Latest / The Punch List / Why smaller AI models might beat Claude and OpenAI

Transcript

Why smaller AI models might beat Claude and OpenAI · The Punch List

• [00:00:00] Tanmaya Kala: Water consumption is minimal in data centers, the net water loss is near zero. [00:00:05] Kanav Hasija: If you ask AI, do X and take care of Y, Z, and A and B as well then it does a better job. It's the same way we humans interact with humans. do this job for me," and, "What do you mean by this job?" [00:00:21] Jon Wright: All right. Welcome to The Punch List podcast, the number one podcast for real estate, architecture, engineering and construction. I'm joined today by TK and Kanav, wonderful co-hosts. And today we're gonna talk about three things. One, AI consumption and the push for return on investment. What that, what does that mean for construction? Two, data center construction. TK is gonna take us through a detailed, outline of cooling and water use and all things related to that. And three, we are gonna talk about the federal government's latest grant from HUD to accelerate plan review and using AI grant opportunity for cities and our view on what that means for the industry. So I think we'll start today, Kunav, if you wanna talk about, um, AI consumption and what that means for demand. [00:01:15] Kanav Hasija: the two announcements that came out last week, both are very interesting. One is what Brian Armstrong, wrote down. he's a very prolific, CEO of Coinbase, uh, o-one of the largest crypto platforms in the US, and he's, he's pretty deep in the tech side as well. What he's saying essentially is, and this graph kind of tells that story. The bars you see on this graph are going up every month, and that's what every enterprise has been seeing as coding became more important and more useful with AI. The spend on the AI consumption, these bars just kept on going up and up and up. but for the last few weeks, every organization is realizing they don't need to spend that much money on these proprietary coding models like Claude or, or Gemini or even Codex. Even for 75% the task or somewhere in that range, you can use the open source models as well, which are way cheaper. So by using the same tokens, which is the black line here, they have reduced the cost by half, which is the bars here. they still use proprietary tokens of Claude and, and OpenAI, but they use them less now. They use more open source tokens, which means the cost have gone down. That was one part of the story with that we are seeing. The other part of the story we are seeing is what, Bridgewater did with the Thinking Machines. So Thinking Machines is launched by Mira Murati, who was one of the top technologists at OpenAI. She was the CTO as well and, uh- [00:02:45] Jon Wright: Bridgewater Capital is one of the largest hedge funds in the world [00:02:49] Kanav Hasija: One of the largest hedge funds went with one of the large AI labs. What they proved are two things. Uh, one is, um- Is a prompt naively written down by someone who is not an expert in the field gives you less accuracy than the prompt by someone who's an expert in the field, right? So experts have a say in how the intelligence works in organizations. And what they also found out is that if you train your own model on the organization's proprietary data, that small model can have so little cost compared to Claude or Gemini, but have really high accuracy. What that means is the proprietary data of an organization, the decisions they make inside the organizations, the meetings they do, the notes they take, uh, all of that becomes so important and will become more important going forward. And both those-- If you combine both those arguments and say there's gonna be less proprietary token used, there's gonna be more open source use as well, and there's gonna be more, custom proprietary models built by organization enterprises on their own data, which will be way more accurate than the large general intelligence. So the question is, is the growth of general intelligence of Anthropic, OpenAI, and Gemini gonna sustain the same trajectory or it's gonna slow down? And what does that mean for construction? [00:04:23] Jon Wright: So good for white collar a- workers, maybe not so good for data center demand or maybe a blip or maybe a reevaluation coming in terms of the Frontier Labs planned CapEx spend if this trend of open model use using your organization's proprietary information continues. TK, what do you think? [00:04:48] Tanmaya Kala: I couldn't help but think or draw a comparison rather, and I don't know if this is a fair comparison or not. But back in the heyday of software, like, a company that built Java or, you know, C++, right? What you're essentially saying is in the software days, anyone can write a software, but a software engineer who's an expert, who's trained can do a lot better job, right? Versus someone who's not good at it. And this kind of harkens back to this prompt engineering