Project Flux
Project Flux is a pioneering podcast that explores the transformative impact of Artificial Intelligence across various industries, including construction, music, infrastructure, and life sciences, with a focus on revolutionising the project delivery profession. Each episode brings to light how AI is redefining efficiency, innovation, and strategic decision-making in project management. Through engaging conversations with industry leaders, technologists, and forward-thinkers, the podcast offers a comprehensive look at the challenges and opportunities AI introduces, the nuances of its integration, and its potential to reshape project delivery on a global scale.
Project Flux is the go-to resource for professionals seeking to navigate the evolving landscape of AI in project delivery, offering insights, strategies, and inspiration to harness the power of AI for a more innovative and effective future.
Project Flux
How Pascall + Watson Navigates the AI Revolution in Architectural Practice
What happens when a leading architectural firm takes a methodical approach to AI adoption? In this thought-provoking conversation with Pavan Birdi, Senior Associate at Pascall + Watson, we explore the strategic implementation of artificial intelligence in architectural practice—particularly within risk-sensitive sectors like aviation and defense.
Pavan reveals why Pascall + Watson chose a more cautious and strategic approach than many competitors in embracing AI, carefully weighing environmental impacts, cost considerations, and risk factors. "We've been quite deliberate in our adoption because we've just been looking at the landscape, reviewing what's worked, what hasn't worked," he explains. This measured approach has now led them to implement Omnichat, an expert multi-model platform leveraging various AI models, with a phased rollout starting with key individuals across departments.
The discussion moves beyond superficial applications of AI in architecture to explore genuine productivity improvements. Pavan shares how he reduced fee proposal creation from 3-4 days to a single afternoon using AI tools, while still maintaining the critical human judgment necessary for realistic project planning. This pragmatic approach stands in stark contrast to firms primarily using AI for generating visually impressive but often impractical design concepts.
We delve into how architectural practices can differentiate themselves through their unique implementation of AI, with Pavan suggesting that "the quality of your service will be all based on your data that you're putting into your LLM." Each firm's specialized knowledge—whether in aviation, workplace design, or other sectors—creates opportunities for competitive advantage when properly integrated with AI systems.
Perhaps most compelling is Pavan's perspective on the human element in architectural practice. With his experience working on complex projects across multiple stakeholders, he argues that AI will struggle with "the irrational behaviors of humans and the unpredictability of us," ensuring that while roles may change, AI won't fully replace architects and project managers. Instead, he envisions a future where professionals manage both human teams and AI agents, focusing their human attention on relationship-building and creative problem-solving.
Ready to rethink how AI fits into architectural practice? Listen now to gain insights from someone at the forefront of implementing these technologies in meaningful, strategic ways rather than simply chasing the latest trend.
https://www.linkedin.com/in/pavanbirdi/
Subscribe to the Project Flux Newsletter!
All our podcasts commentate on the weeks hottest news brought to you by the Project Flux Newsletter!
Get a step ahead on the world of AI and tech, subscribe and get the weekly newsletter to your inbox every Monday!
Hello everyone and welcome back to the Project Flux podcast. Today we're joined by Pav and Birdie, Senior Associate at Pascal Watson. With a rich background that includes stints at Fossas Partners and Kengo Kuma Associates in Tokyo, Pavan brings a wealth of experience to architecture and he's now the poster boy for AI and implementation at Pascal Watson. Pavan, how are you?
Speaker 2:I'm very well. Thank you, honoured to be here, james, how are you?
Speaker 3:Good, really good, and it's great to have Pavan on. I know we've been trying for quite some time to get you on the show, so it's a pleasure to have you here, pavan. There's loads of great experience you're going to tell us about.
Speaker 1:So we'll cover a lot of ground today. We'll touch into project management design with AI as well, and we'll also look into some deeper concepts. So, pavan, there's so much going on in AI and across the industry. People are facing loads of different challenges, and I know that you're facing different challenges yourself, you know, as using AI, but also within Pascal's. So what's been on your mind lately?
Speaker 2:So I would say one of the biggest challenges we have is how we're integrating AI into our workflow and we, unlike you, know there's so many different types of businesses, different types of architecture firms that have a different attitude towards risk. For example, that's associated with AI. We do a lot of aviation projects. That's our main sector. We do defense, so our risk attitude is quite it's very low. We're very, very conscious about it. So we've been quite slow in our adoption because we've just been looking at the landscape, we've been just reviewing what's worked, what hasn't worked.
