KP Unpacked

Your Edge Case Is Someone Else's Use Case

• KP Reddy

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 46:06

What if the detail that seems trivial to you is the constraint keeping the entire project from moving forward?

In this episode of KP Unpacked, KP Reddy sits down with Dr. Barry Clark, CTO of Zero RFI, to unpack why construction projects fail on details nobody thought mattered. A structural beam seems simple: read the line on the drawing, spec the size, done. But the client needs the longest span possible without custom manufacturing (adds cost). The superintendent needs to know when the truck leaves to avoid traffic (adds delays). The permitting team worries about wide-load requirements (adds 90 days). The building supplier tracks lead times and availability. Same beam. Five different perspectives. All mission-critical. The edge case you dismiss is someone else's everyday constraint.

Barry explains why AI's real unlock isn't automating standardized workflows (McDonald's already perfected that). It's mass customization at scale. Every persona on a project looks at the same drawings and sees different risks. AI can now hold all those perspectives simultaneously and optimize for all of them. The conversation also reveals why companies are having a "Facebook moment" with AI (deployed it everywhere, now realizing they don't understand privacy), the three-tier consulting model emerging (billable hours get worst talent, equity gets best), why programming got easy and that's actually good, and why Zero's training spends two-thirds of its time on mental models instead of AI mechanics.

Key questions answered:

  • Why do construction projects fail on edge cases nobody thought were important?
  • What's the structural beam example that shows five different perspectives on the same detail?
  • How does AI enable mass customization instead of McDonald's-style standardization?
  • What's the corporate "Facebook moment" happening with AI deployment right now?
  • Should you go deep on one AI technology or broad across all of them?
  • What are supply chain attacks, and how should executives test their IT teams?
  • What are the three tiers of AI consulting: billable hours, risk fees, or equity?
  • Why did one consulting firm charge $5M but generate $500M in client outcomes?
  • Do employees own their skills files when they leave, or does the company?
  • Why did some software engineers quit when their companies adopted AI coding?
  • What's the difference between LLMs, VLMs, and physics-informed neural networks?
  • Why does Zero's training curriculum focus on thinking frameworks instead of tool mechanics?

If you're an engineer dismissing client requests as edge cases, a project manager wondering why small details derail schedules, or trying to understand why AI matters more for customization than standardization, this episode will show you that everyone's edge case is equally critical to project success.

Listen now.

Duckweed, Guest Intro, And Setup

SPEAKER_00

All right. So this is not Nick Durham. He failed on me. No, actually. Nick is in LA for the week doing a bunch of meetings. We invested in a company called Western Chemical that does takes duckweed from wastewater treatment plants and turns it into fuel, which is like wild. 30 days ago I didn't know what duckweed is. I still don't know, think I know what duckweed is. But anyway, Nickel, maybe Nick will talk about duckweed next week. It'll be the duckweed episode. But today, special treat, Dr. Barry Clark, uh CTO of Zero RFI. And uh we're we're joking around because we have all these great conversations on the phone. Yes, on the phone, not Zoom, not Slack cuddles, on the actual phone that we think is like the best pet podcast content. And we had a couple of those calls this morning. I'm like, did we just do our podcast? So welcome, Barry.

SPEAKER_01

Thanks, KP. I'm glad to be here.

SPEAKER_00

So we got we went deep this morning and on skills files and MCP and all this other stuff. But I think it'd be really one of the things I think that's very valuable to the audience, to people in general. I talked to someone this morning, they're like, I can't keep up every day. There's something new coming out, right? And I think our perspective is very cool, right? Like Claude, whatever, and all that. How do how do you think if if I'm the CEO of an engineering company, how should I think about keeping like how would you keep up with this stuff? Is it more like just let it let it be and I'll get to it in 30 days as it settles in? Because Mythos

Keeping Up By Going Deep

SPEAKER_00

apparently is taking over the planet and it's gonna hack all our systems. Like, should I care?

SPEAKER_01

I I mean, I don't have a good answer to that question, but what I would say is I think something that's healthy is instead of trying to go broad and go right, pick a certain area, right? Like I think you did this with like MCP to start with, right? And now skills, and rather than trying to keep up with all the brands, the implementations, right? Pick a technology or you know, uh kind of agnostic spec and go deep on that and really understand that, understand how it applies to your business. And then that instead of always feeling like you're drowning in new companies, right? Or or new ideas, then I think that helps build some connective tissue for you to actually understand, right? And then if somebody says, Oh, look at this new company, you don't feel like you're missing out, you can pretty quickly understand like how it relates, like once you've done that for a couple of different topics. That that's how I try to approach it.

