บทถอดความฉบับเต็มของ Ep. 56 - Your Data Doesn't Need to Be Perfect: How Contractors Should Actually Start with AI

comes, a version of it comes for free, 00:04:01.363 --> 00:04:02.594 so people have played with it. 00:04:02.966 --> 00:04:05.084 Most, you know, most companies are on the Microsoft stack. 00:04:05.711 --> 00:04:10.903 The difference between different companies and it obviously it does follow how 00:04:10.903 --> 00:04:15.952 big the company is broadly but man you can the difference between some companies and others 00:04:16.359 --> 00:04:20.933 and some of that's just luck like you know one one company that we've talked to they they're they're, 00:04:21.380 --> 00:04:23.713 the founder of the of the company since sold but, 00:04:24.195 --> 00:04:29.933 was was very focused on process and and standardization which really sets them 00:04:29.933 --> 00:04:33.981 up for where they are now way ahead of where they you would think given their revenue new number. 00:04:34.643 --> 00:04:38.283 A lot of other companies that are pretty big have just grown super fast and 00:04:38.283 --> 00:04:40.651 their process isn't that great and their data is not that great. 00:04:41.237 --> 00:04:46.212 So it isn't as clean as how big they are. You know that. You've met a bunch of contractors. 00:04:47.203 --> 00:04:50.269 Some of it's luck. Some of it's what leadership focused on early. 00:04:50.691 --> 00:04:55.193 Also, I've run into a number of contractors where people that are excellent 00:04:55.193 --> 00:04:58.187 at construction also are like weekend coders. 00:04:58.923 --> 00:05:01.973 You can't really hire for that. Like, you're like, did you hire them? 00:05:01.973 --> 00:05:04.273 They're like, no, they've been in the company 20 years. They just do this on 00:05:04.273 --> 00:05:07.997 the weekend. Like, okay, you know, I build model cars and this guy codes Python. Yeah. 00:05:08.853 --> 00:05:13.888 So you can't control for that. But what's cool is this has been in market enough, 00:05:14.208 --> 00:05:17.728 long enough, and there's enough hype that there are definitely, 00:05:17.728 --> 00:05:21.212 in every contractor, there are pockets of people doing pretty wild stuff. 00:05:21.788 --> 00:05:25.718 And finding that and saying, what do we do with that? This person can't be the 00:05:25.718 --> 00:05:27.197 CIO. They're in the safety team. 00:05:27.568 --> 00:05:31.408 But maybe we can make sure that the things they're doing, we turn into something 00:05:31.408 --> 00:05:34.370 other people can learn from without having to learn as deeply as they did. 00:05:35.277 --> 00:05:41.570 Are you finding that the CEO or COO is delegating this task of, 00:05:41.970 --> 00:05:45.230 hey, we need an AI strategy, we need an associated data strategy? 00:05:45.230 --> 00:05:46.806 Are they taking it on themselves? 00:05:47.576 --> 00:05:53.340 That is dependent on size. Yeah. Yeah. Smaller companies, and by smaller, 00:05:53.340 --> 00:05:55.053 I still mean $500 million companies. 00:05:55.630 --> 00:05:58.623 Right. You know, the CEO is almost definitely leading

it. 00:05:58.966 --> 00:06:02.430 And I'll tell you, there was a really interesting study from KPMG like two weeks 00:06:02.430 --> 00:06:07.630 ago, where they found that when the CEO owns AI, it's like three times more 00:06:07.630 --> 00:06:08.817 effective than when not. 00:06:09.107 --> 00:06:14.843 So honestly, you know, if you're an $8 billion, $10 billion contractor, the CEO can't do that. 00:06:15.168 --> 00:06:19.240 But it has to be someone seriously senior that's doing it. Someone with the 00:06:19.240 --> 00:06:22.523 organizational, you know, ability to write an email and stuff happens. 00:06:22.894 --> 00:06:25.657 Because the reality is when you're trying to actually do this, 00:06:26.017 --> 00:06:28.263 you get to a place where people are on projects. 00:06:28.750 --> 00:06:32.080 This can't be their priority. So it won't and it won't be right. 00:06:32.080 --> 00:06:34.640 So when you have someone senior enough, you still get their attention. 00:06:34.640 --> 00:06:38.110 Otherwise, it is really hard to get meetings with people. And I don't begrudge 00:06:38.110 --> 00:06:40.883 them that. Like, they got a building to pull out of the ground. 00:06:42.370 --> 00:06:46.740 Are you finding those first interactions are with a group or a committee? 00:06:46.740 --> 00:06:52.350 Or are you part of the process of kind of helping them around how they should 00:06:52.350 --> 00:06:54.339 approach building that consensus? 00:06:54.995 --> 00:06:58.470 Both happen. And I mentioned before that sometimes it starts with a leadership 00:06:58.470 --> 00:06:59.824 briefing or a board briefing. 00:07:00.190 --> 00:07:03.030 That would be where we say, hey, it makes sense for you to stand up a committee 00:07:03.370 --> 00:07:07.006 or a group who meets regularly just to focus on this topic. 00:07:07.407 --> 00:07:09.218 I find that most companies already have one. 00:07:09.723 --> 00:07:12.893 I find that by now most companies have an AI committee of some kind. 00:07:14.014 --> 00:07:18.354 So you're, say we're kind of in the middle of this implementation, 00:07:18.734 --> 00:07:22.384 you're interacting with leadership team, you're starting to develop like what 00:07:22.384 --> 00:07:23.917 the path forward is looking like. 00:07:24.934 --> 00:07:28.434 What is, how do you leverage some of your product background, 00:07:28.774 --> 00:07:31.263 your understanding of the back end of software? 00:07:31.710 --> 00:07:34.914 You mentioned prototyping. How does that come into the engagement? 00:07:35.512 --> 00:07:40.484 Yeah, that's a big deal. You know, the problem with AI has been from the beginning 00:07:40.484 --> 00:07:41.691 that it gives great demos. 00:07:42.150 --> 00:07:45.974 And people that, you know, saw someone say something in a conference kind of 00:07:45.974 --> 00:07:49.534 make their own inferences and assume that you can do this because they can do 00:07:49.534 --> 00:07:54.456 that. Right. And really grounding it in how AI actually works. 00:07:55.066 --> 00:07:59.424 One of the first things that we do is give a presentation that just kind of 00:07:59.424 --> 00:08:01.634 grounds them in how it works. 00:08:01.974 --> 00:08:05.684 And that isn't prompting at all. I mean, of course, you talk about that if necessary, 00:08:05.684 --> 00:08:09.014 but understanding that AI fundamentally, the AI we're using, 00:08:09.014 --> 00:08:12.084 there's also a number of flavors of it, but the AI that everybody's using is 00:08:12.084 --> 00:08:13.822 a matching algorithm and that 00:08:14.257 --> 00:08:17.444 drives everything else. It makes some things really easier to do. 00:08:17.694 --> 00:08:20.654 It makes some things more limited than you wish they were, you know, 00:08:20.654 --> 00:08:22.520 and it means that things like technology. 00:08:23.608 --> 00:08:28.401 How to think about your knowledge and the expertise of your business as an asset you can capture. 00:08:28.872 --> 00:08:33.376 Okay, well, what does that mean? And how do you organize things so that you can do that? 00:08:33.661 --> 00:08:36.882 An example is, I don't love the term human in the loop. 00:08:37.224 --> 00:08:41.028 That's a throwaway that it's nothing. Look, it made sense when people started 00:08:41.028 --> 00:08:43.538 using it, but like everything, it starts becoming a throwaway. 00:08:43.538 --> 00:08:45.708 Oh, we'll keep human in the loop. What do you mean by that? Well, 00:08:45.708 --> 00:08:48.878 usually they mean, we're going to automate something and then we're going to 00:08:48.878 --> 00:08:50.686 have an approval at the end. 00:08:51.411 --> 00:08:54.530 Fair enough. First of all, you can't automate as much as you think you can. 00:08:54.924 --> 00:08:59.478 And there's a lot of value people add in the process. When you try to automate 00:08:59.478 --> 00:09:01.368 a thing, you realize how smart people are. 00:09:01.763 --> 00:09:05.198 Even the guy that you don't think is that bright is adding a ton of value in 00:09:05.198 --> 00:09:10.448 the process. So, you know, if you pull people out of it, it isn't just that 00:09:10.448 --> 00:09:13.808 you're going to have a little bit of pain later when you think you automated 00:09:13.808 --> 00:09:14.498 something you couldn't. 00:09:14.898 --> 00:09:19.878 I mean, humans in a loop are destroyed souls, right? Like no one aspires to 00:09:19.878 --> 00:09:22.748 be a human in a loop. Yeah, that's right. And so I think that's absolutely to 00:09:22.748 --> 00:09:25.329 your point. Like that is the opposite of purpose. 00:09:25.898 --> 00:09:29.061 That's, I love that you brought it back to purpose. That's right. 00:09:29.450 --> 00:09:33.525 So, you know, you have a collection of 300 to a couple thousand people. 00:09:34.124 --> 00:09:37.618 How do we elevate them? And you hear people hint at this, like, 00:09:37.618 --> 00:09:39.633 you know, automate the drudgery. Totally. 00:09:40.086 --> 00:09:45.103 But why? Like, is it just so