Speaker 2:So I think there's that attitude towards risk. And then there's also the environmental factors that are associated with AI. We've worked really hard to become a net zero company carbon neutral. You know, we've got a sustainability team that have done really well. But you know, there's quite a lot of controversy behind the demand of using AI. You know Google, meta they've been literally, you know, bulldozing their way into building massive data centers, you know, because they want to keep up with the demand and they really want to be prominent in this space and, I think, to fulfill that demand and build these massive data centers they're not the most sustainable projects that they're doing. So we need to be aware of what we're contributing when we're using AI. So that's why we need to be very strategic with its implementation.
Speaker 2:And then also cost we haven't dived in straight away. We know that there's corporations who have spent a lot of money and invested building their own language models. Uh, so we, when we had a look at that and we're looking at the costs associated to doing that, um, we thought it was too early and there may not be a benefit of doing that straight away. And then, for example, we discussed this previously is, uh, you know, when deep seek came out with their language model, you know NVIDIA's stock just plummeted right Because they realized oh okay, it's not just these massive organizations who are trying to monopolize this. Actually, you can make a language model a lot cheaper. You know they did it for a fraction of the price, so you know there was a cost saving there by not rushing into it as well.
Speaker 3:Although I would say NVIDIA, nvidia stocks back up. I think to where?
Speaker 2:yeah, no, it would be yeah, because they're doing so many other amazing things. Yeah, no, absolutely no, true, so, yeah. So I think you know we've, we've taken, we've, we've been a bit slow, but it's only because we've been reviewing the market. But so it's understanding how we're going to bring that into our workflow, our day-to-day, and how can we do that in an ethical way, you know, transparent way. Um, I think it's important, you know, for architects to, and project managers to get involved in this, to help define what ai is doing in practice, before some generic platform comes about and starts defining that for us. That's the worry, isn't it? Yeah, absolutely so, I think.
Speaker 1:I think that's where I am with that's quite interesting that you say that, because when, when chat gbt came out, and you know the few months after it or the year after it, when a lot of businesses were starting to adopt ai, you know, initially at a superficial level, there was so much going on and everyone wanted to jump on to the bandwagon. Right. Um, you guys have taken a much more reserved, conservative approach now. Had you started with ai back in, you know, 2023, do you think you would have done things differently, having done it now? Because it was all chat gpt back then, you wouldn't even heard of deep seek until a few like a year in or so yeah, no, it's a good, that's a good question.
Speaker 2:I think people were playing with it and I don't think at then people would say, oh, this is great, I can ask certain questions. People weren't aware of the risks. So you say, well, I know, I wasn't, I was still learning a lot then. And then we started noticing that people were using it like they'd use Google, and the risk back then, with 2023, was well, it wasn't that. No one was fact-checking it. Then were they. It wasn't something that you could really rely on in terms of the source of material, but, yeah, we noticed people were using it. We never really thought of it as an integrated part of our workflow. People were more focused on generative AI of imagery. So we know the people we were focusing more on at that time where people were playing with mid-gen and the diffusion models that's interesting because you as an architect would have been looking at generative ai from a design standpoint.
Speaker 1:right, and I guess early, early, early concepts of design can kind of favor hallucinations because you can get things that are quite richer, less homogenous. Equally, you've now delved into project delivery, so the use cases that you have within Pascal's are both from a design perspective and a project perspective. Could you speak more into the differences and to how you use AI within both of those?
Speaker 2:more into the differences, into how you use ai within both of those. Yeah, sure, yeah, so why there was? There was, uh, you know there's a group of guys, um, who were playing around with diffusion models and doing imagery and say, oh look, you know, we're, we're producing this. I I started messing around with um, with uh, the language, chat gpt and you know was doing fee proposals. I wasn't putting any client information in there or any information about the states or anything like that, but I was messing around and I was saying how long would it take me to do a fee proposal? So I was basically having conversation with it to say a fee proposal where I'm going back and forth with my director, I'm getting more information from the client and it would take me maybe three, four days to do one that we're really happy with. I could do that in an afternoon. I realized that I could do do that quite well.
Speaker 2:But that was only because I'm quite experienced in doing free proposals, understanding a project program and saying that's realistic. You know there was a lot of going back and fine-tuning it with chat gpt and saying, yeah, this that's unrealistic. You've put, you know, three weeks for a concept stage for this size project. That's unrealistic. So there's a lot back and forth. Um, I enjoyed that kind of interaction with someone, almost like making someone else do it and telling them what to do. And then I started fine-t specific channel on ChatGPT and now I've got a pretty good one where I can go in there. I've primed it enough that I can pretty much set the briefs, the scope and everything and get a pretty decent, realistic program and a fee against that in terms of resource profile as well, and use that.
