From Smart to Agentic: The Next Evolution of IoT

Jake Gilbert (00:00.891)
Good, cool. So welcome back to the IoT podcast. I’m your host, Jake Gilbert, head of 5V Tech. And today we’ve got an exciting episode, which I’ve been looking forward to recording for a while actually, since we’ve been speaking about it, which is exploring the intersection between AI and the physical world, but specifically the potential of agentic AI in IoT and physical systems. Now there’s been a meteoric rise as everyone can see in AI and digital environments over the last

12 plus months, but a significant shift has been toward systems that don’t just generate content and insights, but are also making decisions and taking action in the real world. So to dive into this topic with me today, I’m excited to be joined by Dan Levine, VP of AI, Data and IoT Platforms at Signify, and Hans Nicol, Head of IoT, also at Signify. So, Dan, Hans, welcome. It’s great to have you here. And just as a bit of a introduction, I’m sure,

listeners are aware of Signify, but if not, Signify operate at significant scale in smart home and commercial lighting and building smart systems with, I think when we discussed over 150 million or possibly more connected devices across consumer, commercial, public environments from Phillips Hue in the home, all the way through to buildings and even wider scale and cities. So interesting thing about Signify is you’re operating across the full stack from device to cloud. So

Hopefully we have some good insight on where that’s affecting with the Gentic AI today. So before we kind of break things down and go into it, it’d be great to start with a quick introduction from both Dan Hands and maybe just some initial thoughts on priming up the conversation we’re going to have today.

Dan Levine (01:45.658)
When it enough then. Well, thank you, Jake, for having us. My name is Dan Levine. I lead, as mentioned, AI data and IoT platforms here at Signify. I’ve been in the IoT space for over two decades now. And most recently, especially with the advent of AI and real large scale data infrastructure, taking all of that really interesting data that these sensors and

connected devices I’ve been producing and delivering real outcomes and value has been my top focus. So very excited to be here and looking forward to see where this conversation goes. Great. So Nicole. I’ve been with the company for 30 years. So I outnumber you by one decade. And so today we are Signify. We used to be Phillips Lighting, one of the large three lighting companies.

Now you would think why is lighting an interesting vertical in IoT? Well, I would say it’s three things. It’s plenty. So I think by any counts, you have probably 50, 60 billion light points in this world and counting. So, you know, if you only connected a fraction of that, it would be a very substantial IoT vertical. The other one is the position. IoT is obviously all about sensing and later on AI has to make use of the sensing data to

provide information. And lighting is usually uniquely positioned outdoor, indoor, you to capture that. And the last one is power, right? So obviously you want to continuously feed data, sensing data, and for that needs to be powered and it is lighting typically all the time. And yeah, so I think everybody, know, most listeners I would assume know Philips Hue, which is our product and has quite some widespread consumer application. And so here we are.

Jake Gilbert (03:43.868)
Fantastic. Well, it’s great to have you both. And I thought maybe to start with, it would be good to prime things by getting a few definitions out there. So edge AI, physical AI has become obviously a huge, huge buzzword. guess the technology is in its simple form being around for quite some time with TinyML and even before that kind of really basic single model inference on devices. I guess to start with, when we look at systems like the kinds that Signify is working at, what

What really does agentic AI mean in that context and in the context of signifying Philips Hue?

Dan Levine (04:23.876)
Ben, give it a first try. So from my point of view, I think there’s a few different dimensions that it covers. We have our physical products in the various domains, very impactful in the home, being that kind of the dream has been for these systems that are ultra personalized and respond to your needs.

without having to have so much overhead and configuration and tweaking with AI and being able to do that on the edge where it’s private and secure really enables those types of experiences that companies like ours have been trying to deliver and delivering on for a long time now.

In the commercial side of things, there’s huge opportunities. The biggest issue in commercial is the variations and permutations of embedded systems that are running. is really, every building is a custom configuration. And being able to use AI to understand and contextualize and add semantics to the data that’s coming out.

and then reason on top of that data is also a big aspiration for anybody who’s been in the IoT space. that’s two. And then the third is really around our internal operations, how we serve customers. So making it easier to work with us and find the right products and get the support that our customers need to help them.

enjoy and experience the capabilities that we are delivering.