SPEAKER_00

Yeah. Which by the way, that that is a word that you use a lot that's become my favorite word that I parrot to other people, which is understand, building an understanding and understand. I think in this world of AI, like I think that and that what like that's what we're trying to do, that's what AI is trying to do. It's just a great word that you use frequently. And now I just started parroting you because I don't have any original ideas, I must have just parrot other people's. I have a platform so I can pair every take other people's great ideas, and then I parrot it out. And people are like, KP, you're so smart. I'm like, Yeah, I guess so. But I think what's I think what's really interesting about some of that, right? I think we're having what I'm kind of calling a Facebook moment when it comes to AI. And and you probably remember like everybody was on Facebook, like your parents, everybody's on Facebook, right? And like this is great. And then they realized they weren't paying anything, right? Right. And then the narrative came out like, if you're not paying, if you're not paying for the product, you are the product.

unknown

Right.

SPEAKER_00

And I think corporate America is having a little bit of one of those Facebook moments. It's like, oh, I just like clawed into my enterprise.

The Enterprise Facebook Moment

SPEAKER_00

Yeah, I don't know what privacy looks like, I don't know what they're using. Hey, are they are they gonna take over my like I I've let the you know, the the wolf in the hen house, so to speak. Yeah, and I think I've just seen in the last 30 days, and I think Mythos was probably a little bit of not a catalyst, but probably uh a narrative that got people's attention, right? But we're seeing companies that full claw deployment, and now they're back on co-pilot because they got they just they had that Facebook moment, like, oh, it turns out like I mean, how do we we we talk about privacy security? I mean, for quote unquote a startup, we are very like enterprise grade by design. I don't I don't think you know how to build stuff that's not enterprise grade, it's not in your nature, exact opposite of me that it wants to hack something and deploy. Like, what do you do do you feel the same way? Like maybe we're having this weird Facebook moment.

SPEAKER_01

100%, yeah. I I think the thing that maybe is different, maybe not, right? So the the the the tech community, meaning like hardcore coders have been aware of this for maybe like six to eight months earlier, right? And the reason for that is all the supply chain attacks that happen with the actual software packages, right? And so I think when you combine that with the fact that some of the analytic tooling that's provided by the large platforms is not, you know, what corporate America is used to. Yeah. Right. You know, so our token estimations are exactly that, you know, that's the the warning, like, hey, this is kind of an estimate, like, which is kind of crazy, you know, to think it's guessing at what you're spending and where. But I I think the large platforms, I generally believe they're terms of service, right? And so I do believe they're secure. And you know, we have some really interesting ideas about how you could make them more secure, or like how you would change some of the boundary layers on the security itself. Yeah, but I think the problem is actually what they enable, yeah, right. And so, like the whole idea of vibe coding creation, and then you get the injection. So it's it's like that problematic portion, it's just like one layer down or one layer up, depending on how you think about it. But I totally agree with you that yeah, we had this sort of flash point, and then all of a sudden there's something that no one had really thought of.

SPEAKER_03

Yeah.

SPEAKER_01

Can you explain what a supply chain attack is? Yeah. So I think I always, you know, I think of things kind of statistically sometimes, right? So when you do software development, there have always been bad actors that would inject viruses or or bugs into third-party software that you use to do the development. Generally speaking, that didn't happen at a particularly high rate, but you still had to be very careful about it. Because coding is so much easier, right?

Supply Chain Attacks In Plain English

SPEAKER_01

And you're sort of 10x faster and about 4x more productive, there's infinitely more bugs, right? Basically, right? And the bugs are better is the other problem, right? And so a lot of these sort of open source packages that people have begun to rely on, if they aren't incredibly well maintained, and even sometimes when they are, these bugs can sneak in. Typically, they are you know stealing credentials, they're you know grabbing data maybe off your local machine or potentially out of your CI pipelines. Yeah, so that's that's the thing that I think your average vibe coder may not be aware of, right? And I think is at least one was one of the precursors to some of the problems with how fast we're moving.

SPEAKER_03

Yeah.

SPEAKER_00

So if I'm an executive, I should go talk to my IT department, it's like, hey, tell me what you think about supply chain attack. Like, I'm this is a good nugget for them, right? I always say, like, if you're an executive, you should have good nuggets to test your the knowledge and capability of your IT leadership. And if you go to your IT leader and say, Hey, tell us what what do you think about all these supply chain attacks? And they're like, What? Yeah, yeah.

SPEAKER_01

I mean, the uh conversely, you could ask, like, you know, what what can our platform access and what can it access? Right.

SPEAKER_00

Oh, yeah. Yeah, and I think like the other thing a lot of people don't realize is like GitHub and like what they they just they don't understand, like what I call like the community factor of GitHub.

