“There's never really been an API for buildings. And this is an impediment to every single corporation, every startup, every investor, the whole industry."
The Nexus podcast (Apple | Spotify | YouTube | Other apps) is our chance to explore and learn with the brightest in our industry—together. The project is directly funded by listeners like you who have joined the Nexus Pro membership community.
You can join Nexus Pro to get a weekly-ish deep dive, access to the Nexus Vendor Landscape, and invites to exclusive events with a community of smart buildings nerds.
Episode 65 is a conversation with Troy Harvey, CEO of PassiveLogic. This is Troy and I's second podcast episode together so check out episode 5, if you haven't yet.
Troy gave me an update this time on where PassiveLogic is in their journey after raising a Series A and more than quadrupling their headcount.
Then we took a deep, deep, deep, deep dive into digital twins, The Quantum Digital Twin standard, how it relates to other ontology efforts, why it's needed, what it enables, what types of AI it's using, and more.
Without further ado, please enjoy Nexus Podcast episode 65.
Mentions and Links
- PassiveLogic (0:35)
- Nexus Podcast 5 with Troy Harvey (0:39)
- Nexus Newsletter #65: Google's decarbonization metrics & a new data layer startup (1:35)
- Nexus Newsletter #45: the best resource on HVAC+COVID (1:35)
- Nexus #14: Exploring Passive Logic's industry-disrupting ideas (1:35)
- Larry Webber (3:34)
- Mike Luscombe (3:43)
- Kevin Vigor (3:49)
- Jay Herron (4:04)
- Cory Mosiman (4:13)
You can find Troy Harvey on LinkedIn.
- An update on PassiveLogic (1:43)
- Where we need to go and how Quantum helps get us there (9:29)
- What is the Quantum Digital Twin Standard? (13:43)
- Why Quantum was needed when we have Haystack, Brick, etc (19:00)
- What it means to embed Quantum into devices (25:16)
- The new concepts Quantum brings to the ontology effort (29:05)
- Revisiting the levels of autonomy and how Quantum helps (35:52)
- Deep, nerdy, over-my-head explanation of what "Typed AI" is (39:57)
Note: transcript was created using an imperfect machine learning tool and lightly edited by a human (so you can get the gist). Please forgive errors!
[00:00:03] James Dice: hello friends, welcome to the nexus podcast. I'm your host James dice each week. I fire questions that the leaders of the smart buildings industry to try to figure out where we're headed and how we can get there faster without all the marketing fluff. I'm pushing my learning to the limit. And I'm so glad to have you here following along.
[00:00:31] James Dice: Episode 65 is a conversation with Troy Harvey, CEO of PassiveLogic. This is Troy and I's second podcast episode together so check out episode 5, if you haven't yet, it's for a mid-2020.
Troy gave me an update this time on where PassiveLogic is in their journey after raising a Series A and more than quadrupling their headcount.
Then we took a deep, deep, deep, deep dive into digital twins, The Quantum Digital Twin standard, how it relates to other ontology efforts, why it's needed, what it enables, what types of AI it's using, and more.
Without further ado, please enjoy Nexus Podcast episode 65. All right. Hello, Troy. Welcome back to the nexus podcast. I'm so happy to have you, can you reintroduce yourself for those that haven't met you before
[00:01:17] Troy Harvey: shared gains? And I appreciate you having me back. Yeah. I'm the CEO of PassiveLogic. We're based in Salt Lake City area. We have people operating all over the world.
[00:01:27] James Dice: Yeah. And we did episode five, I believe it was. So we'll put links in the show notes to that, and I've done several different newsletter updates on the company over the past year and a half or so. So put those in the show notes as well. How about you just start with an update on PassiveLogic.
So since I think it was last may that the episode five came out, so what's been new since
[00:01:50] Troy Harvey: then. Yeah, I think, you know, just to remind you know, maybe your audiences, I think what passive logic is uniquely, you know, going out. We, we see this opportunity to really change the way automation and, and more broadly autonomy, how that works in a generalized sense.
And when we talked last, I think we were pre our series eight race. And since then we raised a series a with a lot of lead investors and strategic investors that are well known in industry. We um, have really grown our team quite a lot. So we've probably five times to our, our, our teams.
And, and really been building more deeply in our stack. And we operate at the compiler level where we're innovating a whole new compiler technologies and language technologies at the heterogeneous and on that level new technologies in AI around that's. At this deep physics level and at the product level.
And so we've really been building out that team. We've had a lot of really great folks come in from Google Brain and Facebook who joined the team recently had a guy who was leading cloud, who was leading the Walmarts infrastructure and, and really kind of putting together the team that we think.
Is going to finalize this technology, but bring it to market as we're working strategics, you know, over last year on that, those goals. So that's a little bit of data w where we are and what we're working on.
[00:03:17] James Dice: Yeah. Those in the nexus community I've seen, there's been several members of nexus pro who have gone over to passive logic.
Do you want to share any of the hires that you've announced or are there too many?
[00:03:31] Troy Harvey: I mean, sure. You know, most recently, just today we announced Larry Weber who was the general manager of JCI's Asia Pacific region. And prior to that general manager up to me while still controls. Mike Luscombe, living commercialization, he came over from Tridium on the tech side Kevin Vigor came from baseball.
