Jakson Hugh Seaton täydellinen litterointi

there's a bit of a through line in really loving to create the conditions for people to do what they do best. 00:05:01.889 --> 00:05:04.276 Yeah, let's talk about your current startup, the link. 00:05:04.276 --> 00:05:15.554 So it's, from what I understand, a construction specifications management platform that pulls data and insights from construction specifications for project teams and, as you mentioned, it seems to be very well connected to your past. 00:05:15.554 --> 00:05:22.413 Can you walk us through a specific example or case study where the link has significantly streamlined the construction process for a project team? 00:05:23.647 --> 00:05:26.616 Yeah, let me talk about the way it all started. 00:05:26.616 --> 00:05:48.809 So I was lucky enough to be recruited to go work for CSI Actually, the company that held my W-2 is supplier to CSI, a company called Sympatry, an amazing group and I was asked to go and help them to grow and continue to create their technology and their software offering. 00:05:48.809 --> 00:05:51.831 That gave me a lot of visibility into a few things. 00:05:51.831 --> 00:06:04.498 One of them is some of the standards that they use, but the other one was how specifications have evolved over time and how they kind of haven't evolved as other things have. 00:06:04.498 --> 00:06:08.275 So specifications in construction right now are still documents. 00:06:08.275 --> 00:06:13.528 Fundamentally they behave the same way they did in the 70s when they were being couriered around to each other. 00:06:13.528 --> 00:06:22.665 Yes, they're PDFs, yes, they get emailed and so on and so forth, but there isn't the same reason for them to be all in one document quite like there was. 00:06:22.665 --> 00:06:32.574 For a number of reasons it's going to stay that way for a bit, and one of them is just legal they're all in one place that someone can stamp and put their reputation behind. 00:06:34.107 --> 00:06:37.795 There are ways in software to do it differently than that, but that's one thing at a time. 00:06:37.795 --> 00:06:50.490 So, to answer your question, you wind up getting especially in a classic design bid build project, but even in design, build and a number of other project delivery methods, you get a bunch of specs all at once. 00:06:50.490 --> 00:06:59.811 The project team has to onboard these specs really quickly understand them, start to write contracts for subcontractors and so on, and it's really not that easy. 00:06:59.811 --> 00:07:09.632 It's talking about 2,500 to 4,000, sometimes more pages, a pretty dense, legalese text not always legalese, but certainly technical text. 00:07:09.632 --> 00:07:11.689 That's hard for people to navigate. 00:07:11.689 --> 00:07:15.514 It's hard for them to get through it in a way that is coherent. 00:07:15.514 --> 00:07:22.551 They manage, but they usually manage by putting bodies on the problem and by getting it, getting through it pretty painfully. 00:07:22.551 --> 00:07:25.150 So

what we do is the very first thing is just to do. 00:07:25.150 --> 00:07:31.653 Some of the outcomes early in the project that you want are like a submittal log and some other things. 00:07:31.653 --> 00:07:37.498 So the first thing that we do is make the onboarding of the spec faster and easier. 00:07:37.665 --> 00:07:48.857 So the next piece, though, mark, that more clearly answers your question, is the idea that after the beginning of the project, people tend to look at the spec a lot less than they did In the beginning. 00:07:48.857 --> 00:07:49.458 You don't have a choice. 00:07:49.458 --> 00:07:54.107 You have to go through and understand what you're required to do and what you're on the hook for. 00:07:54.107 --> 00:07:59.290 But later, once you're building, people have a tendency to look at the drawings a lot more than they do the specs. 00:07:59.290 --> 00:08:05.229 The problem is someone said this once and I love it is lawyers don't read drawings, they read specs. 00:08:05.229 --> 00:08:10.470 So at the end of the day, if there's a claim or a problem, they're going to go to the specs, not the drawings. 00:08:12.425 --> 00:08:24.757 So what we do now is we help people to go from this older kind of mentality where it's almost like an old map right, like a Rand McNally map that you only use if you're going really far or if you're lost in the middle of a desert. 00:08:24.757 --> 00:08:29.956 And the analogy in construction is if it's a big deal or it's a disaster, okay, we'll go through the specs. 00:08:29.956 --> 00:08:35.189 But otherwise people on the project know how to build, they know how to pour concrete, they know how to install windows. 00:08:35.189 --> 00:08:45.235 So the temptation is to go with muscle memory and then things get caught expensively and redone with rework or they don't get caught and then you have a claim or even a lawsuit. 00:08:45.235 --> 00:08:58.076 So what we look to do, and what we've been doing, is going from that paradigm of an old map to Google Maps Now it's so easy or ways or whatever you name the digital map. 00:08:58.144 --> 00:09:03.807 But the point is, if you're even a little unsure, you pull it up and go check and that's the big shift that you see. 00:09:03.807 --> 00:09:08.036 In for project teams is a superintendent, a PE, the PM. 00:09:08.036 --> 00:09:16.568 If they're even a little unsure, they're like yeah, I'm pretty sure I know this, but the barrier is so low to just ask, because our system will then go into the specs, find the answer, get it to you in. 00:09:16.568 --> 00:09:49.089 You know two seconds right, and what that allows you to do is really reduce the risk of people going on autopilot or going, more accurately, on muscle memory, and instead they can go consult the contractual document and say, all right, we're sure we're right, we're covered. 00:09:52.211 --> 00:09:52.652 That's awesome. 00:09:52.652 --> 00:09:54.539 I'm so curious about it. 00:09:54.539 --> 00:10:02.502 Can you speak a little bit about you know training the model or using LLMs, and you know the data set and all of that, yeah. 00:10:03.194 --> 00:10:04.278 And we do use LLMs. 00:10:04.278 --> 00:10:11.541 Specifically right now we're using OpenAI GPT-4 plus, but you know, every 15 minutes there seems to be another one. 00:10:11.541 --> 00:10:12.744 But that's okay. 00:10:12.744 --> 00:10:15.223 What we didn't do is train a model. 