SPEAKER_03

Yeah.

SPEAKER_00

You can explain a little bit of that, like how like what why you know, I'm sure a lot of these companies are using GitHub. The executives have no idea what that is, right? Yeah, but like how those can be a source for like how people go out and like download stuff, right?

SPEAKER_01

Yeah, so I think so so Git, which is the underlying technology or software behind something like GitHub or GitLab, allows you to effectively version your code and collaborate with other developers so that you can write in parallel. And as you might imagine, it's as if you were trying to write, you know, work on the same book or same, you know, essay at the same time, but not like on the same screen. You know, that's kind of how these tools work. And so Git is a tool that kind of allows that to be possible and combine things easily.

How GitHub Becomes A Risk

SPEAKER_01

And then GitHub and GitLab and Bitbucket and these other tools allow you to then put that on the internet, right? Cloud hosted, so it's easy for everyone to access regardless of where they are. And the really cool thing about GitHub is that over the years, it's it's it's not really, I mean, it is enterprise grade software, but it is a community in a lot of ways, you know. So people like to use their personal account so that they can keep track of like what all they're committing to, even if it's like private repositories or private code bases. But in order to access some of these open source codes or open source code repositories, all you have to do is download, right? So if you if you have access to GitHub and your your IT team allows you to download things, you can go download this code, compile it on your machine, and and you're off to the races, right? So, you know, on one hand, I I'm a firm believer that AI plus some of these really kind of classic software ideals makes everything more powerful, right? So the idea that you code fast with AI, but have human in the loop peer review, I think that's like an amazing kind of force multiplier. But you you do have to, there is some level of experience and understanding that makes it um makes it that way.

SPEAKER_00

Yeah, I think what's kind of super powerful with some of the vibe coding stuff is if you're in a company and you're trying to convince leadership of a software idea. I I think we've entered the show me, don't tell me era. It used to be, let me put a document together and requirements gathering and put a deck together and go in front of leadership and say, Hey, can I go build this? Can I get a hundred grand to hire an outside development team or internal resources? I want to go build this

Vibe Coding Brings Demos Fast

SPEAKER_00

proof of concept. And you know, I consider myself somewhat of an executive. They just we don't get it sometimes. It's like, oh, it's just a lot of words, and I don't get it. Like, what are you saying? You're gonna build a GIS and gesture, like I don't like, no, I'm not spending a hundred grand, stop it, right? Like, that's where a lot of this enough, but now because of vibe coding and like this show me, don't tell me era, people are showing up with prototypes, yeah. And executives, other than dashboards, a good demo gets an executive excited, right? Like, other than like, here's a great dashboard report on a tableau or BI or whatever, a demo around anything gets them excited, right? Yeah, they're like, Yeah, let's run with it. And so that's a pro it's a good thing, right? I love it. I love that my uh my fellow nerds, which I say affectionately, are in the room in the C-suite getting their time, getting their seat at the table, and are being effective with having their seat at the table. However, that's also like very problematic. A friend of mine just took over a company as CTO, and the first thing he did was shut down all the vibe coding.

SPEAKER_02

Right.

SPEAKER_00

It was like, what are we doing here? Like, there's 50 pro who approved 50 projects? What is going on here, right? And I mean, I think the the teams that were they did a good job of showing, you know, let me show you. And maybe we need to start telling, right, as well.

SPEAKER_01

100%. Yeah, I think you know, the the the class classical development process or the you know, a lot a lot of classical things that have a lot of thought put in them are valuable, and you have to understand what to kill and what not to kill, right? And so, you know, if you talk about product management, I think the way in which we do product management is changed forever because the ability to do four, five, six, ten prototypes and test ideas right quickly isn't you know incredible. But then you can't make the argument then that we no longer need product management, right? It's just like how we do product management is forever changed, not the fact that the actual skill set, right, isn't required. And I think in the same way, having non-coders, non-technical folks build POCs, I think is incredibly valuable, right? Because they may not know exactly what they want or they may not be able to explain it in technical terms, right? So being able to cut through this iteration cycle, but it is really important and very valuable, but you have to then understand that okay, we're gonna go back to the drawing board, right? We're gonna take now we really intuit what you what you need. Now let's build it in a way that works with the rest of our infrastructure so that we don't get this massive bloat. And the same thing is true, you know, like uh there are three or four key examples in my career of you know, two or three really bright, very opinionated technical people who could never agree, you know. And if the if the team wasn't ran well, I think it'd be a real problem and just go on for months. And now there is a much more concrete way to solve these problems, right? And and I think that that will be helpful as time goes on.

