Latest / 4kd.ai / Ep. 4 | Navigating Change Management in AI Adoption

Transcript

Ep. 4 | Navigating Change Management in AI Adoption · 4kd.ai

• Craig (00:00) Hey, welcome back to the fifth episode of forked.ai. I'm Craig Fuhrer joined again by the great David Moses and Mr. Jack BeVier Jack, everybody. Dave, it's great to see you today. Producer Kyle not on screen, but good to have you here as well, sir. So guys, we got a lot to talk about today, but I thought I would kick it off with just some, you know, I feel like every now and again, we feel a shift in the whole AI landscape. And you guys, I wanted to get your thoughts on three things that have come across the radar over the past week that I feel like we're kind of in a way moving away from the agentic layer more to the application layer. so like model may be not be the most important thing, but what are you putting on top of the model? And so a couple of news stories that came out this week that I thought were interesting. One, SpaceX, looking to buy cursor for $60 billion. I don't know why I didn't write that program, Jack. I don't know why you didn't either. It seems like an easy program to write $60 billion. And by the way, they weren't the only ones courting cursor. Microsoft wanted to buy cursor for a minute, but it looks like SpaceX is going to win. Second thing, a great news story that came out this week is Claude design. And, you know, I don't know how much occasion the three of us will have to be working in. Jack BeVier (00:57) my God. Craig (01:17) ⁓ Claude design, but it's basically a killer to two or three other great apps that people have heralded over the past several months, which would be Figma, lovable and replet. if you're using Claude design right now, I can tell you in many ways it's in rev one far better than the three of those. And then the next one was, ⁓ chat, chibit images, 2.0 guys. If you haven't had a chance to see this or just play with it, Jack tonight, just Jack BeVier (01:42) Mm-mm. Craig (01:43) just open it up and have some fun. Like you would be shocked at how much better it is than chat 1.5 and even nano banana. I've got a whole bunch of other bullet points of adoption and it is one of the fastest growing adopted application layer things in the, or applications I should say over the last 12 months. lots of crazy stuff happening in the landscape. Thought I'd just open it up with that real quick guys and we can talk about that and then jump into some stuff that we're all working on. Jack BeVier (02:14) So what's images do? it like nano banana? Craig (02:15) So images is much more of a thinking model than it is just an image model. And so I saw a couple really wild use cases where a guy is like, hey, here's six pictures of me. Put me in a comic strip and use, you know, use what you know about me to make the comic strip and within a couple of minutes, it made a beautiful comic strip. But the more important part that it got right that a lot of image models haven't gotten right in the past is the verbiage. Can't trust the verbiage. Well, it was spot on. And then he took a picture of himself in the room that he was in the studio and it had sort of signage and back of him books and stuff. And it even, and he said that book behind me is, I think it was called Get Done. And he was like, make sure that the ISBN on that book is correct. And it put the ISBN with the, with a barcode on it that was correct on the, on the spine of the book. And then he was like, you know, and while you're at it, you know, make sure I'm in a kind of a burnt, a burnt rust shirt and put the, put like a logo on the front that is, my logo. And so that was a really, really difficult thing in the past. You would get one image with nano banana. Jack BeVier (03:10) That's sick. David Moses (03:17) you Craig (03:25) But then as you tried to iterate on it, it never really did get it right. It was always trying to get it right, but it never, but then it would break something else in the picture. And so yeah, magazine quality layouts, infographics with unbelievable accurate data, multi-format ad campaigns, just all kinds of great things that you can do with the new. And then when you couple, So let's say Jack, then you got all of these great images and maybe let's say that like what you were building was great landing pages. So you give it 20 images of the most amazing landing pages that you've ever seen and then you put all of that into Claude Design and now you've got something pretty powerful. Jack BeVier (04:03) That's super interesting. Have you played with it at all? Craig (04:05) I played with Claw Design, haven't had a chance to really, don't really have a use case right now that's burning for me to play with the image prompt as much as I love that kind of stuff. So no, I have not had a chance play with ⁓ the new Chat 2.0 image. Jack BeVier (04:18) Sounds fun. Sounds fun. Craig (04:20) Yeah. So, so what do you guys think though? Like, so here we have Claude 4.7 that came out recently and, ⁓ we've been banging on that pretty hard. David, I think you, you mentioned as well that, yesterday when you were talking, you said, yeah, it's, it's like a club 4.7. It's 4.6 before it took a right? Like it's so like, I don't notice any great new David Moses (04:39) Hehehehehe Craig (04:44) crazy functionality with 4.7 that I was missing in 4.6. Just wanted to get your thoughts on that. Jack, I know you've been working pretty hard with Claude lately as well. David Moses (04:54) Four points, exactly my opinion yesterday is my opinion today. It seems to me to be 4.6 before I took a it's interesting that chat is, you know, I don't know if chat's just kind of like flexing the muscle with this, with the new image generation thing, because you know, the reason Anthropic does not do image generation in my estimation is the compute that's required. you know, that's, yeah, I know they want to focus on the business customer, but in reality, I just, they could kill ChatGPT in that, the way they kill them in other things, but they just don't. And I think it's really just a compute thing. And I think that's what I see with Anthropic. It's a superior product, but they just don't have, they can't service the demand. That's kind of what I'm feeling. That's why I think 4.6 took a I could be totally off with that, but, you know, I haven't read anything that definitively I trust that says that. Craig (05:40) Yeah. Yeah, Jack, you look like you're Jack BeVier (05:50) No, I just, I can't tell the difference between, um, like I, I've had a frustrating past week and a half building stuff. Like I feel like I, uh, set my sights maybe a little bit too ambitious Lee and, you know, I had success building a couple apps and then I was like, I'm going to go build a castle and I built like, and, I got, I got like 70 % done. And then I, Switched models to 4.7, but I also like just like the as I added features and re-architected some things to accommodate those features. Like the thing just kind of like did a slow right turn and crashed into the mountain and like it works still, but it's not, it's slow, it's cluegy. I'm getting errors like with every little update I do. Like I can tell I'm getting a lesson in like You don't know how to write software, Jack. Like, don't forget, you don't know how to write software. And like, I've got a lot to learn. It's been a humbling reminder as to how early in this, you know, skill set I am, frankly. So I can't, I have no idea with the difference between 4.6 and 4.7 because I'm pretty sure I'm the result. I'm the one to blame, not the model that I switched over to. So. Craig (06:45) Yeah. David Moses (07:07) I don't know if I agree with that. mean, I think there seems to me to be some dilution going on so they can, you know, so it no longer... when we were using Opus or Sonnet a few months ago and it was absolutely killing it, it was killing it until the session died. And then like, okay, I had to open an entirely new chat and then I had to... Craig (07:08) Dave, was, I go ahead, go ahead. David Moses (07:31) somehow summarize what was in the old chat, which I really can't get out because I can't even ask it to summarize what's in the chat. I mean, that was just a few months ago and they've gotten that fixed, but at what cost? It's gotta be, they gotta be compute starved. That's the only, because I don't know how much it's jacked and how much it is that, you know, some of the answers are just, we're not gonna devote enough compute to this question because we don't think we need to. know, 4.7 is trying to be adaptive the way chat, you know, tries to be efficient. I don't know. I mean, it's, it's, it's, I don't even know if they know because now with Jack