thing, which at some point was like there were a bunch of classes on prompt engineering. Then it became, "Oh, no, large language models are so good that you're wasting your time if you think you need to do prompt engineering." And now I'm looking at this study where it's saying, "Well, if an expert writes a prompt, they're like twice as good as a guy off the street," right? So it seems like it's going back how software, uh, industry started. So that's one thing that comes to mind. Second thing that comes to my mind is, uh, in the heyday of software, it wasn't the Javas and like the people who were building a methodology to code were not the ones who were the big winners. I mean, they did well. Now I apply that to Claude, where essentially it's an LLM, it's a means to an end, it's a place where you write the language. Will they become more like in the background and the companies that are able to leverage those models, harness them together, like Bridgewater is saying, will they become the big winners? That's the second thing, and I'm gonna throw the third thing at you. Uh, just, you know, feel free to take notes, I guess. But the third thing is I find the open source model conversation really, really interesting. As someone who likes the little guy to win, I feel like if you have all these open source models that they're never gonna be as good as a Claude. But if they get good enough where companies like Cursor and all are-- who are utilizing them can better utilize them, those small shops might start doing well. You might see more of those small shops come up. [00:06:53] Kanav Hasija: There is vertical intelligence and there's general intelligence. general i-intelligence is where Claude and Gemini and OpenAI come in. And, and the argument is, is specialized intelligence better than general intelligence? Or can general intelligence become so awesome in the next five or 10 years that it can do any specialized intelligence as well? Today, what we're hearing is specialized intelligence has a place and general intelligence is not there. That's what Bridgewater and Thinking Machines said, right? Within general intelligence i-intelligence, we are saying Anthropic and OpenAI have the, have the most funding, and they have the most feedback loop from so many users that they are gonna always be better than the open source models. Sure. But what the world is seeing is, do you need to throw in these big models at all the tasks? You don't need to. Many of the tasks can be done by these open source models as well, while the, the proprietary models will always be much smarter than the open source models, until the, the feedback loop reverses, like the open source gets more feedback loop than the, than the closed source. that's why I think all the organizations who are thinking that they're gonna lose against Claude and OpenAI have a fighting chance now. And they're like, "Hey, we can train our own special models, small models on our proprietary data and still win, and have a specialized i-intelligence than a general intelligence." Right? And to your point about TK about that expert prompt versus a knife prompt, [00:08:24] Tanmaya Kala: Mm-hmm [00:08:24] Kanav Hasija: I think that's not even a prompt engineering problem. That's the alignment problem. How do humans to humans align versus how do humans to AI align? Like if, if you ask AI, do X, and that X is a very broad statement, You're not aligning with AI. But if you ask AI, do X and take care of Y, Z, and A and B as well in that, then it does a better job. What is those X, Y, A and B is what an expert knows more than the person on the street knows, it's the same way we humans interact with humans. It's like, "Hey, do this job for me," and, "What do you mean by this job?" Like, like, "Define this more," right? [00:08:59] Jon Wright: what comes to mind for me is, how do you apply this and, uh, to your own company? one, I would reference what Satya, the CEO of Microsoft, said a couple weeks ago that your internal IP, your internal documentation, decision-making, still is gonna become the key source of value in this world of, you know, increasing generalized intelligence. I think this supports that. So my company, I have the luxury of founding and starting a real estate development and investment company as kind of somewhat AI native, post-frontier models. So we've been very intentional about, documenting everything good file and decision hygiene, if you will. But also just the advent of these note takers, uh, on Zooms, uh, which are incredibly powerful at summarizing, key conversations and decisions, and those are stored. So the-- for us, making sure that those are periodically synthesized and developed into some sort of operating procedure or markdown document that governs how our AI systems think and, and drive output is just reinforcing how important that is and how these models, whether open or closed, are just major force multipliers and leverage tools for industry