SPEAKER_00

Yeah, I mean, so we have our small plug, we have our hackathon coming up July 21st and SF, which is now like it's become a series. We do these quarterly hackathons. And by the way, if you've never coded, never anything, you come in the night before, we do a little happy hour gig, we do a tech check, we get you up to speed on like we'll show you what GitHub is, you know, those kind of things. But it's meant to be for non-coders. If you're a coder, you're welcome. We might have you help some of the non-coders out, and we build teams and we do this thing, you know,

Hackathons And Shipping Toward Production

SPEAKER_00

and then the next day we build stuff, and then we have a show and tell at the end of the at the end of the day. And when I wanted to plug that, because I think we have like 10 spots. We we can only do 50, I think, but we have like 10 spots left. So go to kpr.co and sign up. However, the reason I bring it up is we did our first vibe on. I was thinking it was like your first like maybe week or two early, right? It was early, even though you were kind of helping us a little bit advising, but that was back in October, right? And when we did that, when we did the ViBathon, there were people using all kinds of tool sets, right? And then the end product looked a certain way, but was call it X percent away from being production code enabled. Most of it was probably a rewrite to get into production. And then we did our next one, and we just did one in Atlanta a few weeks ago. Like it feels like the gap between what I can vibe code in a day to getting it into production is starting to get closer and closer. And I think the uh the the attendees asked you, like, hey, can you set up a GitHub repository for us? Because my company doesn't even know what that is, yeah, and they won't and they won't let me, right? They won't let me. In fact, then we we had to have a few people borrow some of our engineer because you have your whole engineering team there, they had to borrow some of our machines because the machines were just like locked down, right? Yeah, yeah. So so we are looking for a sponsor to provide machines. If anyone from Apple or NVIDIA out there and you want to we're gonna keep doing these quarterlies, but I we we don't have budget to buy everybody a machine. I would like it if people could just show up and the machines were set, that'd make our lives easier. But I mean, how are we seeing that trajectory from roll wave code to like getting that much closer to maybe hey, you could put it in production in a week?

SPEAKER_01

The the reason I use the word understanding a lot is I think that's the gap. More uh, I think we're we're getting closer and closer to the point where the gap is the user, not the tool, right? Meaning if you have, you know, so our team did a nice job, shout out to Kevin, right? We we started we're starting to standardize on a lot of our configurations, markdown skills that we use to actually develop the code, right? And so if you have that set up in a way where your tool really understands what your standards are, how you deploy things, how you want things written, uh, then I I agree. I mean, I think you can go production grade stuff and days, hours, you know, depending upon the scope. I I think that if you don't know what you're doing, right, I think the tool almost makes it even harder, right? Because of bloat. I mean, it will it'll pull all the latest and the greatest, you know, but it may conflate a little bit, you know, AWS with GCP, with Neo4, JS, you know, like whatever the case may be, right? So but I do think that in the right hands, I agree that that that window to put production code in place is very, very small.

SPEAKER_00

Yeah, I think I went through that because you you have our internal skills file. And the thing I didn't realize is I thought this was I think earlier this year, like, hey, let me quote unquote load the skills file. Like, here's our tech stack, right? And then I thought I was done. And then I didn't realize that Claude might change its mind. It's it changed its mind. I was like, hey, reminder, dude, like this is our tech stack. What are you doing over here doing this other thing? I told you it's Postgres. Like, what are you doing?

SPEAKER_01

Right, yeah. Yeah, I mean, I I think I don't remember when it was, it might have been right after the holidays we were

Standards, Skills Files, And Human Slop

SPEAKER_01

talking, and I was like, you know, KP, I no longer believe in AI slop. You know, like and the reason for that is not that I don't believe that AI produces things that are wrong or but I I think it's largely user error, right? At this point in time where human human slop. Yeah, it's human slot, right? I mean, you know, I I don't actually I can't recite them off the top of my head, but the four D's of AI uh education, right? I think it's the discernment piece, right? That's the most important. So yeah, yeah.

SPEAKER_00

Yeah, because we had a lot of people from the hackathon say, like, hey, can you guys help us put this into production? Because my IT department's not gonna let me. Yeah, and could you help me get it into production and host it for me? And we're like, I mean, that's not our business, but I mean, maybe it should be. Even like these hack, we had so many people ask us, can you come do a hackathon at our office, like for our team? And I was like, Hey Barry, I need 50 laptops. Which by the way, like every hackathon we've had, we've always had a Wi-Fi problem. It's always a problem, right? Because remember, like one of them, I mean, at Hensel Phelps, we did one, yeah, and poor Ty, he had to go get hot spots from T Mobile, yeah, because he couldn't get around the corporate firewall for us to do anything. It was just so locked down. It's like, I guess we have to show up with like laptops and a Starlink, and like maybe we'll have like you know, I've been wanting I'm wanting a business excuse to buy a Sprinter van. There you go. This might be mobile account. I mean, it's yeah. So because I can't just buy a Sprinter van, I have to have like some reason to. So that might be it. Um what do you you know? We've been talking a lot lately, you know. I I think that this fun and the fun in games of AI is still fun, and they're still fun in games, but as we've been deploying with customers and stuff like that, like I feel like we're spending a lot more time talking about privacy, security, opting in. And I don't know that in many ways, I think because of the nature of construction that everybody works for different