that they can get home early, which is not a bad goal on its own? 00:09:45.329 --> 00:09:48.228 Yeah. Everyone knows construction hours are longer than they should be. 00:09:48.228 --> 00:09:52.876 And, you know, but really you need to be thinking about what is the expertise 00:09:53.160 --> 00:09:55.218 and how do I preserve that? 00:09:55.218 --> 00:09:58.108 You know, how do I, how do I give people more time to think, 00:09:58.108 --> 00:10:02.058 more time to talk to each other, more time to be on the site or go to a sub 00:10:02.058 --> 00:10:06.523 or whatever it is that is really, really human and really complex. 00:10:06.953 --> 00:10:10.878 You know, it's hard to know if somebody is really going to show up with the 00:10:10.878 --> 00:10:13.698 four people they said they would instead of three until you know the person 00:10:13.698 --> 00:10:14.679 and you spent time with them. 00:10:15.010 --> 00:10:18.668 You know, they've heard your voice saying, you're bringing four, right? 00:10:19.677 --> 00:10:23.059 So that takes time, though. You know, like not every phone call goes through. 00:10:23.059 --> 00:10:25.539 It takes a minute to get to people. I mean, I'm getting into the weeds, 00:10:25.539 --> 00:10:32.159 but whatever you can do to elevate the things that humans need to be there for. 00:10:32.459 --> 00:10:33.626 And there's another part to that. 00:10:34.479 --> 00:10:38.909 Some of what you might capture when an agent is running or when a system is 00:10:38.909 --> 00:10:43.459 running can also be recycled as helping juniors to learn, too. 00:10:43.459 --> 00:10:47.959 That's a really big one that comes up, is everybody has realized that nobody 00:10:47.959 --> 00:10:50.435 can take turn-by-turn directions anymore. 00:10:50.760 --> 00:10:53.349 Nobody remembers any phone number except maybe their moms. 00:10:54.259 --> 00:10:59.444 So they're all like, oh, my God, I learned I am as good as I am as a 40-year-old 00:10:59.752 --> 00:11:01.552 because I'm not 40. I wish I were. 00:11:01.952 --> 00:11:06.599 But as someone who's in my career but mid-career, I'm as good as I am because 00:11:07.019 --> 00:11:10.773 I spent all that time grinding through specs to do submittal review. 00:11:11.331 --> 00:11:15.229 Well, okay, do you need to grind that much? Probably not. But the reality is 00:11:15.229 --> 00:11:19.659 some of that stuff is the foundation, as painful as it is, of later expertise. 00:11:20.030 --> 00:11:25.156 So how do you respect that and say, we want the benefits of automating the drudgery, 00:11:25.615 --> 00:11:29.669 but we want to also preserve enough of that and maybe make it into something 00:11:29.669 --> 00:11:34.573 where the senior people know that we want to grow our – they do this anyway. 00:11:34.770 --> 00:11:39.549 We want to grow our people. So how do we create, you know, implementations or 00:11:39.549 --> 00:11:43.336 what's called implementations so that we're capturing things, 00:11:43.551 --> 00:11:46.587 we're feeding it back to the juniors so they have a little more context? 00:11:46.923 --> 00:11:49.729 People in their 20s are very often pretty eager to learn this stuff. 00:11:49.729 --> 00:11:52.618 And they're in construction because they like to know how things work. 00:11:53.106 --> 00:11:57.989 So I think the other thing I'd say real quick is there are grumpy people out 00:11:57.989 --> 00:12:00.426 there, but almost everybody loves to be a mentor. 00:12:00.960 --> 00:12:04.279 I mean, it depends on the moment, right? Like when there's time pressure, 00:12:04.279 --> 00:12:05.349 no one wants to be a mentor. 00:12:05.674 --> 00:12:10.051 But most people like that they're an expert in something and somebody's listening to them. 00:12:10.353 --> 00:12:14.009 So if you give them that opportunity, you know, a lot of people will benefit 00:12:14.009 --> 00:12:15.941 from it. You're going to find people that don't. You always do. 00:12:16.434 --> 00:12:21.287 But I find that that construction especially is the kind of industry where people 00:12:21.677 --> 00:12:23.313 like that. They like being a mentor. 00:12:23.956 --> 00:12:29.954 Yeah, there's definitely more, there's more of a handoff or a desire for handoff 00:12:29.954 --> 00:12:33.365 of kind of legacy information, even if it isn't explicit. 00:12:33.528 --> 00:12:36.094 There's kind of this understanding that through osmosis, there's going to be 00:12:36.094 --> 00:12:39.298 some kind of handoff of what my experience to the next generation. 00:12:39.809 --> 00:12:44.054 So you know like palantir is as obviously in the headlines they signed a big 00:12:44.054 --> 00:12:46.734 deal with mccarthy we've talked about it on a previous episode 00:12:47.146 --> 00:12:51.404 they you know the more you kind of pull back the cart the curtain on palantir 00:12:51.404 --> 00:12:55.746 it's it's really a consulting agency right that kind of builds products around, 00:12:56.541 --> 00:12:58.277 an organization that they're working with. 00:12:59.264 --> 00:13:04.474 What would you say from the outside looking in domain expertise i think is is 00:13:04.474 --> 00:13:08.018 really important obviously they don't have that maybe they've hired some of that internally. 00:13:08.794 --> 00:13:17.832 But what is, what do you think is their approach as they look to kind of layer over data? 00:13:18.053 --> 00:13:23.463 And is there a juxtaposition to how you're interacting with companies to Palantir? 00:13:23.835 --> 00:13:27.207 Or are we still learning about what Palantir is actually doing in our industry? 00:13:27.654 --> 00:13:33.414 Yeah, I think there's a few layers to this. One of them is Palantir was founded 00:13:33.414 --> 00:13:36.363 by a number of people, but the soul of the company is Alex Karp. 00:13:36.764 --> 00:13:39.914 And he's a philosopher first. I think even that was his degree, 00:13:39.914 --> 00:13:43.817 but you can kind of hear it when he's not jumping on a chair and

screaming like Tom Cruise. 00:13:44.595 --> 00:13:50.415 You can hear it in how he talks. And that matters because they talk about this ontology all the time. 00:13:50.415 --> 00:13:54.075 And for people who've heard that word and haven't worked with one before, 00:13:54.075 --> 00:14:00.195 all it is, it's a taxonomy like master format that also has relationships between things. 00:14:00.195 --> 00:14:04.545 And honestly, if anybody's actually ever read the documents of the 2020 master 00:14:04.545 --> 00:14:08.115 format, obviously, I was at CSI, so I know this. They do that in there, too. 00:14:08.775 --> 00:14:12.208 They don't do it for that reason. They'll connect. This thing is related to that and so on. 00:14:12.499 --> 00:14:17.785 Now, that's really very master format specific, but ontologies are just a way 00:14:17.785 --> 00:14:23.425 of making data connected to itself in a way that is very AI appropriate. 00:14:23.425 --> 00:14:26.106 It wasn't made for AI. Ontologies have been around forever. 00:14:26.466 --> 00:14:31.395 But it's a really good way for an LLM to reason over data that's disconnected, 00:14:31.395 --> 00:14:37.245 whereas a table-based database really struggles with things that are two hops away. 00:14:37.245 --> 00:14:41.085 You know, like you change this, that means the electrical sub has to do this, 00:14:41.085 --> 00:14:43.105 which means that we're going to have to change the schedule, 00:14:43.105 --> 00:14:45.146 which means the, you know, you know what I mean? 00:14:45.494 --> 00:14:48.905 And construction is loaded with like multi-step implications. 00:14:48.905 --> 00:14:52.165 You think about how many times have you heard somebody or yourself when you 00:14:52.165 --> 00:14:54.645 were a PM, you hear something's going to change. And in your mind, 00:14:54.645 --> 00:14:57.621 you're immediately like, all right, then we got to call this guy, this guy, this guy. 00:14:58.091 --> 00:15:01.645 Like it's loaded with that. And that's the kind of stuff that AI can help with. 00:15:01.645 --> 00:15:05.175 It can't do what a human can do because you're going to think deeper and you're 00:15:05.175 --> 00:15:09.085 going to, again, know that company, know more about the context than AI will. 00:15:09.398 --> 00:15:13.065 But it can help you with some things and can give you better answers if something's 00:15:13.065 --> 00:15:14.199 organized by an ontology. 00:15:14.425 --> 00:15:17.895 So back to your question about how to think about whether it's Palantir. 00:15:17.895 --> 00:15:20.735 Look, we do want everybody does ontologies, or at least they should. 00:15:21.078 --> 00:15:24.805 It's just it's not that hard to stand a good one up and then you can refine 00:15:24.805 --> 00:15:27.347 it over time. We're not talking about rocket science. 00:15:27.666 --> 00:15:34.148 And I think that that speaks to whether it's a Palantir or somebody likes or somebody else. 00:15:34.729 --> 00:15:41.848 I think that they bring

an enormous amount of experience having developed and 00:15:41.848 --> 00:15:44.808 deployed ontologies in really complicated environments. And obviously, 00:15:44.808 --> 00:15:47.746 the military is what has gotten the headlines, and that's where they cut their teeth. 00:15:48.315 --> 00:15:51.223 I don't know that you need a Ferrari to deliver milk, though. 00:15:51.548 --> 00:15:55.518 So that's kind of point one, is do you need to pay for all of that depth? 00:15:55.918 --> 00:16:00.188 I would question that. The other one that you brought up, which I think is excellent, 00:16:00.188 --> 00:16:03.318 is knowing how construction works. 00:16:03.638 --> 00:16:07.958 Kind of at a pretty deep detail, is I think really important. 00:16:08.298 --> 00:16:10.359 It speeds things up. And there's another part here. 00:16:10.765 --> 00:16:16.106 One of the first things everyone asks in construction is, what is your construction experience? 00:16:16.257 --> 00:16:20.538 So it isn't just doing the work, it's also getting their time and their attention 00:16:20.538 --> 00:16:25.878 and their confidence that you're going to think about what I need and propose 00:16:25.878 --> 00:16:27.687 things that I didn't know to ask for. 00:16:27.994 --> 00:16:31.068 And that happens all the time. I'll give you an example. A contractor we're 00:16:31.068 --> 00:16:35.678 working with right now, I was listening to them and I said, you guys need some 00:16:35.678 --> 00:16:37.685 baseline training you could just give people. 00:16:38.062 --> 00:16:40.998 So we just, as part of the, they didn't pay for this. 00:16:40.998 --> 00:16:44.168 It was just part of how we help them get done with it. We did eight videos that 00:16:44.168 --> 00:16:48.748 was branded to them, that was for them, that was specific to the type of project 00:16:48.748 --> 00:16:52.288 they work on, that started with, here's how you log in, and then got all the 00:16:52.288 --> 00:16:55.048 way to, here's how to think about agents. You see what I mean? 00:16:55.524 --> 00:17:02.048 So that's a little example of a broader point that I think that actually getting 00:17:02.048 --> 00:17:07.427 the job done, being able to speak and people hear you, that you know what they're talking about. 00:17:08.018 --> 00:17:10.771 Starts to really matter because otherwise you sound like a suit. 00:17:11.201 --> 00:17:17.240 And look, nobody in any industry likes that except them. Yeah, right. Yeah. 00:17:17.942 --> 00:17:22.420 And I think, you know, there's a number of reasons why Palantir, 00:17:22.760 --> 00:17:25.227 you know, it's a big frontier, obviously, construction. 00:17:25.396 --> 00:17:30.610 And it does kind of go into data. And I think there could be listeners who think 00:17:30.610 --> 00:17:33.447 through kind of everything that's just been said and think, well, 00:17:33.854 --> 00:17:38.740 we're maybe not yet ready or positioned well from a data standpoint to maybe 00:17:38.740 --> 00:17:40.466 interact with you or the link. 00:17:41.110 --> 00:17:45.446 How do you, like, what's your response to the CEO who says, hey, our data is a mess. 00:17:45.853 --> 00:17:52.004 AI is maybe like a Q1 2027 initiative for us. We're really trying to think through maybe an ontology now. 00:17:52.880 --> 00:17:57.890 How do you respond to that? Yeah. So this has been the big blocker for a lot 00:17:57.890 --> 00:18:02.110 of people for a lot of things, not just, I mean, analytics came before this. 00:18:02.110 --> 00:18:05.060 Everybody was excited about Power BI and they realized they can't trust the 00:18:05.060 --> 00:18:07.851 data that or the dashboards or the reports that come out of it. 00:18:08.542 --> 00:18:11.872 What I would say is that there's a few parts to this. One is people are already 00:18:11.872 --> 00:18:14.787 using AI right now, and they're getting good things done with it. 00:18:15.159 --> 00:18:18.742 The data within a project is probably consistent enough that you can do good 00:18:18.742 --> 00:18:22.622 stuff. The question is enterprise level, and almost nobody has that right. 00:18:23.322 --> 00:18:27.822 And that's kind of point one. Point two is they're often islands of pretty darn 00:18:27.822 --> 00:18:31.352 good data because you're using an Autodesk or a Procore, and they kind of enforce 00:18:31.352 --> 00:18:34.599 you to do certain things. I mean, project setup can differ, for sure. 00:18:35.018 --> 00:18:37.722 And so there's still variability you might need to lock down. 00:18:38.112 --> 00:18:42.672 So a couple of points. One of them is, like I just said, you can get a lot done 00:18:42.672 --> 00:18:46.662 with the documents that you have without having enterprise data that's clean 00:18:46.662 --> 00:18:47.640 and conformed and all that. 00:18:48.186 --> 00:18:52.732 The other one is, we talk about automation as if that's the only thing we want 00:18:52.732 --> 00:18:57.842 AI for. And yet everybody's using it for non-automation things every time they use ChatGPT or Copilot. 00:18:58.312 --> 00:19:03.332 You can get a lot done by just making the problem smaller and supporting someone's judgment. 00:19:03.332 --> 00:19:07.062 My point before about elevating expertise, that's actually a lot of what we 00:19:07.062 --> 00:19:11.082 want to be doing anyway, is giving them the ability to find what they need so 00:19:11.082 --> 00:19:14.578 that they then go the last 80% or whatever it is, percent, 00:19:15.001 --> 00:19:19.616 with their judgment and knowledge of the project and the 20 years they've been doing this. 00:19:19.982 --> 00:19:24.612 So you can get an awful lot done with your data as it is and then say, 00:19:24.612 --> 00:19:29.560 okay, I also want you to tell me what's a playbook? How do we get ourselves better? 00:19:29.949 --> 00:19:35.002 So that, you know, in a year or 18 months, we've now gone project by project 00:19:35.002 --> 00:19:39.092 to get to a place where our data is in a place where we

