Wes Edmiston:  

Gentlemen, welcome to the show.  
 

Dylan Sumiskum:  

Thanks.  

 

Josh Napoli: 

Thank you.  

 

Wes Edmiston:  

Yeah, I’m happy to have you guys here. I work with you guys all the time and you both are extremely knowledgeable in the space that we’re going to be talking about today, which is a real hot topic, but you’re very different on the way that you approach the situation.  

So I think this is going to be a really fun conversation. At least I’m going to enjoy it.  

 

Josh Napoli:  

It’s a really fun topic.  

 

Wes Edmiston:   

It is. And it’s really interesting to dive into it. I guess Josh, people are right now hearing a lot about AI.  

So, there’s this flood of information that people might have been hearing for the first time. What are some of these terms with AI that people might be hearing and what do they mean?  

 

Josh Napoli:  

Well, the conversation this year is about generative AI. So this is software that is designed to generate creative content, like a story or a picture or a piece of music. It’s different than older types of AI because of this sort of creativity, this understanding of human language in a way that wasn’t possible before.  

Older AI was usually about detecting a feature or making a supporting decision. At Cumulus, we’re most interested in large language models, or LLMs. These are models that read and write documents. The cutting edge architecture is called GPT.  

Right. Now Generative, Pretrained, Transformers, you hear that a lot from OpenAI because they are the ones that first invented it and made it practical. But also Palm two from Google and Claude from Anthropic and Llama from Facebook.  

These are all based pretty much in the same GPT architecture.  

 

Wes Edmiston:  

Yeah. That’s very informative billing. Is there anything you would add to that and can you provide I guess you hear large language model.  

What is a large language model?  

 

Dylan Sumiskum:  

Yeah, I mean, Wes, there’s a lot of technical terms out there, right. But ultimately for the non technical people, for the user. Right. We’re talking about AI that has been around for some time.  

There’s products certainly out there already that’s utilizing language models like Siri, Amazon Alexa, or Google Assistant. Right. And the way I like to compare them is these are like the early smartphones.  

They are functional, but the interaction just doesn’t feel great. And then the iPhone came out and everyone immediately was just like, wow, jumped on it, right? And I think it’s because the multitouch gesture, along with all the other things that they design so well, just made it so what many people call limbic resonance, or just like where your user just has a really deep level of connection with the technology.  

And I think this is exactly what’s happening with GPT and LLM right now. What they have done with the AI interface is essentially make it so that you have now this deep level of connection with the technology.  

And you can see this, evidently, in the adoption. You look at Chat GPT, it sets a record for, I think, the fastest growing user base since Facebook or Instagram. And so what does this mean to a lot of people?  

In non technical terms, I think, is that what an LLM is, in my opinion, is the next great UI revolution moment similar to that of the iPhone, and it’s just up to your horizon. And I think a lot of companies are going to be able to leverage this to create some really interesting applications.  

Wes Edmiston:  

Yeah, you touch on this aspect that it’s really been this massive tidal wave that came in since what, November is, really whenever everybody started talking about the OpenAI, because that’s when I believe November was when GPT-⁠3 got released, right?  

Yeah, it’s been this massive tidal wave ever since November. So people are getting kind of they’re drowning in AI right now. But with that, I think that because how quickly this all came on, there are a lot of people that don’t really know how to handle AI in general.  

They’re threatened by it. But most of that, I think, comes from kind of like a lack of understanding. And we can definitely debate on whether or not. AI is going to end up leading to the end of humanity, or if everybody’s going to end up losing their jobs because a computer is going to be doing it.  

But kind of in the interim between now and then, there are a lot of other, we’ll say, lower level concerns that are very important to these different companies, like data security or intellectual property when integrating with some of these large language models or these AI systems.  

Is there any, I guess, weight to that? Is there concern with, we’ll say, lack of stringency around intellectual property or potential to lose intellectual property when working with some of these large language models?  