Speaker 3:Has that had an impact, any measurable impact on the success of your bids? Are you able to measure that?
Speaker 2:So while I was doing that on an individual level, I know it's been beneficial for me in terms of time. Now we're looking at how do we roll this out as part of our workflow. Right so we aren't using ChatGPT right now. We've reviewed a series of LLMs and we've actually settled on one called Omnichat, which, yoshi, I mentioned to you on Crest before. It's like an expert multi-modeled platform, right so it has access to Claude and to Gemini and to chat GPT-4, I think it takes kind of like the best of both, depending on what you're doing. So we've literally just launched this and we're experimenting with it, and we've basically done it in a phase rollout.
Speaker 3:Is that to everyone in the company?
Speaker 2:Not to everyone. No, no, ideally, we'd get it to a point, maybe in the next six months, that we would roll it out to everyone, though we also need to determine who would actually benefit from it, depending on what their role is, what their individual goals are. But we've, um, yeah, we've been just launched omni chat, um, and what we're starting to do now is we've selected 15 people who we know would probably benefit from it the most, and so, over the next month or so, they'll be using it a lot in their workflows and basically give us feedback just just on that.
Speaker 1:On that, how did you identify who those people are or were?
Speaker 2:oh so okay, so that's a good question. So we selected basically a key individual in every department. Firstly, um, so we may have a technical director, you know, who's reviewing the q&A, the quality, the construction quality in our drawings going out. A project director. We've got our design director, who's got access to it. We've got some of our associates and we've got some people who are purely design focused. How can the language model help them? While they have access to the diffusion model that we've implemented, can they optimize an LLM as well, in that sense, to help with design?
Speaker 2:So there's key individuals, and because we didn't want to restrict it and say, no, only you can use this, only you're going to benefit from it, because we don't know. So it's best to select key individuals. The bidding team, for example. We know the bidding team are, um, I'm going to massively benefit from this. So I think it really depends on who we as an organization wanted to give access to, and so what we said is everyone. Let's give it to every, every type of person sorry, let me make that clear every type of person, because they may see benefits in it that we never initially saw.
Speaker 3:That's why best, way you get to experiment like that and it's one thing I'm a big proponent of is is experimenting, because you, just at the moment, you can't tell exactly what the benefits or the risks are going to be until you've had had a go effectively exactly, yeah it's like having a champions, because they always say you know, snowball sampling is always a great way of spreading the message, right, you start with someone who's a champion in a certain department.
Speaker 1:They see the utility of it. They then spread it across the business. When it, when it comes from those who are managing the ai, it's like well, that's your job, right, you're trying to push ai for everyone, so that's an interesting point. When you use omni chat, though, because it does plug into the different models, do you have a kind of a hypothesis as to which model you think would come out on top most?
Speaker 2:most of the time, um, it's a good question. So when, when we had to look at it, there was a table where it basically showed the accuracy in terms of fact checking of the different models, and I think gemini or was it claude, one of those two was seen as, in terms of the information it gives, was was most accurate, because that's that's the thing.
Speaker 3:It's interesting. You say that there was a an article that I just read, um, I think from the weekend, where a UK university actually had all the models uh, take a test alongside a bunch of students and chat. Gpt failed and Claude failed, and the one that became the highest, although it didn't beat the best student, was um gemini that backstop. They're doing really well gemini 2.5 pros yeah it's.
Speaker 2:It's scary how confidently wrong their responses are. Like if you put some stuff in chat gpt, especially the unpaid one, it will make up facts and it will give you a history, but as to why it's correct, yeah so, yeah. So for those untrained or those are just taking a face value, it's quite, it's quite dangerous. So I think that's one thing that you know, you've got to be aware of when you're using this is to to constantly factor, and and that's the benefit of omni, omni chat as well.
Speaker 1:It's a bit like people, because people aren't always right, but the ones that are persuasively are quite persuasive, with the one you tend to and the people with bias and experience particularly.
Speaker 3:You know they'll, they'll tell you. I've always done this on my project and my gut feel is this but they don't tell you the bit that they're always wrong. So how?