AI Policy For Multi Company Job Sites

SPEAKER_00

companies and they show up, right? If you show up to a job site and you say, Hey, we're using AI, you've got one person saying, like, what's AI? And then another person's like, Oh, I got chat GPT. Like, which by the way, my my latest narrative is like when people tell me they use chat GPT, that reminds me of during the internet days when people said they had AOL. Yeah, not exactly close, but not exactly, right? Yeah, I use chat GPT all the time. I'm like, okay. But if you showed up to a job trailer and you got people from 10 different companies there, the level of knowledge and what they're signing up for and how their data is being used, or how they can use other people's data, it's like all over the place. And so now what's happening? Once again, we're having this Facebook moment, we're seeing jobs where people are just like, Nope, nope. AI is allowed it. I'm like, well, I don't know how that works if you're using an iPhone. Like, I don't know how that's enforced. But you know, we've had several of our owners ask us like, hey, on my project, I'm running an airport project. I have to care about all these things. What should my policy and governance and how do I do that? We've been getting a lot of inbound on that. Like, how do we do that? I mean, is it just our industry that's that complicated around this, or is it just because other industries, they have their four walls and they can control everything, so to speak?

SPEAKER_01

I think I think AEC takes it to a different level, right? Because of the like sheer number of entities involved, right? But no, I don't think it's just our industry, right? Any, you know, there are, you know, you can think of manufacturing, defense, like there are a number of different areas where you have contractors, subcontractors, the kind of privacy concerns aren't contained to, you know, to a single enterprise. And as, you know, as you know, we've been doing a lot of of product and technology work on how we can do this for the AEC industry to understand, you know, sort of the what we hinted at earlier, kind of have have trace tracing native AI systems at a project level or at a level that makes sense for the customer, right? So that you're not sort of estimating your costs, you're not estimating what happened, but really using kind of best in class tooling to track all of the decisions and give you know as much insight as is possible, right, to effectively like a neural network, right?

SPEAKER_00

You brought up tokens, and uh, I think we're done with token maxing. Is that is is that out of the zeitgeist now?

SPEAKER_01

I you know, I think at least the early adopters are done with token maxing. I don't know about the rest of them.

SPEAKER_00

But it I think Devin posted something on Slack the other day about Uber has now limited like tokens per employee or something.

SPEAKER_01

Yeah, it was a little a little more than our limit, I would say, but not much. So I was like, I don't know, we're in the right, you know, actually felt pretty good about where we're at. So yeah.

Token Limits And Outcome Metrics

SPEAKER_00

What did I tell you the other day? I was being I was being a nerd. I was like, hey, I think we should do like a distribution of token usage and look at like mean and median, and like when mean and median lines up, that's probably the right point.

SPEAKER_01

100%.

SPEAKER_00

I think you're the only one that laughed at my joke.

SPEAKER_01

Yeah, yeah. I told you like my first job I got because I I answered the question, what is a more meaningful metric, mean or median? Right. I answered that correctly. So yeah. Yeah, I think you know it's funny. You sent me your your Substack to read last week, and within 24 hours, two or three other people that I really respect on LinkedIn posted very similar ideas, right? Kind of the idea that for early adopters, okay, we've spent the last 12 months experimenting, prototyping. Now we need to apply our known good practices to measuring outcomes, right? And so I think you know, the the LinkedIn article that I I really liked that reflected what you said was, you know, programming is easy. It's already a very metric-based thing, right? You have PRs, you can measure output very easily through GitHub, GitLab. And while they aren't perfect metrics, you know, they typically move in the same direction as code quality and productivity. I think for other squishier fields, you know, I think it's it's harder to be able to do that. But you know, in the same way, you know, we talked about our OKRs for Q3, right? The important thing is to have them, right? To have your metrics and to measure them and then to iterate, right? So yeah, I think that's that's the direction that everyone, again, I you know, the early adopters at least, that that's the direction people are beginning to go. And my guess is there's going to be more and more tooling, and the the the front runner platforms that are doing the best with the actual intelligence will start to build that functionality in to make it easier for people to see.