can actually now start 00:19:39.092 --> 00:19:41.822 to apply, you know, enterprise-wide questions. 00:19:41.822 --> 00:19:46.512 And the reality is you think about enterprise as only leadership needs that, 00:19:46.512 --> 00:19:48.982 but that's not true. Wouldn't it be great if a PM could say, 00:19:48.982 --> 00:19:51.592 what other projects have had this problem? Or what, you know what I mean? 00:19:51.592 --> 00:19:52.409 Like reference all that. 00:19:52.862 --> 00:19:56.502 But they can still do their job really well without that, by not having that 00:19:56.502 --> 00:20:00.478 level. So think of it as a process. 00:20:00.978 --> 00:20:05.060 And the other thing, and this also relates a little to the Palantir question, is. 00:20:05.691 --> 00:20:11.078 We're on a path, like the industry, not just construction, but AI is like early. 00:20:11.398 --> 00:20:15.678 There's a lot coming and a lot of stuff that'll make some of the process and 00:20:15.678 --> 00:20:17.753 some of the data easier over time. 00:20:18.043 --> 00:20:25.010 Definitely the infrastructure for really dealing with lots and lots of data is still in its infancy. 00:20:25.242 --> 00:20:27.848 That's been a big focus for folks like Databricks. You know, 00:20:27.848 --> 00:20:30.721 that's why Microsoft has Fabric, which is made for that. 00:20:31.216 --> 00:20:33.265 That stuff's still early and it's still being worked out. 00:20:34.227 --> 00:20:39.295 Well, and I think also another response to that CEO who's maybe thinking they 00:20:39.295 --> 00:20:42.969 need to start from a data foundation standpoint is also. 00:20:43.765 --> 00:20:47.975 Building a solution without a real clear understanding of the end or the outcome 00:20:47.975 --> 00:20:51.975 or the use cases you're trying to drive to is a little flawed, right? 00:20:51.975 --> 00:20:56.035 And so by having a conversation with you and understanding what that outcome 00:20:56.035 --> 00:20:59.545 is that you're trying to drive to, it's really going to change the structure, 00:20:59.545 --> 00:21:01.306 I'd imagine, of that data strategy. 00:21:01.933 --> 00:21:04.375 Is definitely going to change the strategy. You're absolutely right. 00:21:04.375 --> 00:21:05.944 It's also going to change what you do when. 00:21:06.368 --> 00:21:11.185 Yeah. Right? Like if you ask somebody who doesn't understand data because it's 00:21:11.185 --> 00:21:15.965 not their job, they're going to think of it as this big black cloud of, I don't understand it. 00:21:15.965 --> 00:21:19.065 And it sounds like, you know, a half a million or $2 million that I'm going 00:21:19.065 --> 00:21:22.575 to need to spend. And frankly, I can't picture what it's going to do. 00:21:23.133 --> 00:21:27.395 As opposed to saying, what if we just clean up how you operate in a couple of 00:21:27.395 --> 00:21:29.054 systems of record and start there? 00:21:29.495 --> 00:21:32.519 And say, you know, we want to have your ERP and your PM system. 00:21:32.734 --> 00:21:36.145 Those two, let's focus on those and maybe timesheets too. That's another one 00:21:36.145 --> 00:21:38.539 that often feeds into really important questions. 00:21:38.974 --> 00:21:42.775 But my point is you don't have to say, I want to go root and branch and change 00:21:42.775 --> 00:21:46.505 what everybody does. If you can do that, that's wonderful. But that's a real 00:21:46.505 --> 00:21:50.405 commitment that most companies just struggle with because what's the trade-off? 00:21:50.405 --> 00:21:51.474 Everything is a trade-off. 00:21:51.875 --> 00:21:55.410 If I'm going to spend management time and attention because it needs to be. 00:21:55.625 --> 00:22:01.383 So the CEO or the lead is saying, I'm going to spend 10% to 15% of my time just on this. 00:22:01.737 --> 00:22:06.445 That's 10% to 15% of their time they can't do with clients, with customers they 00:22:06.445 --> 00:22:09.025 work with over and over again and all of the other things that they get paid 00:22:09.025 --> 00:22:10.921 for. So thinking about... 00:22:11.580 --> 00:22:15.015 Where you might use it and then say, great, well, then where do we start and 00:22:15.015 --> 00:22:17.995 how does that enable the next step and the next step and the next step? 00:22:17.995 --> 00:22:21.595 And as I say, often it's a good idea to say, rather than try to clean everything 00:22:21.595 --> 00:22:26.185 up, let's clean up pieces and then start to see value and then go from there. 00:22:26.568 --> 00:22:29.250 Because the other thing is there's a long tail of pain. 00:22:30.220 --> 00:22:33.445 There's stuff in the middle because it's in a system of record that's much easier 00:22:33.445 --> 00:22:35.328 to do than some of the other stuff. 00:22:35.653 --> 00:22:38.575 Like, you know, again, timesheets tend to be one of them because people still, 00:22:38.575 --> 00:22:41.535 big companies still use, like, you'd be surprised. Well, you wouldn't be, 00:22:41.535 --> 00:22:43.665 but people can be sometimes surprised. They're like, oh my God, 00:22:43.665 --> 00:22:49.311 it's a billion dollar company. Surely everything's robotics and automation. It's like, no. No. 00:22:49.816 --> 00:22:52.745 All right. Sponsor break. Lars, real quick. When was the last time you got a 00:22:52.745 --> 00:22:57.157 new set of drawings from an architect and they didn't cloud anything? 00:22:57.540 --> 00:23:02.480 Well, the one that cost me a couple hundred grand was back, I think, in 2017. So- 00:23:03.650 --> 00:23:07.243 Lesson learned there. Yeah. So, you know, what I like about what Trunk Tools 00:23:07.243 --> 00:23:13.953 is building is they're looking at tools to help kind of remove the mundane from the job site. 00:23:13.953 --> 00:23:18.603 And, you know, same as you, every time there was a bulletin or an addendum, 00:23:18.923 --> 00:23:22.463 I would go back to designers and ask them to at least highlight or give me a 00:23:22.463 --> 00:23:23.503 cover letter on

the changes. 00:23:24.163 --> 00:23:27.253 And that was always like pulling my teeth, but Trunk Tools is now automating 00:23:27.253 --> 00:23:31.258 that. So you get a bulletin come in, it automatically highlights the changes. 00:23:31.647 --> 00:23:32.471 You don't have to do a thing. 00:23:32.976 --> 00:23:36.223 Well, it's also hitting RFIs, right? It's also hitting submittals. 00:23:36.703 --> 00:23:40.818 And when we think about standardizing ops, right, and providing that, 00:23:41.317 --> 00:23:47.403 you know, sort of resource to those field operators, what I see it as, you know, 00:23:48.085 --> 00:23:52.773 it's sort of like this, like, mentor within the company that structures, 00:23:52.773 --> 00:23:56.643 standardizes, keeps it moving in the right direction, looks out for risk. 00:23:57.683 --> 00:24:01.923 And just this last week, you know, I had a QC manager just pass along the submittal 00:24:01.923 --> 00:24:05.836 to an owner without doing a review because he was too busy and it was wrong. 00:24:06.983 --> 00:24:11.270 And, you know, the classic owner statement is, did you even review this? Right. 00:24:11.869 --> 00:24:16.019 So this to me is carte blanche. We absolutely did. 00:24:16.391 --> 00:24:22.563 And think about getting that first pass before your field manager ever even 00:24:22.563 --> 00:24:24.843 actually has to do the first pass. Someone's doing it for them. 00:24:24.843 --> 00:24:30.404 So to me, it's about speed. It increases accuracy and helps standardize ops out in the field. Right. 00:24:30.874 --> 00:24:33.633 Yeah. And this is the kind of kind of AI that we've we've talked about on the 00:24:33.633 --> 00:24:37.303 podcast. Really appreciating. Right. Which is is looking at ways to elevate 00:24:37.303 --> 00:24:39.683 the craftsman at the heart of project delivery. 00:24:40.003 --> 00:24:44.346 Give me time back so I can be out in the field managing risk. Right. 00:24:45.159 --> 00:24:50.524 So maybe let's scenario play this. You're the CEO of a $500 million contractor. 00:24:51.355 --> 00:24:55.285 First step, obviously, is you hire yourself. What are your second through four 00:24:55.285 --> 00:24:58.475 steps that you kind of immediately focus on? Yeah. 00:24:59.555 --> 00:25:04.385 What do you want this to look like at the end? And how does your business run 00:25:04.385 --> 00:25:06.264 in a way that we can point this? 00:25:06.636 --> 00:25:10.715 So an example is if you find you're competing against big dogs a lot because 00:25:10.715 --> 00:25:14.386 of the type of project you do, let's think about that. And what does that mean? 00:25:14.885 --> 00:25:18.849 You know, are you doing a lot of mission critical? Okay, let's focus on that. 00:25:19.232 --> 00:25:21.275 You see what I mean? So the first thing is, what are your goals? 00:25:21.275 --> 00:25:25.175 The thing about strategy is people go there first, but you really need to do 00:25:25.175 --> 00:25:30.615 your best to define goals. And most