experts who are already in the firm. It-- Look, there's a lot of conversation about, are these gonna wipe out entry-level employees? maybe to some degree, but there's still going to be a lane for specialization and expertise, even if you're early in your career using these powerful models, and you can now use these powerful models to become that expert. [00:10:42] Jon Wright: I think this is great, uh, for the white collar wipeout narrative, if you will. So let's think about this. How do we quantify this? How is this going to translate into what our listeners care about, which is people building data centers, people, uh, developing data centers? What do we think, you know, forecast in terms of demand that's already stated out there in terms of that trillion-dollar CapEx demand that we're hearing from the hyperscalers and the labs? What reduction are we gonna see if this becomes a growing trend? [00:11:14] Kanav Hasija: What do you think, there are, the cloud hyperscalers, which is your Google, Amazon, Microsoft. There are the neo-hyperscalers or the neo-scalers, the neo clouds being formed every day, they're not building a gigawatt data centers. They're building like, uh, hundreds of megawatts of data out of, you know, close to where we live. And then there are these AI labs who want their own data centers, right? The first question is will the overall CapEx investment into data centers k-keep on continuing the same trajectory as we are seeing before or not? We saw the CapEx investments going up from three hundred billion to six hundred and fifty billion this year. Someone's talking about it might go up to one trillion next year. I think It will not go from six hundred billion to like one point two trillion, but also it, it won't go stay at six hundred billion. It'll be somewhere in the middle. Uh, my guess is eight hundred billion to nine hundred billion would be the right sweet spot of the CapEx increasing. So there'll be a slight slowness in the CapEx increase [00:12:18] Jon Wright: What's the multiplier on that CapEx for construct or what's the percentage that goes to direct hard costs for construction? [00:12:24] Kanav Hasija: So rough r-rule of thumb is construction is about thirty-five percent of the CapEx. so this year we spent six fifty billion of CapEx, uh, the hyperscalers did. about two hundred billion of that is, is in building data centers, Does it mean that's the annual construction volume? No. If I'm building a, a data center for two years, that's a commitment for two years of two hundred billion, right? I think this year the annual construction run rate of data centers is about one twenty billion, so that one twenty will go to like one fifty billion-ish, is my sense, which is the annual construction. It will not go to two fifty billion that they were projecting, right? So there'll be some slight growth, slow stoppage in growth there. Growth is still there. The acceleration of growth will slow but if you see the mix between them, which is how much of the demand has been consumed by Anthropic and OpenAI versus Google and M-and Microsoft and Amazon versus the neo clouds. I would say the neo clouds and the Googles will, will consume more, than the Anthropic and the OpenAI. These open source models also need to be hosted somewhere. They're good, but they need to be hosted somewhere. Uh, we can't use them into our machines today, yet. Hardware in our laptops are still not good to have a big open source model onto our laptops yet. So they still need to be hosted, and be neo clouds and Google and Amazon and M-and Microsoft who will host them. So that's gonna be the mix, I would say. But there's still growth in CapEx. It's just not gonna grow as fast as we thought. [00:13:49] Jon Wright: So we're still bullish on data center development. We're just speculating on how much, is it gonna come down from these kind of really high projections. And that's, that's not just a mix of proliferation of open source, it's also many things. It's political resistance, it's public market scrutiny as some of these large, uh, labs go public. It's a, it's a variety of things. But overall net, we're still bullish. This is a huge boom for the construction, and development industry over the next several years, and that's not stopping [00:14:23] Tanmaya Kala: Jon, you asked predict the future and, if I may, I'll take a stab at that too. I think CapEx will, will in-- on data centers will keep going up. I, again, cannot help but compare this to the, the last big technological revolution, right? So one thing I would, uh, invite all our listeners to do as well is to look up, uh, GPT. When I say GPT, I don't mean ChatGPT. I mean, this is a more historical term. Um, it's called purpose technology. Uh, some historians argue computer was that. Um, all historians would agree that steam engine was definitely a general purpose technology. So what did that do, right? If you look at it, it completely reorganized the society in