And is there any way of, I guess, circumventing that or avoiding some of these concerns?  

 

Dylan Sumiskum:  

Yeah, first of all, it certainly seems like from the doomsday scenario and all that, a lot of many people just watch too many Sci-⁠Fi movies.  

It’s good business for the media to hype it all with all kinds of emotions. It’s profitable. But for us people, we really have to be objective about it and try to assess the opportunities and the threats.  

Right. From a business perspective, and certainly the things you brought up, there are concerns. There are concerns about data security, about IP, concerns that need to be addressed, especially if you’re a company that wants to use B2B SaaS.  

And there’s also concerns about copyright, right. That’s sort of like the ongoing debate right now is basically any content that the AI generates, is that copyright or not? Right? Like, is it violating copyright?  

And I think Japan just released a new law. Josh, correct me if I’m wrong, but I think they’re saying that. Basically, it’s not copyright material, right? Like there’s something like that. Really the question really just comes down to specifically for B2B SaaS definitely the regulations need to play catch up, but there are a lot of opportunities to be had by integrating private data with LLM. So the question there is, well, how do I make sure my data stays private and doesn’t leak to the Internet?  

Right? And so when building applications using one of the popular LLMs, like, say, OpenAI, they’re basically privacy protecting by default. They don’t use your prompts or your conversation history to train and improve their models.  

However, if you use ChatGPT or Bard, which is their front facing application, not their API, you can read on their terms of use and they do use and collect your information to improve their product.  

And they’re very explicit about and warn you about sharing confidential information, which is why many companies ban their employees, such as Apple. They ban their employees from using ChatGPT because they don’t want their employees to be sharing confidential information.  

But just to be clear, this is not the API the developers and companies use to build application. This is their own user facing applications, either in the web browser or their mobile version, which I think ChatGPT just released.  

Now, going even further for application builders who want to build software that integrates with LLMs and want to be extra safe with their customer data, or their own data for that matter, they can definitely explore other options, such as using one of the open source models that are available right now. They’re actually pretty good. They might not be as good as the pretrained LLMs from OpenAI or Google, but definitely they can be trained to be good at specific domains, such as being really good at construction or industrial LLM.  

And definitely by doing that, we can also host it ourselves so that we can assuage, like, any fears of privacy or data concerns, right?  

 

Josh Napoli: 

Yeah. So ChatGPT, Bard, Bing, they’re kind of free. They’re pretty attractive to you because they can do some useful stuff for you.  

But like anything else on the Internet, you really have to think about this. If you paste your company’s confidential information into the random box of the Internet.  

 

Wes Edmiston:  

This is probably the most common sense thing that I’ve heard in a long time of don’t copy and paste basically your sensitive data into some random box on the Internet.  

I think that’s a wise lesson to live by. But I guess the thing to reduce a down that I’m hearing really from both of you is while there are in certain areas, certainly some of these data security risks, that’s predominantly to kind of the chat GPT or the Bard that we’re all familiar with, which is you pull it up directly online and you talk to effectively the chat bot.  

That’s right there. But if you’re communicating with it via an API, so, like, many of these software solutions are that are popping up, that is not training LLM, and you don’t have that threat of losing intellectual property aside from the typical data security for that SaaS company as it is.  

Is that pretty accurate?  

 

Josh Napoli: 

Yeah, you can use Bard for free, use Bing for free, but you kind of pay with sending your content to Microsoft, I guess.  

 

Wes Edmiston:  

Yeah. How free is free? Right? So with that, it sounds like there are opportunities for companies to be working with some of these emergent technologies in a way that is secure fundamentally that is protecting their data and their intellectual property.  

So how is it, Josh, that a company can begin working with one of these AIS or integrating an AI into their service?  