00:10:15.223 --> 00:10:25.366 What we do is take the model as it is and fine tune it, because that space is moving so fast and there's no way we're going to I mean, they're raising billions of dollars. 00:10:25.366 --> 00:10:32.247 So the state of the art in AI models is not something a startup at our size is going to engage in. 00:10:32.714 --> 00:10:45.907 What we're engaging in is how do we take these emerging capabilities and make them useful, make them into products, because chat GPT or, yeah, chat GPT is interesting and it's a generic problem, like Google is. 00:10:45.907 --> 00:10:53.582 But if you want a specific product, something that has security, something that has workflows around it, you need to build actual software. 00:10:53.582 --> 00:10:54.838 So that's what we focus on. 00:10:54.838 --> 00:11:03.041 An example of that is most PMs and superintendents often reference the last job or two jobs ago when they're just thinking about their current job. 00:11:03.041 --> 00:11:12.825 What we allow them to do is upload those specs to and point spec GPT at that also, so they can find the thing that there's on the tip of their tongue or what did we use in that last one. 00:11:12.825 --> 00:11:22.103 Well, they can go find it like that and then continue to use past experience to influence and really help them to execute the project at the right minute level. 00:11:23.674 --> 00:11:26.723 I love that you and I'm super excited about diving even deeper with you. 00:11:26.723 --> 00:11:32.718 Would you highlight some unique features that the links, that's apart from other specification management platforms? 00:11:32.718 --> 00:11:34.343 If any others exist, I'd love to hear. 00:11:35.035 --> 00:11:39.475 Well, actually, let me start with the last half of your question, because that's a really good place to start. 00:11:39.475 --> 00:11:47.935 People have spent a lot of time making drawings better, like whether it's planned grid or it's pro-core. 00:11:47.935 --> 00:11:54.588 A bunch of things have been done on the drawing and on the modeling side, but specs have really not had a lot of attention. 00:11:54.588 --> 00:12:00.866 There's a couple of pretty good spec writing platforms, but they're really more about writing them and not so much about managing. 00:12:00.866 --> 00:12:10.493 They're really more for the architect side of the contracts, so to speak, as different project delivery methods have grown. 00:12:10.493 --> 00:12:17.705 The fact is, you don't always get all of the specs at once and they may change and things you

know grow, design build being an easy example of that. 00:12:17.705 --> 00:12:24.246 You get about a third of your drawings and you get or of your specs, excuse me and you get going and then you fill it in over time. 00:12:24.246 --> 00:12:50.455 So one of the things that we think is absolutely missing is the ability to manage specs as they change, as they get amended, but also just the fact that accessing them and effectively using them it's a really archaic system that people have right now, whether it's a viewer, whether it's even if you get a submittal log, that's not the whole spec. 00:12:50.455 --> 00:12:51.339 That's just the submittals. 00:12:51.339 --> 00:12:54.269 I'm getting a little deep into my own story here. 00:12:54.269 --> 00:13:08.751 So the net net is we think that it's really a ripe time for not just a platform that can manage specs on the contractor side, but also really do we need to have them as documents? 00:13:08.751 --> 00:13:16.145 These are collections of data that relate to each other and we've been doing databases, you know, everywhere for 40 years now. 00:13:16.145 --> 00:13:36.874 Turning those documents into data is sort of the fundamental thing that we look to do at the link, and we actually spent a bunch of time with a guy who works for the Air Force I believe excuse me, his PhD is being funded by the Air Force who does taxonomies and ontologies. 00:13:36.874 --> 00:13:42.764 What that means is he helps us to look at specs and start to create buckets so you can break them down into data. 00:13:42.764 --> 00:13:48.960 Whenever the project is, it'll all be the same kind of buckets Took a lot of work. 00:13:48.960 --> 00:14:00.124 We put thousands of specs through it and that's the beginning of us starting the process of changing what is fundamentally a document into fundamentally, a database, because that's what it should be. 00:14:00.620 --> 00:14:08.027 If you think about what BIM is, bim was originally meant to be it still is meant to be a database of information about a building. 00:14:08.027 --> 00:14:28.803 That is, of course, related to a geometric model, but the point is it was meant to be data, not just geometry, and similarly, specs are supposed to be the narrative of the project what you're building, how you're installing it, how you're preparing before installation, how you're inspecting it, how you're paying for it, you know, how you're meeting about it, all this sort of thing. 00:14:28.803 --> 00:14:35.207 There's no reason that needs to be all in a document, and one of the problems we have I mentioned before these are long documents. 00:14:35.207 --> 00:14:47.846 Part of the problem is you need to include everything that's going to be needed for a two-year project into the same document, which means, at any given moment, you only need you know 5% of what's in that doc, probably less than that. 00:14:47.846 --> 00:14:50.748 Again, it's a very archaic way of doing things. 00:14:50.748 --> 00:14:55.891