Dan Levine (06:18.554)
And maybe let me add to that because you mentioned Edge AI versus, let’s say, AI in the cloud. I think if you look at this massive scale of deployed devices, right, and even if a fraction of them is IoT, you can imagine if you have constant sensor feeds, let’s say you have an outdoor lighting situation like we have a parking lot or a street. And even if the sensor is fairly simple, it creates such an abundance of data. And if you think, if only a

of that needs to be transported to a network, it’s a total overload very quickly. So having AI in the edge to compress and just send the necessary useful information. It’s not just even compression, right, but already making sense at the edge. Removing the noise. Removing the noise, right, which probably most cases, you you have 90 % noise, if not more.

and just send the relevant ping basically. I think that’s a huge gain and I think will help to dramatically advance the number of use cases that can be addressed moving forward.

Jake Gilbert (07:22.747)
Well, it’s interesting. I think Dan, as you mentioned, when we talk about what’s been a kind of aspiration of IoT systems now becoming possible, what do you both think is probably the main material change that’s happened in the last 12 months technologically that’s actually making agentic AI in these systems a reality?

Dan Levine (07:46.96)
Maybe if I may, just quickly, for me really it’s, you know, if you had this first…

how to say, flash moment of chat GBT, all of us remember back in 23. But that was nice to have, and all of a sudden you had this brain. For me, agentic AI is really another one, probably even more impactful. Because now, before it’s like, OK, you could argue it’s a bit of another search engine, right, and gives you the tailored answer. But now you have agents that can actually do something, right? They can perform actions in a logical sequence.

And of course that is already used today in a lot of enterprise applications. But if you think this forward, that each of these edge nodes eventually can transport that intelligence that can do something, act on it. I think the potential is just enormous. And for me that’s really also eventually the productivity gain that can be accomplished only with agents.

I agree. think the reasoning ability, the fact that you can kind of bake intelligence into these agents, and if you create, say, microagents, as you will, that handle specific outcomes that they’re looking to achieve, like keep this room comfortable. That type of request involves a few different things. One, what kind of room is it?

is anybody in the room and what does really comfortable mean? the fact that it used to, prior to really agentic and kind of these language models.

Dan Levine (09:38.352)
To write that type of control code would take a very extensive amount of time. And then the tweaking and the adjusting, and once you have it locked in, that’s the reality. It’s locked in. There’s no changing. So enabling a of a non-deterministic ability to make a room comfortable versus a step-by-step logic sequence, it’s a whole new paradigm.

Jake Gilbert (10:09.435)
And when it comes to, like you said, the difference now between having to kind of program things in closed logic step by step to now it being a bit more of a new kind of decision making, I guess also brings into focus what’s where there may have been a bit of a ceiling hit in the traditional kind of dashboard analytics and insights. So today, when we think more about how it works, what are the kinds of things that

agents are going to be able to enable building management or building automation systems to do that wasn’t possible in the kind of approach we had up until recently.

Dan Levine (10:48.004)
You triggered a good word, I may, dashboard, right? So if you look at building management, you know, the standard practice, I would say, for the last five to 10 years was, you you observe, you know, functions in the building and it displayed on the dashboard. Then you have an operator sitting in the building or remote, right? And it has, I don’t know how many screens, you know, observing all of these dashboards. People hate dashboards. I think at least that’s our finding, right? And, you know, first of all, the dashboard just displays information.

Jake Gilbert (11:11.931)
Yeah.

Dan Levine (11:17.938)
and it’s up upon the operator to act on the information displayed on the dashboard. So it also comes with a lot of information trading. And now if you have an agent, first of all, you take away the dashboard and you go immediately into action. There’s no reason why an agent shouldn’t perform a simple action that before was displayed on the dashboard, and now you can immediately act upon

And many, think, you mentioned, let’s say, well-being in a building, that’s a summary of many, you know, it’s temperature, it’s humidity, it’s noise, it’s lighting, obviously. Many of these actions, if you compartmentalize them, are fairly simple actions you can tune to make it comfortable. All of them can be done by agents. And I would say from…

the cognitive load that a building manager, building operator, there’s these buildings, they’re in many cases are aging and they have these systems that were built that have issues. And diagnosing those issues can be a real challenge. having agents that are augmenting and looking into problems and coming up with potential fixes and even as Hans mentioned, taking action, even

calling in the right people to come repair the issue versus a disaster happening, I think is going to have a really strong place. Security is another area. For security operators, it’s really hard to observe and keep track of everything going on. having these agents that are able to look for anomalous behavior or something that is not

Jake Gilbert (12:45.755)
Hmm.