Speaker 1:how do you manage that? How do you manage that at Pascal's? You know, because you're you're aware of some of the hallucinations, you're aware of some of the times, that the AI will typically try to give you a response even though there was no response to give, because that's what it's been trained to do. How do you manage that when you have a user at Pascal's who is not aware of all these emergent the diffusion models, whether that's creating visuals or videos, animations.
Speaker 2:I think for us is we see it as part of the concept stage, brainstorming, being able to visualize your ideas quickly, having workshops, discussions. That's quite useful for the design team. They want to be able to quickly prompt something and say this is what I mean, I think, around design, because I don't get involved with that aspect of it that much. I think for me it's more focused on the project management side and delivery side. But in terms of just diffusion in general, the benefits that the design guys have come back with is being able to generate lots and lots of ideas. But then I have a problem with that. Me personally, you know, to me I think one of the issues and this is something I've discussed with my design director as well is, while this is fun and this is great, if you look at other industries right, you look like aerospace, rolls-royce, I think they're using AI. You know to look at the design of the engine at a molecular level. You know to look at the design of the engine at a molecular level. You know energy efficiency, maintenance, making it more monolithic, so there's less components and stuff like that. You know that's a scientific challenge that they're trying to improve the industry, improve their product. I don't know something random. Like zoologists, I read an article in National Geographic how they're using AIs to map, you know, behavior, environmental data to kind of decipher how animals communicate, the relationships they form. You know this is all scientific backing with challenges For me.
Speaker 2:When I think about architects, I get worried when they're purely focusing on producing these dreamy, amazing imagery which a lot of them come up with flaws. You know you look at a lot of architectural imagery that's generated and I may alienate some of your audience here if I say that a lot of it's unrealistic and we don't want to be distracted too much down this path. You're actually creating more problems, um, more challenges that you know the other disciplines, like the engineers and stuff, will, will, will, challenge. You know, but I that's what I said, I think and and my design director got this as well he said we should only really be focusing on this in the concept stage. You know to brainstorm internally to understand what that means, um, what you don't want you to constantly communicate what your value is. You know, as the architect, you shouldn't be relying on, you know, ai models to make decisions for you. Yeah, I agree with that. So when I was at a previous practice I won't mention any of these names because I don't want to get in trouble I remember we were designing a skyscraper and you know we had quite a big budget.
Speaker 2:So I think we 3D printed like close to 50 miniature models of this tower and we laid them out on the table, we lined them all up and we had visuals and stuff like that. And when the client and the stakeholders came around, were walking them through and we showed them this table which looked impressive. But you had so many variations and we're saying you know, we could go down this route, we could go down this route it was overkill to me.
Speaker 3:Sorry, it was overkill, was it?
Speaker 2:well it's, it's not even that it was overkill. You're effectively communicating to the clients. We don't know which one to go with. You should be saying, look, based on your budget, based on the site constraints, the way your company culture is, your work, the challenges that you have in this environment. You should be saying, look, we've got one to two to three options here. This is the way to go and it's all backed up by science. Or, you know, there's a technical drive behind it. There are, and I think that's probably one of the reasons as well that I I kind of nestled into, uh, workplace design, uh, because I found it easier to communicate to the client that there's a business response to this design decision. So I remember we were working on an investment bank project, so they were doing office refurb. They had a decent budget.
Speaker 2:There's a massive light, well, a beautiful light, well in the middle of the office space, across multiple floors, and we had pushed for a feature stair that connected all the floors into the middle. And I remember the client saying you know well why, why are we going to do that? We've got the the stair calls around the perimeter. We've got three of those. Why, why do we need this one in the middle, we were able to simulate. This was before ai. We were able to simulate movement and user flow.
Speaker 2:We do that a lot in aviation aviation as well. When we're designing large terminals, we think about passenger experience, the different users of an airport terminal and how they move through the space. We were able to do that in an office as well and we were able to demonstrate. Well, actually, what you're doing is you're encouraging people to use that main feature stair and you're increasing sightlines. You're bumping into people you wouldn't normally bump into from different departments by using that as your main circulation. When you increase sightlines to people in the workplace, you increase communication. If you increase communication, you improve business performance. So these are the things that ai can actually plug into. These are the things that we should be using ai for in architecture to to define our um design decisions yeah and I think architects can get distracted.
Speaker 2:You know they can do the fun stuff, but ultimately we should be learning from other industries that do how they use ai to justify their design decisions.
Speaker 3:It's really interesting because, like you say, people get kind of very distracted by the whiz-bang stuff and you know the fact that you can generate pictures or you can generate music or whatever it may be.