It assumes that you have a set of documents that get stamped and that's the legal agreement. 00:14:56.460 --> 00:14:57.424 Well, you can do things. 00:14:57.424 --> 00:14:58.609 You don't even need blockchain. 00:14:58.609 --> 00:15:01.440 Just normal, you know, just normal software. 00:15:01.440 --> 00:15:03.305 You can see who did what. 00:15:03.305 --> 00:15:12.177 You can audit what was changed when it was changed, way more efficiently than a PDF that has you know markup on it, because you can't do a lot with a markup. 00:15:12.177 --> 00:15:13.683 You certainly can't do it at scale. 00:15:13.683 --> 00:15:15.019 Someone has to physically look at it. 00:15:15.919 --> 00:15:31.190 So the point I'm making is you know you started this with a great question about spec management platforms and the first thing to do to make these things more manageable is to kind of change the way you interact with them, and that's the first thing we did. 00:15:31.190 --> 00:15:33.830 So we pointed in LLM at them, and now you can. 00:15:33.830 --> 00:15:38.307 The way you interact is now much more conversational, much more directed, much more intelligent. 00:15:38.307 --> 00:15:48.246 But then, downstream, what you need to do is say great, these documents, for reasons that are beyond this podcast, they're probably going to stay being produced the way they're produced for the next little while. 00:15:48.246 --> 00:15:53.671 But once they're handed over to the project team, why do they have to stay a document? 00:15:53.671 --> 00:16:07.570 Why can't we turn them into data and then do all the wonderful things we know how to do with data, whether you're searching, recombining things, networking them with other information like standards and RFIs, and change orders and product information and so on. 00:16:08.439 --> 00:16:17.889 We have a whole network of information that goes into the building of a product or a building of a building, and specs often reference other things. 00:16:17.889 --> 00:16:21.004 But if we turn them into data, you don't need to just do the reference anymore. 00:16:21.004 --> 00:16:25.168 You can have it all in one place, and that's a long-winded way of saying that. 00:16:25.168 --> 00:16:36.669 We're on the beginning of a journey where we think there's a new way to think about specs and the sort of building requirements that they're a part of, so that it's a little bit less siloed. 00:16:36.669 --> 00:16:46.609 You're asking a lot less of people when they're busy to find things and execute against them, but also, over time, what happens when regulations change? 00:16:46.609 --> 00:16:51.129 What happens when you're being asked what's the carbon content of this building? 00:16:51.679 --> 00:16:54.167 Right now it's hard to do because you have to go hunt through data. 00:16:54.167 --> 00:16:56.044 Sorry, you have to go hunt through documents. 00:16:56.044 --> 00:17:01.129 Well, that should be a quick search that should be very fast to say here we

have this EPD. 00:17:01.129 --> 00:17:08.672 These are documents that come from the environmental EPA as well as building product manufacturers. 00:17:08.672 --> 00:17:16.185 I ought to be able to reference EPDs automatically just based on the fact that I have data and I have this EPD coming in. 00:17:16.185 --> 00:17:18.547 It shouldn't need to be document against document. 00:17:18.547 --> 00:17:22.508 That's just one example of how we see this changing. 00:17:22.508 --> 00:17:33.866 Again, once you have things in the form of data, it unlocks a ton of downstream innovation and ability to do smart things with your project. 00:17:35.059 --> 00:17:57.357 Yeah, really, I can just see so many use cases and simplifying the interaction with the data or the specs and where it makes so much sense to create a spec GPT using an LLM like chat GPT and something the audience may be aware of chat GPT has been. 00:17:57.357 --> 00:18:02.511 There's a common word called hallucinations, where it may just create an output on an input. 00:18:02.511 --> 00:18:06.770 So I'm wondering how are you tackling hallucinations when it comes to spec GPT? 00:18:07.640 --> 00:18:09.045 Yeah, that's a great question. 00:18:09.045 --> 00:18:12.728 There's a number of ways that you answer that. 00:18:12.728 --> 00:18:17.391 The very first one is the underlying models themselves are getting better all the time. 00:18:17.391 --> 00:18:29.349 So the OpenAI's and the Metas and the Mistros in France and Google, of course, they were aware of this time last year, so there's been a lot of work to make the models better and better. 00:18:29.349 --> 00:18:36.682 In fact, what isn't commonly talked about is prior to November 2022, when chat GPT really broke out into the scene. 00:18:37.063 --> 00:18:39.210 The underlying models have been around for a couple of years. 00:18:39.210 --> 00:18:41.446 Gpt-3 had been around for at least a year. 00:18:41.446 --> 00:18:46.511 The problem was it produced a range of answers that weren't that useful. 00:18:46.511 --> 00:18:55.992 The big innovation that made all this happen is OpenAI, among others, innovated a way of training them to only give you the best answer. 00:18:55.992 --> 00:18:59.448 So an extension is called reinforcement learning with human input. 00:18:59.448 --> 00:19:13.109 The extension of that is to continue to train the models to give answers that they know came from somewhere, as opposed to they're just filling in what looks like the best answer. 00:19:13.109 --> 00:19:17.771 So the first line of defense is the underlying models continue to get better. 00:19:17.771 --> 00:19:21.368 In fact, gpt-4 plus or turbo I believe it's turbo. 00:19:21.368 --> 00:19:25.947 It was out last week, so, and if this podcast comes out in a month. 00:19:25.947 --> 00:19:28.173 That's probably still going to be true because it came out with another one. 00:19:28.173 --> 00:19:29.324 They're just doing it so quickly. 00:19:29.324 --> 00:19:30.801 That's

point one. 00:19:31.302 --> 00:19:34.311 Point two is asking better questions. 00:19:34.311 --> 00:19:46.630 So if you treat spec, gpt, chat, gpt, any one of the LLMs as if you're talking to Google, you're going to get much more variable answers than if you ask a better structured question. 00:19:46.630 --> 00:19:53.149 Typically, you want to be separating out parts of the question like who are you asking as? 00:19:53.149 --> 00:19:54.351 What's the context? 00:19:54.351 --> 00:19:55.153 What are you looking for? 00:19:55.153 --> 00:19:57.323 Are there any examples? 00:19:57.323 --> 00:19:59.482 Depending on what you're asking for, we bake that. 00:19:59.482 --> 00:20:01.088 You don't see that as the user. 00:20:01.088 --> 00:20:04.590 Well, we bake that into the way your question gets asked. 00:20:04.590 --> 00:20:10.708 So you type in your question but we fill in some of the other things so it'll answer the problem in a more narrow way. 00:20:11.439 --> 00:20:25.985 A general rule of thumb for all AI, whether it's deep learning or whether it's LLMs or even the vision models the smaller you make a problem, the narrower you define the problem, the better it's going to be able to answer the question and, by the way, that's true for humans too. 00:20:25.985 --> 00:20:33.126 One of the most famous admin ever said give me the freedom of a tight brief, and his point was the better you define this is David Ogily. 