BeVier (08:06) I want you to be right. I definitely want you to be right. I want to not suck at this. So. David Moses (08:10) Hey Craig (08:11) Shaq, you, can you, go ahead, go ahead Dave. David Moses (08:14) No, I always, I guess my litmus test is I will take something that I think Claude could do before and I will dump it into chat 5.4 or I'll dump it into Grok or I'll dump it into Gemini. And I know if Claude got it wrong, if Opus got it wrong and one of those other models got it right, you know, It's yeah, very well could be that I've I've framed it for chat as You know, this is screwed up fix it for me, you know versus build it from it's always easier to poke holes and someone else's Work than it is to actually invent it from scratch. But at the same time it just you know if there are other times where Opus will just Kill it and I'll be like, wow, that was that was absolutely incredible, you know Where have you been the last two weeks? Jack BeVier (09:01) So you switching models and the stuff that you're building, you changing up? David Moses (09:05) I'm still, I play with the models, but generally speaking, most of the workflows that I build don't need the models I'm using. So it's really just chasing the most reasonable cost. It didn't matter when I was, it didn't really matter when I was building things that were, like analyze a tax return. Okay, maybe it's hundred pages, maybe it's 200 pages, whatever it is, but it's just, Craig (09:18) Yeah, but- David Moses (09:29) a few times that I'm doing it. But now that I'm literally processing and analyzing every single piece of paper that comes to our office, mean, the tokens actually do start to add up. it is looking for... Craig (09:40) not it's not the maintenance model that I was that that I think Jack was mentioning it's the build model right so once you build ⁓ a project and it's it's it's running on its own I'm okay with using some lesser model to do that work but it's the build right it's the plan and it's the build and I think that's what Jack was running into what Jack what's your what's your take on was it So like, maybe you can just explain quickly what it is that you built initially, and then what you were trying to add on to it. Jack BeVier (10:13) Yeah, I was building a, ⁓ or I built a risk manager, basically risk analyst. And it was taking the, our borrowers, taking information about our borrowers and doing some public records research, also doing an internet search and basically compiling like a risk analysis intelligence report and excellent work product was like phenomenal, like really, really good. And then I wanted to add in like the, I wanted to make it SOC compliant from a security perspective because I wanted to add in some financial information, like, you know, give it access to a location that included financial information. But I wanted it to pull out all the non-financial information and incorporate that into the analysis, but keep a Chinese wall between the financial information to make sure that that no financial information ever got used to search on the web. And then I wanted to layer in like servicing data to like really make it, you know, add some like really private data, you know, that only them and you only the customer and us know about and basically turn it into a, to a risk analyst. And so like, I got two thirds through that and it was excellent, like really, really good. But then as I continued to, but then as I started to like, well, pull in large quantities of data into that analysis and try to structure it or try to try to put it into workflows. I just started getting, I just started getting more and more errors. And I think this is all I I'm not blaming the models at all. Like I think I just my my vision for what I was building migrated about 23 degrees to the right. And that was too much. You know, like if I had now having screwed it up or having built it and screwed it up, I now know what I need to go build in. know, and I bet I, you know, I bet I damn damn near one shot it. But I'm literally going to go start from scratch because I don't think that I think that this, you know, that the context window is full, is full, full, full. So like trying to debug it, trying to go in there and debug it is just going to be working at, frankly, just to me, you know, work more time than it would just be to start from scratch. Cause I got like 65 % of the way there and like eight hours, you know? So like, I bet I'm just going to go, you know, get over this and go back, get back in there. you know, put on a pot of coffee and rebuild the thing but with you know, where I you know, with the ending point of work as the vision. So David Moses (12:35) Where are you building it? Where is it going? I Claude's building it, but where is it building it? Jack BeVier (12:38) It's just building it in a it's building the code base and then we're going to host it on our server. And I'm giving it access to all the different layers of information that company information, just giving it a bunch access to the different databases that that employees can access. Craig (12:52) Jack, it a... So when you go to rebuild it now, how would you do it differently this time? Would it be more structured? there parts that are like, hey, I'm gonna build this one as like a separate skill? Like, how do you do it now in a way that you don't think it will break once you get to all the features that you're thinking? Jack BeVier (12:53) It'll be cool. Yeah, it was, you know, this was all totally foreseeable, right? Like this is completely user error. It was the, know, I was describing all the features that I wanted it to have. We spoke, we talked for a long time for me to like give it all the features and give it the whole vision. And then, and then it built, but it wasn't like a hundred percent. that, that conversation was like 85 % correct. The problem is then at one shot, like built damn near that vision. But then I tried to steer the ship from 85 to a hundred. And the architecture underneath the beginning, like, wouldn't steer, you know, like I built a steel foundation for this vision that was 15 % wrong. And you just can't shift this, you know, the foundation that much to the right was what I found. by the end of it, I was just like, yeah, you're running into blocks about, you know, for, for issues that I really, you know, for, for features that they made part of the foundation that shouldn't have been, that's all. Craig (14:00) All right. All right. David Moses (14:12) I think you just described, I think you just used the perfect analogy for large institutions trying to institute AI in general. It's just the ship is just too big to steer. Jack BeVier (14:18) Yeah. Yeah. Yeah. Yeah. Exactly. Yeah, yeah, I mean, I think you can end and I'll get you know, it'll be better this time. Right. And then maybe and maybe this one will break too, because I get ninety five. But then the process like I could like in retrospect, I'm like, I could have seen all this coming. I just I don't know. I wasn't smart enough. I didn't. And so like but the process of building it and then trying to steer the ship, I now see where I really wanted to go. And now I'm like, ah, I should have just described this. And I bet you if I just described this, then it will get it right the first time. David Moses (14:53) Yeah. Craig (14:54) So when so so I think I think David and I have a some concept of how you would describe it. But tell those who are maybe haven't worked as much with these with these tools as us. Are you going to talk to it any more structured way this time? Will you will you give it a prompt and say like hey you're just building a skill at this point you're not I'm not asking you to like. I'm not going to talk to you for 20 minutes and then ask you to go build off of that prompt. Like that's where I think that's where I think you run into a lot. Like I'm not going to talk for 15 minutes in a prompt and then expect all of that context to live in your build context as well. It's just too much. Jack BeVier (15:31) Yes, so I think that like, if my previous use cases, I had done those things before, right? Like, and so I knew how to do that job. And so I could describe perfectly how to do that job. And the AI did a fantastic job of building the thing that I knew how to do because I'd done it before. The the the risk analysis, the risk analyst project was something that didn't exist before. It was based off of it was an idea. Craig (15:49) Right. Jack BeVier (15:58) born out of the idea that, I've never been able to connect these databases before in a practical way. It costs too many human hours to do this level of analysis. And so no one's ever done that here before. But if