professional services companies don't think 00:25:30.615 --> 00:25:33.615 that way because it's very relationship based. 00:25:34.135 --> 00:25:36.965 It isn't really head-to-head competition. I mean, it's part of it, 00:25:36.965 --> 00:25:38.067 but a lot of it's not that. 00:25:38.467 --> 00:25:41.345 You know, a lot of it's about people and so on. So what are your goals? 00:25:41.345 --> 00:25:45.635 And then from there, okay, how will AI fit in there? And that isn't a question 00:25:45.635 --> 00:25:47.942 I ask so much as we get to that answer together. 00:25:48.303 --> 00:25:52.565 Then we go talk to the second tier of leadership and what are their teams doing? 00:25:52.565 --> 00:25:53.463 What are they actually... 00:25:53.731 --> 00:25:57.245 A lot of this also is discovery of what AI they're currently doing because we 00:25:57.245 --> 00:25:58.369 are at that point now where... 00:25:59.055 --> 00:26:03.785 And again, you'll find a GM that, like I said, writes Python. 00:26:03.785 --> 00:26:07.545 So he's built his own agents. And are those agents reliable? 00:26:07.545 --> 00:26:09.956 Does he know how to do that yet, even though he knew how to wire them up? 00:26:10.292 --> 00:26:14.879 So maybe not. And so the next piece after that is going into the organization. 00:26:15.047 --> 00:26:20.425 One of the things that you hear a lot about that I think is a bit of a misconception 00:26:20.425 --> 00:26:21.958 is then we got to map your workflows. 00:26:22.540 --> 00:26:25.826 Okay. Every time anyone's looked at any business, not just construction, 00:26:26.081 --> 00:26:28.571 people don't have step-by-step SOPs. 00:26:28.821 --> 00:26:32.238 They do for some things because you have to, especially if it's audited or it's regulatory. 00:26:32.586 --> 00:26:34.380 But for most things- And they're only six years old. 00:26:35.760 --> 00:26:38.690 The SOPs that are in place are just, yeah, they're never dynamic. 00:26:39.184 --> 00:26:42.550 And are they really followed or are people saying, okay, I use them as a guideline, 00:26:42.550 --> 00:26:46.841 but I really know where I need to end up. So what do I have to hand off to the next step? 00:26:47.102 --> 00:26:50.480 What are my guardrails? So what can I not do? What do I have to, what are my inputs? 00:26:50.945 --> 00:26:53.679 And what are some rough steps in between? You see what I mean? 00:26:54.050 --> 00:26:56.650 Now, to automate something, you do need to turn it into a map. 00:26:56.650 --> 00:26:59.690 But that doesn't mean you're asking somebody, okay, you know, 00:26:59.690 --> 00:27:02.119 give me boxes for what you do and in what order. 00:27:02.421 --> 00:27:05.650 Sometimes you got to, usually you need to actually create that and say, 00:27:05.650 --> 00:27:08.668 hey, would this work? And this may not be how you always do it, but would this work? 00:27:09.086 --> 00:27:13.108 And then you can start to say, well, all right, of those pieces, what can we automate? 00:27:13.300 --> 00:27:17.200 And the automation question is, it's not as hard as people sound. 00:27:17.200 --> 00:27:20.323 It's just not as, you can't automate as much as people think you can. 00:27:21.543 --> 00:27:27.471 Yeah, nor do I think we always want to or should need to or I don't know, some version of that. 00:27:27.831 --> 00:27:31.931 I think there is a problem of at times, and we've talked about this before, 00:27:32.271 --> 00:27:36.292 we've set ourselves up as if like AI is this inevitable engine. 00:27:36.623 --> 00:27:40.451 And we just kind of are like sitting at its mercy. Like we do still have the 00:27:40.451 --> 00:27:44.221 opportunity to say, what do we want to do versus what do we want to automate? 00:27:44.221 --> 00:27:47.012 I don't think we always start with that, which I think is a problem. 00:27:47.575 --> 00:27:51.255 But maybe back to the Palantir side of this. 00:27:51.534 --> 00:27:57.181 So Carr had some comments recently raising some concerns about token usage of 00:27:57.521 --> 00:28:02.569 open AI and Anthropic and essentially kind of calling out the fact that it's a little bit of a, 00:28:03.221 --> 00:28:09.541 not extortion, but the business practice needs to be maybe thought through by CEOs. 00:28:10.481 --> 00:28:15.901 How do you think about that as you advise these companies? Do you start there? 00:28:15.901 --> 00:28:20.652 Is that something that you kind of do some quarterly check-ins around spend? 00:28:21.311 --> 00:28:26.375 I was talking to a buddy who worked for a SaaS company that does $75 million in revenue. 00:28:26.868 --> 00:28:31.797 And their spend is already forecasted to be beyond $2 million for the year. 00:28:32.238 --> 00:28:35.999 And their kind of conversation as a leadership team is like, 00:28:36.272 --> 00:28:39.181 how do we map all this ROI? And it's very hard to do, obviously. 00:28:40.094 --> 00:28:45.835 Yeah, I think software companies are different from a contractor because they're 00:28:45.835 --> 00:28:47.152 doing really heavy coding. 00:28:47.767 --> 00:28:51.905 So hearing about big token use, there's a few things to this. 00:28:51.905 --> 00:28:55.925 One of them is, you know, I've heard people say, are they, you know, 00:28:55.925 --> 00:29:00.395 they're subsidizing the tokens so they can lock us in and then, 00:29:00.395 --> 00:29:02.072 you know, ratchet it up later. 00:29:02.583 --> 00:29:07.716 And I don't love that kind of going right to the other guy is evil, 00:29:08.523 --> 00:29:10.886 or this is some strategy that we've seen before. 00:29:11.136 --> 00:29:16.915 And I think it just, it slows down or stops you from thinking about what's really 00:29:16.915 --> 00:29:21.745 happening over time. So think about what happened in, like in April, 00:29:21.745 --> 00:29:24.495 everyone is certainly on the West Coast was talking about token maxing. 00:29:24.780 --> 00:29:27.645 Like, I want everyone to use as many tokens as they can, because I want them

00:29:27.645 --> 00:29:30.925 exploring how this might work. And then like six weeks later, 00:29:30.925 --> 00:29:32.425 everyone's like, oh, my God, don't do that. 00:29:33.359 --> 00:29:37.625 Right. Like dumb incentive. Now, that's not really a risk at most contractors, 00:29:37.625 --> 00:29:40.714 but it's still a good question. And I want to stick with it for a minute. 00:29:41.277 --> 00:29:43.865 But what happened then is, oh, my God, we got to be token efficient. 00:29:43.865 --> 00:29:46.902 And then a little while later, like right now, oh, we should look at open models. 00:29:47.338 --> 00:29:51.635 And then you have, you know, Alex Karp, who kind of jumped on that bandwagon 00:29:51.635 --> 00:29:53.712 because it's right now it's a thing. Right. 00:29:54.378 --> 00:29:58.842 History has shown that every infrastructure build out, which is what we're in 00:29:58.842 --> 00:30:01.951 the middle of, and tokens are the result of that, right? 00:30:02.578 --> 00:30:07.007 They always overshoot every single time. The railroads, broadband, 00:30:07.321 --> 00:30:09.522 you know, wireless, everybody overshoots. 00:30:09.842 --> 00:30:13.032 Chips do this. Every 20 years, there's a chip cycle where they build too much 00:30:13.032 --> 00:30:15.319 because there's so, and then there's a glut, then there's a blah, blah, blah. 00:30:15.650 --> 00:30:18.024 And the reason for that is you just can't time it right. 00:30:18.471 --> 00:30:21.622 When it takes six months to two years to build something, you're never going 00:30:21.622 --> 00:30:25.315 to time it right. and there's plenty of money around to say, you know, I'll invest. 00:30:25.872 --> 00:30:29.862 That's going to happen here. There's just, I do not believe that we're going 00:30:29.862 --> 00:30:34.752 to buck the trend and that won't happen. So what history shows is the lock-in 00:30:34.752 --> 00:30:38.722 risk isn't usually at the infrastructure, isn't ever actually at the infrastructure 00:30:38.722 --> 00:30:40.721 layer. It's when you're closer to the customer. 00:30:41.157 --> 00:30:45.022 So I would be more worried about a Palantir and I'd be more worried about a 00:30:45.022 --> 00:30:46.927 Microsoft as who's going to lock you in. 00:30:47.322 --> 00:30:50.462 And frankly, at scale, it's Microsoft, not Palantir. 00:30:50.462 --> 00:30:54.602 They're too expensive for most contractors, and they're pretty invasive and so on. 00:30:54.602 --> 00:30:58.742 So I don't know that you're going to find Palantir's penetration gets that deep. 00:30:58.742 --> 00:31:01.462 I think you're going to find, you know, McCarthy scale companies. 00:31:01.462 --> 00:31:05.962 And remember also, McCarthy has been pretty data forward for like years. 00:31:05.962 --> 00:31:10.542 I remember five years ago, I saw them talking about pretty sophisticated data 00:31:10.542 --> 00:31:14.512 lakes. So they've been really focused on this problem for a while. 00:31:14.512 --> 00:31:19.079 So they're much

more likely to say, we want that, you know, Ferrari to deliver our milk. 00:31:20.342 --> 00:31:24.182 They did glean also earlier than everybody else. And it's interesting they got 00:31:24.182 --> 00:31:26.179 off it, I think, kind of quickly. 00:31:27.674 --> 00:31:34.811 So just to finish the point, yes, you should be thinking about the cost this is going to be. 00:31:35.380 --> 00:31:40.114 Governance is going to need to be part of it anyway, because what works on Monday 00:31:40.114 --> 00:31:44.651 might not work in three weeks or three months. And by that, I mean an agent can drift. 00:31:45.494 --> 00:31:49.214 The tools that you're using, the documents you're using, the reference documents, 00:31:50.154 --> 00:31:51.472 all that can drift over time. 00:31:51.814 --> 00:31:55.424 So unlike SAS, it isn't like install it, train people, and then don't think 00:31:55.424 --> 00:31:57.770 about it again. You're going to need to govern this forever. 00:31:58.734 --> 00:32:03.076 It's a new skill that the CIO's office or the IT team is going to have to have. 00:32:03.325 --> 00:32:05.375 That's going to include what you just talked about. 00:32:06.674 --> 00:32:09.794 Are we using open models and are we even hosting some on-prem? 00:32:09.794 --> 00:32:12.281 Because I think that's the other thing that's going to really change this is, 00:32:12.833 --> 00:32:16.107 you're going to find that AI is available in a lot of different places. 00:32:16.548 --> 00:32:21.923 It's not just going to be an API to claw it or whatever. It's going to be stuff that you host. 00:32:22.173 --> 00:32:25.014 It's going to be stuff that people host for you. It's going to maybe on-prem 00:32:25.014 --> 00:32:28.584 if you've got – I think on-prem is going to make more sense now than it maybe 00:32:28.584 --> 00:32:32.437 did a year ago for certain data types and certain project types. 00:32:33.627 --> 00:32:37.494 Well, and I think we'll maybe get into the model landscape a little bit later, 00:32:37.494 --> 00:32:44.254 but I think there is even a lot of work to be done in understanding what tasks 00:32:44.254 --> 00:32:46.584 kind of map to different models, right? 00:32:46.584 --> 00:32:51.794 Like not every task is going to need the fable, the model in terms of compute 00:32:51.794 --> 00:32:57.214 and understanding maybe an open source model might be able to kind of be the 00:32:57.214 --> 00:33:01.574 underlying foundation for this group of tasks and being able to differentiate that. 00:33:01.574 --> 00:33:07.851 But I wanted to start with this kind of question of, again, putting ourselves in the seat of a CEO. 00:33:08.296 --> 00:33:11.404 Give us kind of a plain English test, right? 00:33:11.404 --> 00:33:16.404 How does a CEO recognize a task that's actually automatable versus one that 00:33:16.404 --> 00:33:17.284 just kind of looks like it? 00:33:17.284 --> 00:33:21.864 Because I think if they're looking at the landscape of the workflow that enables 00:33:21.864 --> 00:33:26.064 a contractor to put work in place, both at a project level and at an enterprise 00:33:26.064 --> 00:33:29.811 level, there's a there's a lot of potential candidates for automation. 00:33:30.136 --> 00:33:33.564 But I think there's also not always a really good understanding of that full 00:33:33.564 --> 00:33:38.074 lifecycle of a workflow where they're going to miss the human that is actually 00:33:38.074 --> 00:33:41.864 driving the value in a in a value stream or in a workflow. So how would how 00:33:41.864 --> 00:33:42.588 would you respond to that? 00:33:43.221 --> 00:33:47.474 Yeah, the classic way and classic to your point, it's not six years, 00:33:47.474 --> 00:33:51.274 it's more like six months. But the way to think about what you can automate, 00:33:51.274 --> 00:33:53.218 there's really three easy tests. 00:33:53.711 --> 00:33:57.544 One is, could I list all of the inputs that are part of this? 00:33:57.544 --> 00:34:00.074 Doesn't mean every time you run it, it's the same two or three things, 00:34:00.074 --> 00:34:03.028 but maybe it's only ever five, six, ten, whatever. 00:34:03.336 --> 00:34:06.196 So if I can list all of the inputs, it's what we call bounded. 00:34:06.353 --> 00:34:08.158 So there's a limit to it, right? 00:34:08.722 --> 00:34:11.834 The next one is, obviously, is it repeated? If we don't do this a lot, 00:34:11.834 --> 00:34:12.750 it doesn't make any sense. 00:34:13.203 --> 00:34:18.451 The really hard part is, can I verify it? Can I tell you this is what good looks like? 00:34:18.840 --> 00:34:21.824 Because I mentioned before, and this is why understanding how AI works, 00:34:21.824 --> 00:34:25.521 it's not that hard to get the fundamental idea of how it works, kind of a theory theory. 00:34:26.690 --> 00:34:30.460 It's a matching algorithm. So if I can't give it something, it can match its 00:34:30.460 --> 00:34:32.361 output against and say, that looks right. 00:34:32.965 --> 00:34:39.286 And it's complete. It's not almost there. It's 100%. Like, yes, this works. 00:34:39.646 --> 00:34:42.550 You'll hear the term evaluations. You'll hear the term evals, 00:34:42.550 --> 00:34:44.290 which, you know, why not make it shorter, right? 00:34:44.702 --> 00:34:47.910 That's what that means is can I write an eval or a set of them? 00:34:47.910 --> 00:34:51.670 It doesn't only have to be one that says, yes, the AI can go and say, 00:34:52.030 --> 00:34:56.438 I've gone and automated this thing and here's my output. And now I want to check it against that eval. 00:34:57.030 --> 00:34:59.694 And if I can't do that, then I can't automate it. 00:35:00.135 --> 00:35:03.478 But again, it doesn't mean you don't do something. It just means you say, 00:35:03.948 --> 00:35:06.470 maybe it's not an eval. Maybe it's just a filter. And it says, 00:35:06.470 --> 00:35:09.753 okay, it's mostly right. You know, see and have a look. And here's what I found. 00:35:10.037 --> 00:35:12.139 And AI is