00:20:33.126 --> 00:20:39.830 The better you define a problem, the more I'm able to put resources against solving that instead of trying to figure out what you meant. 00:20:39.830 --> 00:20:48.025 So the same thing is true with LLMs If you can make the problem smaller by defining it, by saying who's the persona, what's the thing you're looking for? 00:20:48.025 --> 00:20:49.586 Is there any context that helps? 00:20:49.586 --> 00:20:52.663 Are there any constraints I should know about, and do you have an example? 00:20:52.663 --> 00:20:57.069 That's actually a framework called the Rice Framework R-I-C-C-E. 00:20:58.099 --> 00:21:01.126 That is just a better way of writing prompts. 00:21:01.126 --> 00:21:31.628 In two weeks I'm actually doing the whole webinar just on how to write prompts, because I think the industry and just everyone in the economy is on a path of learning how to interact with LLMs more effectively, not as if you're talking to software and not as if you're talking to Google, but an entirely new class of cognitive systems that are a little way more powerful than what we're used to, but they can also be a lot more variable than what we're used to. 00:21:31.628 --> 00:21:36.184 So, just like if you were talking to a human, the better you brief them, the better your answer is going to be. 00:21:36.184 --> 00:21:39.047 So hallucinations are just like a human. 00:21:39.047 --> 00:21:46.031 They will still make mistakes, but the degree to which those mistakes are made is becoming much less than it was. 00:21:46.031 --> 00:21:51.751

But the final way we deal with that is by giving people access to where the answer came from. 00:21:51.751 --> 00:21:56.310 So in our interface you can click on a button and see the part of the spec that that answer came from. 00:21:56.310 --> 00:21:59.287 So that's the ultimate point is go look for yourself. 00:22:01.019 --> 00:22:01.501 That's brilliant. 00:22:01.501 --> 00:22:01.923 Thank you, Hugh. 00:22:01.923 --> 00:22:09.407 I appreciate that and this is certainly a shared passion between you and I, and I'm sure you and I can and we'll talk for hours on this. 00:22:09.407 --> 00:22:16.888 To transition out of the link specifically and out of spec GPT, We'd just love to talk a little bit more broader sense. 00:22:16.888 --> 00:22:25.466 What do you feel are the current technological challenges in the construction space, above and beyond the link that you've been looking at? 00:22:26.159 --> 00:22:34.232 Yeah, I think there's some things everybody knows about, and that is BIM should be more a part of how projects are managed. 00:22:34.232 --> 00:22:37.148 I think ProCore has done a pretty good job of bringing that into the field. 00:22:37.148 --> 00:22:40.928 They're not the only ones to have done it, they're just seeing a lot more of that. 00:22:40.928 --> 00:23:03.567 I think construction has had a pretty heavy couple of years where a lot of money and a lot of options have been presented to contractors and along the way they got pretty good at assessing and integrating and internally marketing new ideas. 00:23:03.567 --> 00:23:10.032 So you've seen, even since I wrote the book, the growth of innovation teams. 00:23:10.032 --> 00:23:17.085 They're not always called that, but let's just for the moment say that you know a team whose job it is to survey what's out there but also survey what's needed by their. 00:23:17.085 --> 00:23:38.326 So constantly talking to project teams to find out what they're missing, what they could do with, what processes they could do better and trying to match those and then, once they've done a couple of pilots and worked out kinks, then figuring out how to market it internally so that it gets adopted, because it's pretty infrequent that some central authority tells everybody you absolutely have to use this. 00:23:38.326 --> 00:23:47.289 That happens for things that you don't have a choice, like the accounting software or your project management software, but usually that those are most of the time those are the only two. 00:23:47.289 --> 00:23:49.326 I mean, sometimes it depends on the contractor. 00:23:49.326 --> 00:23:53.289 So I think that you know a lot of things are going better than they were. 00:23:53.289 --> 00:23:55.242 I will tell you. 00:23:55.362 --> 00:24:10.361 I think the biggest opportunity for the industry totally is for IT teams to get elevated to a strategic role, so to have a real CIO who is C-suite, not just the IT director with a better title, and that there's. 00:24:10.361 --> 00:24:11.125

You're seeing that. 00:24:11.125 --> 00:24:14.791 You know Albert Ricci did a great job pretty early. 00:24:14.791 --> 00:24:16.759 Actually, they're not at all alone. 00:24:16.759 --> 00:24:18.646 Dpr is famous for being good at this. 00:24:18.646 --> 00:24:21.168 You know Suffolk has been fantastic at this. 00:24:21.168 --> 00:24:27.189 I'm going to forget to mention some folks, our friends at Ellis Don, who I've got a lot of affinity for. 00:24:27.189 --> 00:24:29.964 They were wise enough to choose us as part of their accelerator. 00:24:29.964 --> 00:24:36.003 The degree to which they have a fully functional internal software team is surprising. 00:24:36.003 --> 00:24:40.919 It's pretty amazing, and you know they spent a lot of time and a lot of money getting there. 00:24:40.919 --> 00:25:02.809 But the point I'm making is I think across the industry, contractors need to spend more money on IT, because that's how they're going to unlock a lot of the value that things like LLMs and the link or you name it, because if we are alone helping a process, that's great, but if we're part of an integrated set of processes, the value we can add is literally exponentially more than it was. 00:25:02.809 --> 00:25:04.806 So I think that's going to be the big one. 00:25:05.420 --> 00:25:12.569 In all honesty, cyber security is such a thing right now that I'll bet you a lot of them are breathlessly trying to keep up with the threats they face. 00:25:12.569 --> 00:25:20.659 So, as much as they probably want to be more strategic than they are, I understand why it's slower than everybody wishes it were. 00:25:20.659 --> 00:25:32.251 You know, the problem with the contractor is unlike some other fields of endeavor or other professions if your documents get locked up, it's really really hard to proceed. 