someone had my god, we'd be it'd be such a better work. It'd be such a better stream. It'd be so additive to the company. So like I'm literally describing the risk analysts job for the first time. And hey, guess what? I got his job description wrong. Right. Like He should open this database, then that database, then that database. And if you can't find somebody in that second database, he needs to go to this other one. And just the flow of information, the flow of searching for information, I got wrong. but then when you but it built it that way, right? It built it in this order and it aggregated data here and then it added features at this layer. And then it went to the next step. And then I realized, ⁓ I got the steps out of order. I should have put this one over here and then the thing started to break. Just because I got things that I built it out of order. I started hanging drywall before the electrical rough-in was done. That's what I did. David Moses (17:00) That is, I mean, it's so impactful what you're saying. So I'm thinking about as we do, I build workflows in N8n. Why? Because I can visualize it. I can see what's going in and where it's coming from. And I can see where the errors are. And I can see where, you know, context just isn't making it to the agent. But I had this workflow that was working well for creating, you know, these, phone call recaps, right? So it's like a, it's pretty simple. It's an email that goes after every phone call to the person in the company that was on the call. There's an internal recap and because it knows who, what, where, when, why, how, it gives a dynamite recap. Then it has like a recap. You can just click a button, like an external facing recap. can click a button and send it to the person you were talking to. It'll add to do's. It'll suggest to do's that you can just click and add and it'll. add appointments, right? And it was working, right? It was working well. But I wanted to make it better. And I also wanted to make it truly agentic, where it's not just here I'm going to feed you all this information from all of these different databases, all these different collections of data, and then have you eat up a giant amount of tokens, many of which you don't even need to use, and then reason through it and then spit out an answer, I'm going to just tell, I'm just going to build the workflow from scratch and I'm going to tell it, you know, this, here are the different data repositories you can look in and here's the result I'm trying to get and let it figure out which repository to go to first, which one to go through second, which one to go to third and just let it kind of figure it out. So, and it would reason through it and it failed miserably and I'm saying it failed miserably not because it did it in the wrong order, not because it just forgot to look in certain databases, and not because it took the data out and didn't use it correctly. It's all of those things. It did all of those things and not consistently. Sometimes it did, sometimes it didn't. And what you're describing is like, for me, the most difficult part of building is deciding, do I want to build it so it just fricking works now? Jack BeVier (18:37) Mm-hmm. David Moses (19:03) Or do I want to build it so maybe it's cutting edge, maybe it's not going to work right now, but as the models get smarter and as I get better at prompting it and telling it, then maybe it'll be better. Because it solves the problem of where to look first, where to look second, because now an agent's deciding where to You haven't built such a fixed structure that now you have to... completely change everything around in order to get it to change the order in which it ingests and analyzes the data. Just let it decide how to get it. And hopefully, you can add some recursion in there that will allow it to, over time, continue to improve itself or, as the models improve, whatever, whatever. So that's my dilemma. I... I don't know the answer. I can tell people, go build something. Just do it. Jack BeVier (19:56) You know what, I probably, you know, probably like, think the N8N approach like is probably a best practice from the perspective of like what you just said, right? That you can visualize the process. You know, it got the information flow correct. Whereas when I go in and I spend two hours describing my dream software and then it one shots it, that's amazing. I get a huge dopamine rush out of that, right? It's like, my God, like 30 minutes later, this thing's like, built, like that was so fun. The problem is there's complete opacity as to what exactly it built there. like, yeah, exactly. Yeah. And so I don't find out until like four days later that like some of the fundamental architecture was built in a way where it's slightly misunderstood what I was going for. And because I got no visual representation of what's going on in that code. And like you said, yeah, and I'm not Neo, like I just can't do that. Like I probably, maybe, maybe the thing is to like, David Moses (20:29) And you're not Neo. Craig (20:29) Yeah. Jack BeVier (20:51) Hey, while you're building, put in a screen over here, a visual representation of the data flow of what you are building so that I can catch you up the information flow in real time. Because it got too big. By the time I realized that there was a freaking, the foundation was built like, had some screwy thing in it. like trying to go back in there and rip out that piece of the foundation or replace it, it was just like functionally impossible. like the, I love the dopamine rush of the one shot, but the opacity is a problem from an enterprise software perspective. Craig (21:32) Yeah, agreed. Agreed. David Moses (21:33) Yeah, and if you're not a programmer, you're not a programmer, you know? And you don't want to bring a programmer in, right? Because their instinct is going to be to go to that code line by line and fix it. And that's not what you want. Jack BeVier (21:36) Yeah. Yeah. Yeah. And I could just tell as I was like making corrections, it's just like writing 30 new lines of code and like literally like, you know, going in and finding in the workflow, like, you know, carriage return, carriage return, carriage return, Like, hey, does it work now? And I'm like, yeah. And, you know, 12 of those in, I'm like, this is I built a spaghetti bowl. Like this is this is garbage. Like this is no longer enterprise grade. So. Craig (22:10) asked. Jack BeVier (22:12) Try again, try, try again. David Moses (22:13) That's right. Craig (22:14) that is ⁓ a that's that's that's a tough one to come to been there know exactly where you are there. Not a good feeling. Jack BeVier (22:20) It's cool. It's, it's worth it. I'll redo it. It's worth it. Craig (22:23) It's all learning experience. David, ⁓ I understand that you're that you might be freelancing now creating creating AI solutions for local banks. So David Moses (22:33) Well, that's kind of a premature way to say it, a bank that I do a lot of business with called me in just to kind meet and talk to one of their team members who is on this AI team. So the owner of the bank really believes in AI and they don't want to be the dinosaur bank that's like the last one to the party. and they're I mean, I love that they're that they're looking at it, right that they're that they're thinking about it. know that so far they're only allowed to use copilot, ⁓ which you know, somewhat of a closed system, you know, and you could tell, you know, he was, he knows there's better out there, right? no, no, no, it would be like the opposite closed claw. Craig (23:11) Not exactly open claw. David Moses (23:16) ⁓ But it is interesting. And they asked me at the end, he said, you know, would you come in and maybe freelance and help us with something if we needed down the road? I said, I think that would be really cool to think about it from that perspective. Because I'm a business owner and when I want to make a mistake or not when I want to make a mistake, when I want to do something that has a high risk of failure, that's a risk I can take, right? But, you know, obviously a bank is not going to take that risk. They can't, right? So this is interesting to see that dichotomy. But I applaud them for, you know, certainly none of my other banks that I've spoken to are... that's not anywhere near on their radar to actually use AI to handle some of their... Craig (24:00) you out. David Moses (24:03) thoughtful processes. Craig (24:05) Yeah, I was just wondering like, give us, you know, without being too, too specific, but like, just a couple of use cases that you could see for them immediately. David Moses (24:13) ⁓ so for, like my first instinct for any business with AI is to