some ways, so first of all, even if that engine more and more e-efficient, like from steam engine, where we went to internal combustion engine and so on and so forth, we didn't have less factories, we had more factories. Like the expenditure on factories just kept going up because the society just moved to a different thing. Second thing is, with that coming about, it led to urbanization. People lived far apart, and you needed to have all the skills in one place because getting goods from point A to point B took too long. Like you, you need to be growing your stuff, you need to have blacksmith working on the iron. Like y-y-you needed all the skills in one place. Whereas with a steam engine, you could now have railroads which could transport stuff from point A to point B, which completely restructured our, uh, our society that we live in today. You could have these urban centers, you can consolidate factories, you can get raw material from somewhere else, um, so on and so forth. Like you didn't need to be next to a river to transport your stuff. You could be inland. All this happened because of this general purpose technology. Those are the positive sides of it, where humanity became way more productive. Let's not forget the, negative side of it as well. with that also came communism, workers of the world unite and all that. It was a direct, uh, reaction to what that GPT did with consolidation of these factories. There were few winners who were making a lot of money. A lot of people were kind of working in like the Dickensian world of working in factories and slogging in there. And, you know, you got a few people making a lot of money, other people not making as much, uh, until, you know, that problem was solved in the West. You had Soviet Union come out. Like, uh, it just came out a direct reaction to what GPT did last time. So those are, I think, things we need to keep in mind, but the short answer is, I think we're gonna build more-- I mean, that's the factory of the future, I feel, where you're gonna have more data center, more and more world will run on tokens. We'll probably get more efficient at utilizing it, but then more and more people will consume a lot more of it. So on balance, we're gonna be using it more [00:17:16] Kanav Hasija: This also, stems to one more point for construction. [00:17:20] Jon Wright: the Engineering News-Record, uh, maintains a list of top 400 GCs every year. They've been doing it for the last, I don't know, one or two decades. That list came out last month. So they track the revenues of each GCs. and if you look at the top 400 GCs, their revenue, combined revenue of the 400 GCs increased by 11.8%. But if you look at the top 20 who are involved in building the data centers more, the revenue increased by 20% to 25% in one year, uh, which is unprecedented. usually goes up by like 5% or 2% every year. It went up by 20% to 25%, for the top 20 GCs. So that's the kind of growth of their centers that's kind of growing every year. [00:18:03] Kanav Hasija: so we are still bullish on that trend. [00:18:06] Jon Wright: Excellent. Great points, TK and Kanav. Um, so segue into data center construction. There's a lot of commentary out there today about data centers, I would say. It has become the political hot potato across the country, Middle America, et cetera. There's a lot of information and misinformation out there around the inside guts of a data center, what it's doing, how it's constructed. So we want to take a, just a quick look, TK, at how data centers operate, how they're cooled, how much water they use, and I think you're gonna address a little bit of the, misconceptions around that. So take it away [00:18:49] Tanmaya Kala: one thing that's the easy part that I'll quickly address is the water consumption is minimal in data centers, and I'll, I'll go over that because first of all, there are no humans living in the data center, like permanently who's consuming water and taking showers and all that. Most of the water that is used for is cooling. And what data centers do very well is they use closed loop systems. Um, and, you know, uh, I'll say it's not for humanitarian reasons they do that. Just the OPEX cost is a lot lower in using a, a closed loop system compared to an open loop system, which many of us see in our buildings. so like, like on buildings we build, a lot of them were on open loop system, the cooling demand is not as high. Data centers have very, very heavy cooling demand. So a closed loop system, what it does is you have water or glycol, depending on the data center type, circulating within the data center. and there is the net water loss is near zero. So that's one myth that can be busted. So what you're seeing on the left is at a very simple level, it's a closed loop system. So the big picture is in any data center on the left you have a cooling tower. The job of the cooling tower is to basically cool the water down that is used by the chiller plant, which