SPEAKER_00

Yeah. I think I did the thing today because when we worked in the at the robotics company, I used to walk in and ask the engineering team like, hey, how many how many lines of right code did you write yesterday? And they thought it was serious. It was so funny. Like they thought it was serious. They didn't know it was like, hey, I'm just messing with you. I did that to some of the interns when I was in Atlanta. And I think they were like, What are we supposed to be tracking how many lines it?

SPEAKER_01

Well, so historically, right, the the idea of dorometrics, which isn't exactly lines of code, but is sort of how frequently you deploy and how many issues come from it, right? I mean, it there's something real, right, to to those types of things. If you look at the top code producers, at least within their their level, let's say, right, it typically aligns pretty well with uh the top performers.

SPEAKER_00

So yeah. You know, you said something that I think is interesting coming from you, right? Which is programming's gotten easy, right? And I have a lot of friends that are really struggling with that. They define like the artisanal part of like creating code and programming. And I've had a few friends just like opt out, like I don't want any part of this, and I'm just gonna go do something different.

SPEAKER_03

Yeah.

SPEAKER_00

Like, like their their personal ethos is so tied to this. And

When Programming Gets Easy

SPEAKER_00

I think it's interesting. Like, I of course I picked out you said that, and a lot of people don't want to say those words out loud.

SPEAKER_01

Yeah. So there's a really good article in the New York Times about this, which is sort of interesting that you you bring up your friends who are less interested in coding, and it it actually contrasted coders who've adopted AI heavily versus say like artisans or or craftspeople, right? And the what the distinction that they drew, which I thought was really interesting, is that the fun part of coding is the ideation, right? That is, you know, I think they called it the soulful part, right? So the the that is the creation, the typing of the code was like not the best experience for most people, right? Versus, you know, if you want to create a graphic, right, that painting, that you know, manipulation of pixels is the soulful part. And so in that way, you can think of like AI haven't taken kind of the hard part out of coding, but left sort of the joyful part, yeah, right. And it's not necessarily doing that for every field, right? So for me, I don't really care about a logo all that much. So I actually really like it, right? Like I'm like, hey, can you can you give me a graphic? I just you know, so I yeah, I don't know. That's interesting. I but I for me, I think it allows your your it a lot, I think it lets people shine, right? It lets their like skills shine and kind of covers up maybe the more difficult or more annoying parts of it.

SPEAKER_00

So well, that's why last year you kind of motivated me, like, hey, dude, get back on the keyboard, it's fun now. Yeah, and I was like, I'm gonna keyboard in a minute, and you're like, it's fun. And then next thing you know, I'm like up every morning building stuff, right? But no, but I think it's really interesting. I'd I published this other article about the idea of you know, and people like to misinterpret a lot of the things that I say, but I was basically on a we're on a team call, and I said something like, I feel like process and workflows, et cetera, like over over-optimizing towards process and workflows is a great path to mediocrity. And someone on the call, when one of our team members said, like, oh my god, like, what are you saying? We shouldn't have any process or I'm like,

Process, Consistency, And Mass Customization

SPEAKER_00

I'm just saying that one of the best process engineering companies, and this was my my article I posted a couple weeks ago, is McDonald's. They're very good at delivering. If you want a big Mac in most places, anywhere over the country, you will get the same big Mac. That is highly optimized processes, workflows, playbooks, all of those things, which is great. It's it's a high quality burger if you define quality by consistency and quantity and quantity. And I was like, so my feeling is like that's what happens. We'll all be like just creating big Macs. I mean, by the way, we're starting to see it. Like, we see a lot, like we know when some of these websites and pitch decks and I mean the user interfaces, you can immediately tell, oh, you're cloud coding, right? Yeah, because people aren't spending the time to to think about those things, which in some cases, if you're a startup, it doesn't matter, right? Like get get it working the right way. But I I think what's interesting about that is I feel like the unlock with AI is kind of like what we were doing at the robotics company was the idea of mass customization, right? If I'm just gonna use AI to build a set of standard set of process workflows, playbooks, and then operate along that, then we don't need people, right? Whereas if you can say, no, I want to deploy 1500 processes, workflows, playbooks, and I'm gonna have AI like massively customize them so that I can, you know, what was the it's probably before your time, maybe, but you know, Burger King had an ad campaign of having it your way. Oh, yeah, I remember. That was their big differentiator, right? Because it you get a big Mac one way. Don't ask them like to not put the pickles on it, right? They'll they'll get you, right? But I I think Burger King, I like I want to have it my way, and I kind of feel like with AI, like I get to have you know, the Burger King, I I get to have it my way, right? At scale, though, at scale I get to do it, yeah. So I I think it's really interesting when people think that AI is taking the creative out of stuff, it's more like you're choosing.