Dan Levine (13:05.744)
correct and ultimately, historically, coding that in logic, in line coding, saying, I’m looking for this was really not effective. And you just never got it right. having an agent that you could say, look at this environment. Make sure that the temperature is always correct, that the lights are on when people are in the room, and that there’s no

nothing that is out of ordinary from a security standpoint. You can give it that kind of instruction and it’ll, as Hans mentioned, it’ll ping you when if it finds an issue and other than that it’s quiet. So I think that’s really an area that’s going to become more more prevalent.

Jake Gilbert (13:56.172)
And I guess like you mentioned Hans, the sustainability aspect comes out in that from energy consumption, where rather than having to constantly feed back a continuous stream for a decision to be made more centrally, it can actually, like you said, transmit what’s actually needed. And aside from that, you maybe don’t hear as much. And that maybe comes back to an interesting point around the hierarchy in this. So where is it that agents actually sit in the system?

Is it that you are effectively swapping out the overall building operator with an agent, or is it actually that they’re on maybe the device level or the zone level within a building or a whole building or even actually being coordinated on a wider level? Do you think at each of those layers, there’s going to be an equal opportunity for kind of agents to play a role or do think it might be higher up in the chain?

Dan Levine (14:52.88)
say at least in the beginning, I think they will take away all of the tedious decisions that I would say an operator just hates to work around, which probably today still takes 90 % of his time or her time.

which are simple things like error messages that happen and they often are wrong, then just need to reset or it’s a real thing that could have an easy fix and could be fixed by an agent. So many of these noisy things, if that already was taken care of, I think would be a huge efficiency gain and would eventually focus the operator on things that really matter, more fundamental decisions where you still need a human brain to comply with.

So I would say at least for the next few years I would see the agents working their way up in the end. Of course you never know where it ends, but for now I think that’s such a vast load of tasks that need to be accomplished. So I think the agents have a while to work around that.

Jake Gilbert (15:59.312)
Yeah, I think I was speaking to someone the other day about that, which is, you know, it’s not just a binary, is there or isn’t there going to be someone sat, you know, in a room looking at stuff, but someone described to me a bit more, you might have a situation like Jarvis from Iron Man, right? So you still have someone that ultimately that’s taken a high level decision, but you’ve got someone that’s taking all the, taking all the background, ultimately taking commands, feeding back information and you know,

more of an interactive kind of co-ownership system, which is interesting and I guess comes down to that question around autonomy and are we talking about something that could be completely fully automated at a building level or something that’s assistive and maybe what the implications are when we talk about it being a black box or who takes the blame if something goes wrong in that decision making.

Dan Levine (16:58.554)
Can anyone take that? So eventually, most of these systems will be highly autonomous. However, it’s going to take time to get to that point. a human in the loop is going to be critical for years to come in many domains. The key.

is really building trust and having that auditability and traceability of all reasoning and decisions that were made, not just to go back and look at why a decision or why an autonomous system made that decision, but to use it to improve its decision making for the future. That feedback loop and that circle around

you know, keeping track of everything that’s being done and then adjusting the way that it’s operating is absolutely critical to get to highly autonomous systems. Maybe let me add to that, I think a very important, you know, boundary condition is security. So if you talk about a building, you know, and it’s increasingly autonomously managed, that also means, you know, it’s, you know, vulnerable.

And to shield that, today, at least in building management, you have already the network, example, LoRaWAN or so, that is completely isolated from any other building function. So at least you have a detached between other networks and the building management network. BACnet, right? But now if you add on top of that intelligence and the autonomy of the decisions that are made on the information that’s being fed out of these networks, you have to have similar.

Jake Gilbert (18:38.779)
Mm.