00:25:32.251 --> 00:25:35.470 So the more digital they are, ironically, the more that's true. 00:25:35.470 --> 00:25:43.928 So they're getting really good at fighting against ransomware and some other attacks, but nevertheless I have a feeling that's slowing things down. 00:25:43.928 --> 00:26:01.663 But I think over the coming years you're going to see the role of IT and CIOs become more strategic than has been true in the past and probably be separated so that there's an infrastructure team who makes sure that the light stay on, the emails work and the ransomware gets taken care of, and then the second half. 00:26:01.663 --> 00:26:08.569 That is really looking forward to how these technologies can help manage risk and deliver projects. 00:26:10.540 --> 00:26:19.250 Thanks, and you did mention some industry titans there that are setting incredible examples to all of us, so testament to their investments and their forward thinking. 00:26:19.250 --> 00:26:40.387 So, looking at your formal position as a general manager of the crosswalk business line at CSI and Symmetry, and having gained the firsthand experience at a fundamental level of the construction industry, can you share specific instances where the industry has given you a unique insight into how companies utilize data to run projects? 00:26:43.639 --> 00:26:47.294 Yes, I think it's a very, very big question. 00:26:47.294 --> 00:26:51.428 I think that there's a few things. 00:26:51.428 --> 00:27:03.371 One insight that maybe is counterintuitive to the question you asked is what construction teaches you is the limits of what data can do versus human intuition and judgment. 00:27:03.371 --> 00:27:13.608 So you really are finding yourself saying the job of data isn't to make decisions, it isn't always even to define what the decision is. 00:27:13.608 --> 00:27:24.691 It's to support the decision for sure and help guide the intuition of someone who's seen this problem 40 times before, and often it's to make sure that the person realizes this isn't the 41st time. 00:27:24.691 --> 00:27:26.606 It's actually either new or it's different. 00:27:26.606 --> 00:27:35.207 So I think that what you see in construction is because of the complexity of what's really going on and all the interconnectedness of it. 00:27:35.207 --> 00:27:46.528 So if there's a thing that I would say was really a revelation for me, is the incredible interconnectedness of construction, and ask anyone who's tried to manage a construction schedule. 00:27:46.528 --> 00:27:48.011 They'll tell you that it's so. 00:27:48.011 --> 00:27:53.967 I mean, every day there's another ding on it and it moves here, and it moves there, and, oh my gosh, we have to send it out to the subs, and so on. 00:27:53.967 --> 00:28:02.269 That speaks to the fact that one thing is late because somebody has a flat tire and all of a sudden it has 15 impacts that you have to then manage. 00:28:02.269 --> 00:28:15.348 That sets an upper limit on what certain types of data can help you with, because that complexity means a human has to kind of manage it and make trade-offs, because it isn't just that you move one thing and then everything moves accordingly. 00:28:15.348 --> 00:28:16.984 You're now making trade-offs. 00:28:16.984 --> 00:28:22.091 Am I going to accept a little more risk by having two trades in the same room with, or three or whatever? 00:28:22.091 --> 00:28:24.945 Am I going to accept risk of trade stacking? 00:28:24.945 --> 00:28:28.362 Or am I going to accept risk of delivery time by not doing that? 00:28:28.362 --> 00:28:31.287 There isn't really an equation for that. 00:28:31.287 --> 00:28:32.443 Data doesn't get you there. 00:28:32.443 --> 00:28:42.288 What data does do is start to help with benchmarking, help people to understand how high is up and what is good and what actually we should be doing better. 00:28:42.288 --> 00:28:51.208 Then I think it helps people to understand the problem a little bit better, and definitely less on the decision side than on the operation side. 00:28:51.208 --> 00:28:57.225 The more visibility people have to what's going on, the better they can apply that intuition

and say we're going to do better than this. 00:28:57.225 --> 00:29:01.314 An example of that is a number of companies have gone and started to automate. 00:29:01.314 --> 00:29:05.329 Things like this is a little more on the trade side, but things like the fab shop. 00:29:05.329 --> 00:29:07.707 There's a lot of the same thing going on. 00:29:07.707 --> 00:29:12.891 It's not quite a manufacturing well, it is manufacturing technically, but it's not like an assembly line. 00:29:12.891 --> 00:29:19.788 Nevertheless, there's enough repeatability that you can start to really tighten up how you're doing what and how you're making what decisions. 00:29:19.788 --> 00:29:27.232 I think the number one place that you see data helping construction companies is visibility. 00:29:27.232 --> 00:29:31.859 The next thing after that is starting to really think about process and process redesign. 00:29:31.859 --> 00:29:37.250 Frankly, I don't see a ton of that Seeing it more in the fab shops, a little less elsewhere, but it's happening. 00:29:37.859 --> 00:29:41.650 The final one that is the holy grail is predictive analytics. 00:29:41.650 --> 00:29:47.351 I think that predictive analytics in construction is not impossible. 00:29:47.351 --> 00:29:51.942 It just takes more creativity to think about what you're predicting and what's it really telling you. 00:29:51.942 --> 00:29:56.270 I think what it's not going to do is tell you how things are going to work out. 00:29:56.270 --> 00:29:59.328 It tells you where the risks might be and what you should pay attention to. 00:29:59.328 --> 00:30:02.026 You're seeing some folks that are making some headway there. 00:30:03.220 --> 00:30:11.731 The problem is, as always, finding the same data that you can aggregate together into big enough numbers to run math against it, much less train. 00:30:11.731 --> 00:30:14.322 A deep learning model is always going to be tricky. 00:30:14.322 --> 00:30:18.307 You really have to ask yourself what are we training? 00:30:18.307 --> 00:30:22.147 How complicated is the thing we're trying to predict? 00:30:22.147 --> 00:30:29.105 I mentioned before a general point with AI is the more you can narrow the problem, the better it's going to perform. 