This is like a total superstar. Brad Larson, who came from Google brain he's he's guys driving all of our differential programming technologies that the ComPilot level, and really building up some of the open source technology works. Jay Heron who meet many people in the community might know came, came to us six months ago and he's driving our continent guy.
And Corey you and he know one another and I've worked in real priority, so yeah,
[00:04:18] James Dice: totally. Or cool. So, today, What would like to do is dig into the you mentioned the quantum API dig into the quantum digital twin standard. So on the podcast, I've talked to Corey about haystack.
I've talked to others about other ontologies. This would be fun to unpack, you know, you guys announced quantum, I don't know what it was six months ago. Maybe it was last year, even kind of unpack what it is and Sort of how you see the world changing with it. So I guess first, like we should talk about what digital twin is in your mind, because we've also had a lot of episodes around what is a digital twin.
I think if I could maybe compile all the, what is a digital twin to answers and like turn them into their own episode. How do you see the answer to that bad
[00:05:04] Troy Harvey: question? Yeah, I think, I think there's. Evolution that happens here early in technologies, as people are trying to sort out what terminology means.
And sometimes they're not even really related. But they're they're terminology that are easy for people to grasp. What, what is digital twins? We used to speak in more academic terms. When we, early on got going and realized bras digital twin was very descriptive for people, but, but at the same time, I think it's gotten a little muddy as people are talking about digital twin in their own framework.
Right? So, so I'd say, you know, when I look broadly at the industry, I've seen most of the things people are calling digital trends, I would say best are described as. BIM or building information management wrapped into sort of more of a cloud description, language and, and analytics. So where do you take, you know, maybe CAD for building in your attaching data points to, to building?
I guess, you know, there's a bunch of things you can ask. Is that, is that just them or is it actually, you know, formulating something new that we might call digital plan? I think there's also a market question is, is this just analytics replayed? And I think we, we all understand how the market around analytics had some challenges that it's like a little bit.
So what, you know, like you have analytics through a lot of work. Now you need somebody to interpret it, you know, is that enough? Is that meaningful? Right. So, so I think there there's some, a little bit that we can dig into as, as a marketplace on. W what's what's important. What do we have to accomplish as a community?
But what we mean as digital twin is, is literally a digital plan. And we think you have to have some fundamentals of what we would call the word I'm tology what does that mean? And bras it's the physics of what things are. And, and so this describes what it is. Who it is, what something's purpose in the universe is as opposed to maybe how I say it, or some analytics or some description.
And we can get a little deeper on, on what that actually means. But I think when you unpack it, it's really distinct. And I guess the last thing I just state this is ontology is really related notion to digital trend and the root meaning of ontology is. Really a existential term, it was developed by the existential plus versus to say like, what is your purpose?
Was my purpose in the universe? And this has been adopted too by AI and machine learning people to say, well, what is technological person? And we see that, that very explicit ontology formulation of digital twin is that every entity, whether that's a pump or a fan or a boiler or a wall or a person, they have a fundamental purpose in the universe and they have a fundamental way in which they connect to the universe, the other equipment around them.
And that it's, it's really the question of how, who am I in the universe? What do I do as opposed to. Where I'm Telogis is used in semantic purposes to talk about Symantec and pathologies of, you know, how do I describe things with words? All right.
[00:08:25] James Dice: So, so back on the digital twin piece, I think one thing that you're getting at here is like,
We need another level here of a definition to get at like Trump solving true problems. Is that what kind of, what I'm hearing? Whereas like,
[00:08:40] Troy Harvey: yeah, I think that's that's right. I think, you know, and, and, you know, we've, we've kind of promoted that the artistry needs to get aligned around some common terminology.
We promoted a notion of levels of the time needs this reason. That we need to come into a conversation, knowing what we're talking about, you know, are we talking about a level one on a like autonomy topic? We're talking about a level five autonomy talk topic. And for us, what we see as where do we need to be as an industry, you know, next year in five years, 10 years, as opposed to.
Say a lot of the activities that I see in the industry are kind of rear view mirror activities of like, given what's happened in the past, how do we collect up all of that into maybe a semantic that we can describe it, but that doesn't necessarily get us to where we need to do. Totally. Okay.
[00:09:29] James Dice: So where do we need to go?
And how does, how does quantum
[00:09:34] Troy Harvey: help us get there?
So. There's a couple of pieces too. Where do we need to go? I'm going to pull back from the buildings market for a moment. I think, you know, everybody's familiar with fully autonomous vehicles. What we basically need across the industry is generalized autonomy and everybody in whatever automation project you have, your project is some form of autonomy.
And often there's things that are unexpected or unique to that project that may not even be considered classical building automation. Maybe you're asked to tie into a manufacturing system, or what about your logistics systems in a warehouse where you're carrying boxes across a conveyor belt, some robotic shelves, or, you know, an IOT network that's trying to frame.
How everything in a space like comes together with one unified point of view. So these kinds of problem sets were never going to be solved by, you know, how automation has been done in the past programmatically. But as we probably talked before, the biggest realization, I think for the industry as a, and I was just having this conversation with Larry, you know, there are projects that any big integrator's going to have that are on the order.
500,000 IO points. On the, on the large side, 500,000 I opened. So it's like take that into like, you know, context for a moment that is way out scale. Anything that anybody automates outside, the building industry, you go to industrial robotics. You're talking about tens of points to autonomous vehicles.