use it to find your bottlenecks. Like before you do anything else, first use it to try to figure out what is slowing you down, what is holding you back. ⁓ cause AI does a really good job of analyzing large amounts of data. and it can, it does a good job of predicting and it has a lot of general knowledge about every industry. And it can, even without feeding it really good data, it can really well predict based on your industry and how you're set up and where you want to start. That's the first place a small business should start. First, let me get the things that I'm paying somebody in Pakistan. six dollars an hour, nothing against Pakistan or any of those places, they're, you know, essentially inexpensive labor where they're doing repetitive, you know, I guess low value tasks. And you look internally and you say, OK, well, now look at my look at my internal staff. Like what is eating up their time? I think that's the first place to look. And I think a bank is a great example because a bank and a lending company, to Jack, to both of you, you're the bottleneck between them and whatever their goal is. They're not, you know, they're a real estate investor or they're whatever it is that they're doing, they're going to the bank to help them, to help leverage, to get them to their goal. So if a bank can speed up that process, if a bank can eliminate you know, time overhead in accomplishing those tasks, or a lender, you know, can do that. It's going to drastically improve the customer experience. So I think it's looking where those bottlenecks are and trying to make things faster. If you're going to give somebody a bad news, give it to them very quickly. And if you're going give them, you know, if you're going to give them good news, then you want to deliver on that in a very predictable timeframe that gives them the best advantage to have an advantage over everybody else who is, you know, demanding that money to accomplish whatever goal they're trying to accomplish. Craig (26:10) Yeah. Jack. David Moses (26:11) That's my feeling. I don't, you but yeah, I think you first look at what's slowing you down, remove the obstacles. AI does a really good job of that. Jack BeVier (26:20) So what are you showing them? Like you showing them like, here's my download claude desktop or like here's my end flows. David Moses (26:26) I showed them a couple of the flows that I have built around document processing. seemed like something that would speak to them. I showed them a little bit about our communication. They're really big on having great service. They want to have a small bank feel, but they want to grow. They want to be a larger bank. So, but they want to maintain that kind of personalized quick response. So I showed them some of the things that we've been working on to, you know, analyze the calls that we're on, to summarize them, to recap them, to add to dues, to, you know, remind us where we've left off on, you know, and haven't gotten to somebody fast enough. That's the kind of thing. Those are two things I showed them and they showed me. Kind of document analysis is really what they're, but really more one-off, like put this document into here, what does it say? And trying to get it to a, you know, to respond consistently. I don't think any chat bot is good with that. I think you need to build some kind of infrastructure for it. I kind of mentioned that to them. Just having it be, you know, you can set up a project and pick your favorite LLM and you can say every time I put this document, which is of this structure, into this chat bot, I want you to analyze it in this way and spit back out to me a document that looks like this in this format. That is fundamentally a, you know, that's a programming thing, right? That's really a programming thing. should be, you know, to have that kind of thing in an LLM project is, I think, know, right for error. know, it analyzed an appraisal and it said, hey, know, this, ⁓ the legal description is correct. And I'm like, come on. There's no, I mean, no offense to copilot, but it's just telling you it's correct. There's no way it knows. It's not connected to any public record database. Certainly not, you know, ⁓ a public record database in Oakland County, Wayne County, Michigan. Like it's not, there's no way. Craig (28:23) I'm just... You're so cynical, Moses. You're so cynical. David Moses (28:28) I, I, I, you know, it's just, I, this is just a hallucination. There's no way this is, you know, there's no way it knows that like if it's, it's probably accurate, right? Because probably the appraiser has software that you put in the address and it pulls the legal from a real public record database. So it's probably accurate, but, at the same time, it's, ⁓ you know, Copilot doesn't know that, you know, and I don't even think if you put it in Opus, Craig (28:32) It's true. 100%. David Moses (28:52) It might not even have the chutzpah to say, yeah, I don't know that answer. I have no idea if this is, there's no way I know. But it's much more likely, Opus is much more likely to say, I don't have access to a database that can confirm that this is an accurately. Jack BeVier (29:08) How many people, this is a related question, promise I'm gonna circle back and make this make sense. How many people are at your company? David Moses (29:11) Okay. Uh, 20 or we're, we're at 28, think we were, we were 35, so we're down to 28. Jack BeVier (29:18) many How many people are building AI tools? Three. David Moses (29:24) Hopefully by the end of next week four. Jack BeVier (29:26) and how long How long do you think it's going to take you guys scope creep aside? Just, you know, put a, what, put a fence around the company right now to AI everything you can, not every last little thing, but like, you know, to get 90 % of your process is AI. How long is it going to take? David Moses (29:43) That's a phenomenal question. To get it done right, think, I would say, know, what's our target is that everybody has the largest part, the largest process in their department, their scope, right, has that automated and agentic by the end of the year. That may not be a very ambitious goal because the largest process or the most time consuming process, you know. may not be appropriate for, there's eight months left in the year, right? That's probably the target. Jack BeVier (30:19) Do you think that 10 to 15, is that, sorry, you know what? Is that fast enough for you? Like is that? David Moses (30:26) Well, you know, when I set the objective at beginning of the year, seemed right. I didn't want to scare everybody, you know, and it seemed right. think it, you know, but I'm an entrepreneur. Like all of this stuff should happen tomorrow. And there's really no reason why it hasn't happened yet. Other than the fact that I haven't done it yet. entering, know, inserting some reality into it. Everybody has to kind of, they have to learn it, they have to go through it, they have to fail forward. Yeah, they're starting from a, they're never gonna have the same frustrations that we have now because they're gonna be starting with a much smarter base. no, it's not gonna be fast enough. I think that everything agentic by the end of the year is realistic. I think five years, if it's not, we're not here. Thanks. This is real. Jack BeVier (31:16) Do you think that 10 to 15, you described 10 to 15 % of employees are spending some significant portion of their time thinking and building and retooling with an agent goal. Do you perceive that as the bottleneck? Is that enough for you, 10, 15 % of the team working on that and everybody else is holding down the fort as you guys introduce new tools? Is that the right? What's the right as a business owner, what's the right percentage of change that is fast, is, you know, that is practical, that is manageable, right? Like, cause you can't just like flip the switch tomorrow, even if you, it was all built, you, you literally couldn't flip the switch tomorrow. So like, what's the change management philosophy or what's the change management approach for I don't know, maybe it's different between small, medium and large sized businesses, but like, know, how should, how should different size businesses be thinking about how much horsepower they need to be throwing at this and how and what the timeline needs to look like to be moving fast enough to stay ahead of the curve. but also not just like drowning yourself in like capex, right? Like there's a balancing act there, right? Like what's the, what's the, what's the optimal median, you know, David Moses (32:27) There definitely is. I think it really is, I don't know what the percentage of people is, but I think that if you get the right people, the