Dan Levine (18:49.004)
know, precautions, because if you have hacking events or whatever, you know, and you can shut down a building or you can shut down a city, right, or the traffic lights or the lights, right, of course, that has huge implications.

Jake Gilbert (19:03.791)
Yeah, well, this is, you know, really those kind of conversations where it starts to get into what we probably would have called sci-fi, you know, in movies not so long ago and now actually being, you know, reality. Those kind of horror situations where you could be stuck, trapped by a rogue AI or some kind of, you know, presence that’s actually got physical control of every part of a building. But it might be the air, it might be airflow, air control, elevators.

Dan Levine (19:15.728)
Yeah.

Jake Gilbert (19:32.324)
Not to turn people off to the prospects of building management systems, always an interesting thing to think about. I think, you know, one of the final things I was going to bring back on when it comes to us talking under the hoods, kind of how these things are working is, and I forgot to mention it earlier, but when you said, damn, moving from the need to have to program step by step, kind of control logic where you might actually be relying on learned behavior over time and more of a kind of evolving system.

Dan Levine (19:35.824)
You

Jake Gilbert (20:02.363)
What impact do you think that’s going to have on the people or the whole section of engineers at the moment who really specialize in being able to build those systems on the assumption that we can’t just have a system learning?

Dan Levine (20:19.652)
Then I think that’s a great one for you because you have a large engineering team. Yeah. It’s, it’s, it’s a really interesting question because at this point for the foreseeable future until, cause really we’re at the beginning of a lot of this technology, even though for those that are in it, it feels like it’s everywhere. The reality is that the actual deployment footprint is quite low and to, to get to a higher deployment.

it’s going to require the engineers, the people with the actual domain expertise to say, train and, and iterate on, on these, these tailored solutions. And I think, I honestly think that those that embrace it really, become force multipliers, right? It’s really, it multiplies their abilities. It does create a lot more work for them too though.

you know, have engineers that are now using pure agentic development and it, it, creates 10 times the amount of code and it’s 10 times the amount of work. and, and it’s the barrier to delivering outcomes is, is shorter now. I mean, it’s, it’s less, there’s less of a barrier, but those that are trained and that understand these, these domains, whichever domain of engineering that is, there’s a really

a high value for them to leverage these technologies and really make a bigger impact on the work they’re doing.

Jake Gilbert (21:55.035)
Sorry Hans, I didn’t know.

Dan Levine (21:57.2)
Maybe we add what I’m excited about is also if you look on the application side, you know, take in Europe, for example, this EPBD, that’s a mandate from the European Union, you know, to have energy saving in also small and medium sized buildings, you know, it’s rolling out in steps now 26 and 2030. So ever smaller buildings should participate. If you look and we are looking in that space, right, this the variety across the European Union, you know, each country has different flavors of that.

Jake Gilbert (22:07.291)
Yes.

Dan Levine (22:26.704)
law. Then you have an endless variety of these buildings. So if you imagine eventually you want to tailor all of these agent processes to all of these variations, I think you can only do it there. so I think, that means a lot of programming to tailor these agents to these various variations. So I think we are not going to run out of work for that anytime soon.

Jake Gilbert (22:53.307)
Yeah, no, it’s, think, like you said, the rate of these actually being deployed on a commercial scale is still quite far behind what people see, you know, in generative tools, right? And things will evolve. I think, I don’t know quite where the statistic comes from. And, you know, everyone says how much you can’t always trust it, but I think it was something like 80 % of the job titles that we have these days didn’t exist in the 1980s or the 1970s. So.

People, I don’t know if, don’t quote me on that, but it’s something you can kind of see where I’m going with that. And I think you will see the same thing, right? People think, how are jobs going to be impacted today by AI? Well, it’s also going to create a lot more trust-based roles, I think, where we have to have a human in the loop doing a slightly different job than the hands-on work. Great. Well.

Dan Levine (23:46.724)
It’s George Jetson where you’re just basically human just pressing a single button.

Jake Gilbert (23:50.94)
Yeah, yeah, yeah, that’ll take a… Well, then you got the whole disutility of work issue, And then do people even need to? Well, that brings us nicely on to impacts and I guess some use cases as well. So we spoke a lot about this when we were leading up to this episode. out of all the potential kind of use cases that this technology has, particularly in the context of whether it’s Philips Hue and in…

Dan Levine (23:58.138)
You

Jake Gilbert (24:19.131)
consumer homes or whether it’s on larger scale, what do you think are the ones that have the biggest potential, the biggest impact?