You're talking about. Dozens of points, right? 500,000 points. So once you realize the scale of that problem and why buildings don't work, you need to solve it with something new, right? You need to solve it with something that can actually bring to autonomy to that system, to be able to deal with the complexity in real time and not expect somebody who's going to be able to digest that in one setting at one commission stage, because it's not possible.
We know it's not possible. So
[00:11:44] James Dice: can you kind of define generalized autonomy for real
[00:11:47] Troy Harvey: quick? Yeah. Generalize, autonomy. Yeah. So, so I think where we need to go for just not only the building sector, but all industrial control. If you first frame, what is say an autonomous vehicle, what's the shape of that.
And it speaks to what is the shape of AI and deep learning today and it's monolithic, right? So, so what happens if you're a, an autonomous vehicle manufacturers, you're training up millions of miles of cards and you're looking under different scenarios and you're training one neural net that is monolithically does this one thing, which fundamentally is actually kind of simple, right?
You're going to steer brake and accelerate. This is not that like, conceptually that part. Vision is a problem, but, but that is monolithic. There's no way to pull steering out. There's no way to pull breaking out. Right. It's it's all one thing. And so we look at the future of like, how do we make all systems autonomous?
It needs to be compositional. It needs to be. In the hands of the average user that they can say, well, I'm going to take this piece and connect to that piece, connected to this other piece. And once I've connected all the pieces together, the system knows what each of those pieces are, how they work together and how to control that.
And that's not a building's problem. That's an everything problem. And to do that, you need some new foundations in AI because deep learning will get you to that compositional. Type of AI scenario that, that you can, as an average user, can shape the AI yourself and generate that platform for whatever problem it is or you're solving for.
And so that's generalized. Tommy is how do you make a system able to a ton Ensley control, any kind of system? It doesn't matter what it is, or any kind of combined set of systems that are interacting with one another. And then how do you make that accessible to the average user?
[00:13:43] James Dice: Do you want to introduce the quantum standard?
[00:13:46] Troy Harvey: so let's talk about that. So, so what was missing for us in order to, as an autonomous systems company, we had to have an Italian system description language, and we built quantum and D one several years ago. And this, this bringing quantum to a public spotlight. So that there is not just positive logic, it's its own standard backed by the department of energy, backed by a bunch of strategic partners that we'll be announcing here in the next few months.
Is, is. To power an autonomous platform. So that's its first goal is you need a description language for this generalized autonomy in order to make that work. But it then has these other really important attributes. One it's the first real, fully descriptive language of a building. And that means everything in and around that building, whether you're talking about the building itself, the, the it systems that are, you know, that's your networking infrastructure.
It could be your logistics systems or manufacturing systems, your HVAC systems, your like all that, that information in it's fully interconnected, real, you know, how it's physics actually works as an interconnected effect. So that solves a, how do we describe the buildings as a whole? And then that solves another really key piece is how do we provide an API for buildings?
Because there's never really been an ETF for buildings. And this is an impediment to every single corporation, every startup in this space, every investor, you know, looking at this space is like, holy crap. The most trivial thing. Require, you know, let's go back to, you know, something as simple as comfy back in the day where you had, you know, warmer, colder, two buttons, you know, that started as like $25,000 per building integrations.
And that's the problem for the whole industry. If you don't provide a real API, you're never going to solve, how do we get these services plugged into buildings so that those are some of the fundamental pieces, but there's some things we can break down. You know, quantum in, you know, it has a lot of faces that go into how you think about like networking of all these devices.
How does everything in a track?
[00:16:00] James Dice: Totally. Well, I want to ask you about the API piece. Sounds like Jay, my friend Jay is working on that.
[00:16:04] Troy Harvey: What
[00:16:05] James Dice: separates the API? Well, I mean, obviously besides the huge integration costs that would be removed, what are the keys to building this first sort of unified API for
[00:16:15] Troy Harvey: buildings?
Yeah. So really when you think about, okay, to get a positive logic control of the grounds, you need quantum to describe it, to make it operate. We can talk about how that gets set up and why that's a really powerful thing for the marketplace. So then maybe we'll, let's step setback. So what you need is two things, right?
You have all this complexity and industry, and we gotta make that. Controllable and automateable and make systems autonomous enough that they can do it in real time to solve the real problems. On the other hand, you also need this to be simple. You need to be democratized and you need to get it out into the hands of not just the the biggest, you know, baddest programmers and the best integrators in the world, but we need to get in the hands of the average guys on, on the.
So one of the key pieces that quantum does is it describes what things are. And that physicality is exactly what are our heroes here at pathologic, our boiler, Bob and HVAC, Harry. Th these heroes, they understand they may not be programmers, but they understand how the boilers connected to the buffer tank and the buffer things connect to a pump.
And that's connected to distribution in the chiller plant and the VAVs and they understand those things. So now having the description language of what things are makes it so that, or their Bob, our hero can go and draw and make that. Just a diagram like he normally does. And what he's done in that process is described what is actually the AI for the system for it to control that description now provides this API by its nature of it.
Is this what we call a compute graph or a data graph of the building so that you could walk from the campus level, to the building level, to the floor level, to the zone level, to the sensor or to the point. To all the variables in the pump, whether or not those variables are exposed through, say backnet or something else, they still exist.