SPEAKER_01

Yeah, I mean we we were talking about this in a different context the other day, but like thinking about things not as like a production grade process, but a production grade artifact, right? And you can kind of get there however you want, you know. So if I'm trying to analyze a data set, or if I'm trying to produce a white paper, or if I need to, you know, put a PowerPoint together for my 4 p.m. meeting, I think AI can help you do that faster, better, and in your own way, right? And I think you don't always have to over-index on the the speed aspect because honestly, cloud makes everything faster, anyways, right? OpenAI makes everything faster, anyways. So yeah, I totally agree. And I, you know, I do think there, you know, there's always going to be value in human creation, right? In some ways, I think AI makes it even more valuable. But when we talk about the types of industries we work with, right? I think you you still add value by using the tools, right? I don't think that's ever in doubt. I think the tools are just allowing you to do the work faster, right? And give you insights at a speed which would have taken much, much longer in the past.

SPEAKER_00

Yeah. What do you think? And and part of the the the listenership here is a lot of AC people, right? A lot of startups, but there's high fixation on LLMs, right? For good reason.

SPEAKER_03

Yeah.

SPEAKER_00

And then you know, we're doing some work in kind of what I would call the VLM space. And then you see, you know, one of a big fan of Fei Fei and what she's doing at World Labs, right? Um, and so the the physical AI and you know that world, right? So maybe like just kind of

Beyond LLMs Into VLMs And Physics

SPEAKER_00

break it down, like how people are thinking about this. So because I I think everybody thinks when they say AI, they think LLM's purely Yeah, sure.

SPEAKER_01

I mean, I think so fundamentally, right? LLM large language language model, right? What is it? It's a word or a token predictor, right? So you put a set of words in, they're broken up, and then pass through a transformer or a set of you know make-believe neurons, and then out comes a bunch of other tokens which are reassembled into a set of words, right? And so that's in it's been shown that that's incredibly powerful, right? And so VLM, right, doing something similar, but is able to also take in images, right, and process those in conjunction with the words and have an output. And then, you know, we're going, you know, further than that in terms of physics-informed models, right? And so the goal there is that not everything has to look you know like an activation, right? So neural networks get their name because basically we're convolving a bunch of mathematical functions that sort of look like the activation spike of a neuron, right? And so that doesn't fit well with things like F equals MA, right? Like gravitational pull. And so the hope of you know, kind of the next generation of AI is to be able to better invoke math in conjunction, right, with LLMs, or at least that's a very simplistic way to think about it. And so for us, I think you know we go pre-LLM, right? We use machine learning for image processing on very specific tasks where it makes sense, right? We're obviously always using and evaluating the top shelf LLMs and VLMs. We're training our own kind of VLMs, beginning to get into that space based upon things we've learned. Uh, and then we're trying to stay to the best of our ability up to date with the state of the art because we know that for us, right? I always, you know, I would say my passion is software at the intersection of the physical world, that idea of physics-informed neural networks is going to be central to kind of how we progress forward.

SPEAKER_00

Yeah. Yeah, I think we're we're walking through kind of an example. If if you read drawings on a framing plan, it's it's there's here's a line of a beam, and here's the the section, you know, what what type of beam it is based on the dimensions on it. That's more of like a little bit of VLM, but also like LLM, right? It's pretty simplistic. So it'll say, like, hey, that's a beam. Cool. But then the next level is like, well, it's just inferred from a line, right? Now I'm going to create a 3D representation of a beam and apply its material properties, and also understand that there's a moment connection there, and that this is actually this being this line actually represents this physical thing. And so then you have a model that understands that, and now I can run FEA, whatever. Yeah, I can do simulation around, like you know, earthquake impact and all that, but you you can't do that from a line, right? It's funny, like I used to work with this company up in Canada, they just they still do wind tunnels, they create my physical models of buildings and they put them in a wind tunnel. I'm like, this is bizarre to me in this day and age. Like, what are we doing? Which led into a whole nother conversation. It's like, well, a building sitting in a wind tunnel does not represent the variables of it sitting in in the actual space with other buildings around it, right? And and modeling that, right? So I went into down this road of like, well, you to do multi-physics simulation, you can't isolate a building by itself, it doesn't exist by itself, right? It has to exist in this world, which is the how Feifei thinks of it, right? Like, how does this thing, this object, have its physics uh you know, uh, you know, angle to it, but also like it has to exist in a world where there are other variables, which is like what I love about all that is that that gets us closer to thinking about robotics, yeah, things like that, but it's going so fast. It's I mean, I think that's where people think LLMs are going fast. Like you're not if you're not paying attention to this other stuff, it's fascinating, right? It's like it's why I don't sleep hardly anymore.