you see in the middle, which chills that water, and then that water is pumped into the data center where it's used to cool the racks directly. And there are many different ways of cooling the racks. I'll touch on them a little bit. But what happens is the warm water comes back to the chiller plant and, uh, from the chiller plant gets pumped into the cooling tower where it's cooled down and used again. So it's a circuit, right? So cold water going out, warm water coming in. [00:20:37] Tanmaya Kala: So if I now open up the cooling tower itself, which is the slide, uh, you're seeing here. On the left is what you're seeing is the closed loop system, this is the chiller plant, right? It's pumping the water in there. This leg is actually warm water, and it is using, water evaporation through forced air cooling to bring the water temperature down as it coils through it. The water is rejecting heat, which is then coming down, pumped back up, and then this cycle keeps getting used to cool the water which is in these coils, which is pumped back into the chiller. Right. So this is the closed loop system. And then, uh, I'll show you what happens from the chiller downstream towards the data center. So, so first part is you got your cooling tower, you got the chiller plant, and then you have the data center itself. Open circuit is, uh, different. Uh, so same thing happening. You got your chiller, you got your cooling tower. Here, it's an open circuit system, so as the wa-- hot water is dropping through, uh, this mesh-like structure, lot of the water gets evaporated out and the cool water basically comes down below, which is put back into theright? Which then supplies to the data center. So there's a lot of water losses here. Uh, we see this in buildings a lot more. Data centers don't use this [00:22:01] Jon Wright: So on the right, you got your cooling towers, then you got your chillers, and then you got the data center racks themselves where cooling is happening. [00:22:13] Tanmaya Kala: So, uh, again, in the closed loop system, whether it's a closed loop or open loop system, everything downside of the chiller remains approximately the same. It's the cooling tower that's different when you do closed or open loop. That's why this slide was critical to understand. but sake of the audience, I'll go over what happens downstream of chiller. So, so you went from chiller, pump the cool water down to the data center racks. three means of, uh, cooling. One is you have cold plates that are getting cool water directly to them, uh, which are next to the processors. The most efficient is you immerse the whole thing in dielectric fluid. In that case, you don't have water moving around, but you have the dielectric fluid moving around. This is the most efficient heat transfers, and this is what you use when the density is really high, kilowatt per rack, I mean. And then this is a more traditional old school one, uh, that we used to use a lot, where you have a rear rack door which is cooled. you have a rack and the rear door is what's, what's being cooled by sending cool water in and then the hot water [00:23:25] Jon Wright: In the middle single phase, are those, uh, vertical lines racks? [00:23:30] Tanmaya Kala: They are actually the, the boards themselves, like the chip boards. So you're looking inside the rack. Whereas on the right, the rear door exchanger, um, that is the rack, and then you're seeing the rear door behind it. So [00:23:44] Jon Wright: are the chip boards actually exposed to the fluid or are they just adjacent to it? [00:23:49] Tanmaya Kala: No, no, no, they're exposed to They're, they're immersed in it. That's why it needs to be dielectric fluid and not water so in that case, you're pumping dielectric fluid, [00:23:58] Kanav Hasija: Just a quick question here. So in the-- The biggest difference is in the open loop system, when the wa- the, the water heats up, water va-vapors go out in the open air, and that's why we need to always keep on recirculating new water supply from the natural sources, right? [00:24:15] Tanmaya Kala: Yes. And what you're not seeing here is there is a makeup water line which keeps on making up the water that's lost. [00:24:21] Kanav Hasija: Sure. But in closed loop, what we're saying is, the water wave vapor that gets up because of the heat, it's absorbed back into the cooling, right? So we-- we never lose the water. [00:24:31] Tanmaya Kala: So you're, you're having two water loops. One is the water loop that's cooling and heating the building, which is this coil one, the second one is the water loop that is used to cool the coil itself. So the coil is closed loop and the water that is used to cool the coils is also closed loop within the cooling tower. [00:24:50] Jon Wright: I assume we're gonna go see the pros and cons here because why would you choose the open circuit if you've got ongoing water cost? [00:24:58] Kanav Hasija: Yeah, that's my question. Like why do buildings use open