Dan Levine (24:31.6)
Well, maybe let’s start with Philips Hue for a moment. anybody who thinks, Jake, I know you have a Philips Hue set at home, I believe. So it starts with you acquire, let’s say, something like Philips Hue, you want to install it. It’s quite tedious still. I think Hue did a fantastic job to make it as simple as possible. It still is. It’s just not like that.

Jake Gilbert (24:55.799)
it is quite a process.

Dan Levine (24:58.864)
And I think Dan can talk about it, because how many devices do have in your Over 200. Over 200. But Dan doesn’t only have Hue, he has many other IT devices. So I think you had a first attempt, say, with Alexa, where you had a simplified interface. say, Alexa, turn on the light. But that was 2015 or so, believe. And it kind of got stuck there. So it’s still tedious.

Jake Gilbert (25:04.699)
Yeah.

Dan Levine (25:27.3)
What I find really exciting is finally to make it much simpler to first of set up the lights and then control the lights in conjunction with other IT devices. And all of this will be underneath done by agents. And I think we were really woken up. Who started just using, in this case it was Claude Code, to basically, with vibe coding, so to speak, code a simple sequence to commission uLights. So that’s possible today, right? And it’s just the very beginning.

I think it covers the whole spectrum of Signify. In Hue, there’s a ton of impacts. Hans was mentioning, I’m a big home automation, home assistant user. if you look at what you can do by tying into Agenda.ai to both create new interfaces, create automations, support an issue when there’s a problem.

very, very impactful for the consumer, especially those that are not geeks like us. the commercial side, there’s, as we were discussing, for buildings and the occupants experience within them. But another big area for Signify is really focusing on customers and how do we make it easier to work with Signify.

Jake Gilbert (26:34.491)
Okay.

Dan Levine (26:57.422)
You know, we’re a company in many cases by acquisition. We’ve acquired a lot of brands and sometimes it’s really hard to find a product from us. And we want to make that easier. And when customers need support, being able to find it right away and get that support. And more importantly, providing the best products for the right experience and outcome that our customers are driving for.

So making that match, being able to do site audits with AI and commission systems with AI, all to deliver the experience that our engineers and product managers spend years of their life perfecting, being able to replicate that in our customers’ homes and businesses and work life.

Jake Gilbert (27:54.22)
And with, I guess across that full spectrum, right, where we’ve got home, kind of commercial spaces, even wider in cities. You we talked before this around, you know, occupancy analysis and all these, you know, quite surprising things that could be done with a couple of lights and sensors, you know, maybe from each of you, what’s maybe been one of the most surprising or unexpected use cases that we might be able to unlock with

kind of agents operating in the system that may not have been possible before.

Dan Levine (28:28.314)
Well, one, maybe the consumer space is what we launched at last. was Motion Aware by Hume.

So basically, just screw in your light bulbs or mount your luminaires. And you don’t need a dedicated sensor anymore because the light or the network itself becomes a sensor. That means you mount them, let’s say, throughout your home. And this is an interesting use because for security, for example, you might have a security camera at your entry door. But the rest of the home, let’s say, the rooms in the back where you have maybe windows, could also be security issues. They are uncovered. Now if your whole home is basically one network and is one

security mesh if you want. You can then detect events, even the camera is only one participant in this security scene, and to classify events, to say, okay, there is maybe an incident, you absolutely need AI, because otherwise you are drowned in probably clutter of wrong, false positives. And that’s where I really see a major step forward, for example, in this security

vertical.

I building on that, you know, obviously anonymous occupancy detection is and has very big potential implications to ensure that an environment is comfortable for an occupant, but also that energy is not just being wasted and considering the changes over the past five to 10 years, especially with people working remotely.

Dan Levine (30:07.662)
with the amount of power requirements for data centers, energy in the grid is really being taxed right now and it will continue to. So being able to fuse these types of problem statements and solve these problems by using agents to do so, I think is going to really have a big play in the near future.