 

Josh Napoli: 

So whether it is we’re talking about a company that is a retailer or somebody who is like our main customer base, which is some of these construction companies, well, the most obvious way to use an LLM is to use it for what it was designed for originally to generate some text. A common example is, well, we’re starting a new project, we have a lot of job openings for it and it’s really easy to generate job descriptions using the chat system.  

And so if it’s a chat system right now, you give it a lot of examples of good and bad chats, good bad users and prompts. And that kind of destroys the ability to know, to be confident in being wrong because giving the best answer that has confidently turns out to be higher ranked.  

 

Wes Edmiston:  

Yeah, that’s actually kind of funny to think about though, as far as kind of it being very confident but wrong, I imagine have all been around those people before, maybe a good portion of the time they’re right.  

But they’ve just been so convinced that since they’re right all the time, that they’re right all the time, so they just espouse whatever garbage that they think makes the most sense. So it sounds like that there’s a technical risk of doing that for these GPT models or whichever LLM, whether we’re talking about Bard or Chat GPT or any of these others, it’s kind of funny to think about.  

Dylan, I wanted to ask in what ways have you seen people using these LLMs to either augment or replace work in the various industries? And I guess how is it that you all are using this most recent development in AI in order to aid the construction industry?  

 

Dylan Sumiskum: 

Sure, yeah. I don’t know if many people know about or what a Reg X is as a programmer, it’s something that it’s useful, but I’m just too lazy to think about it or learn it. There are some people out there who are really good at it, but I’m just not.  

Now, it turns out that Chat GPT is incredibly good at it, and ever since then, I have been using it to build my Reg X’s. This is just one example of how LLMs have augmented my work as a programmer, and other examples like it’s really knowledgeable with AWS.  

So if companies that use AWS as their infrastructure can lately augment their programmers by leveraging Chat GPT now, in regards to the construction industry, we’ve been working on a prototype to infuse our workflow application with AI.  

And it’s certainly exciting because it can help our customers to easily build workflows. But ultimately it is like still using public internet knowledge. And where we see this moving forward in aiding people in the construction industry is bridging the non-public information with the technology so that they can get amazing insights into their work as well as seamlessly be able to digitize their procedures, which are mostly paper based.  

Right. For instance, you can turn your document into a digital workflow and then ask the question, well, is this the best thing? Like what could be done to improve this, say, to minimize the number of rework? 

So these are just one of many ways that you can really empower the construction worker as well as the project managers in the industry. Yeah, I can see that being really valuable in the sense of really for a couple of different reasons.  

But one problem that you end up seeing on various projects is differences in interpretation of information. Right. So you would think that either a procedure or code or whatever specification that it’s written down in text, black and white, everybody will be able to read this and we all understand the exact same thing because it’s spelled out clearly, but that’s oftentimes just not the case.  

So you have these large bodies of text that may be a particular line everybody will interpret slightly differently. So being able to push this through, we’ll say honestly, through a funnel, to be able to reduce this down to the nuts and bolts of what matters in the procedure and to get a clear instruction of what needs to happen.  

 

Wes Edmiston: 

I think is extremely valuable in the sense of it’ll eliminate a lot of that that kind of debate, for one. But also I’d like to hear what else you guys have been seeing that maybe other companies are doing in using this emergent AI technology for construction.  

But the thing that I see is and it’s not even just with AI, it’s with a lot of different technologies. Everybody is pretty well focused on, say, administrative duties or managerial duties, but not very many people are involved in and really where the work happens, which is really the problem of a lot of these big projects, especially where we’re talking about schedule and cost being well outside of what was allocated for that individual site.  

But it’s because nobody’s doing anything to actually go toward where the work is happening. Right. So with what you’re doing, it sounds like you’re able to really take all of this body of knowledge, reduce it down, and put it in the hands of the people that are actually doing the work to be able to actually improve what they’re doing and to pseudo automate automate what it is that they’re doing.  

Is that pretty accurate? 