pretty good at that, right? It's pretty good at saying, 00:35:12.970 --> 00:35:17.160 I took what you did. I did what you told me, and I checked it against other references. 00:35:17.491 --> 00:35:19.950 Here's what I've come up with. Here's what I think. 00:35:20.330 --> 00:35:23.190 Like, it can often guide you and say, I didn't know about this. 00:35:23.190 --> 00:35:26.830 This looks like something I couldn't verify. Does that make sense? Yes. 00:35:27.406 --> 00:35:31.700 But then when you think of it in those terms, it's like, how many real workflows 00:35:31.700 --> 00:35:35.350 could you actually write an eval for the entire workflow? And the answer is 00:35:35.350 --> 00:35:36.595 almost none. Maybe none. 00:35:36.926 --> 00:35:40.090 Right? Maybe some accounting ones because they're so numbers-based. 00:35:40.502 --> 00:35:43.550 So then you say, what tasks within it? Then it becomes a lot easier. 00:35:43.550 --> 00:35:46.990 Then it's like, okay, yeah, I can verify that and then hand it over to you to 00:35:46.990 --> 00:35:49.918 do what you need to do. And then you hand it back and so on. 00:35:51.032 --> 00:35:55.344 Well, and I think it also, AI is going to expose a number of different things. 00:35:55.344 --> 00:36:02.247 I think one of it is going to expose how little of our operational reality maybe we understand. 00:36:02.399 --> 00:36:06.414 Like, I just think there's a lot of companies that don't have their workflows mapped, right? 00:36:06.774 --> 00:36:11.914 And so I think maybe that's even part of your implementation is like, 00:36:12.395 --> 00:36:13.706 yeah, you need to go learn your company. 00:36:14.474 --> 00:36:18.104 I think so. I would balance that though a little bit. And, you know, 00:36:18.104 --> 00:36:21.894 I think one of the reasons people who are in construction like to stay in construction 00:36:21.894 --> 00:36:23.205 is they like to figure things out. 00:36:23.834 --> 00:36:28.524 So preserving that creativity, preserving that, you know, figuring it out and, 00:36:28.524 --> 00:36:31.404 you know, we don't have the, we have three of the four things that we need and 00:36:31.404 --> 00:36:32.894 let's just, you know, we'll figure it out. 00:36:32.894 --> 00:36:35.944 You know what I mean? Like, or we're waiting, you know, we're waiting for the 00:36:35.944 --> 00:36:40.164 safety guy, but can we get started, you know, looking up what probably is what 00:36:40.164 --> 00:36:41.738 we need to do for this situation. 00:36:42.162 --> 00:36:44.635 You see what I mean? Like, I think that... 00:36:45.567 --> 00:36:51.119 Yeah, you need to know your company. I think that they probably do, you know, pretty well. 00:36:51.119 --> 00:36:55.109 Like the question is, do they need to really know it more or do they need to 00:36:55.109 --> 00:37:00.799 say at the kind of operating level, you need to at least understand where your 00:37:00.799 --> 00:37:03.399 risks are and so on. And again, I think people know that.

00:37:04.162 --> 00:37:07.259 I think the mapping thing gets overstated a little bit. 00:37:07.434 --> 00:37:10.979 You know, what matters is, are they doing things, obviously, 00:37:10.979 --> 00:37:17.656 so that the quality is good, but also the big on lock and the thing that people are not so great at is, 00:37:18.271 --> 00:37:23.349 are we making sure that the way we enter information is uniform so that we can 00:37:23.349 --> 00:37:26.099 aggregate it into something we can analyze? 00:37:26.099 --> 00:37:29.309 That is the thing that, that, that is the really big change. 00:37:29.309 --> 00:37:33.569 So, you know, we talked about data before the biggest problem isn't technology. 00:37:33.569 --> 00:37:36.829 It isn't where things are sitting. I mean, that's usually pretty easy to fix. 00:37:36.829 --> 00:37:38.380 It's money, but it's pretty easy to fix. 00:37:38.757 --> 00:37:42.479 The hard part is getting 300 people across your company to say, 00:37:42.839 --> 00:37:44.498 you have to enter it this way. 00:37:45.139 --> 00:37:48.329 And then they say, well, I can't because, you know, this is a different project 00:37:48.329 --> 00:37:51.829 type. It's like, okay, then 90% has to be the same and you have to, 00:37:51.829 --> 00:37:55.297 you know what I mean? Like, how do we manage it when, when weirdness happens? Cause it does. 00:37:56.129 --> 00:38:03.369 So, you know, one of my maybe operating theses is that there's maybe there's 00:38:03.369 --> 00:38:07.339 like two parallels that are there's two things in motion happening, 00:38:07.339 --> 00:38:10.699 maybe in this kind of AI landscape at the individual level. 00:38:11.179 --> 00:38:17.819 It's everyone like me figuring out how I can use AI to be a better professional, 00:38:18.259 --> 00:38:20.236 to be all the different things. Right. 00:38:20.595 --> 00:38:23.839 Like this podcast, I really wouldn't be able to do the podcast without some 00:38:23.839 --> 00:38:27.014 version of AI just with all the other things going on. 00:38:28.199 --> 00:38:32.179 And we have this window of time where individuals like me can do all amazing 00:38:32.179 --> 00:38:36.259 things. Like I prototyped an interesting like field memory app over all weekend 00:38:36.259 --> 00:38:40.058 to like think through how we could do better use like voice to text for something. 00:38:41.553 --> 00:38:45.153 In parallel, you've got organizations trying to think through how they're going 00:38:45.153 --> 00:38:50.818 to, you know, at the end of the day, use AI to increase productivity from workers like me. 00:38:51.509 --> 00:38:56.283 And so if I'm trying to think through how I build a second digital brain, 00:38:56.283 --> 00:38:58.114 maybe using Notion or Obsidian, 00:38:59.345 --> 00:39:03.433 and my company is also thinking about how to extract more productivity from 00:39:03.433 --> 00:39:11.094 me and kind of leverage my expertise, I feel like there's a coming collision of some variety where, 00:39:11.717 --> 00:39:15.963 the enterprise is going to have to put some parameters and maybe this is just 00:39:15.963 --> 00:39:21.913 like everyone's going to have a like a very clear like data contract of like 00:39:21.913 --> 00:39:23.894 what is mine versus what is the company's, 00:39:24.772 --> 00:39:30.603 but you know for a podcast that is mainly kind of oriented to field operations 00:39:30.603 --> 00:39:34.460 leads superintendents pms you know our message often is. 00:39:35.793 --> 00:39:39.333 Leverage your expertise make sure you're documenting your expertise so you can 00:39:39.333 --> 00:39:40.212 leverage it in different ways. 00:39:41.145 --> 00:39:46.997 Do you see that collision happening? I think that's a really good place to go. 00:39:47.745 --> 00:39:53.483 And it's about to be made worse because Apple with Siri AI, which is still buggy 00:39:53.483 --> 00:39:58.783 and whatever, but they're going to, because they're using Gemini and like, 00:39:58.783 --> 00:40:01.579 so the AI quality is going to be there. And that was always the gap. 00:40:02.090 --> 00:40:04.963 They're going to make it useful enough that people are going to get used to 00:40:04.963 --> 00:40:08.678 having a digital personal assistant, whether you call it second brain. 00:40:08.974 --> 00:40:11.433 And I think there's almost two different things there. The second brain is, 00:40:11.433 --> 00:40:15.894 is a little richer, but, and, and takes more, more thought and more effort, 00:40:16.277 --> 00:40:19.863 but they're going to get used to being able to get really good answers about 00:40:19.863 --> 00:40:21.153 their lives and about things. 00:40:21.153 --> 00:40:23.903 And you already are. I mean, the number of times I hear, you know, 00:40:23.903 --> 00:40:28.173 a CEO or somebody who you could be surprised, they're like, you know, 00:40:28.173 --> 00:40:31.533 I, I did this because I had knee surgery and I learned all about what I'm going 00:40:31.533 --> 00:40:33.576 to need to do to rehab or whatever. Does that make sense? 00:40:33.843 --> 00:40:39.292 Yeah. So I think the personal assistant is going to make even people who are not that focused. 00:40:40.302 --> 00:40:46.372 On AI really good at just thinking about it now and really integrating it into their lives. 00:40:46.372 --> 00:40:49.412 And the next step after that is what you just said is, hey, I know all this 00:40:49.412 --> 00:40:52.892 stuff and I have these thoughts and I can just speak them into my phone, 00:40:53.135 --> 00:40:56.677 capture them, kind of convert them into something that's like a document. 00:40:57.042 --> 00:41:01.773 And over time, I'm going to have my own really rich lessons learned that I'm just a spreadsheet. 00:41:02.336 --> 00:41:08.472 How do I think about my expertise that is my IP versus is what the company wants. 00:41:08.472 --> 00:41:10.202 And no one has figured that out. 00:41:10.782 --> 00:41:15.906 I don't know that it's going to be, I don't think