And that, that brain, all the data in a navigable way, that's unified building the building makes it. So now you can build applications once and every building becomes just a one click, one, click, one, click, you just add it from our app store and you don't have to like integrate. With each new building as a different entity.
[00:18:44] James Dice: It sounds like it's also more than other description languages or ontologies, whatever we want to turn, we want to use, it sounds like it's describing more about the building than these other efforts.
[00:18:58] Troy Harvey: Yeah. So, I mean, what, why didn't you need this, I guess is a good question. So we have some initiatives to, you know, to, to work together.
the different existing semantics, like we're happy to see things like haystack bloom, because that provides more labeling people can embed that later. At the same time, we are working on embedding. Quantum might like a full con quantum description. Let's, let's talk about like what the differences are here, why they're compatible, why they work together And why why you need both. So what I would say is first the existing semantics that we have primarily haystack, I think a brick is in that orbit as well. They're meant to do something very different and they kind of emerged that right. You start with backnet or something like that. Nick KNX, whatever.
and you have a very low level description here. I've got a point. It might be analog. That's all I know about it. It's not enough. Let's put a label on top of it. At that point has a certain name. So that's a labeling process and labeling an AI is helpful because we now know what things are. We don't have to discover what they are.
But that's, that just says in some linguistic terms, what it is, it doesn't say who it is, why it is, what it is like. It just, it just is a it's, it's a, it's just a little bit of language. And so I often use the example of, to separate these concepts. No back nets at this low level of the protocol, the bits of the protocol, the Adams of the protocol, haystack or Breck would add this next level of, of labeling.
And then what we're doing is adding what is this true ontology, this like existential piece. And if you put this in robotic terms, if I had a robot navigate around in my office, that bottom level, it's kind of like the robot navigating with its sensors to know, Hey, I'm not coming bounced off the walls.
And it's like having rain, getting around the building without wanting to end the dunes the next level, because you have cameras and you can say, Hey, that's a door, that's a window. That's a person. That's a dog. That's a Bobcat. So now things are labeled, but you still don't know what a door is, right? Like that a door doesn't tell you anything.
So the existential nature of the doors, the door is actually a router. A door allows you to route from one to the next room and it hasn't been calibrated. It screens on hinges. And that, that physicality is the fundamental description of what to do functionally. and so that's where I say this, this, this, the differentiation is between what quantum does is it describes that existential piece that is missing from just the labeling piece.
Now it turns out if you have the extension piece, it's often easy to work back to the labels. But the labels can feed into the existential piece too, to help inform what things are at commission town.
[00:21:57] James Dice: So, where do you see all of these, like longterm? We have all these separate efforts, right? Where do you see that going?
Do you see a convergence? Do you see them just like being translated automatically between each other, by some other middle tool? How do you see this going with all the different efforts out
[00:22:15] Troy Harvey: there? Yeah. Well, I, I, one, what I see is the Rema remarkable amount of excitement we get from large. Entities in the equipment and in the, the space on quantum to solve their overall problems, but that's not incompatible with making back net smarter that making Baton that smarter will always make things easier.
So efforts like to dudes. That that's the right thing to be doing. Haystack, the more people like do labeling or embed quantum, the better off everybody is. So the thing that quantum does that is, is different. What we're not looking to do is invent another standard for semantics. There's plenty of those.
What we're looking is that Metta piece, that is the, is the singular, like, what is things. Ontology that you can connect all those into and the more you have the richer it becomes. And so it really doesn't matter which things, you know, does matter. Start taking a pole position, start moving into commercial, right?
That gives us more information. Is, is haystack going to do that rate that gives us more information is, or we're going to see more direct embedment of quantum? Well, terrific. We have the full, like, you know, description, but all of those efforts can work together under the singular. Umbrella of bringing that into a singular, like, well, descriptive ontology of the X distinctions of what things are.
Is that too abstract? I mean, you know, we can, we could kind of try to break down like that, that abstraction of like the existential, what things are, if it's, if it's not. Yeah. I'll,
[00:23:59] James Dice: I'll see if anyone reaches out to me, but I feel like I get it with the robot example. I think that was, that was helpful.
Hey guys, just another quick note from our sponsor nexus labs. And then we'll get back to the show. This episode is brought to you by nexus foundations, our introductory course on the smart buildings industry. If you're new to the industry, this course is for you. If you're an industry vet, but want to understand how technology is changing things.
This course is also for you. The alumni are raving about the content, which they say pulls it all together, and they also love getting to meet the other students on the weekly zoom calls and in the private chat room, you can find out more about the firstname.lastname@example.org lab. Start online. All right, back to the interview.
You, you mentioned embedding this into devices. I, I have struggled with. You know, following haystack and following others in that there hasn't been enough adoption from device manufacturers in self-describing using these standards. I don't want to throw Stevens under the bus, but like Siemens is very active in the haystack community, sponsoring everything, doing all this.
But how many of their devices actually self-described using the haystack standard, right? I'm sorry, but there aren't very many, like if you go to the branch office in my local market, they're probably not using it. So what do you mean by these partnerships that are growing, where they're embedding quantum into their devices?
Is that, is that what you mean?
[00:25:25] Troy Harvey: Yeah. So let's talk about on a three, three levels and help quantify. Is powerful and effective for the end user, regardless of whether or not they embed it. So first with quantum, when you're on your hive and you're in town, any studio and you're dragging and dropping your different parts into your diagram, we know what those parts are right in the hive.