real leaders, the people who other people look at and say, well, they're doing it, so either I better or I can, whatever it is that, I think if get the right people doing it, think 100%. buy-in is required to be a real viable company in five plus years. I think 100 % buy-in. Whether they all need to be building things, maybe not, but they have to understand how it works. maybe they're just envisioning what we can do. Like there's two camps, right? There's the, can we do that we do now? significantly faster and cheaper so that our lunch doesn't get taken. And then the opposite side of that is what could we not do before without these tools that we can do now and eat everybody else's lunch. That I think are, those are the two, and I think every one of the company needs to fall into one of those camps. You're either helping us push forward with things that just weren't possible before. or you're helping us get rid of antiquated, slow, and expensive human processes? Is that a good answer? Okay. Jack BeVier (33:51) Yep. Yes. Do you? OK, so now this is my circle back. Do you think that you're wasting your you don't but you don't think you're wasting your time talking to this bank? You think the bank is you think the bank is capable? Knowing everything that you know about banks and you know a lot, you know, I think more than just about anybody I know about how banks actually, you know, go about. David Moses (34:11) dysfunction. I think that if there is a bank that's going to succeed, the type of bank would fit this profile. It's a bank that they bought. They bought a hundred year old bank that was maybe a 20 or 30 million dollar bank. They've, you know, more than 10x that in six years. That's the kind of, you know, that's the kind of CEO, that's the kind of like Jack BeVier (34:17) Mm-hmm. Mid-size. David Moses (34:34) They're not like, ⁓ I'm fifth third bank. What do we do today? Let's go merge with Comerica. That's not going to change how banking works. They're not gonna be the ones that make that happen. I think it's gonna be young, smaller entrepreneurial banks that get it done. And... You know, I think as long as they can, you know, I think they'll, they'll, you know, they'll do it thoughtfully. But who are they up against? Right. They're up against regulators. Regulators are, you know, the biggest pain in their asses. Right. Actually, that's probably a great use case is just, you know, here's, here's, here's what I am and what I do. Paint this picture in a way the regulators will, you know, Jack BeVier (35:15) Yeah, get what? I can get away with it, yeah. David Moses (35:16) damn, you know, just, yeah, nothing, there's clearly no risk here. They really know what they're doing. We don't even need to be here. We should, we should go on and move on. You know, that's really, I think a great, you know, cause, cause AI does a great job of taking, you know, there's so many ways, there's so many ways to take the facts and use them to tell a story, you know, and I think the same song and dance we do or borrowers do when they go to a bank. It's the same song and dance the banks do when they go to the regulators. It's like, here's why this was a good decision. Here's why this is not risky. Here's why. Why, why, why? And I think AI is going to be a great tool for that. But wherever it shakes out, I think there will be banks in 10 years. I think there will be banks in 20 years. I'm not one of those people who thinks that banking as an industry is completely dead and there's going to be some way in the future that we will completely decouple this massive beastly process between capital and those who need capital. The medium? Maybe medium. Craig (36:24) depends on the depends on the type of capital. Jack BeVier (36:24) I hear you. Craig (36:26) So, Jack, what so the question I had is I was listening to all of that as is, you know, you're talking about giving access to claude code claude desktop to most of the people in the company. And there has to be a part of you that says to yourself, I saw how much Craig fear was banging on that thing for you know, for the last six months, how much work was he getting done? And, and like, I love the idea of each stakeholder in the company, saying, I mean, I just I got a use case here that I think was is brilliant, that would save hours ⁓ off of my weekly toils. Like, how do you Jack, how do you think about that as as the guy who owns a company like still got to keep our foot on the pedal with sales and operations and all of those things. But if I've got Half of my staff now that's playing with AI all day, how productive are we? Jack BeVier (37:17) I think that my token bill is my favorite money to spend. ⁓ Craig (37:22) Jack, Jack coming from the Jensen Huang standpoint. Yes. Okay, I pay my guys 500 grand if they're not spending $150,000 a year in tokens, I'm going to I'm going to sit them down and find out why not. David Moses (37:25) That's right, that's right. He's got the token leaderboard. Jack BeVier (37:27) Yeah, yeah. Yeah, well, there's I mean, there's some combination, though, of like, like David, to your point, like everybody needs to be using it. Everyone needs to know how it works. Everyone needs to be get comfortable with being able to call like in spot a hallucination, because that's not how that works. Like that's that's Like everyone needs to learn how to use these new tools and figure out what they're they're good for and what they're not good for. And you're just you're only going to do that banging on banging on tokens, you know. David Moses (37:52) you Jack BeVier (38:03) And so I'm either paying for their learning curve, which is cheap, you know, man, by the time you spend a thousand dollars in tokens, you probably know something like that. That's a lot. You know, that's a, that's a lot of banging. and then, but then I, then you have these edge case outliers where a manager creates a software that I would have happily paid $50,000 a year for. And I probably have like four of them so far. So from a dude, from a value perspective, Craig (38:24) Yeah. Jack BeVier (38:30) I haven't spent, you can't, know, the, got four piece, you know, maybe, maybe we've got three or four pieces of enterprise software that I would have happily paid 50 grand to pop for that's $200,000 a year, capitalize that at any reasonable level. I haven't spent a 20th of that number in tokens yet. So like, to me, this is just like the biggest ROI thing I could possibly do. Like we're, we're in the middle of it. Still. It hasn't like the snowball is rolling. It hasn't caused me to fully. take off screaming away from my competitors yet. But I don't know, but I see that kind of value being created. know, the and that, know, we were talking about last week, and revenue is moving up a little bit and expenses are moving down a little bit, but that's not really the needle mover. A lot of it's like the, but the business is stronger because I have these four pieces of software so far that like I didn't have before. So I know I just sleep better knowing that there's a better business. ⁓ waking, waiting for me when I wake up. And so from an enterprise value perspective, I love paying for tokens. I think you have to take a little bit longer view of it. And if you look at the month over month spend and say like, Hey, what have you done for me this in the past 30 days? Like I need an output for the, know, for the, you know, for your $700 bill this month, that's a little bit too draconian. And I think it's a little bit too short sighted, but also, Hey, it's been 90 days. push some code, know, like, you know, like, do you know, show me show me something, you know, David Moses (39:52) Yeah. Craig (39:55) To the larger point, 200 people work for Dominion Financial Services, all 200 get access. Like, again, I can't have an entire department getting so excited about AI that they're addicted to playing with AI all day. know, so like, don't know how you start to put guardrails around that. Jack BeVier (40:13) yeah, you still gotta do work. David Moses (40:15) Let's just, two points that Jack, you know, to what you just said, looking backwards on the token spend and comparing it to what you would have spent if you had paid some enterprise or, you know, you know, hired a programmer to actually program it manually. I think that's one way of looking at it. But another way to look at it is all of that token spend, like This is as dumb and as expensive as these tools will ever be. Right? Like, will never, you know, with the midweek or midmonth taking that Opus 4.6 data aside, generally it is going to be... ⁓ But it's still, ⁓ this is as dumb, and so it's like you're learning this now. Jack BeVier (40:38) Mm-hmm. Mm-hmm. Mm-hmm. Craig (40:42) Right. Still so angry. He's still so angry about it, Jack. David Moses (41:01) and these things may cost a fraction of what they cost in the future. How does that work on the enterprise software end? Is that how it goes? Like, do you pay for an enterprise software and buy into it and then it just gets cheaper later? Because that's not been my experience with paying for enterprise software. It just keeps getting more expensive and they just keep finding new ways to deliver me things that they are selling me that I need. And, you know, Now I can say it's not just do I need that thing or do I not need that thing. It's do I do I want this thing and if so should I pay them for it or should I build it myself. Jack BeVier (41:39) Yeah, yeah, 1000%. Yeah, yeah, totally agree. David Moses (41:41) And you need the people like that. If I were a small company, I am a small company, but if I were like giving other companies advice, I would say find everyone in your company that knows how to play a musical instrument. I would say, because that that same part of the brain that will like, you know, have you ever written a piece of music before? Like to me, that is I'm not a programmer, right? But I have spent, I have sat down for an hour and written a great song and I have sat for four days and have nothing to show for it. But in both senses, you have the you you have the gall to keep going because you just love it. And I think there is something about, you know, you know, I, that's, I think that's a, I would love to do a study on that and actually figure out like how many people who are in. Craig (42:31) would too. Jack BeVier (42:32) The correlation. Craig (42:32) I would too. think that's brilliant. I played an instrument for a very brief time, but I guess singing doesn't really count as playing an interest instrument. So but I have written music and it is ⁓ I totally get what you're saying there. I said to you, I think it was at the Christmas party, Jack, we're like, I think there's just going to be like, there's going to be that group of people that like, like Dave said, they have that creative curiosity. They see they see what you know, how cool it is to be able to think something, speak some English and build it. And then I think there'll be the people that just use it, right? And from them, hopefully you get, well, you've been using this great new tool that we've built for you. How would you make it different? And I think ⁓ empowering them to make it different at that point, to either do it themselves or then just come to the AI team to have it done is where you'll get a lot of use out of the extra probably hundred and 75 people in the company who don't really want to sit and bang with tokens all day, you know, like, and and I think, Dave, we last night, I don't know if you met john McClellan before john's our, our resident AI God. And he built something last night that I thought was just the coolest thing. And we've gotten to a point as a company now where we have a whole bunch of people making really cool stuff. And you have to believe that a lot of stuff has been rebuilt and rebuilt and rebuilt because I need to I need to connect to the same super base database. I need to connect to Salesforce. I need to all of those things. And what John did last night was basically put together. a really cool connector, if you will, to all of those things that we would want to connect to here at Dominion. it just and and then programmed in some given all of that access, right to those to the to the tools that we need to do our jobs. What could you what could you what what reports what what functionality could you give me at that point if I was Jack, you'll have to forgive me. I don't know Andy's position in the company. But he built it for a general one of our managing directors in the company. And it is brilliant. showed it off on our call this morning. And I and he called me afterwards. And he was like, How'd you like to play with this real quick? And I said, Great. And he's like, Here are the files, put it on your hard drive, go into clog code and just hit go. Dude. Jack BeVier (44:49) Yeah, it's like the idea is like it's like a like a company starter kit, right? It's like so like you've got it's it's hey, it's it's got all the the connections, the API connections. It's been told in MD files how like what what this basic you know what the basic processes are, how these databases work together with each other. And so and and then he built a you know. Craig (44:53) Yes. Jack BeVier (45:12) All the, all the also, you know, plugs in all to the, all the external services, like not only the internal databases, but the external services that we use on a regular basis. And he just put it all into a starter kit where you can just drag it and then point it, you know, point Claude to that. And now it knows a ton. It's like a really solid starting point probably saves you the first five days, right? Like the first five days are just like setting up your APIs and like putting your secret token, you know, putting your, your keys into your ENV file. Craig (45:34) yeah. David Moses (45:39) That's right. Jack BeVier (45:41) And he just. Yeah. Craig (45:42) setting up a file structure, like, you know, like all of those things. This is, hey, here's, here's a group of files and it all, by the way, it all runs locally. So you can have your own access and your own, you can make your own changes. And, ⁓ right. And so that's what I did. And, ⁓ I think within a matter of, Jack, I'm, I think you give me, give me like a week or two to really work with John on building out like the loan officer sort of portal. Jack BeVier (45:53) Yeah, and then you can fork from there. Craig (46:08) that I think guys would want like access. Like this is all just everything's right at your fingertips now. Like I don't have to open Salesforce. I don't have to open RingCentral. I don't have to open PhoneBurner or any of the other tools that we use. So ⁓ it was really interesting. And the reason I brought it up, Dave, was, you you mentioned that you're going to be giving these tools to folks in your company. And I just thought, man, what a great way to to get people introduced. and not have to have them go through all the minutiae of, you know, setting it up and getting all your connections and all of those things. You're just off and running to telling the thing, hey, based on everything that you know about me and my job, how would you change this? And that's what I did. And it went to work. David Moses (46:49) And that's an interesting... So you load this on your computer, right? You load these files and... Jack BeVier (46:54) Well, and by the way, my argument is like, why don't we just host it? Like we'll just host it. And as we get new services, do new subscriptions, we'll just update the MD files up there. And then we'll just push. We'll just be like, hey everybody, there's an update to the thing. Pull from the repo and pull this new code in. And then all of sudden this new feature set shows up on the sidebar. It's like, sweet, I got a new set of tools that I can start playing with. And we didn't have to. individually person by person by person by person by person. You know, hey, can you send me an API key? Hey, what's the you know, like, you know, where's that API creds link again, like you don't have to just save all of that, you know, it's infrastructure, right? Like it's it's company infrastructure hosted somewhere at make changes to the infrastructure, push it down, and then let everybody customize that base core set of infrastructure for their particular for their particular use case. ⁓ David Moses (47:47) But you need the recursion. You need that somebody's building something, they run into an issue, right? Or somebody discovers something that is an easier way to do this. You you want that feedback, right? Like that's you, that's where, that's how ideally those MD files get changed and updated. Jack BeVier (47:48) And then it. Craig (48:07) So John, so let's say Jim over in processing is playing with these tools. He comes up with something great, some new tool. You're saying, how do we get John's up in the end of production, right? What's the feedback loop on that? Or if John found something that's broken or is that what you're saying? Jack BeVier (48:18) has to be a feedback loop. David Moses (48:26) Yeah, I mean, whether it's a bug or a feature, you know, if they have access to these tools. You know, just having, like Jack, you made the point when you specifically said, I knew how to do this, right? I've done this before. So when I told it how to build this, it was coming from the experience of someone who's done this before. And