it's going to come to a head in 2027, sorry, 2026. 00:41:16.365 --> 00:41:19.430 I do think it's going to start to matter next year. I think it's going to take, 00:41:20.011 --> 00:41:22.852 I can't really know this, but let's say six to nine months for people to get 00:41:22.852 --> 00:41:28.762 really used to what Siri can do, Siri AI can do as just a thing that's right there, right? 00:41:28.762 --> 00:41:32.182 And ChatGPT and Claude are both doing the same thing. They're just not as good 00:41:32.182 --> 00:41:34.036 at the interface like Apple is. 00:41:34.782 --> 00:41:37.922 It's going to come to a head. The other thing that early on somebody from one 00:41:37.922 --> 00:41:41.582 of the big HR companies, I think it was Challenger Gray, said, 00:41:41.582 --> 00:41:43.973 you're going to wind up hiring people plus their agents. 00:41:45.382 --> 00:41:48.342 He was talking in abstract and he didn't really know. I think it might not be 00:41:48.342 --> 00:41:52.065 agents so much as the documents and expertise that any agent can use. 00:41:52.401 --> 00:41:54.259 Because the agents will be a little bit of a commodity. 00:41:54.613 --> 00:41:57.382 It's more like, what are the instructions you load up into them and so on? 00:41:57.382 --> 00:42:00.923 And we don't really know that yet. Because the other thing, Steve and I hinted at before. 00:42:01.440 --> 00:42:08.829 Imagine what it would be like trying to create your cloud strategy in 2002 when there wasn't a cloud. 00:42:09.114 --> 00:42:12.312 That's what AI is like right now. We're not at the broadband. 00:42:12.312 --> 00:42:14.118 We haven't crossed into broadband yet. 00:42:14.443 --> 00:42:19.072 That was like 2005. That was a real moment when, well, actually, 00:42:19.072 --> 00:42:22.082 you look back in 06 and we're like, oh my God, we have broadband. 00:42:22.959 --> 00:42:26.482 That's going to happen with AI, and we can't know what it's going to look like 00:42:26.482 --> 00:42:29.942 because real companies have to actually put products in the market and they're 00:42:29.942 --> 00:42:31.611 going to make choices and trade-offs. 00:42:31.855 --> 00:42:34.176 Yeah. An example, I may have given in an earlier one, 00:42:34.808 --> 00:42:39.870 We were going to have smartphones no matter what, but we have iPhone-like smartphones because of Apple. 00:42:40.010 --> 00:42:42.735 Otherwise, they might have looked a little bit more like BlackBerry, 00:42:43.037 --> 00:42:46.996 which would have been pretty different and would have been much more boring. Yeah, yeah, for sure. 00:42:47.472 --> 00:42:50.302 I mean, I do like that, that analogy, because it's a good picture. 00:42:50.302 --> 00:42:51.291 And I think you're right. 00:42:51.675 --> 00:42:56.022 I think, you know, I've written an article about this, of like the job set of 2030. 00:42:56.022 --> 00:42:58.879 And obviously, everyone's kind of taking a stab at that. But like, 00:42:59.233 --> 00:43:05.062 I do think to your point, I imagine, you know, I did a bunch of DoD construction 00:43:05.062 --> 00:43:07.542 back in the day when I was working at RQ. 00:43:08.509 --> 00:43:13.102 And, you know, instead of RQ having to, or another company that's maybe getting 00:43:13.102 --> 00:43:17.622 into DoD work for the first time, having to hire someone, they license the data 00:43:17.622 --> 00:43:21.051 of someone who's worked in that sector, in that project type of vertical. 00:43:21.521 --> 00:43:26.165 And that helps, you know, understand the supply chain network in that region, 00:43:26.577 --> 00:43:28.876 you know, constructability around certain project types. 00:43:29.556 --> 00:43:36.549 So it is this call to, hey, okay, everything you know in your head is potentially licensable. 00:43:37.442 --> 00:43:42.712 And how do you put a strategy by your, you know, for your person together in 00:43:42.712 --> 00:43:46.413 parallel with whatever your company's doing maybe switching the question around 00:43:46.784 --> 00:43:51.752 for the pm or super who is maybe paying attention to what's out there is is 00:43:51.752 --> 00:43:57.500 using claude and you know to quite a great degree of success and and maybe, 00:43:58.282 --> 00:44:01.142 looking at their company as a little bit of a laggard, 00:44:02.054 --> 00:44:07.042 what should they do at a project level to help raise hey this is what we're 00:44:07.042 --> 00:44:11.046 able to do and and pitch maybe some ideas up up the ladder. 00:44:12.364 --> 00:44:16.311 The pitching up the ladder is a very company-specific thing. 00:44:16.671 --> 00:44:20.081 Some companies, there's a pathway for that. And others, you get people who get 00:44:20.081 --> 00:44:24.981 really frustrated because they're off in some other office with great ideas 00:44:24.981 --> 00:44:29.366 and nobody's there to listen to it. So that's going to really be dependent on the company. 00:44:29.991 --> 00:44:34.711 That said, most people I know, especially because when it comes to AI, 00:44:35.251 --> 00:44:36.799 there's a lot of anxiety in the industry. 00:44:37.194 --> 00:44:41.001 So leaders, I think, really appreciate when they hear that someone's doing a 00:44:41.001 --> 00:44:42.720 thing. They may not know what to do with it yet. 00:44:43.080 --> 00:44:47.731 They may not have a mechanism to say, great, we're going to host an event where 00:44:47.731 --> 00:44:49.889 you and two or three others can share it with everybody. 00:44:50.591 --> 00:44:53.959 So it really depends on the company and what they've chosen to do. 00:44:54.335 --> 00:44:58.751 I think using it, there's a few pieces to your point about someone at the PM level. 00:44:59.321 --> 00:45:01.695 I want to go back a little bit to your point about expertise. 00:45:02.305 --> 00:45:05.231 Claude, all of them are, but Claude is particularly good at, 00:45:05.231 --> 00:45:09.480 and Chachapiti's not bad either, at helping you to take what you think, 00:45:10.031 --> 00:45:13.761 even

if it's not really well formed yet, and restate it back to you, 00:45:13.761 --> 00:45:17.101 that you can then work with it to say, hey, I'm going to talk to you a little bit. 00:45:17.101 --> 00:45:22.503 And a way of thinking of that is to say, I'm going to go through a project from 00:45:22.805 --> 00:45:25.981 division one all the way through to the end of the project. 00:45:25.981 --> 00:45:30.565 So master format is one way to look at it. obviously the schedule is another but between those two, 00:45:31.134 --> 00:45:34.981 you can really and even depending on what the project is even the the drawings 00:45:34.981 --> 00:45:39.181 but that gets really detailed my point is using that as a as almost a table 00:45:39.181 --> 00:45:43.720 of contents to say on a sunday afternoon you know kids are out playing i'm going to spend, 00:45:44.230 --> 00:45:47.971 two hours and i'm going to go through a project and i'm going to tell stories 00:45:47.971 --> 00:45:51.091 about what mattered here and what you know how we saved the day here you know 00:45:51.091 --> 00:45:55.131 we had to do this we had to do that think about that as as claude interviewing 00:45:55.131 --> 00:45:57.797 you right and then you criticize you you kind of. 00:45:58.427 --> 00:46:04.675 Take that and say, all right, turn this into a reusable list of things that keep the context. 00:46:04.675 --> 00:46:10.645 So the hard part about lessons learned always are without context, they may not make sense. 00:46:10.645 --> 00:46:14.065 They may not be reusable again, right? So that has to be part of, 00:46:14.065 --> 00:46:17.985 just keep that in mind, is that that matching thing, something has to be able 00:46:17.985 --> 00:46:20.115 to look at what you did and match it to something else. 00:46:20.115 --> 00:46:22.755 So that's going to mean we need to know what project type, you're probably going 00:46:22.755 --> 00:46:25.906 to need to know some of the specifics of the materials or whatever it is. 00:46:26.475 --> 00:46:30.125 But the point, that's not hard to do. You know, and you also probably can upload 00:46:30.125 --> 00:46:33.545 some of that stuff. I want to be careful about what your company policy is to 00:46:33.545 --> 00:46:37.345 say, capture these in a way that I could use them again later, 00:46:37.345 --> 00:46:40.563 or someone that works for me could use them with, with an AI. 00:46:41.092 --> 00:46:45.695 I think that's a good thing to do for two reasons. One of them is just the capturing 00:46:45.695 --> 00:46:50.605 is smart, but also you learn things about AI when you go into longer projects 00:46:50.605 --> 00:46:53.733 that you don't, when you're just doing, you know, quick little prompts. 00:46:54.064 --> 00:46:58.265 When you try to do something that's like 15 turns long, you start to understand 00:46:58.265 --> 00:47:00.505 what drift means. Because it's like, dude, we're not doing that. 00:47:00.505 --> 00:47:03.085 We're doing this. How'd you forget what we're here for? 00:47:04.055 --> 00:47:07.787 And all that stuff only really happens when you're kind of deep into some sessions. 00:47:08.095 --> 00:47:10.202 So that would be one way to think about it. 00:47:10.798 --> 00:47:14.725 The other one, back to your point about how could they either sell something 00:47:14.725 --> 00:47:17.509 in or use it for their own projects... 00:47:18.201 --> 00:47:21.943 Definitely keep an eye on what the company's policies are just to make sure 00:47:21.943 --> 00:47:23.993 you're not using a tool you shouldn't use and so on. 00:47:24.428 --> 00:47:30.698 But after that, I mean, I think it's a really good idea to think about what could we spin up? 00:47:31.663 --> 00:47:34.790 What could I get a PE and say, hey, I need you to take two hours, 00:47:35.075 --> 00:47:37.675 go into co-pilot and see if you can automate that thing. 00:47:38.105 --> 00:47:40.613 And if he says, I don't know how to do that, it's like, that's the point of 00:47:40.613 --> 00:47:44.386 AI. Tell it what you want to do and tell it to give you step-by-step directions. 00:47:44.757 --> 00:47:49.210 So A, you're learning as the PM, but B, you're growing your PE to be better at that stuff. 00:47:49.639 --> 00:47:51.643 You know what I mean? And then you can say, we're going to have a meeting and 00:47:51.643 --> 00:47:53.443 I want to know what you found. What didn't work? 00:47:54.289 --> 00:47:56.733 What did work? What was easier than you thought? So on and so forth. 00:47:57.163 --> 00:48:01.104 And that shouldn't take an enormous amount of time. I mean, no one has a lot of time to spare. 00:48:01.493 --> 00:48:05.237 But that's a good way of thinking about it because I think you're going to find, 00:48:05.690 --> 00:48:11.803 maybe it isn't efficiency per se, but less time on answering silly questions. 00:48:12.128 --> 00:48:19.278 Like you can spend more time on the hard stuff and less on just catching obvious mistakes. 00:48:19.953 --> 00:48:25.898 Yeah, I think the whole, like, ask the model what you're trying to learn about 00:48:25.898 --> 00:48:30.378 the model is a really underrated approach to, you know, first embracing AI. 00:48:30.398 --> 00:48:34.208 Yeah, I think it's a great point. Well, it's such a weird idea, right? 00:48:34.208 --> 00:48:37.578 Like, you can't ask your car how to fix itself. Like, the fact that it's the 00:48:37.578 --> 00:48:41.269 first technology in history that can do an okay – now, I will tell you, 00:48:41.687 --> 00:48:44.368 be careful because AI will tell you it can do things it can't, 00:48:44.368 --> 00:48:46.748 and it'll tell you other software can do things it can't. 00:48:46.748 --> 00:48:52.078 It will always overstate and it really doesn't appreciate how limited AI itself is. 00:48:52.385 --> 00:48:55.708 So it kind of, it buys the hype that you shouldn't buy. You know what I mean? 00:48:55.708 --> 00:49:01.088 So it'll say it can do things that, that absolutely underestimate how