So it owns who something is as a proxy. So a device doesn't need to know their quantum like model on their own, because what you're doing is defining, Hey, that's a pumper about. At the point that you've dragged and dropped it, and now we go connect out to it. We know that we're connecting to a pump. And if that pump is just backnet well, we can get you part way there.
You're going to have to do some of the visual wiring of, of the points yourself. If it has haystack that can help it further so that we can automatically wire up more in the commissioning phase, if it has quantum it can fully describe itself and a little bit more detailed that it's not just AP.
But it, and, and we can have with our manufacturer partners, we're going to have a very rich library of manufactured models. So not just a pump, but manufacturer AEs pump with model XYZ still we're proxying that for the pump. But if we ask that pump who it is, it can have a quantum model, not for its manufacturing.
But for that exact pump, as it came off the validation of their, their manufacturing line. So it provides a level of detail and information that wouldn't otherwise be feasible. So, firstly, I just want to set that aside because I think that's, that's an important difference here to pass the logic. It's not crucial that it's embedded.
We can approximate those things and we make the dumb, the dumb sensors and the dumb controls, the equivalent of smart sensors and ING devices by proxy. But when it comes to, why would a manufacturer want to embed quantum first, you got to understand that quantum actually solves problems at the point of manufacture.
What we just talked about is you have unique information per device. That's coming off the assembly line that you need to store. And this gives a common format. That's solving problems, that manufacturing, that then solves problems at configuration and control that then solves problems at runtime. And so that whole arc of how does quantum help you in your ecosystem is the device manufacturer is, is very unique to, to what quantity.
The other reason, I guess, that I would say that we're seeing uptake here is a couple of things. One, I think most manufacturers have come to this con, this, this realization that there may be coming to the end of how much innovation they can have within the. Of just their one component that they sell.
They've maybe innovated that to a point that they're not that differentiated from their competitors. So what makes them differentiate is how well they work with controls and with passive logic, really leading that market edge of this is the future of controls. There's a lot of reasons why manufacturers one or regress to that.
Yeah. And I'd say haystack got a little bit at the same, there was other efforts similar to haystack that that didn't go anywhere. Why did people start using haystack? Well, they had a killer application for it, right? The sky spark. Yeah. So we have a killer application for quantum and that's, that's the passive logic, you know, control ecosystem.
[00:28:58] James Dice: Very cool. So can you talk a little bit about the, kind of the, for, for the ontology nerds out there? What are the, like the new concepts that quantum brings? To the ontology
[00:29:10] Troy Harvey: effort. Yeah. It's so beyond being a fully descriptive on Taji of everything, the building and everything in and around it we bring these new concepts that are really crucial to how autonomy has to do.
And how this new type AI we've been working on functions and that's actors, quantum behaviors. Those are the key three key pieces. And there's two subtler pieces that are similar to what people have experienced before, particularly properties. That's more similar to other semantics. We focus on the S on the physicality there and then computed properties.
Well, I'll focus on those top three, because they're really important to this new kind of AI that we call typed AI or a heterogeneous on that. So first, we spent a lot of time looking at all the equipment and all systems, not just building systems, but you know, logistics sets, doms it systems manufacturing systems, process control.
And we found that there were actually only nine actors in, in the universe of systems. And this was somewhat surprising that there are only nine roles that things play. For instance transport, transport. A Quanta, then we'll get to quantum in a minute, but it's, his job is to transfer quantum from one place to another.
So that could be a pump, moving water or fan moving air or a conveyor belt moving back. A router. We talked about a moment ago about outdoor is a router while it's from one room to another. Well, that's the same notion as a valve routing water from one part of a hydraulic system, another, or a damper and an air system, or a, a literal it router that just routes, network traffic.
Those are all doing the same job. So this notion of actors really important to solve a key thing that quantum has to solve as this existential on ontology. Who am I? So we often in, in, in semantics talk about like, well, I've got a machine to person communication, then I need things to be more machine-readable and things like haystack try to be more human readable.
So it's mixed sense. You have machine to machine communication. That's, there's a few projects like that. Like a one and two end, the tries to solve machines. But there's this fundamental thing. If you're going to have a ton of meat, you need this machine to self, like, who am I? Right. If you can't self who am I, you can never control yourself.
And, and so actors are really important notion of who am I? What is, what's my job in the universe. It's counterpart is Quanta. What is the thing? That's the currency that actors act upon and it turns out again, there's a limited set of currencies. And so fluid Kwanzaa are are one Quanta mechanical quants.
And now these have sub types, like, so Quanta could be a phase change. Material could be water. It could be air. Those are all different kinds of fluid points. A mechanical Quanta could be a translational mechanical Quanta or rotational or 2d vehicle. These, these are all like, how do I interact with, you know, my actor?
What does it do with the substance that it's the currency between them? So these two concepts on the high level are required for understanding yourself to autonomous as a control. On the low level is the foundation for a new kind of AI called typed or heterogeneous neural nets. And it
solves this problem, which is we talked moments ago about that.
[00:32:37] Troy Harvey: AI has tended to be a monolithic. You train a car to drive. It only knows about driving a car and there's no way to tear out just the steering from that. And then compose things together. Quanta are the types by which actors connect to each other. That means you can develop your own digital twin for a bow.