loop at all? [00:25:02] Tanmaya Kala: No, no, because open loop is cheaper. [00:25:04] Jon Wright: First [00:25:05] Tanmaya Kala: look, day, day one cost is cheaper yeah. [00:25:08] Jon Wright: So if you have cheap water, you're incentivized to use open water, open circuit [00:25:14] Tanmaya Kala: Yeah, if you're not paying a lot for water and, uh, your day one cost is a lot lower or you're a developer who's gonna pass the cost of water to tenant anyways. [00:25:25] Jon Wright: What's the rough order cost delta between these systems? [00:25:28] Tanmaya Kala: it depends on the size. So, so this is the thing, why it makes sense for data centers is, uh, their, uh, their cooling demand is humongous. Like, it's like 100 times a commercial building, right? Per square foot. so if you're using an open source system, First of all, you need a lot bigger open source system, uh, cooling tower, and you're gonna lose a lot more water, like 100 times more. Whereas a commercial building cares less about it because their biggest water usage is just people using it in restrooms and toilets and, you know, if it's a multi-family, you know, showers and stuff [00:26:03] Kanav Hasija: And does closed loop consume more power than open loop? [00:26:07] Tanmaya Kala: No, actually it consumes less power that's a good point because you're recirculating the same water. So, uh, whereas here you're losing water, then makeup water comes in, so you have to pump the makeup water too. closed loops uses slightly less power. N-not enough to say that data centers use less power than a traditional building. we can have a separate conversation on the power side. but closed loop is basically near, zero water consumption. Energy efficiency is way higher, uh, with a closed loop system. the biggest one is it has a higher upfront capital expenditure. It just costs more. And then, um, closed loop systems are newer, so the skill set is still being developed, uh, on people who are able to build these systems, right? [00:26:54] Kanav Hasija: So the CapEx of a closed loop system is higher than open loop. OPEX is lower, and closed loop doesn't consume any much water from the outside sources. It's just looping in whatever it's h- it's heating up back. Uh, [00:27:09] Tanmaya Kala: And the main thing you have to watch out for is last thing you want in a closed loop system is a leak, because it, it doesn't handle that very well. that's the other part where you have to keep monitoring it for leaks [00:27:21] Kanav Hasija: So all the myths we were hearing about, which is the water consumption goes off the charts in a data center is all a myth [00:27:27] Tanmaya Kala: yeah. So the water consumption is not really an issue. and traditional system, which is used in commercial building, the main reason is lower upfront cost, like I mentioned, and there's a large talent pool of vendors and the equipment is not as sophisticated, so it's easier to procure as well. [00:27:43] Jon Wright: All right, third topic, last topic. HUD is offering roughly three million dollars in a grant to cities who are-- want to experiment with, AI adoption in terms of plan review. comments on that kind of thing [00:27:57] Kanav Hasija: I think it's a great move by, by HUD. I applaud their, uh, cause for this. plan reviews are always a bottleneck for every developer, contractor, architect I've ever heard. Uh, they all think it's a black box. They don't know what happens there. and it can be sped up as much with AI as possible. Can AI help do the whole plan review by itself within one day? We can get there eventually. We don't need to get there day one. Even if we can cut off half the time or 80% of, of the time in plan reviews using AI, uh, there's nothing wrong with that. if you take inspiration from some other places, they're not as big as and, and as democratic as US is, but if we look at Singapore, every developer has to submit a 3D BIM model of the building and that gets code checked or plan checked through a software automatically within one day. can US reach there? It should sometime soon. But, uh, government spending money in supporting to incubate these technologies is a great move. I, I applaud that and hopefully we see something good coming out of this, uh, which can reduce the plan reviews by 50 to 80%. [00:29:08] Jon Wright: 100%. I mean, this excellent. This is the role of government research money to incubate technologies to help, uh, jurisdictions come along. I mean, when I look at this, I get excited, but at the same time, just being a practitioner out there in multiple jurisdictions, there is a wide gulf and gap between the capabilities and the processes from one city to the next. Small, large, uh, sometimes places haven't had a large development wave, and they're really behind. there are, cities that are very, uh, thoughtful and progressive on this, and I assume a lot of California jurisdictions are gonna jump on this, some other, uh, et cetera. But then you've got