bad the 00:49:01.088 --> 00:49:03.432 inputs are or how variable things can be. 00:49:03.688 --> 00:49:06.998 So take it as a guide and just discount it a little bit and really try it out 00:49:06.998 --> 00:49:08.228 and actually call it out. 00:49:08.228 --> 00:49:11.938 The other thing is to say, I need you to make sure that, that we wouldn't run 00:49:11.938 --> 00:49:15.848 into, you know, you know, errors or hallucinations or confabulations. 00:49:15.848 --> 00:49:19.158 That's the other thing is sometimes it looks right, but it's taken two ideas 00:49:19.158 --> 00:49:22.218 and combined them or it inferred that's true. That must be true. 00:49:22.218 --> 00:49:25.197 So it doesn't read like a bad fact, but it's still not right. 00:49:25.874 --> 00:49:29.018 Right. But I think even that's what's interesting about where we're at now is, 00:49:29.018 --> 00:49:32.956 is you can not just ask it kind of theory, hey, what can you do? 00:49:33.362 --> 00:49:36.688 Then go do it and then see how well it does. Right. Like that's the thing that's 00:49:36.688 --> 00:49:40.438 kind of blowing my mind with cloud code right now is, you know, 00:49:40.438 --> 00:49:45.088 it has a really deep underground understanding of, of its abilities to integrate 00:49:45.088 --> 00:49:46.220 with almost everything. 00:49:46.714 --> 00:49:50.928 And it'll go try and it'll tell you exactly what broke. It'll tell you exactly 00:49:50.928 --> 00:49:52.002 how to potentially fix it. 00:49:52.344 --> 00:49:56.118 So, yeah, I think the whole brokerage of trying it, I think, is underrated. 00:49:56.623 --> 00:50:00.938 That is, though, back to the token thing. What you just talked about is a different 00:50:00.938 --> 00:50:03.832 level from what I was talking about, which is a little bit more words. 00:50:05.078 --> 00:50:07.868 You need to know that if people are going to go try to spin things up, 00:50:07.868 --> 00:50:11.292 that is where tokens start to matter because that is where tokens get burned a lot. 00:50:11.785 --> 00:50:14.838 And it's okay. I think it's just you want to budget it. You want to say, 00:50:14.838 --> 00:50:18.698 all right, you have a $300 a month budget to go try things. And if you need 00:50:18.698 --> 00:50:22.169 more than that, tell us. you know, pitch us for why you need a thousand bucks or something. 00:50:22.852 --> 00:50:28.837 Yeah, you know, via Technic, we have kind of an approach where it's like we, you have access. 00:50:29.317 --> 00:50:32.901 And if you need more access, you put kind of a business case together. 00:50:34.037 --> 00:50:36.897 And it doesn't have to be long. And you can use AI to put it together, 00:50:37.297 --> 00:50:38.681 but explain what you're trying to do. 00:50:39.052 --> 00:50:44.674 So, you know, KP Reddy kind of had the comment of the kind of token usage as the new timesheet. 00:50:45.092 --> 00:50:49.811 And I'd be interested to get an update from him if he still thinks that. 00:50:50.468 --> 00:50:57.117 But it's this idea of like your output is being measured by token usage and 00:50:57.117 --> 00:50:59.739 your kind of impact on the business is associated to that. 00:51:00.192 --> 00:51:04.077 So I think it makes a lot of sense in the businesses that he's in where there's 00:51:04.077 --> 00:51:08.190 a lot of analysis. There's a lot of code. There's a lot of presentation and so on. 00:51:08.602 --> 00:51:13.097 I'm not sure that fits for construction. So much of the value is two people 00:51:13.097 --> 00:51:15.967 talking to each other, you know, or thinking and so on. 00:51:15.967 --> 00:51:19.697 So I don't know that it's not, I think it's an interesting thing to keep an 00:51:19.697 --> 00:51:25.277 eye on and say, hey, and I think if anything, token usage is a way of thinking 00:51:25.277 --> 00:51:29.377 about where your pockets of like self-taught experts are, right? 00:51:29.377 --> 00:51:32.937 As you're transforming the company, nothing is better than somebody who is a 00:51:32.937 --> 00:51:34.707 peer saying, I can do this. 00:51:35.032 --> 00:51:38.457 That's right. Yeah, I do training all the time. And when we do it, 00:51:38.457 --> 00:51:42.167 we include a workshop as most of the, actually most of the session is a little 00:51:42.167 --> 00:51:45.365 of me talking, but then it's people doing things and then talking to each other. 00:51:45.754 --> 00:51:50.781 Because this is, by the way, one of the oldest findings in technology adoption 00:51:51.199 --> 00:51:54.237 is that listening to peers is critical. 00:51:54.617 --> 00:51:56.487 Nothing is more important than that. Right. 00:51:58.152 --> 00:52:02.726 Here's a question. So you're advising CEO, president, leadership team. 00:52:03.422 --> 00:52:06.764 Do you guys get into orchestructure and what it is now versus what it should be? 00:52:07.816 --> 00:52:12.803 Yes, to a degree. You know, thinking about, 00:52:13.357 --> 00:52:17.786 accountability, thinking about who's going to own what of the things we propose 00:52:18.088 --> 00:52:21.465 and how that fits with what they're already doing, because they usually have, 00:52:21.465 --> 00:52:24.649 as I say, an AI committee. They obviously already have a CEO. 00:52:25.234 --> 00:52:28.415 They often will have a standards and process committee. 00:52:29.115 --> 00:52:33.505 It may not meet that much, but they have one. So there's already pieces in place. 00:52:33.505 --> 00:52:35.114 The question is, how do they knit together? 00:52:35.526 --> 00:52:40.365 And what is the ultimate accountability for AI? Now, ultimately, 00:52:40.365 --> 00:52:42.515 ultimately, that's the CEO, but that can't be the day job. 00:52:43.815 --> 00:52:47.535 Someone has to report to the CEO. And who is that? So often, 00:52:48.575 --> 00:52:54.236 whether it's a CIO, but the problem with CIOs is they are always a staff function. 00:52:54.585 --> 00:52:57.933 Which means that they're not from the field and they have an authority gap,

00:52:58.347 --> 00:53:01.829 no matter how smart they are, no matter how good they are, even if they started in the field. 00:53:03.015 --> 00:53:05.805 And that's true. Again, that's not a construction problem. That's true in every 00:53:05.805 --> 00:53:09.264 industry is when you're in a support position, you have an authority gap. 00:53:09.618 --> 00:53:14.055 So when possible, you want to pair the CIO with someone who has authority like 00:53:14.055 --> 00:53:17.455 that, who also, by the way, is usually closer to the work itself anyway, 00:53:17.455 --> 00:53:21.233 so can really translate and bring a level of detail that it's hard for a CIO. 00:53:21.553 --> 00:53:25.385 Because you actually want a CIO. It's great that they have construction industry 00:53:25.385 --> 00:53:28.623 experience, but what you really need is that they know what scale looks like. 00:53:28.884 --> 00:53:31.444 They know what happens the 300th time something runs. 00:53:31.717 --> 00:53:36.054 They think like a systems thinker, not a process thinker. 00:53:36.472 --> 00:53:40.915 They need to do process too, but the thing that makes CIO different from other 00:53:40.915 --> 00:53:45.905 people is they're worried about what is going to happen across 300, 1,000, whatever. 00:53:46.201 --> 00:53:48.775 And that's not something that construction teaches you per se. 00:53:48.775 --> 00:53:54.438 It's something that you need to have learned as a CIO or somewhere else or in that role. 00:53:55.600 --> 00:54:00.597 So you don't just walk in with a blinking orb and say, this is your boss now? Well, we did try that. 00:54:02.668 --> 00:54:07.961 We had adoption problems. But no, the org chart thing is, as I say, 00:54:08.175 --> 00:54:13.268 that's as far as we're likely to go, except maybe thinking about train the trainers, 00:54:13.268 --> 00:54:15.437 which is not really org. It's more like implementation. 00:54:15.780 --> 00:54:19.028 Beyond that, we can't know their business well enough, and we look like idiots 00:54:19.028 --> 00:54:20.648 telling them that you should move people around. 00:54:20.648 --> 00:54:24.758 Like, you need to be really humble about anything org-related because they're 00:54:24.758 --> 00:54:28.428 thinking of the people that they have lunch with, you know? Like, 00:54:28.428 --> 00:54:31.024 you got to fire that guy. He's not that good. No, I don't get to say that. 00:54:31.245 --> 00:54:33.015 Yeah. Yeah, I think that's fair. 00:54:33.607 --> 00:54:38.348 So maybe just a short kind of segue to talk about the models real quick. 00:54:38.348 --> 00:54:43.128 I think I listened to a podcast called Limitless, and it does a fairly good job. Great podcast. 00:54:43.468 --> 00:54:47.098 Yeah, keeping track with all the things that are happening in the AI category. 00:54:47.568 --> 00:54:52.428 But, you know, Fable 5 kind of came back released. It's like an iteration or 00:54:52.848 --> 00:54:57.647 a derivative of the kind of infamous mythos model

from Anthropic. 00:54:58.459 --> 00:55:02.708 But I think someone not tracking these things can see all these different models 00:55:02.708 --> 00:55:07.278 dropping throughout the months and just kind of wonder how applicable is this to me? 00:55:07.278 --> 00:55:10.918 So what actually changes? Like if you're a contractor that has maybe an enterprise 00:55:10.918 --> 00:55:16.078 license with Claude or OpenAI, these models drop, what does that mean to the 00:55:16.078 --> 00:55:19.982 folks on the day-to-day using these things? Does anything change? 00:55:20.727 --> 00:55:24.682 Rarely. And I think it's important not to get swept up in that stuff. 00:55:25.182 --> 00:55:28.152 We are in a really rapid moment of change. 00:55:28.889 --> 00:55:33.712 Much more important than model changes is the system that the model is a part of. 00:55:33.712 --> 00:55:37.695 So whether you call it a harness or it's in co-pilot or it's in co-work or whatever, 00:55:38.241 --> 00:55:42.542 the system and the things that it's connected to, the reference documents that 00:55:42.842 --> 00:55:46.012 you've created that allow it to check whether that's right or wrong, 00:55:46.012 --> 00:55:49.382 or you can do things like, here's how I want my emails to be written. 00:55:49.382 --> 00:55:54.662 So all of that is more important than the models from a little increment to 00:55:54.662 --> 00:55:58.812 little increment. Now, that said, there are moments when truly new things happen. 00:55:59.068 --> 00:56:03.632 So when OpenAI's ImageGen came out, that was different. That was the first time 00:56:03.632 --> 00:56:06.962 you could make images that have text that's right. And then Gemini did a really 00:56:06.962 --> 00:56:09.642 good job. They really did better. And then with Nano, Banana, 00:56:09.642 --> 00:56:12.466 and both of them have an updated version of it. 00:56:12.953 --> 00:56:16.372 That's actually really interesting because now you can make signs for people. 00:56:16.372 --> 00:56:18.509 And the number of people I know, that's really cool. 00:56:18.811 --> 00:56:21.242 The number of superintendents and safety people that are like, 00:56:21.242 --> 00:56:23.362 yeah, I can't really do a lot with AI. And I'm like, really? 00:56:23.362 --> 00:56:27.302 And then you talk to them like, well, I made a sign and then I translated it into Spanish. 00:56:28.203 --> 00:56:32.748 So they're doing really useful stuff. That matters because it's a genuinely new capability. 00:56:33.102 --> 00:56:38.262 Fable is different enough that whether it is better, that's true, 00:56:38.262 --> 00:56:42.672 but it's also, it behaves differently because of this additional intelligence. 00:56:42.672 --> 00:56:47.882 It tends to summarize things and it tends to make references that you have to go then look up. 00:56:47.882 --> 00:56:51.771 You have to tell it. I literally have a skill in Claude that says human readable, 00:56:52.224 --> 00:56:56.431 because it'll do this goofiness where it speaks in like code almost because 00:56:56.699 --> 00:57:00.592 it's so good. I mean, you know, so that, that, but that's just more like a style 00:57:00.592 --> 00:57:02.045 thing that you just have to correct for. 00:57:02.312 --> 00:57:06.412 But it also means that the evaluations that you created for an earlier model 00:57:06.412 --> 00:57:10.805 definitely need to be looked at again because it might behave differently against them. 00:57:11.165 --> 00:57:16.406 My personal sense is, and again, I'm not alone in this. This is kind of best practices. 00:57:17.072 --> 00:57:20.735 Don't swap in new models if what you have right now is working. 00:57:21.332 --> 00:57:25.185 Just keep doing it. Then look at the new model and say, this is unlocked things 00:57:25.185 --> 00:57:26.278 that we couldn't do before. 00:57:26.667 --> 00:57:29.975 So like Fable can handle more complexity and it's better at, 00:57:29.975 --> 00:57:31.792 you know, keeping track of things and all that. 00:57:32.083 --> 00:57:36.005 So it might allow you to do something you couldn't do before. 00:57:36.005 --> 00:57:39.685 Maybe look at more of the project at once. Maybe give you some thoughts on schedule 00:57:39.685 --> 00:57:41.550 that are just beyond the earlier models. 00:57:41.875 --> 00:57:44.360 Schedule is a tougher one than people, I think, think. 00:57:44.789 --> 00:57:47.365 I don't think construction people think it's not hard. I'm just saying people 00:57:47.365 --> 00:57:51.206 that come to, hey, we've got a schedule optimization. Oh, do you? 00:57:51.723 --> 00:57:54.345 Yeah. That's pretty tough to do. And I know you've spent a lot of time with 00:57:54.345 --> 00:57:58.225 Alice, so you understand how hard that is to do. And also how hard it is to 00:57:58.225 --> 00:58:03.089 offer up something that isn't like, yeah, I know, or what, right? 00:58:03.252 --> 00:58:07.803 That's one of the problems with AI where something like Fable does do a better job is. 00:58:08.600 --> 00:58:12.416 Giving you a shorter list of things that are more likely to be relevant to you. 00:58:12.736 --> 00:58:15.906 That's where models get better. So as you're thinking about all the lessons 00:58:15.906 --> 00:58:20.036 learned and all that, better models will do a better job of choosing from, 00:58:20.036 --> 00:58:21.464 because they're looking at more at once. 00:58:21.603 --> 00:58:24.488 You know what I mean? They're saying, considering 15 factors, not five. 00:58:24.755 --> 00:58:28.236 So I'm going to look at your list of 100 or 1,000 lessons learned, 00:58:28.236 --> 00:58:30.229 and I'm just going to pull the three that look like they're relevant. 00:58:30.409 --> 00:58:31.593 That's where models will get better. 00:58:31.918 --> 00:58:34.985 But my point is, everything's changing so fast. 00:58:35.774 --> 00:58:39.366 I think it's fine if you want to pay attention to that. like on LinkedIn or 00:58:39.366 --> 00:58:43.316 on Twitter or X, whatever, just don't