And I can develop my own digital twin for our pump. Both of those have the Quanta of a liquid and they just connect together. And I don't need to learn about a pump and a valve together as a whole new entity. They can compose themselves in real time. So now this enables people to make their own custom autonomous systems and gives you that generalized autonomy at the AI level that now because of types or because of Quanta interfaces, you have this composability of different fragments of model.
Is that right? Does that generally read? Yeah.
[00:33:39] James Dice: I mean, It makes sense to me.
[00:33:43] Troy Harvey: So then, then there's one other piece there that we didn't cover, which is behaviors. Okay.
So an actor, if you're a, for instance let's say you are a transport actor. You will have a certain set of required behaviors that you have to do.
[00:34:02] Troy Harvey: And so when somebody is making their own digital twin, you're saying, okay, this is an actor that I'm making a pump. It is a transport actor with a fluid or subtype liquid Quanta. Well, that has certain behavior requirements that you're going to have to fulfill that will then fulfill what fluid quantum need to know in order to move and understand pressure and flow and all that stuff.
So behaviors are then these sort of known notions and I'll take a simple one. Resistance or resistance behavior would be pretty much any hydronic or liquid actor is going to have to fulfill the resistance behavior. Well, the resistance behavior is also true of any other kind of energy type quantity.
So. In the energy type quantities, you have mechanical Quanta, you have energy, quantity of liquid one, and you're the only nuts. So that it's true for the autonomous system that I don't really have to care about the details at many levels of how you work resistance, resistance in a meta way across all of these different types of systems.
And so what you're doing is you're providing these higher level notions of behaviors of this. We're talking about, like that robot again. The rate robot has behaviors. It it's going to have the behavior, let's say opening that door, that is a rider. And that opening behavior is a behavior on the robot that you, you can fulfill what that means.
So these three concepts are really new to not just the building space butts ontologies in general, how do you build AI in generalized autonomy? On top of these core notions of what things are ecstasy. Cool.
[00:35:47] James Dice: And then I guess my question then is then
[00:35:49] Troy Harvey: how
[00:35:50] James Dice: maybe we can just bring it home and like revisit the levels of autonomy, autonomy.
Like where does that get us now that we can describe a building in that way? We're now allowed to do what? Maybe kind of like wrapping a bow.
[00:36:05] Troy Harvey: Yeah. Yeah. So you know where, let, let's kind of like talk about where the industry is today. So the vast, vast majority of everything we see out there is this level one autonomy where it's just manual control.
We're doing PID, we're doing sat points. That's all like hand program. You see a few efforts. There's probably half a dozen companies in what I'd call level one economy that says same dumb, controlling the building and same kid. Since that point, you put some kind of machine learning AI on top, whether it's local in the cloud, that's just tweaking and tuning the variables of dumb control.
We believe just going through the set of logic that it's not possible to get that stack of technology to level two autonomy and, and sort of breaking these down. So I think, you know, like let's go to the top answer, work backwards. So level five, autonomy is fully autonomous. The building is making its own control decisions, meaning you are coming out with however you want to frame that your own sequences in real time.
There's no, pre-canned, you're saying for this given moment what's happening in this building, all the inputs that I see am I. Able to generate a sequence that that generation of sequence is really the fulfillment of level four autonomy of like, can you generate sequences? Relative. So you're, you're basically saying this is the path to take the level five autonomy is, can you bring it all together as a bow where I am taking all of the systems in coordination and the building solving it as a whole system and coming up with the sequences for all of them.
Fully autonomously. There's no assistance. There's no human assistance in the loop. And so this takes us to that point of pure, true autonomy and not just in buildings, but it could be, you know, again, a process control or manufacturing or whatever kind of system that you have to manage. We've built on top of that level.
One to five, where we see the building industry has more opportunity where the system can do even more than just run itself, but help guidance on install, help guide its own permission. And those are also these interesting outcomes of this approach. Is that by definition, if you define the system, now you can then guide the person through the install because the system system's supposed to look like, and you can look at the depths of the deltas of how the system does, the reality doesn't match the design.
And that's true for commission as well. And so it sort of unlocks these other steps as well. Got it. Is that, that makes sense.
[00:38:37] James Dice: Yeah. And I've seen, I've seen the levels before as well, so that I think those were six and seven. Those two that are added on top
[00:38:46] Troy Harvey: six and seven. Okay. Look, that's our proposal for industry.
I think as industry, this should be an active conversation and I'd love to see instant come together and settle down. Like, Hey, we all agree on that. So that we're not just like digital twin AI, you know, like that people are talking about.
[00:39:05] James Dice: Yeah. Do you guys get any feedback? I I'd imagine that most people aren't thinking at that level and therefore you're not getting a lot of feedback, but what's
[00:39:14] Troy Harvey: your experience been?
No, I mean, yeah, one in this industry, I think most people aren't thinking at that level and most people are just like, Possibly it solves all these problems. That's really awesome. And they don't really see any, anything else in that category. But I think as we see more things emerge in the marketplace and we certainly will, we should have a framework for talking about clearly because at the high level, if we're talking on investors and we're talking, you know, in an ashtray conferences, we should have some specificity to, you know, what are the experts talk to you about?