 

Dylan Sumiskum: 

It’s basically more or less a Cumulus story, right? Like, yes, we’re using AI, but it is still Cumulus. We’re very good at capturing or observing the truth on the field, whether it be construction or maintenance work.  

And that’s a great point that you brought up. It’s like, well, a lot of these software solutions, they’re mainly like replacing Excel, right? Like, yeah, you have a great database, but, well, where does your data come from?  

So really the complexities here are in what you said, the workers on the field that are actually trying to perform. Those procedures that are captured in the documents. And so what we have done, at least what we try to do, is transform that experience.  

So you can go from your paper processes seamlessly into digital workflows that gets enforced and is giving you clear data transparency over what’s going on with your project.  

 

Wes Edmiston: 

Is there anything else that you’re seeing that people are using AI for that you think is really an interesting idea or that is promising to aid again for construction and maintenance?  

 

Josh Napoli: 

There’s an interesting startup named Hypar that has a rule based generator for building systems. Those is a tool for architects. It generates not just the shape of offices, but like the ductwork and electricals and the routes of plumbing.  

All of the building systems, it empowers them to make accurate cost analysis of the building at an earlier stage of the design.  

 

Wes Edmiston:  

Yeah, I mean, honestly, that could realistically save days, if not weeks during detailed design and or in project estimating as well.  

That massive opportunity in order to really improve some really inefficient areas in the construction industry, which typically just go straight to overhead. Dylan, is there anything that you would add into that?  

 

Dylan Sumiskum: 

There’s definitely going to be an explosion in because it’s generative AI. Right. So there’s going to be an explosion of tools, especially in the construction space or maybe like modeling your systems, where essentially you’re engaging in this creative process alongside the AI.  

Helping you do that. So that’s certainly going to be the case. And not just in construction, in Arts as well, design like Adobe, they’re kind of aggressively pursuing this direction or even just your basic tools like Microsoft Word or Excel.  

Microsoft is releasing Copilot, which I think will heavily leverage their AI capabilities. But really on the field there’s not a lot of examples there, which is kind of like an interesting opportunity.  

One of the themes that we learned just developing our workflow product is that the workers on the field, they don’t want to click many buttons. They’ll complain that this is too many clicks.  

So really what they want, or just in general, what the UI should be is that it shouldn’t be intrusive to the main work that’s being done. Right. And so this creates an interesting so the advent of AI creates an interesting opportunity for this type of UI development because now you can sort of create a different kind of user experience that’s more AI first, if you want to call it, where the user experience is supply.  

Guiding you through the process in the background doesn’t intrude your main work, but nonetheless is very useful. So I think this is also a great example of what the future could be like and certainly like it’s, it’s an interesting way of solving problems in the, in the construction space.  

Wes Edmiston: 

Yeah, I mean, it sounds like there’s, there’s a lot of really, really some large opportunities that are coming up and. Especially because we all know in the construction industry there are big problems that need to be tackled.  

Softwares and different solutions have been popping up for ages now. And while there’s no shortage of software solutions, there’s also still no shortage of problems in the construction industry.  

We can cite statistics from McKinsey or really any of these other large consulting firms that go to show that projects are consistently behind schedule over budget. And there are companies that go out of business all the time because they’re just not able to flow all the way through the project.  

So with that, there are definitely problems and sounds like AI is going to be able to help out, at least make things more efficient in particular areas. Josh, I wanted to ask, before we wrap this stuff up, I guess, what is next in this AI race?  

Do you think that we’re going to continue to see expansion of these LLMs, or are there additional forms of AI technology that should be pursued to make the most complete AI tool out there?  

 

Josh Napoli: 

Well, I’m most interested in the LLMs.  

You can see Facebook keeps coming up with really cool things that are language understanders or music generators or different kinds of visual tools. Yeah, open AI, I think kind of like stumbled onto a business.  