Speaker 3:But actually the real benefit and we've talked about this quite a few times everyone is using the AI, as your thought partner, as Jeff Woods said, and, from an architectural point of view, it's using it to supercharge your powers to be able to communicate those kind of concepts you were talking about to clients. And that's where I think the real power that a lot of people I think are missing with AI at the moment is. And you said something really important at the beginning, pavamavan, which was too many people are using it like google, which is true, so they use it and they use the whiz bang. They generate some pretty cool images, which which is impressive, but then, when it comes to actually using it to really help your day-to-day work, they're using it like google still. So how do you, yeah, how are you sort of training or developing people so that they're using it so it can help them in exactly the kind of tasks you just spoke about?
Speaker 2:so I think, um, think I'm thinking of it as uh. So if we, if we take an llm, for example, thinking of it as, like you mentioned, the partner um, or your colleague, you know you're working, you, you've got a back and forth with it to tackle a task, rather than a Q&A session. What I've found is, if you use it in that way, it's a lot more beneficial. It responds a lot better. You're absolutely right, and I think that's one of the great things about why it's picked up the way it has is because it's just based on human language. The conversation you'd have with a colleague if you have that same conversation with someone where you can perform the task, can perform better. If you have a conversation, you're working collectively together towards a task, rather than asking a question with no context and no background, um, and it's just going to give you a generalized response. If you were to just tell the colleague okay, can you do X, y and Z, what is this, what is that? Well, they're going to come back with lots of questions that they need in order to do that task. It's exactly the same. You need to have that kind of conversation with them. So I think, yeah, when we've rolled it out.
Speaker 2:We have had some internal CPDs and we've said look, this is how we want to communicate, this is how you use it. Sorry, have that conversation with them. Just experiment with it. See what works for you as well. Again, like I said, we're in phase one. It's quite early days, so we're experimenting and different people will have different ways of using it and different approaches, which will be more effective than what we initially thought it would be. That's the great thing about this experimental phase is that we're learning a lot from each other, and it's about knowledge sharing.
Speaker 1:And it's been really insightful, because I mean, coming from a place where I'm not an architect and I haven't been in construction for very long, I mean, people tell me how would an architect use AI in the design phase. I'll tell you, it can give you many different variations of a design, and it would have to be really early on, because you can't really connect data points into that generated piece of an asset. Right, it's an image, it gives you inspiration. Within that, though, you've just spoken about something a lot more in depth, right, understanding human connection, ergonomics, human factors, all these different elements, and you said that an AI can plug into that. Quite well, and I've found that really interesting, because what we're now doing is is we're using an ai to help us to understand human connection, when you would naturally think that a human would better understand what it would be like to use a space. Actually, well, an ai-solving point can do that quite effectively.
Speaker 2:Yes and no. I hear what you're saying. I think so. This is one of the issues that I've always experienced is, you know, in this industry which is any industry is people. People are incredibly complex, they're irrational, you know. I used to think that you know.
Speaker 2:You look at the REBA stages. A project is quite a linear process. You've got all these parties involved and they're all working towards the same goal. But that actually is far from the truth. Projects change completely right Through the lifecycle of the project. Different stakeholders have different agendas, goals. They all need to do their job really well. You may have a cost consultant who needs to ensure that the cost is within the projects, within budget. You've got contractors. You've got project managers. You've got the architects who are trying to push the boundaries, be innovative. So everyone's kind of um, pushing, pushing the boundaries and trying to come to a neutral ground, right, um, and then you may have a project that you know.
Speaker 2:Certain stakeholders leave the project, especially these longer, lot, more large, complex projects. They go on for they can go on for years and in that time agendas can change. So when you've got new stakeholders who come in, they may have a whole different strategy of how they're going to tackle it. So I think AI is really going to struggle with the irrational behaviors of humans and the unpredictability of us, and I think that's why I don't think AI will ever actually properly replace us. I can imagine that the roles will change. I can imagine AI agents being implemented into the workflow and project managers now not only managing people and risk and stuff like that, but they're also managing different agents, agents that are integrated into the project, and then you have varying degrees of skills level for that. You'll have people who's so comfortable managing ai agents and being really efficient, and then you'll have people who are oh, I'll just focus on on the end stakeholder, on the clientele. Is that all the consultants?
Speaker 1:is that a world that you welcome because you're in that, you're in the project management space as well, from an architectural perspective, is that a world that you'd welcome? The integration of ai agents yeah, so as a pm, you're managing not just the schedule, the costs, the people, but you're managing a new dimension which is going to be ai, and I guess to some extent you have to then manage the interaction of the ai with other stakeholders and and the people yeah, no, absolutely I think.