[00:39:49] James Dice: Cool. So there's a couple of concepts that you've talked about that I want to make sure I understand before we move on one is you you've said typed AI, what do you, we just walked through actors and Quanta. What do you mean by type.
[00:40:05] Troy Harvey: Yeah. So, when we started this and we said, wow, deep learning is not going to do it.
You can do certain things with you, deep learning, right. You can have dumb controls and you can learn with deep learning, some pattern recognition of how to tweak them controls to adjust them. But there's no way to get deep learning as it stands to ever control it. Got it. And the key, the key fact here that, you know, people should kind of get clear in the mind is one of the big differentiators between cars and almost every other thing that you might want to automate is that almost all the other things are one-off unique projects.
There's no duplication, right? So if you're doing cars, you can, you can train that neural net and it's monitored. And you can sell a hundred thousand of them. Okay. But in buildings and factories and all these different scenarios that we have to in industrial automation, everyone's kind of one-off even the ones that are seem like they're patterns.
Every other customer I talked to is like, I thought I was going to do a VRF and then there's always like this one extra thing. Yeah. So, given that everyone's unique, if you're taking a deep learning approach, you'd have to learn on that unique building because deep learning is monolithic. It's not compostable.
And let's think about how slow buildings are. They take a whole year to go through a cycle. To get enough cycles of data that you've got, you know, a good scenario of all the different, you know, situations and states and people activity and weather me, you're talking hundreds of years and that's just not practical, right?
Like all commission you're building in a hundred years. so how do you solve this? So we talked about the composability Cades and that's now built on a new kind of framework for AI to understand. What this is all about. It's important to understand that positive logic work we're literally inventing new language and compiler infrastructure to, to drive our AI ambitions around our product.
And you know, Brad here who came from Google brain is, is driving quite a good deal of that. And the key piece of information here that's that is maybe new to most people, even most developers.
It's differentiable programming. So here's the kind of underlying, you know, sort of breakthroughs is differential.
[00:42:27] Troy Harvey: Programming is a way in which you can run code backwards effectively. So I'll make the simple, why this matters for the history of all computer science. There's been what we call timescale problems or often referred to as intractable problems. We, we, we stop short of condoms. Because they're technically possible.
They just may take so much time. That's not feasible. And I'll give you example of this, you know, in neural nets and what unlocked things in deep learning was differential programming, which is if I've got a known that that recognizes cats or whatever thing you're using your deep learning for, you might have a million neurons each with a variable that you can tweak.
If you have a million inputs to get an output, is the cap or not. It turns out trying to find the right tweaking and tuning of a million variables is an exponential time problem that may take you thousands of years in the fastest time. And that time problem is if it takes one C like one minute to solve a one variable problem.
If you have 10 variables, it's one minute to the 10th power. If you have a million variables, it's one minute to the millions power and all of a sudden you're in trouble. And this, this has like held back whole swaths of computer science. There's tons of problems that fit in these categories, these timescale problems.
Okay. Through a tick, a trick of return reversing the problem. That's not just the same problem in verse yet. You're saying I've got one thing to tweak and I now get a million things out and you turn what has been an exponential, tiny problem to a flat constant time problem. And this is what's unlocked, deep learning to happen.
Now that's, it's not because the neural nets are new. They've been around for 50 years, but in, in, in deep learning differential, programming was. Really kind of a hack because you only had one function in all of the neurons that you are differentiating, which meant what is its derivative function? Well, you only had to define it once.
And so that was an easy thing to hack and what we call back propagate in deep learning. Well, Our work over the last couple of years with a couple other actors in the open community is to have been to let's generalize differentiable programs. And if every code, you know, chunk of, you know, functions, whatever you're doing, whatever you're writing, everything you wrote, you would automatically get the reverse derivative of that code.
Now, all of a sudden, all kinds of things are possible and it solves a broad swath, the CS problems that have, how did people do that? Well, that's what we've been doing on the component level. It's specifically with the swift language work with apple and Google, and that we've been, you know, recently collaborating with MIT.
There's a very small handful of, of ComPilot engineers who are, you know, really understand how the step work down deep and swift we've started really evolving that language because that's some unique features around what we call types which is like types of. That the language itself is know what a type like a float or an ant is and the relationships to one another, as opposed to, they're just like hard coded magic values.
Okay. Well that piece, so that that's a little deep, but that keys okay. And allows us now to have generalized differentiable programming. It's no longer linked to deep learning in the simple use case. We can do it in all kinds of use cases. And then we can build quantum on top to describe these types of Quanta, these types of actors.
And they can all work with that same machine learning power that we've seen in deep learning. But now with things that actually understand their own physics and their own interfaces to one another. And so that's really like what, what quantum on the bottom side is really enabling is, is it is really a description language for the AI on this whole new kind of, of differential partners.
So hopefully that stuff that was described well enough, that it wasn't too deep, but that that's like kind of the underlying, you know, technology,
[00:46:39] James Dice: I'm going to make it my goal to replay that section back to myself until I, I understand what you just said,
[00:46:47] Troy Harvey: aren't to lose all your, all your listeners in the process.
You know, really to try to make these deep technology things as tangible to people as possible. And then through our product at the end of the day, people don't have to worry about that. Right. It's it's all enabled to make it like people, right? Like w R I, I tell our team, our job is to make deep AI, turn that into people.