And so we’re going to keep seeing sort of breathless investment in language models for the medium term. To be able to have just the most accurate LLM would certainly serve the industry in general and really aid in the development of a lot of these additional tools, engineering related tools, estimating tools, or anything else that are going to be emergent.  

 

Wes Edmiston:  

And like Dylan, you were saying, with being able to integrate really the AI tool. Into the applications to be able to allow the users to define what their experience is going to be. I think for software in general, Josh, you’re right.  

That’s just going to make everything more flexible, and it’s going to help kind of tailor to suit what the individual wants. Tailor these applications to suit what the individual wants. Dylan, is there anything you would add on kind of what the future of AI looks like?  

 

Dylan Sumiskum: 

Yeah, I mean, there’s definitely going to be a lot of improvements in the large language models, but really, the one thing that people don’t talk much about is really the application layer. Right. I think that’s also going to be the next stage of the AI race.  

There’s going to be a lot of investments there in the companies that are building applications on top of these large language models. Certainly the hint is there in the term it’s a language model, right.  

It’s not a model of the world or the physical world or the model of problem that you’re trying to solve, like, in your particular domain. Right. So while the LLM truly is great at predicting a lot of or just predicting human language in general right, but ultimately, it’s not the right tool to predict the physical world.  

Right. But that’s where all these applications can come into play, is using LLMs to get there faster. And so it’s easy for us to mistake kind of, like, performance with competence when you look at what the LLM can do and return all these, like, seemingly really intelligent responses.  

Right. We immediately equate that performance with some level of competence, as if, like, it’s a human being that’s like, well, if you know this, you must know other things. Right. But I mean, that’s not the case with LLM, and I don’t think that’s something that we can truly, that shouldn’t be the mindset we should treat with this LLMs and that’s why I think it is important that we also enable all these different applications that are actually trying to solve for these physical world processes by leveraging LLM.  

I think that’s also where the race is going to be in the next chapter.  

 

Wes Edmiston: 

Yeah, I think it’s especially valuable, like you’re saying, to be able to interface with really the boots on the ground in getting work done and getting work done right.  

A little plug for the title of the show there. But if somebody wants to get involved and either check out or test out the work that you all are doing on this workflow builder application, how can people find this and get involved?  

 

Josh Napoli: 

Sign up for our beta list at cumulusds.com/ai-beta and we’ll be in touch with you.  

 

Wes Edmiston: 

We will provide a link to that in the show note captions just to make sure that everybody can get involved and check this out.  

I’m interested in trying it out and being able to see just kind of where all of this goes. There’s so much opportunity to be able to just improve processes, standardized things, and coming from a quality background that is right up my alley, to just have standard process that we have documentation all the way through cradle to grave and to be able to just get things done, get it done one time.  

So we can stop wasting time and start finishing projects.  

 

 

Rapid Fire Questions 

 

Wes Edmiston:  

Guys, we’re running right up on time. I’m going to ask a few last minute rapid fire questions to get to know Dylan and Josh as people, not just as professionals.  

So first question that’s going to pop up. Where is your favorite place that you’ve traveled? Josh. 

 

Josh Napoli:  

Favorite place I’ve traveled is Hawaii. We went to the Big Island is my favorite there.  

 

Wes Edmiston: 

Yeah, Hawaii is beautiful. Dylan, what is your favorite book? 

 

Dylan Sumiskum:  

I would say principles by Ray Dalio.  

 

Wes Edmiston:  

Yeah, it’s definitely good. If anybody hasn’t given it a read yet, I highly recommend there’s a free app that people can download as well that gives access to the full text of the book.  

Josh, what is your favorite quote?  

 

Josh Napoli:  

Oh, make no small plans because they’ll fail to stir hearts. I don’t remember the quote exactly like that.  

 

Wes Edmiston:   

That was pretty accurate. Very close. Definitely captures the essence of the message.  

Dylan, what’s your favorite or what is your dream job? 

 

Dylan Sumiskum: 

My dream job? Right now.