00:30:29.105 --> 00:30:40.592 Frankly, if you think about things like surveys, if I want a survey to be valid, I want to have a big group of people and I want to have as few questions, as few cross tabs as possible. 00:30:40.592 --> 00:30:46.550 It's a little like that the more of the same thing you have against the narrower problem, the more accurate it's going to be. 00:30:46.550 --> 00:30:48.945 Well, that's a challenge for data for us. 00:30:48.945 --> 00:30:52.444 Just to recap again, I think the first thing is visibility. 00:30:52.444 --> 00:30:57.926 That makes the existing intuition and experience that much more valuable. 00:30:57.926 --> 00:31:02.442 Then, after that, you get the process redesigned and you ultimately get to some predictive analytics. 00:31:02.442 --> 00:31:04.284 It's all coming one way or the other. 00:31:05.682 --> 00:31:06.405 Yeah, thanks so much. 00:31:06.405 --> 00:31:07.621 You're definitely highlighting. 00:31:07.621 --> 00:31:17.222 One of the largest challenges in terms of training the LLM, or predictive analytics, is how to find enough data and how to make sure that you're training the model. 00:31:17.222 --> 00:31:22.191 How do you handle finding the data and ownership of data and transporting the data? 00:31:23.320 --> 00:31:24.785 It's a bunch of things in there. 00:31:24.785 --> 00:31:25.628 I'll tease apart. 00:31:25.628 --> 00:31:27.627 The first one is the value of an LLM. 00:31:27.627 --> 00:31:32.522 Why are they so transformative is you don't need to train the model, you need to fine tune it. 00:31:32.522 --> 00:31:34.568 So it goes and finds just the thing you want. 00:31:34.568 --> 00:31:38.728 It comes to you with literally trillions of parameters already. 00:31:38.728 --> 00:31:45.251 That's really transformative, because now, all of a sudden, it's going to be useful for construction. 00:31:45.251 --> 00:31:52.567 The way you fine tune it is you have people go through the exercises you want and point out to it it should have been this, it should have been this, it should have been this. 00:31:52.567 --> 00:31:58.101 Now you're talking about hundreds and thousands of data points, which, while annoying, is absolutely doable. 00:31:58.101 --> 00:32:03.389 That's point one is you want to make sure you're selecting the right model. 00:32:03.389 --> 00:32:06.307 The truth is that the top end of them are all pretty good. 00:32:06.307 --> 00:32:09.927 The next is do I need to fine tune it? 00:32:09.927 --> 00:32:15.672 If so, how do I create a repeatable process for people to train it? 00:32:15.672 --> 00:32:17.523 There's a couple of ways around that. 00:32:22.765 --> 00:32:25.992 You also talked about storing data and how do you handle it. 00:32:25.992 --> 00:32:31.390 It's a really big point, and I think you were also asking about who owns it. 00:32:31.390 --> 00:32:32.663 That's actually a bigger point. 00:32:32.663 --> 00:32:35.704 Contracts right now are not written for this. 00:32:35.704 --> 00:32:45.947 It's really something that people are starting to get smarter about, but the classic AIA contract didn't need to deal with this Not really, because they didn't think about derivative data. 00:32:45.947 --> 00:32:52.883 The way the magic really matters right now isn't the project data, the plans and the specs and all that. 00:32:52.883 --> 00:32:55.819 That's pretty clear cut. 00:32:55.819 --> 00:32:59.369 It's funny that it often is the architect that owns it, but that's not what you're asking about. 00:32:59.369 --> 00:33:03.869 While operating on that, I'm creating new data and metadata and so on. 00:33:03.869 --> 00:33:05.684 That's still being figured out. 00:33:06.019 --> 00:33:08.028 Frankly, I'm not sure there should be a standard

there. 00:33:08.028 --> 00:33:13.130 I think that should be something that gets negotiated depending on what both parties want. 00:33:13.130 --> 00:33:22.929 I think what has happened, what I've seen, is contractors gave up a lot of data four or five years ago to start-ups who then went on to raise huge rounds. 00:33:22.929 --> 00:33:24.763 They felt a little bit taken in. 00:33:24.763 --> 00:33:27.342 Wait a minute. 00:33:27.342 --> 00:33:30.731 We gave you all this data, which wasn't free for us to organize and produce. 00:33:30.731 --> 00:33:32.886 We put effort and resources into this. 00:33:32.886 --> 00:33:36.627 We got to thank you from you and off you went into the sunset. 00:33:36.627 --> 00:33:40.288 Some of these companies didn't succeed, so it is what it is. 00:33:40.288 --> 00:33:43.427 It did leave a bad taste in contractors' mouths. 00:33:43.427 --> 00:33:48.781 Now they want to know either they own it or there's some compensation for the effort they put in. 00:33:48.781 --> 00:33:50.847 That's more than just paying for hours. 00:33:50.847 --> 00:33:53.246 How do we take part of this upside? 00:33:53.246 --> 00:34:05.586 I think the ownership thing is going to take a little while, not least because, like I said, I think that should be negotiated in the contract, depending on what the different goals of the parties are. 00:34:05.586 --> 00:34:09.369 But the other one is on the contractor side. 00:34:09.480 --> 00:34:18.228 You have very accomplished but very focused lawyers who are really good at construction contracts and have had no reason to be good at IP law. 00:34:18.228 --> 00:34:23.724 So a lot of them are struggling right now to figure out how to write an IP policy and how to think that through. 00:34:23.724 --> 00:34:27.125 And you know they're smart people that really understand their business. 00:34:27.125 --> 00:34:49.268 Probably don't love the idea of bringing an IP specialist in, but the funny thing about IP is, almost anywhere you go, whether you're a startup or you're, you know, a non-construction corporation, IP is so specialized that you very often bring in an IP specialist, even if you're, you know, general counsel at Proctor Gamble, Because it's just a thing that people really get good at. 00:34:49.360 --> 00:35:06.569 So I think the IP question is just starting to get asked and properly thought through and I don't think the I don't even think we understand where the problem is going to be in coming years, because what if you've automated a process and now the automated process is creating metadata? 