Speaker 2:Um, I think, just engaging with any kind of new technology, if I can plug it into my workflow and experiment with it and play with it and constantly push efficiency from my side, I would definitely welcome that. I think the risk of just turning it back to say, no, I'm going to do this traditional way Well, everyone else is going to evolve. Everyone else is going to start to integrate ai into their work streams. You'll you'll either be delivering less efficiently or you'll just completely cut yourself out in terms of competition. So I think I think it's important to engage with it because it's going to happen anyway.
Speaker 1:That's my view on it anyway with a lot of these llms and generative ai itself. They're trained on finite data points made by humans, right, and there's going to be a certain point where a lot of these models are outputting the same thing purely because of similar, prompting a lot more generic inputs, and then the outputs themselves become quite similar, so that might lead to a homogenization of outputs. Now, what do you think that looks like with an architecture, and is there a danger of things looking the same?
Speaker 2:So I think that's it's a good question and I think that just highlights the importance of what individual practices are doing and the experiences of what they're doing. So, for example, at Pascal's, you know, biggest sector is aviation, right, they've been doing it since the 90 the night. So they have a whole lot of data in terms of aviation, how you design an airport, you know the flow of people through terminals against the you know, world-renowned workplace office design. So different practices have their niches and they have their oh sorry, their expertise and they have their experiences on projects, the various projects. They have their experiences on projects, the various projects. They've been working on no two projects, regardless of the brief, the budget, the client, they would never be exactly the same because of the experiences and everything that's happening.
Speaker 2:So if you're able to capture that data, you know, per practice you're going to get different outcomes. And then I think that's where you know it puts pressure and, and I think what's a good thing is, each practice has competition, right, it's competitive data collecting and that intellectual property that you're collating and that you're holding internally to build up your, your data sets or your language model, all of that makes you stand out to your competitors. Um, I think, with all of that information, those experiences, those risks, how you overcame those risks, all of that data that's being fed in, it's going to be experienced different and it's going to be different ways of addressing it. So I think what will happen is, you know we talked about I think we talked about I think we talked about it briefly offline which is about LLMs being, you know, more important than your, than a website. You know a website is your, your identification right. You go to a website, you say, okay, this company does this, blah, blah, blah, blah. But the quality of your service, the quality of what you will do, will be all based on your data that you're putting into your LLM, and I think that's why it's also really important that you're teaching your staff, or you're training your staff, in the quality of data and how to capture that and how to input that.
Speaker 2:Of course, if you're going to have, you know, architects or project managers saying generic stuff and using, you know, doing very simple stuff and not fully optimizing what these tools can do for you, they're going to use it generically, so the quality of the data that they're putting into it isn't going to be great. So it can happen, definitely, but I think I'm not too worried, me personally, there being a generic output across an industry of what's going on. And if there is, it may be that well, that's the best solution, because multiple people are, multiple practices are coming out with the same solution well, there's certain industry, I suppose there's certain sectors where the aesthetic appeal doesn't matter, let's say data centers, for example.
Speaker 3:you know, it's just a box or a shed, but uh, I still believe, um and we wrote about this on the newsletter recently that I still believe, when it comes to art and you take an architecture as an art form, people don't buy the art, they buy the story and that story is a human story about how you came up with certain design decisions, still hopeful that actually what ai will do will take away the grunt work for your sheds that you've got to design and the boring kind of things that are just kind of things that you've got to got to produce and then actually separate the real artistic stuff where you need a signature architect to come in and provide a story around it yeah, no, absolutely.
Speaker 2:You're right. It's the application of ai in in practice, and I think it's not just in construction, it's all industries, right. So if you take healthcare, for example, I went to a talk last week at the Barbican and one of the panelists was talking about a study in healthcare where they developed an AI language model or assistant to take calls from patients, getting appointments, diagnosing them and stuff like that. What they found was the AI was more empathetic to the people in the call compared to the individuals. This is interesting.
Speaker 2:I think this is not so much to do with. It could be worrying from one sense, but I think it's more to do with stress and workload, the the enormous amount of pressure that's on on the existing workforce. So if you can have something like that in place effectively, like you said, james, is it takes pressure off you to deal with the masses or the repetitive or the mundane and to really connect with people, and I think that's what this industry is about right. When I think about the way I approach project management or architecture or stuff I'm working on, it's about dealing with people more than anything, and that's also one reason I think that AI probably isn't a threat to say that it's going to come in, because if you look at a project, previously I used to believe okay, it's a very linear process. The reba stages are perfectly organized. It's a breakdown. It tells you exactly what it is. But what it doesn't put in there is how annoying people are and how unpredictable they are in a project, especially, you know, large infrastructure.