Right? Like make that people both. Yeah.
[00:47:19] James Dice: Yeah. And certainly like the drag and drop interface. I'm going to put this pump here and this, you know, expansion tank there and everything happening in the backend that that's yeah. That's certainly understandable where you're trying to get to for sure. Can you talk a little bit what, like where you're at with that piece of it as well.
So like creating the digital twin and, and really, I'm also curious around like where you guys are at as far as bringing the product to market, like where, where are you at in that cycle of things as well?
[00:47:49] Troy Harvey: Yeah. So all the stuff we just talked about that's very deep. There's, there's you know, a few players that I could have all the differentiable, like compiler engineers, products that have in our conference room worldwide.
You go up the stack to quantum. There's a bunch of industry players that we're continuing to work with. And I, I see that expanding very rapidly over the next year. That, but again, that's a small community. I think we see that in all of the Semantics. And one of the problems with these description languages is they're a little too inaccessible.
So the first pice you know, taking inaccessibility and make it accessible is we're building a piece of software that will be available to our partners in the next month or two called quantum crater. But again, it's own very easy to use. Graphical soft built into the digital trends.
And so this again, takes it out of the hands of the PhDs, the people who like swing no Modelica and energy plus around and puts it in the hands of, you know, what I'd say is more average engineers and technical people. But that's, that's a, a certain band now we've increased. What is you know, often obscure in these semantics to a much less scare, but then how do we get into the millions of people's hands?
And that's where a tiny studio our software. I think that we've talked about in the past. Everybody can do this, define their own town system, draw it and, and go ahead. And that's, that's really your minute millions of people face. So we see what we're doing at multiple levels of how do we engage with different audiences all the way up the stack to the point of the hive controller, which we'll install to your building in this very practical way.
That product right now is going into private. Okay. Won't be gone in a public data in the beginning of next year, and we'll be going into general market sometime next year. As you know, supply chain is kind of a everybody's problem right now. So we'll see, we'll see how supply chain goes. And when, when the exact release date, we'll certainly let you know where I'm going to announce that.
In the meantime, though, we will have these other pieces that we're working with strategics on quantum, quantum creator, and a ton of the studio. Cool.
[00:50:01] James Dice: And do I have it in my memory remembering correctly, in that you can, when you do go to market, it will either be through a supervisory or controller that pulls in the existing stuff, or basically a full stack that you is that kind of
[00:50:17] Troy Harvey: passive logic is like an solution.
So our hive controller is the core product that. Everything we've talked about is down in the engine of that thing with autonomy studio on the screen. And you can run from beginning to end of your project, including the setup, the networking, all of the decisions, making all in one box as, as it as a single point solution.
And then you might put one or 10 or a hundred of those in your building and how big know number of IO points you have to manage. And then in a retrofit scenario, we'll be talking down backnet to your tier two, tier three tier four type control devices. But we're, we're always that, that wrap up at the top.
[00:51:00] James Dice: And then you mentioned app store earlier. That is the direction still to be able to plug anybody's software application into. The API
[00:51:10] Troy Harvey: that's right. So, we on the app store with this quantum API. We have a third-party partners that are going to have their applications. So people are, you know, start up in the space.
We need a certain service that we needed enable, but it's been a real impediment, huge cost hurdles to get into the building. This makes it not only a free market that you can get to a wider audience, but something that you can take away that integration. And then for corporate customers, we have a ability to do corporate applications that only appear to you and your team that allow you to do your corporate integrations to weather analytics solutions, or, you know, corporate management, costing costings, or whatever do you want to hook in?
[00:51:53] James Dice: And then I think this is my last question. The, you mentioned HVAC, Harry and boiler Bob, but you also mentioned the quantum describing literally everything about the building. Do you foresee like expanding beyond HVAC into other siloed systems, traditionally siloed systems in the building? And other, other sort of use cases and different personas besides those two.
[00:52:14] Troy Harvey: Totally. So we that's, that's some of the more exciting partnerships that we've recently signed or in these domains that haven't been the, at BMS in the past. And in some of them they're on a cost sensitive domain in some areas it's industrial domains. Just DMS doesn't add value historically. So let's say your problems.
That is more of an it networking infrastructure problems that, you know, with some IOT, like knowledge around that, or, you know, the logistics or the industrial or the manufacturing, the process control. This is, this is where I see a lot of exciting new emerging opportunities because. It's not just the shell of the building.
It's not just the systems that make the building go. Of course, buildings are full of things and those are interacting with the buildings and the systems. And so it's all one thing and more and more, what we see is people wanting to get that one total view of their building and their businesses. Totally.
[00:53:15] James Dice: All right. I think we should given the depth that we went into. Maybe leave this episode here for now. I know we have other topics that we want to explore in the future, but maybe we can revisit this again and another episode in a couple of months, but thanks for, thanks for coming on the show again.
It's been been fun. I always learn a lot every time I talk
[00:53:36] Troy Harvey: well. I, I appreciate having me on and I look forward to talking further.
[00:53:44] James Dice: All right friends, thanks for listening to this episode of the Nexus Podcast. For more episodes like this and to get the weekly Nexus Newsletter, which by the way, readers have said is the best way to stay up to date on the future of the smart building industry, please subscribe at nexuslabs.online. You can find the show notes for this conversation there as well. Have a great day.