00:35:06.569 --> 00:35:08.380 Who owns that Like? 00:35:08.380 --> 00:35:11.670 So I think that it's going to get a little squirrely, but we'll figure it out. 00:35:11.670 --> 00:35:14.708 And that's where, again, I don't think a standard is a great idea. 00:35:14.708 --> 00:35:16.365 I just think it's going to depend. 00:35:16.365 --> 00:35:19.565 Some contractors are going to care, some owners are not going to care. 00:35:19.565 --> 00:35:21.565 Some owners are going to really care. 00:35:21.565 --> 00:35:26.027 You know, if you're putting up one building a year, you care less about the IP than if you're. 00:35:26.027 --> 00:35:27.170 You know target. 00:35:27.170 --> 00:35:30.407 You put up, you know, 50 projects a year. 00:35:30.407 --> 00:35:31.451 You see what I mean. 00:35:31.451 --> 00:35:38.105 Like, for one group it might be something that really matters for their competitiveness and for another it may not. 00:35:38.105 --> 00:35:41.007 So I think it'll be an area for discussion. 00:35:42.770 --> 00:35:45.366 I love it, hugh, and you're always the best one to start the discussion. 00:35:45.366 --> 00:35:49.505 And, speaking of, you're the host of the Constructed Futures podcast. 00:35:49.505 --> 00:35:50.969 We're on a weekly basis. 00:35:50.969 --> 00:36:00.465 You host two rock stars in the construction industry and you're the author of the Construction Technology Handbook, which provided a primer on technology for Didey's construction professionals. 00:36:00.465 --> 00:36:04.067 We'd love you to tell me a little bit more about the Construction Technology Handbook. 00:36:04.067 --> 00:36:10.068 What was your inspiration and what did you find that helped you transform construction practices along the way? 00:36:10.900 --> 00:36:22.400 Yeah, I think what I found was I wanted to write a book because I wanted to dive deeper and dive deeper into the industry and talk to the most amazing people I could find, and it was kind of early. 00:36:22.400 --> 00:36:24.989 I was still pretty new to the industry. 00:36:24.989 --> 00:36:34.366 I'd done some things in the background, but not in the middle of it, and the cool thing about writing a book is almost no one says no to a request for a book interview, so I got to talk to some pretty amazing folks. 00:36:34.366 --> 00:36:41.119 I'm actually doing a second book, which, not surprisingly, is called AI and Construction, so I'm interviewing a little bit less this time. 00:36:41.119 --> 00:36:42.824 I'm starting from a different basis. 00:36:44.340 --> 00:36:54.268 The point of that book, though, and sort of the core insight of the first one of the Construction Technology Handbook, was people talk over each other too much. 00:36:54.268 --> 00:37:04.688 So folks who spent 20 years figuring out how to pull a building out of the ground probably don't spend a lot of time wondering what API stands for, and yet people are throwing this word at them all the time. 00:37:04.688 --> 00:37:08.429 So I actually spend the first third of the book defining terms. 00:37:08.429 --> 00:37:11.862 I mean, I literally define what Wi-Fi is Understanding. 00:37:11.902 --> 00:37:39.364 Most people know what it means, but have probably never had a discussion about it so that people walk away saying, look, I may have known some of that, but I know all of it now, or at least I've heard all of it now, because I wanted to level the

playing field so that consumers of technology would be able to make better decisions than if they have to kind of shut the brain off and get through the part with all the words they don't know and then have to feel a little stupid asking questions because it's on them to know whether the thing they're buying is what they need. 00:37:40.020 --> 00:37:52.268 But if they're not armed with the right terminology and understanding of what's going on, they're really hands-strong and that slows everything down, because now people are like, look, maybe this would help me, but I don't understand it well enough and I can't take a risk with this project. 00:37:52.268 --> 00:38:08.865 So if you give people at least the table stakes of what these words mean and thinking about how software is built and so on, they're better able to make a decision which helps both sides it helps the contractor and it helps the vendor of software and it helps the people managing the whole thing. 00:38:08.865 --> 00:38:10.302 That was the big deal. 00:38:13.527 --> 00:38:33.048 So, in addition to I mean I'm really excited to hear about this new book you have in addition to the success of the last one and the new book that you're working on, and I'm sure it all ties together to this question, which is what is your vision for the future of construction and how do you see the work you're doing contributing to the evolution of the construction, best construction practices and project management? 00:38:33.940 --> 00:38:39.527 Yeah, I think that it's important not to get focused on one part of what's going on. 00:38:39.527 --> 00:38:47.911 I think there's a lot of change and there's a lot of new ideas and a lot of new ways of doing things that are being worked out and tried out. 00:38:47.911 --> 00:38:56.779 It is such an intricate as I said before such an interconnected endeavor to build a building that it's really hard to change one thing without it impacting others. 00:38:56.779 --> 00:38:58.246 So they're all inching forward. 00:38:58.246 --> 00:39:09.099 But I think what you're going to see and what I, to the degree I have a vision about this is we start to differentiate buildings and think about how to build them in the way that's most appropriate to them. 00:39:09.099 --> 00:39:30.751 So things like industrialized construction, which absolutely relates to specifications and absolutely relates to AI, especially on the generative design side, but some buildings are going to be better for that and some are not, and apartment building is probably better for that, multifamily maybe less so, or a factory definitely less so. 00:39:30.751 --> 00:39:32.610 So you see what I mean. 00:39:32.610 --> 00:39:42.759 Like, I think that we have a tendency sometimes, when we're talking about construction, to try to continue to lumping it together, and I think you're going to find specializations more and more be the case. 00:39:42.759 --> 00:39:48.759 I think that the degree to which a human needs