SPEAKER_01

Yeah, I think I mean, you know, so the problem you described, right? Okay, I have a beam identifying it as a beam, pretty easy, right? But then how do I sort of put it in the larger context of like what it's made of, how it's attached, how it ages, how it deals with different weather conditions, if it's exposed. And there's different ways that you can think about that, right? You can take a very straightforward approach, which is to use existing VLM to identify the beam and connect it with the data extracted from the LLM, and then kind of recreate that in like a standard CAD format, you know. And so that's that's some of the stuff that we're beginning to experiment with. But over time, yeah, it may have a completely different sort of embodiment or representation as these like, you know, quote unquote like world models, right? Or physics informed models gain traction and and and continue to increase in capability. We we may or may not get further and further away from this classical like representation, right? Right. Um, so it'll be interesting to see what happens.

SPEAKER_00

Yeah, because I think for us as humans, like we can't process all the details, right? A line is might be enough for us, right? It's funny, I was on a call the other day, and we're keep going with the beam, right? This was another interesting thing. So this client says what was important uh to the client was to have the the longest span they could have without, you know, they wanted less columns for a lot of reasons, but then they were also at a time commitment issue. So

Edge Cases That Run The Project

SPEAKER_00

two things they couldn't have something custom manufactured, and then the other thing was to get a permit to transport it to the location as a wide load was going to put it 90 days out. So the two the the the two conditions they had to meet was it had to be a standard size and it had to like go on a truck that did not require a permit for delivery. And I thought it was like fast because we've been fixating on because I think a beam is easy for even engineers to understand what that is, right? And I was like, oh my god, that is amazing. Because think about that, like if it's starting to think about supply chain availability and how do I get it there, and maybe even like what time of day, like that truck needs to leave at three before it hits traffic on the 101 kind of thing, right? And I think that's where my mind just like that's why I can't sleep, right? It's just like, oh my god, that's and and it might feel like an edge case until you talk to the building supplier and until you talk to the superintendent. Yeah, it's not an it's their everyday case. Like, so I think it's fascinating. It's like what we might think is an edge case to the person that has to live that life every day, it's their everyday case.

SPEAKER_01

100%. I mean, we I talked about this with Dave earlier today, right? We're building some technology that I think is gonna do really great stuff, but as we deploy it into the field, everything we encounter is an edge case, right? We we didn't think of something, right? Or it's it's a unique data type, right? Or they did something that has us scratching our head, and we have to incorporate into how the model and how the module understands things. So yeah, I think as we embed the intelligence within the system, within the models, its ability to predict what might happen, the edge cases that you know, your point, like a lot of these people live with it every day. But what about the edge cases they didn't even think of, right? Somebody forgot to look at the weather forecast, right? Uh, you know, uh traffic patterns changed because of a big football game, whatever the case may be, the the models and the the systems will do a better job of looking at that.

SPEAKER_00

So yeah, so the punchline did this the person that was telling me this was a little bit more of a senior person. And his point was like to tell me, like, I mean, AI can never do that. It it takes a guy like you know, it takes this old guy that's been there done that to understand this stuff. I was like, I don't think that's true, right? And so I think when we when we look at the stuff we're building and being able to gain an understanding from a set of artifacts, that understanding isn't relevant unless you have the persona, the the the view of the persona, right? Like we talk about you know, the thing we talk about is like you roll out a set of drawings and there's a room full of people, everybody's looking at the same set of artifacts, but has a different point of view of the things that keep them up at night and what they're worried about. Once again, like every it's almost like you know, maybe it's the that's the thing. Everybody's edge case is equally important to the success of the project. And we now can take all those edge cases. It used to be like, I mean, I don't know anything about that, right? Right. And so it's it's interesting when if you look at the critical path of any project and budgets, it's the edge cases that get you.

SPEAKER_01

Yeah. Well, we could even bring that back to what we were you know saying earlier, KP, where like AI adapts to your workflow, right? Like AI adapts to your situation, it's allowing us to build custom software for each person. We can think about you know its value as also adapting to these edge cases within the industry, right? That's what makes it special.

SPEAKER_00

So well, Barry, this was a lot of fun. We're gonna do this more often because I think um people could be a fly on the wall on our multiple daily phone calls with it.

SPEAKER_01

Sounds great, Keep it.

SPEAKER_00

The only way I get steps in is like uh call Barry and go get a coffee and walk down the road. But this is great.

SPEAKER_01

Yeah, great to see you. Thanks for having me, Keepy.