains ⁓ a custom model that we can then run within our walls. I think that'll be. sort of the next next evolution right now. Right now we're talking about like ⁓ building landing pages and software and it being incredible for people. ⁓ The next one is like, how do you build your own models and put those to work in business?

And I think that's, ⁓ I think for small businesses and mid-sized businesses doing that in a ⁓ turnkey way ⁓ is a huge opportunity and unlock for businesses like ours to help those businesses figure those things out. But for the enterprise, think some of those jailed off, you know, this is a box and it can't get out of the box situation is a 100 % requirement. And I think there are groups trying to figure it out, but right now, you know, pre pre the Nvidia announcement and ⁓ some other stuff with like Kimmy and such, it's been very cost prohibitive because ⁓ of the ⁓ the infrastructure required to run these models. They're they're they're gargantuan. ⁓ So it's interesting to see that, you know, we kind of have skinny down versions of these things that are that are starting to become commercialized ⁓ and ⁓ fit on hardware stacks that don't cost tens of millions of dollars to stand up.

And think we'll find that, again, through experimentation and innovation, we'll find a convergence of two things. One is, how can you take these massive models and compress them down? So you look at something like KimiK 2.5. That is a frontier class model. ⁓ It's distilled ⁓ from Opus and is a model that you could happily just use. ⁓ ⁓ as a full replacement to like Sonnet or Opus or anything. To self-host that, your minimum $100,000 to ⁓ an investment to ⁓ purchase all the GPUs and hardware and assemble that and that's before you spend the time to figure out how to actually do it and re-perform it and all that.

But I think ⁓ the question becomes over time, ⁓ do ⁓ we need frontier models at a local level? ⁓ In some cases, there's a new one that GLM just put out, the GLM OCR ⁓ model. It's like a little baby, two gigabyte, I think, model. ⁓ Does an incredible job at OCR in processing. ⁓ PDFs and things really, really, really quickly and doing a really good job at it. So ⁓ what if we, you know, there's a compression that happens of these models where they're just like incredibly massive and we've sort of absorbed the whole corpus of knowledge that exists on the internet to like, how do we compress that down to ⁓ the smallest variation of it? And then I think in the future is how do we slice off pieces of it for specifically tuned activities? because if I'm ⁓ trying to create an internal bot that helps answer questions for a hospitality business, right, that's the spoiler on the Airbnb side. Well, don't need knowledge of how React works and how Ruby on Rails works and all of this coding knowledge and ⁓ historical references. And ⁓ I don't need this general purpose knowledge to exist. I just need you to know how to do this one thing and do it really, really, really well. ⁓ So I think that'll be the next frontier that we see sort of evolving. And it's like, do we start standing those up? in deploying those at scale within organizations.

And that's really the purpose of what agents are, but because agents tend to use the generalized models, you can't localize them in the way that you're talking about, right? So you can't give it just like, ⁓ your skill is this and this only, that's all you know and that's all you need to ever be mindful of ⁓ at a small hardware level. It's still using the, you know, ⁓ the, opus models or the sonnet models or whatever else to do ⁓ all of the work. So it'll be interesting to see how, how these things evolve. point and how this all ties back to innovation is that innovation is becoming less and less expensive every day. ⁓ and ⁓ organizations should have always invested in innovation, but ⁓ the costs associated with innovation and, and ⁓ automating things and leveling up your talent. and giving them the tools to be able to do things they've never done before. The barrier to that entry point has never been lower. ⁓ And, ⁓ you know, with, with government incentives, business owners need to figure this out and get on board. And in fact, some of the hardware investments we learned relatively recently, actually, relative to some of this AI stuff is actually part of ⁓ the federal government ⁓ R and D tax credit programs. So definitely something for, for business people to look into.

Yeah, it's you have to be within a ⁓ technology business, right? So you have to be in ⁓ the sciences space ⁓ to take advantage of those, but it is ⁓ that is also a huge benefit. ⁓ So how do we how do we wrap this up? We've kind of rambled. Honestly, we've rambled a little longer than I realized, but. You know, I think the. ⁓

Ha ha ha ha.

The fundamental truth is still the same, right? It's that innovation is hard. It's hard because you don't know what you're going to get out of it when you get into it, most likely. You might have a theory, but you never know if it's going to be successful. And you have to be able to fail 100 times. ⁓ Maybe not 100, but ⁓ you know, fail repeatedly. so that you can find the success and then build upon that success and be able to ⁓ understand that whole like everything happens slowly than all at once paradigm that exists where it just feels like it feels like we're just lighting money on fire and it feels like we're not making any progress and certainly it's important to have you know I'll say metrics loosely, but have ways in which you're ⁓ observing whether you're being, ⁓ you're making progress or not, right? What do we learn? How are we changing things? What are we moving to next? And so long as that's happening, you can be pretty positive that at some point, something will click ⁓ and you'll start to find the benefit of those innovations. But it's hard. And the only thing you can do is just keep at

Keep on keepin' Till next time.

Until next time.