when you extrapolate that to your team, you're inviting the expertise of people who not just have done it before, but are currently doing it every single day. I think that is, you know, and then you'll find what I found in my company is you'll have, and not from giving them access to a God-like tool like that, but for just from giving them access to new tools that I've built, you'll have the people who will just point out where it's broken, right? And just say, well, this doesn't work. Yeah. And then you'll have the people who, you know, who say like, what if it could do this? Like, what if it could do that? And like both are valuable. Jack BeVier (49:12) Right. it doesn't work. It's broken. Yeah. David Moses (49:22) Right? But one side of that coin, like you won't even need that in six months because even if it can't one shot all these ideas, right, it can certainly test and find the bugs in them looking record by record and check and find everywhere where it made a mistake. You don't need that. You won't need that six months from now. The person who says Jack BeVier (49:29) Mm-hmm. David Moses (49:44) Well, this doesn't work. This is broken. This didn't work. You know, you will that that person who says, well, what if it did this? You know, what if we what if we took this that makes my job so much easier and made a customer facing, make it so that they can do my job for me and make it so they can play with it and have access to things that. Jack BeVier (49:59) Mm-hmm. David Moses (50:05) that maybe they don't have access to at all and it makes their lives better and makes our relationship stronger, whatever it is. Jack BeVier (50:10) Yeah, and to piggyback on that point, so what I'm hoping we'll see literally last night, like John came up with this literally two nights ago. And so like what I'm hoping is that it's the starter kit that gets people where they can they can literally just like load the file, point it and they immediately can start playing. They don't need to do any of the infrastructure work. None of this not sexy stuff. They can go straight to like. sexy, productive, talk about my day stuff. And then, and if that's the case, does it decrease the barriers of entry to the point where I'll use this use case, we were talking about it today on a, on a manager's meeting. In our loan servicing department, Lily handles your extensions. Extensions are really annoying because in order to grant an extension, we don't want to do that just de facto. So we want to see like, Hey, when was the loan originally mature originally originated? How many extensions have been, if any, granted so far? What's the re-underwriting on this? Are they current on their payments? Have they paid their taxes and insurance? There's a bunch of check marks that we want to do and make sure that everything looks good before we allow someone to extend the maturity of the loan because, we collect an extension fee maybe, but we're also taking additional market risk and it's an opportunity to reassess the collateral at that point. So Lily's got to go check around in seven different systems at seven different places. But you know, all this, you know, 12 different pieces of data before she can send them an email and say, Hey, here's your opportunity for an extension. And here's how much it's going to be for six months, four months, whatever it is. And so it's one of those painful jobs that like, there's no extension module, right? That like aggregates all this stuff. And we actually have to go even outside of the Salesforce system. So what you'd really like to do is. David Moses (51:51) It's been a really, really long time. Sorry. Jack BeVier (51:56) have Lily build a custom app for herself that pulls in all the things and the thing is no one is going to build that app like Lily is. She's going to build it perfect the first time. I'm going to it up for four rounds before I finally get it right. But she'd get it right the first time because she knows how to do her job and I don't. And so like can we bring it? Yeah, exactly. And she and she needs to. Craig (52:04) Yeah. David Moses (52:12) That's right. She just has to know what's possible, right? She just has to know, you know, and she has to know that by replacing Craig (52:14) That's so cool. David Moses (52:22) herself, she's not going to lose her job. Jack BeVier (52:23) Yeah, but if I go ask her to go like pull in the API keys and figure out like figure out if she has to go build her own infrastructure. Yeah, if she has to go build her own infrastructure, she might not. Right. But if I don't but if I don't have to ask her to build her infrastructure and she gets to go straight to the sexy stuff of like making her day better, I think I got a pretty damn good shot at conversion, you know, at making that happen. So that's why I'm real excited about that idea. It was real like Craig (52:31) Salesforce mapping. Jack BeVier (52:51) brilliant piece of vision from John, I thought. Craig (52:54) was very cool. I've always. David Moses (52:54) Yeah, it's, I was talking John about that. Can we have John on here to, or is it? Jack BeVier (52:59) For sure, for sure. We don't let him out of the basement. Yeah. No, for sure, for sure. He's been working on this stuff for 18 months now. Craig (52:59) No, he's way too busy. Yeah, just the I David Moses (53:02) started drinking. Craig (53:08) still maintain that within 18 months, we will all just have like a thin OS that overlays all of these horrible bloated legacy apps that make us go searching all day long through that we that we should have at our fingertips. Like I think That's exactly that's what I'm building. And I think that most people are going I think that's what claude desktop is all about. That's what codex desktop is all about. It's all going to just be a thin layer that puts all of everything at your fingertips for you. David Moses (53:41) Have you built an agent in Claude yet? Like a Claude agent in the agent builder? Okay, so they just came out that they're adding memory to it and that memory can... I mean, I think this email came out like two hours ago, but they're adding memory to it and the memory is accessible from the API. ⁓ Craig (53:45) yeah. Yeah. David Moses (54:00) So I wonder how, you know, because I want to be somewhat platform agnostic, but if Claude keeps doing this, it's like, why? know, what's the, if they're always going to be the best, then what's the difference? But it's really super interesting to me, the idea that you can have, you know, this accessible memory from every, so, Pardon me once you just say, everybody gets a Claude agent, build a Claude agent for everybody, connect Claude to the APIs, the MCP servers, and have some kind of shared memory that way. And I don't know. I I literally read the email on my way back to the office to jump on the call with you guys. that to me is a lot more brilliant than building any agent is the idea that you can access the memory from the... ⁓ Craig (54:51) from David Moses (54:51) You're an open-claw, Craig (54:51) the agent. David Moses (54:52) but I mean... Craig (54:52) Yeah, I was thinking when you mentioned it that it was going to be global memory. sort of like a account level memory, not an agentic level memory is what you're suggesting the email said or David Moses (55:06) I, it, to me, it is, what I'm guessing it is, it's session memory that is, you know, probably some combination of a vector database and relational database that we, you know, Craig (55:12) Yeah. David Moses (55:21) that we can access in some meaningful way. don't know how, I don't know. Like I said, I haven't played with it. I literally just read the email and it just felt super interesting to me. Cause I built a claude agent, but I haven't actually used it yet because I'm not done connecting it to everything. So it's still not as fast as just going into claude desktop and just talking. Craig (55:41) So maybe I maybe I misunderstood you and I'd love to get into it on the next episode where I'm I'm Yeah, so building a club so we've all built Claude agents. But you're saying build a Claude agent that would do what connect to everything. David Moses (55:48) Nice hook. Thank Yeah, that would be connected to everything. So it's the same agent and everyone has a session, right? Everybody has a session or everybody has their own agent and the agent and then they can build as many sessions as they want. But at some point you can extract that memory and you can aggregate it in some meaningful way so that you have a consistently updated shared memory. You don't have this. everything you did over there and everything you learned, and let's have a meeting and talk about it. It's like everything you