some, both within California and, and others, I just want them to have a website that you can upload documents and then track where it is, uh, in their own process. That would be a huge leap forward. and certainly adopting AI. this will just continue to widen the gap if you have not, at the city level, gotten your, zoning, planning, building, uh, engineering, uh, teams, everyone's involved, fire who's involved in plan review. If you've not gotten that process, uh, you know, brought to, call it the twenty tens, this is going to further widen the gap of falling behind and make your city less attractive for development. Um, so I'm all for this. I've said this before, uh, that AI is not going to replace the plan reviewer or the architect, but it's certainly going to create, uh, an expectation that they move much faster. Day-to-day, I get frustrated now when my civil engineering firm hasn't ran, their draft plan sets through our standard operating procedure just to check comments. Um, we've talked about the models aren't as good at the visual, but certainly comments and call-outs with errors or omissions or that's just, you know, inexcusable and kind of like a free Chat, ChatGPT or Gemini account or Claude account will, will catch that for you. So I just send it back to them and say, "Fix that." I certainly am ready for the time where the city engineer's doing the same thing, because then we're just gonna really move the timeframe up in terms of their reviews. So this can't come quick enough. [00:31:36] Tanmaya Kala: agreed on all things as someone who's been a victim of not being able to build buildings because it took too long for permit. Uh, you know, people change their mind. the business realities change. one of the best ways to, to turbocharge housing build out, getting people in homes, is the ability to move the permit process faster not reduce quality, not reduce quality of inspections or design. It's just, can we do it faster? And we have the tools to do it. I say this all the time with construction. The, the reason I am very, um, optimistic about AI in construction is it can meet you where you are. Like, like the previous technology, like most of the SaaS tools require you to do a lot of form fills and kind of change your processes to really leverage it. With AI, it's, it's less of that. if there is a certain review process you like, certain things you wanna do, you can have AI do that work for you and become a reviewer rather than doer of the work itself in whatever format you want it to be. It doesn't-- AI doesn't force you to change the format itself, I feel. Um- [00:32:44] Kanav Hasija: Let me just steelman one thing on this, right? if you look at all the plan check departments in every city, many of them don't even have a e-permit application where they can see all the things digitally, right? I still remember I was talking to someone, I won't name the company, but one of the big hyperscalers, they were building a, a data center. And the plan check department did not even have a license to Ac-Acrobat Reader, a PDF reader. this hyperscaler said, "We can purchase one for you if you want." They're like, "No, that's-- government cannot, cannot accept money from any pri-private party." They not have Acrobat Reader, which costs, 20 bucks or bucks a year. For that, they had to send three copies of printed documents. They sent a trailer full of paper [00:33:31] Jon Wright: Yeah, these are tens of thousands of dollars in printing costs. TK and I have gone through this before, And they sit in the b- corner of a city hall for years [00:33:39] Kanav Hasija: Right. that's my only steelman is even if AI does it, like we have to upgrade the infrastructure of all these plan checker departments to get to a level that I can use AI. I hope that's not coming in the way of this when I encounter this, I just think about it as this is all of our problem here. I mean, we're all stakeholders. We're all either members of that community, or we're trying to build something in that community, and I just kind of give them a full embrace to say, "Look, how can we help you?" [00:34:09] Jon Wright: Whether it's, uh, providing the license or, uh, giving you some sort of access. Like, and, and, and what I would say is that response about we can't take it is the-- I don't wanna do it attitude versus, um, a jurisdiction that does want to and does want to solve problems. You're, you're truly in a partnership with your city plan checker as a developer and a contractor. We're all kind of in it together here. Uh, so I'm halfway tempted to fill out this application for a city I'm working in now, obviously with their permission, just to... You know, it would be a blessing, and I think they would be open to it. And I think those in the development community who wanna embrace this would probably, benefit from taking that same approach if you're working in a place to help them, um, adopt this as well [00:34:58] Kanav Hasija: Yep Sounds good.