get spun out by it. 00:58:43.316 --> 00:58:46.766 Like this is a moment where the industry is figuring a lot of things out. 00:58:46.766 --> 00:58:50.686 There's a lot of arguments. There's a lot of people acting as if the current 00:58:50.686 --> 00:58:52.971 moment is going to be forever. 00:58:53.531 --> 00:58:57.376 And that's why it's really important to think that we're on a really fast, 00:58:57.376 --> 00:58:58.785 but we're still on a ramp. 00:58:59.139 --> 00:59:02.958 Things are changing and big things haven't been fully worked out yet. 00:59:03.266 --> 00:59:06.790 So don't get too caught up in what you're hearing on LinkedIn or on X. 00:59:07.144 --> 00:59:10.616 Because a lot of that's just people who I guess have more time than they need, 00:59:10.616 --> 00:59:13.656 you know, like, and they get all wound up about, this is the strategy. 00:59:13.656 --> 00:59:18.446 They're going to screw us. It's like, maybe not the way you think they need to make a lot of money. 00:59:18.446 --> 00:59:23.326 But I think the analogy is like, if, if we're all runners and the models are 00:59:23.326 --> 00:59:28.506 shoes, if you haven't gone for a jog in three months, like it doesn't really 00:59:28.506 --> 00:59:32.956 matter if you buy the latest like cushion Nike max air pro, 00:59:33.376 --> 00:59:35.681 like just like go for a jog, you know? 00:59:36.290 --> 00:59:39.367 That's right. Well, and you know, there's a term that they use. 00:59:39.727 --> 00:59:44.047 It came from the benchmarks where they call it saturated. So the model is so 00:59:44.047 --> 00:59:49.170 good that it's at 95% or 99% of that benchmark. So it's saturated. 00:59:49.530 --> 00:59:54.227 That idea also works for tasks. So you ask yourself, did GPT-4, 00:59:54.552 --> 00:59:56.557 was that good enough for this little thing that I need? 00:59:56.557 --> 00:59:59.727 Great. Then I don't need to pay all that much money. And the model makers know 00:59:59.727 --> 01:00:04.387 that too. That's why they're trying to make these router models that will look 01:00:04.387 --> 01:00:09.974 at your request and your task and route it to the right. It's harder to do than people think. 01:00:10.230 --> 01:00:14.015 They tried it like two years ago and it failed. This was ChatGPT. They did a router. 01:00:14.807 --> 01:00:18.217 So what you will absolutely get to the point I think you're generally saying 01:00:18.217 --> 01:00:22.117 is you're going to get to a place where certainly a company, 01:00:22.117 --> 01:00:25.367 but even an individual, you're going to have like three or four levels of models. 01:00:25.367 --> 01:00:29.467 And either you're doing it, which no one remembers to do, or the model or the 01:00:29.467 --> 01:00:32.527 system itself is going to say, this looks like that kind of task. 01:00:32.527 --> 01:00:36.767 I can route it to sonnet this looks like it's hard until we figure it out we're 01:00:36.767 --> 01:00:39.387 gonna have fable look at it and the other thing you're seeing and again who 01:00:39.387 --> 01:00:44.767 knows what this lasts fable is your in is is the start and then it delegates to smaller models, 01:00:45.307 --> 01:00:48.967 and claude is definitely looking at doing that how exactly it works we'll see, 01:00:49.699 --> 01:00:54.427 so last question you a lot of nuggets in this this conversation so i appreciate it, 01:00:55.527 --> 01:00:59.467 let's let's maybe put ourselves in the shoes of an organizational leader who 01:00:59.467 --> 01:01:04.647 is feeling behind i think everyone generally feels behind right and and whether 01:01:04.647 --> 01:01:08.166 or not they're they're the ceo president or or on the leadership team, 01:01:08.819 --> 01:01:13.747 maybe they're not using it themselves not using any variety of of clod or anything, 01:01:14.206 --> 01:01:18.227 we've spoken to just maybe what that start looks like but but what would be 01:01:18.227 --> 01:01:20.608 the overarching message to that person. 01:01:22.121 --> 01:01:24.888 What's the old adage? The best time to plant a tree was 20 years ago, 01:01:24.888 --> 01:01:26.271 and the second best is right now? 01:01:27.502 --> 01:01:33.298 I think the thing I would say is, yes, you should start if you haven't started. 01:01:33.298 --> 01:01:37.458 Yes, you should accelerate if you have started, but don't get overwhelmed by it. 01:01:37.798 --> 01:01:40.359 And remember, you already know the hard stuff. 01:01:41.978 --> 01:01:45.938 Knowing how to run a construction company is infinitely more complex than what 01:01:45.938 --> 01:01:47.228 they need to know about AI. 01:01:47.228 --> 01:01:50.748 This is just another tool, right? Of course, just complexity behind the scenes, 01:01:50.748 --> 01:01:52.758 but they don't need to know that any more than they need to know how to build 01:01:52.758 --> 01:01:55.968 a drill or a backhoe. You know what I mean? Like, they already know the hard 01:01:55.968 --> 01:02:00.676 stuff. The really valuable information is in their minds and in their employees. 01:02:01.024 --> 01:02:05.978 So, you know, look at this as probably the most important transformation that 01:02:05.978 --> 01:02:09.448 you've done as a company, but it doesn't mean you're root and branch changing 01:02:09.448 --> 01:02:12.508 everything. It means that this is something you absolutely can do in stages. 01:02:12.508 --> 01:02:15.820 It's something you can do in manageable chunks without creating new risk. 01:02:16.162 --> 01:02:18.705 It means that, you know, in six months to a year. 01:02:19.401 --> 01:02:23.180 You're going to be in a place, you can easily be in a place where you're not 01:02:23.180 --> 01:02:26.228 at a competitive disadvantage when that owner comes and says, 01:02:26.576 --> 01:02:29.043 tell me your AI strategy and show me what you mean. 01:02:29.455 --> 01:02:32.730 Whereas Skanska came in and said, we've got 50 bots over here and we've got 01:02:32.730 --> 01:02:35.353 a team of, you know,

white coats that are, you know, in our lab. 01:02:35.678 --> 01:02:39.470 Like there will be a competitive need to do this. that you're going to be at 01:02:39.470 --> 01:02:42.800 a, it's a little bit like BIM, like, you know, for a while BIM was optional 01:02:42.800 --> 01:02:46.300 and now it's really not, you know, and same thing with a PM system for a while. 01:02:46.300 --> 01:02:50.110 That was no one really remembers that, but for a while it was kind of optional and now it's really not. 01:02:50.110 --> 01:02:55.066 And being, having a solid AI implementation, having solid governance, 01:02:55.327 --> 01:02:58.890 being able to say, you know, weird stuff isn't, isn't being sent to you because 01:02:58.890 --> 01:03:03.657 no one checked the AI, having all that in place is going to become a competitive necessity. 01:03:04.058 --> 01:03:09.540 Quite when will depend on what project type you have, If you're in mission critical, 01:03:09.540 --> 01:03:10.995 it's probably essential right now. 01:03:11.395 --> 01:03:18.180 If you're doing local, I don't know, five-story high-rises or retail, maybe you have longer. 01:03:18.180 --> 01:03:22.380 But ultimately, it's going to be something where you're going to be at a disadvantage 01:03:22.380 --> 01:03:26.830 because people will say, you're not going to be as capable or you're charging 01:03:26.830 --> 01:03:28.892 me more than someone who's better at it. 01:03:29.941 --> 01:03:34.271 How do they get in contact with you, Hugh? LinkedIn? Yes. So actually, 01:03:34.271 --> 01:03:38.081 there's a couple things. One is I've actually partnered with a guy, 01:03:38.081 --> 01:03:41.621 Jay Snyder, who half the audience will probably know. So we actually together 01:03:41.621 --> 01:03:44.226 have, so the link is a very technical part of this. 01:03:44.726 --> 01:03:48.841 BBI, his company, has done tech strategy and a bunch of stuff for a long time. 01:03:48.841 --> 01:03:52.769 We're calling that offering AI Edge. So there's actually, and I'll give you the URL for it. 01:03:53.350 --> 01:03:56.362 I can always be found on LinkedIn. I love hearing from people. 01:03:56.687 --> 01:03:59.731 You know that the offering which combines you know all of what i've done and 01:03:59.731 --> 01:04:02.911 all of what what jay's team have done is you know it gives us depth it gives 01:04:02.911 --> 01:04:07.591 us about 30 people but also it means that there's just a lot of a lot of expertise 01:04:07.591 --> 01:04:09.710 we can bring together with this this ai edge thing. 01:04:10.645 --> 01:04:16.181 That's fantastic well i appreciate you coming on you also have a book coming out at some point i do, 01:04:16.681 --> 01:04:21.201 i do and so there's a series of them so the first one will be in august and 01:04:21.201 --> 01:04:24.001 it'll be self-published so that it's relevant instead of waiting the nine months 01:04:24.001 --> 01:04:27.661 to get something like the last one I did, and that's just going to be using AI. 01:04:27.661 --> 01:04:31.691 And it's a series of, it's very tactical. I mean, there's some upfront, 01:04:31.691 --> 01:04:34.301 about 20 pages or so of, you know, here's how to think about AI. 01:04:34.585 --> 01:04:38.631 But then we get into a series of stuff you can do to learn as somebody who, 01:04:38.631 --> 01:04:42.440 and so it's very, very, like my first book, it's very focused on day-to-day users. 01:04:42.776 --> 01:04:45.421 The one after that will be enterprise AI, where I've talked to, 01:04:45.421 --> 01:04:49.631 you know, 20 to 30 CIOs and really talk about some of the questions we talked 01:04:49.631 --> 01:04:51.147 about here. That'll be in Q4. 01:04:51.691 --> 01:04:55.931 Fantastic. Well, that's just a lot of, A lot of things happening in the Hugh 01:04:55.931 --> 01:04:58.645 world, but thanks for coming on and taking some time with us, my man. 01:04:59.110 --> 01:05:02.352 Always a pleasure. I really appreciate it. Thanks, Steve. All right, see you.