to be doing. 00:39:48.759 --> 00:39:53.259 Lots of the minutiae in construction, I think, is going to go away. 00:39:53.259 --> 00:39:56.391 If there were a vision, it would be. 00:39:56.391 --> 00:40:10.940 We need fewer people who are doing more value at, more of the intuition and thinking that goes into making a great building, and less of the moving stuff around and less of the paperwork and less of the stuff that people typically hate now anyway. 00:40:11.202 --> 00:40:20.858 And the interesting thing, though, is often when you say things like that, veterans in the industry will tell you yeah, but how are people going to learn if they don't go through and do 15 submittal logs? 00:40:20.858 --> 00:40:23.777 Well, the answer is the same. 00:40:23.777 --> 00:40:25.795 Technology that automates. 00:40:25.795 --> 00:40:44.039 Some of this is fantastic at creating little micro-trainings, so whereas before there really wasn't a ton of training unless maybe you're talking about safety, and even that is kind of recent you're now able to have the LLM or whatever the software platform is, create little snippets of learning along the way, so you're learning what they call learning in the flow of work. 00:40:44.201 --> 00:41:11.639 So the other thing I think you're going to find is the tools that automate and the tools that make it easier for us to not have to do all the Mimousha that, again, people don't usually love will be the same ones that are the source of learning and the source of upskilling people, not only to where they would have been 20 years ago, but, as things continue to change, they're able to be there to do what in the learning world, we call scaffolding learning, so supporting learning. 00:41:11.659 --> 00:41:19.005 So you're not taking on too much at once, you're just learning a little piece each time, but pretty quickly you get really good at whatever the new thing is. 00:41:19.005 --> 00:41:30.173 So I think, rather than automation, necessarily meaning that our skills go away, I think there's actually opportunity to be thinking and saying these things can also write learning. 00:41:30.173 --> 00:41:35.790 So let's do that, let's have them do that and deliver learning in a way that's optimized for whoever the learner is. 00:41:35.790 --> 00:41:52.483 I think you're going to see a revolution across manufacturing, across construction, across pretty much every service business, where learning in the flow of work it comes from the tools you're using, not necessarily some learning management system that no one pays any attention to. 00:41:52.503 --> 00:41:58.081 Yeah, and these lessons that have been learned are very costly to the business too, right? 00:41:58.081 --> 00:42:04.686 So if there's a way to avoid that the cost in the learning I think that would be revolutionary as well. 00:42:05.319 --> 00:42:07.148 One interesting point really quickly. 00:42:07.148 --> 00:42:10.585 Almost every project manager I've ever talked to has a list of lessons learned. 00:42:10.585 --> 00:42:21.235 They often need a little bit of translation because they're usually reminding themselves or they're delivered by them to someone else, but the point is there are huge repositories across the industry where people have taken notes. 00:42:21.235 --> 00:42:25.027 They've taken notes on projects, they've taken notes on specs, on drawings, so on. 00:42:25.027 --> 00:42:42.947 There's a lot of context out there that these new AI tools have the promise to be able to distill into lessons that make some sense, especially when you combine them with the work that the International Code Council and CSI and ASTM and some others. 00:42:42.947 --> 00:42:50.530 They have amazing repositories of information about how to build a building and I think that those things start to become unlocked. 00:42:50.530 --> 00:42:58.726 We're going to be able to think about buildings in a much more sophisticated, comprehensive way than any one human's brain could handle. 00:43:01.222 --> 00:43:08.695 So, looking into the future, as you and I just were, I would like to transition into the final question of the show and the tradition of our podcast, future Construct. 00:43:08.695 --> 00:43:16.467 If you could project yourself out 25 years and wanted to have any device or technology that would benefit you personally, what would it be and what would it do? 00:43:17.802 --> 00:43:19.807 25 years is an awfully long time. 00:43:19.807 --> 00:43:27.811 I want a spaceship that'll take me to the moon, where I can hang out. 00:43:27.811 --> 00:43:29.394 I'm kind of coming up with that one on the cuff. 00:43:29.394 --> 00:43:41.480 It doesn't really relate just to construction because, frankly, in 25 years I think the way we build buildings will be, some of it will be unrecognizable and some of it will look a lot like right now. 00:43:41.480 --> 00:43:57.277 But me personally, yeah, I like the idea of personal transport that can I don't know about the moon, but that can take you out into space, and I love the idea of experiencing space and I think I'm young enough that that's a reasonable expectation. 00:43:57.277 --> 00:44:00.150 I'm not so sure about Mars. 00:44:00.150 --> 00:44:04.706 I moved to Austin and say I never want to be culled again and I don't know that Mars is going to satisfy that. 00:44:07.061 --> 00:44:07.202 Well. 00:44:07.202 --> 00:44:08.565 Thank you so much for your time. 00:44:08.565 --> 00:44:12.614 It was an incredible conversation with you, hugh. 00:44:12.614 --> 00:44:17.481 Your journey from the 80s in Hong Kong to becoming a construction tech visionary is truly captivating. 00:44:17.481 --> 00:44:25.516 The link specification management platform, especially with spec GPT, seems like a game changer, streamlining processes for project teams. 00:44:25.516 --> 00:44:37.246 Your insights into the industry, challenges, your unique offerings, such as your constructive future podcast and the construction technology handbook and your upcoming book they all provide a rich perspective of looking ahead. 00:44:37.246 --> 00:44:42.583 Your vision for the construction industry and the links role in shaping the future is insanely inspiring. 00:44:42.583 --> 00:44:49.186 Thank you, hugh, for being a thought leader and to our listeners, stay tuned for more insights on the future construct podcast. 00:44:49.186 --> 00:44:51.764 Hugh, it's been an absolute pleasure having you on the show. 00:44:51.764 --> 00:44:52.788 Thank you so much for your time. 00:44:52.788 --> 00:44:54.083 A thanks for having me. 00:44:54.083 --> 00:45:06.677 Absolutely Thank you.