Speaker 3:We're strange creatures, aren't we? That's the thing.
Speaker 2:Yeah, and I think AI relies on logic and systematic approaches and to make those more efficient. The moment you put human nature into that, you know whether it's driven by. You know individual and genders, different roles, different personalities, egos. You know individual and genders, different roles, different personalities, egos. You know it gets really, really difficult to navigate. And that's what project managers are there for they're there to deliver but to manage people, stakeholders, interests of certain businesses and so forth. So I think, ultimately, the application of AI should be exactly like you said, james it should be towards automation, things that we can get done, repetitive tasks so we can focus on human connection and building rapport and those client relationships, and also then to step away and say let's be innovative, let's push boundaries because the data that AI has, they're not going to be innovative. Let's push boundaries because the data that AI has, they're not going to be innovative, not at the moment anyway, until we get to what do you call it?
Speaker 3:AGI. Agi yeah, which is what next week?
Speaker 2:Yeah, maybe, but yeah. So the moment we get to AGI or ASI, I think they say ASI is superior.
Speaker 3:Superintelligence.
Speaker 2:Yeah, but then we've got a whole different problem. We're not just talking about industry, we're just talking about our existence or the issue we have with that. Yeah, exactly.
Speaker 3:I try and break things down to the short or medium term on this podcast, because when we do talk about the long term, you are guessing there's so many directions it could go in.
Speaker 2:Yeah.
Speaker 3:And it depends how optimistic or pessimistic you're feeling on any particular day, particular day. Some days I'm feeling extremely hopeful about the future and what it means for humanity and for architects and people, and for surveyors and for project managers. There's other days I do feel extremely frightened and I know you feel the same. You're sure it's. It's like a bit of a roller coaster of emotions at the moment when we're looking at AI.
Speaker 1:It does. It feels like one week there's a big jump in AI and everyone's screaming AGI and then the next week there's news that these things can't actually do what they're promising to do. So the Apple event or the Apple paper recently about language reasoning models don't actually reason and as problem complexity increases they collapse. And there's some weird emergent properties, right, like some of these models, when they can't solve a task, they collapse, both in their accuracy or performance and the fact they don't actually use all their tokens up. So it's like they just give up. And we don't know if that giving up is because, you know, it's the system actually having some level of learned frugality.
Speaker 3:Right, it knows it can't solve it, so it's gonna, you know, it's gonna be a lot more reserved in terms of what it's using what gives me some hope, though, is or the another study I saw recently was like, when you actually get these kind of multiple agents working together in an energetic workflow and they actually start checking each other, that that improves performance. It's quite interesting. There was a study which showed that it was three. If you get three relatively um unpowerful models, say gpt 3.5 I know that doesn't exist anymore, but this study had 3.5 and you know the the base model of claude and the base model llama the three of them together, checking each other, perform better than the most powerful. I think it was claude's sonic 4.5, which is really interesting because they're actually that. That's the combination and accumulation of models working together in an energetic kind of system yeah, this is where one training technique is having smaller models, moderate, bigger models as they grow exactly so it's an interesting point and also like going kind of corroborating your point, james.
Speaker 1:A lot of these papers they talk about, you know so o3, kind of dropped in performance I think it was deep secret. All of them dropped in performance, right, but they didn't take into account that scaffolding. So, like you said, we're in the age of agents. We're not just relying on pure intelligence of a single model, we're relying on a connected scaffolding of different agents and with different roles and those who have code interpreters or other sorts of tool use right, because we know those strengths using them to their strength.
Speaker 3:So we're in that phase now and I guess, from a perspective or an architectural project delivery, this scaffolding is going to be key, even managing hallucinations, bringing out originality but that's isn't that going to be a really key role of the architect or the project manager, pavan, in that you, you need to understand what model to use for what use case, even within architecture, because if, if you're doing a shed, an area schedule or a door schedule, that that's going to you know they'll probably be a model that's more suited to that than if you're trying to do some blue sky thinking and some creativity. There can be two very different use cases. So part of a role of a well, any professional really is you need to understand these models and understand which one to use for which particular use case yeah, no, exactly um, but I think that's so.