learned, my agent already knows. Craig (56:28) That would be pretty bad ass. If every session that I ran in Claude had memory and then it was like almost a shared memory vector, vector search. ⁓ yeah, that'd be pretty sick. David Moses (56:38) Yeah, mean, we're all kind of, you know, learning from each other, you know, in that sense, or our agents are. Craig (56:46) Jack, any last thoughts, Jack BeVier (56:47) You're like, if you could, if you could like the MD file, right? Like has become really important, you know, like that that reference and like if you could best practices, right? Like, hey, here's all the things that we've because because they change over time, right? Like everyone. who owns a company gets in and things get sloppy, then they go to some seminar and they get told to they need to write down their systems and processes and document everything. And if they haven't documented it, doesn't exist. And we all go we do that. And we spend four months doing it. We've got systems and processes and they're documented. And then a year later, that stale and is no longer. Yeah. Yeah, that stale. No one's looked at it. No one's updated it. It's garbage. Yeah. And like and it's so you know, and David Moses (57:21) A year. I think a month later. Craig (57:27) probably wasn't even that good to begin with. David Moses (57:30) That tongue deadline and then one day before that deadline, it was created. Jack BeVier (57:32) And you hand it to. Yeah, and you use it as the training materials for the next two years and you train people on a bunch of ways we did it two years ago, right? It's like it's a terribly flawed, horrible system, right? So like like an MD file with SOPs, though, that an agent is looking at has a lot of value if and then if you added like this, if you added a recursive, like recursively self-improving or recursively updating MD file that always is like And we learned this thing the other day. so here's the things that were, you the MD, this is, you know, we've seen this pattern happen four times. Turns out this company that we, this counterparty wants to work in a different way. And so we've updated the SOPs for how we're working with that, because we have been working differently with them and something that could recognize changes of systems and processes, and then updated the new MDs with that so that all the other agents were like, always had a real time set of systems and processes that that would be a sexy. little system, right? David Moses (58:29) It would be, and yeah, it seems like it's something that can grow going forward. Like that's a sustainable system because yeah, you're accumulating all this memory, but also as you're accumulating it, the models are getting better and the context windows are getting larger. And so the more memory you have, the more over time, the more memory Craig (58:29) very sick. David Moses (58:52) it can handle. And that to me is a, that's something that you can build that actually, can confidently say that kind of architecture will work into the future, Like significantly into the future. Jack BeVier (59:05) And I feel like if it's just an aggregation of every conversation that was had, and you end up with just terabytes and terabytes and terabytes of history, that's not that useful. But if it's a distillation of best practices, a distillation, you need the distillation process to turn all that stuff that happened, all those chats, all those conversations, all those emails back and forth into, know, did the... Did this comply with the SOPs or did it not comply with the SOPs? If it didn't comply with the SOPs, someone tell the manager that we're not operating based off of the SOPs. And maybe that just means that we need to update the SOPs with the new best practice. ⁓ Would be a nice. Craig (59:47) Speaking deep, David Moses (59:47) But it gets that from sentiment, right? Like you're using the agent, you get frustrated, right? It somehow figures out a resolution so you're no longer frustrated. And that nugget is gold. And everybody can share in that knowledge now instead of just, yeah, I ran into that problem four weeks ago, here's how I solved it. Or maybe you've been telling myself, yeah, I know I ran into that problem sometime in the past. Craig (59:48) sort of. Jack BeVier (59:52) That would be a great way to do it, yeah. David Moses (1:00:14) What did I do to fix it? Now it's just, it's there because you're being, you you're just expressing your frustration, your displeasure, your pleasure with whatever it is that it did. And it's, you know, it's using that to build, you know, to build the database. Craig (1:00:30) I think we're 18 months away from a frickin company suggestion box and these tools are so good. You just pick the best suggestion of the day and that's the new feature of the day. Like that's like, I don't think that you'll need 50 people in a company working on this I think you'll just need a really good Jack. I think to your point, when we first started the podcast, you build a really great foundation. And then as these tools get better, the speed of implementation will be unlike I think we can even fathom at this point. And so yeah, to have more people that play with them now, good thing for sure. Because I think those people ultimately become the heavy adopters. But it's the heavy adopters mixed with what Dave was saying the rest of the company that has access that will be able to implement stuff departmentally very, very quickly, very quickly. Jack BeVier (1:01:12) saying the rest of their company that David Moses (1:01:19) And your team will get you where you want to go. So we always want to give them the tools that'll make them happy and make them, you know, make them want to make our customers happy. But then, you know, at some point you got to turn it to facing the customer and saying, well, what do you think will make this easier, better, faster, whatever for you? How does this, you know, what can we share to make the customer experience better or to do things that, like I said before, Literally, that's the holy grail. That's what the holy grail isn't. I can manage X dollars of revenue with X dollars of overhead. The holy grail is I can do X because of these tools that was not possible before. And if you stay on that bleeding edge, know, your customers will, they will be very loyal. They will stick because they, you know, They don't even, like you may not have the very, very best thing for every single part of your relationship with them, but if you are constantly improving and constantly getting better, there's no reason for them ever to jump ship, right? Because it doesn't matter, there will never be a commoditization of your business if you're always improving what the business is instead of just, what it is that your customers get from you now. Jack BeVier (1:02:36) Mm-hmm. Follow. Craig (1:02:36) Deep, very deep. Hey, what would you guys think of having, I think memory is an issue for me that constantly comes back. I thought I figured it out with Obsidian and SupaBase and sort of clawed, open claw that structure, but I'm more and more inclined to think that before these LLMs get great memory ability. I still think that it's really important to have a structure for both short term and long term. I feel that Obsidian is very much the short term brain, sort of second brain, if you will. And I don't know if you guys have had a chance to play with it or not, what you think of it, but it would be really interesting to have a substantive discussion about it and how I think it can become ⁓ for those who might be using it, can become very much like email. where you're constantly feeling like you're, have to do housekeeping on your files, or you can set it up in a very smart, agentic way where it just kind of takes care of itself and becomes this really great source of search. And so, I don't know if you guys are interested in talking more about that, but, ⁓ love to maybe spend some time on the next podcast to talk about what we have learned about the, announcement from Anthropic today with, with regards to memory and how obsidian might be used to, to help. in that vein. David Moses (1:03:58) For sure, because they're giving you access to it, you still have to put it somewhere. And you have to figure out how use it. Craig (1:04:02) That's exactly right. All right. All right. Jack BeVier (1:04:05) Let's do that. Let's do that next one. I got a jump voice. Craig (1:04:08) Alright guys, always good. Mr. Moses, thank you for your time. Jack, what a great day today, by the way. Great day here at Dominion folks and hope you guys enjoyed this fifth episode of Forked. I know we had a great time as always. We'll see you on the next one.