Speaker 2:We're quite early in our adoption in terms of how we've integrated into the business. So right now we have a generative model with. You know, in our phase one we've only put it towards certain individuals, and then we have a language model, um, who the project managers are using. That mostly, um, just to get their feedback. But what we're seeing is that people are starting to customize them according to the specific tasks that they're doing on the day or that they do daily, and I think that that's what's important is that you give that creative freedom, because people are establishing their own workflows at the moment, rather than having something that's generic, that's put in there and say use it like this, because I think I'm saying this is the. This is the thing that excites me the most about it is that level of customization. Yeah to it.
Speaker 2:So, whether it is um, you know, we've got the, we've got some visualizers who are using in a very specific way, and then we've got marketing uh, who are using a different way, in a completely different way, and then they feed that information to each other as well and they say look, I experimented with this, I got this output yeah and the visualizers doing this. So I think I think that's important to just especially where we are at this stage is to allow it to be completely open and give people that creative freedom so they can experiment.
Speaker 3:That's how you get the best outcomes yeah, exactly pavanth.
Speaker 3:I mean it's I can't believe it's been uh sort of nearly 45 minutes since we've been talking. So we're going to have to bring the podcast to a close, unfortunately. But before we do, we always like to ask sort of a more personal question about you. So my question to you is if you could put one belief of yours and that can be in terms of architecture or in terms of AI or in terms of anything else, so one belief or value on a billboard seen by millions, what would it say?
Speaker 2:oh, these questions. It just makes me find it really difficult to narrow down into one point, one belief you can have more than one if you want yeah, no, no.
Speaker 3:I would say it's a big billboard yeah, it's a really big bill, it's got an essay on it okay.
Speaker 2:So my, my worry is why I get fascinated with ai, for example, why the technology. You know my dad was a techie. He had an it company. He was really fascinated. So I grew up around tech, so I got you know, and so for me to go down this avenue and and to adopt it as much as I have and push it in the business. It's not unusual the way I see it, but I do, at the same time, really worry about our reliance on technology right and um, you know I've seen a generation of.
Speaker 2:You know I came, I came from a generation of online gaming. You know going through that and now I have kids, I have a son and I see how embedded technology is. See how long this billboard is. This is a massive bill we're going to narrow it down to.
Speaker 3:We're going to narrow down to a few words, yeah.
Speaker 2:I would say uh, while is there to to assist us and aid us, um, we, we should be able to escape from it, as well, turn off your device.
Speaker 3:Is that a good?
Speaker 2:device. I know it's a bit cliche and I think, just because I see I see my son and I think just get out um run around, um move exercise.
Speaker 1:I think it's so important. What's that saying? What's that saying?
Speaker 3:there's a certain saying was it a?
Speaker 2:touch grass. Yeah, as simple as that, just one. There's two words there. You go.
Speaker 3:We've said, we've summed it up, but it may be a cliche, but it bears repeating because it we really are getting more and more alone on screens, to the point that it's it's, it is worrying, and you know I there's some days where I find myself and I think I've been on a screen the entire day, since the moment I've woken up to the moment I got to bed. That's not healthy, that's not good. So I think it's, that's a good. But see, we got. We got down in the end to a really good billboard yeah, with some editing.
Speaker 1:Yeah, definitely yeah, I've just got a question for you, pav as well, just before we close up. Slightly different track to james, but, um, you know, based back on you. So in your life, if you had a theme song right now that would describe you, what would it be and why?
Speaker 2:well, okay, yeah, I've got, I've got kids, I've got um, I've got some babies at home and so we're just trying to survive right now. So probably be um, yeah, that survive song.
Speaker 3:I'm a survivor, okay, yeah there's a lot of survivor songs out there. Actually, you could make a playlist of survivor songs exactly, yeah.
Speaker 2:Or staying alive, yeah, staying alive by bg.
Speaker 3:So I think that's another one just trying to stay alive, yeah that might be more apt than than we think in 10 years time, when the agi comes and we are genuinely trying to stay alive. But where can people find more about um, either yourself or pascal's?
Speaker 2:yeah, I mean you can look me up on linkedin if you want to reach out there. Or pascal watson uhcom, you can go to our website and reach out to us awesome, okay.
Speaker 3:Well, we'll put those links in the show notes so people can just uh, click, click along if they want to find out more. But, and all that remains to be said is just thank you very much for your time. It's been a really interesting conversation, amazing.
Speaker 2:Thanks so much, guys. Excellent take care. Excellent Take care.