Jake Gilbert (30:35.395)
And I think as was mentioned in the kind of introduction as well, I guess one of the reasons that this is so, has such good resonance with actually existing smart lighting or just lighting infrastructure in general is like you mentioned that they’re pretty much everywhere already. They have a constant supply of power and also you already have a lot of data that’s coming in through all those different sources. And Hans, you mentioned when we had a previous conversation that, you know,

That data is great, unless it’s properly, unless the quality is there, unless it’s labeled properly, classified properly, you still have those issues. So what I wanted to ask from that perspective is what do you still think are the constraints and things that we need to get over before we can fully realize the potential of this?

Dan Levine (31:19.894)
You mentioned, for example, the labeling, right? You think it’s trivial, right? But it’s far from, right? So we are in a building right now. How many stores do we have here? Four? Four stores?

So if you would, you know, and because you have to see those four stories have been equipped maybe in the very beginning, let’s say with one lighting layout and maybe one layout for HVAC and so on. But over time, you know, there’s renovation going on. So you have a multitude of lighting systems and HVAC and other verticals. So over time, you know, let’s say this building, I don’t know, is maybe 30 years old. You know, it’s vast variety. Nobody keeps track of that. Nobody. Right.

Jake Gilbert (31:57.019)
Yeah.

Dan Levine (31:58.016)
Then you come in, let alone is there a floor plan, right? Maybe the floor plan is 20 years old. In the meantime, it has changed. I don’t know how many times. So all of this mandates that there’s ideally accurate data, right? And that is today one of a really huge threshold. I just come actually from a conference in New York City where there were a lot of architects and lighting designers. And that is one of their major hurdle to implement, to first have reliable data on the ground.

Jake Gilbert (32:26.363)
interesting.

Dan Levine (32:27.578)
to make the right decision and it boils back to what you mentioned, Dan, to eventually decide, okay, I want to renovate this room, I would like to have these type of luminaires, but first I need to have reliable information on the floor plan, on the existing lights, what I want to replace. And again, you would think it’s too far from.

Jake Gilbert (32:49.871)
Yeah. And would you say then that kind of bad data or limited data is probably the biggest bottleneck in this context?

Dan Levine (32:58.722)
say it could be limited or missing data but even

Let’s say the tagging, example, take a very simple example, right? So you’re down here and let’s say you have a coffee area and the lights over the coffee area have been named, you know, coffee area. you know, and then you go in the next floor and the coffee area is espresso area, let’s say, you know, because whoever designed that and labeled that, right, gave it some sort of random name. There is no convention, nothing.

So that’s actually great about an agent who eventually can make an autonomous decision that espresso and coffee means the same. It’s a terrible maybe not, but you know what I mean, right? It’s a terrible example. And data is really the name of the game for any sort of AI scaled solution. It’s the input and the output of AI. And if the core input isn’t of

good quality, doesn’t have context, doesn’t have any semantics around it, the output is going to be very poor and these agents just won’t function. So, it’s, you can’t really have AI without data.

Jake Gilbert (34:15.845)
Yeah, all the right data as well. know, well, that’s interesting. And so I guess in that context, when we’re talking about the availability of data and the crucial nature of that in making any of this work, even though it might seem, you know, not trivial, but ultimately one of the kind of first initial building blocks before you have all this fancy stuff on top of it, what kind of learnings do you think or that you guys see that you take from how

Dan Levine (34:18.832)
Yeah.

Jake Gilbert (34:43.237)
how much of a deployment you have on the consumer side, then looking across to kind of buildings or public infrastructure.

Dan Levine (34:54.842)
Well, would say obviously there is a huge difference. So we have, of course, a massive deployment of consumer devices, right? And those are in the hands of the consumer. And the functionality around these devices is fairly limited. Whereas on the professional side, the deployments are, know, I should say more focused, but the functionality is a lot more critical, as you said, you know, a light in a professional environment, a building, indoor, outdoor parking lot, right? The city infrastructure must function.

in any circumstance and a lot more guardrails that like to function properly. So I think that’s a huge difference in…

requirements, let’s say, in a professional setting versus a consumer setting. And hence, you know, the accuracy and the security of the data surrounding that to make the right decision at a given time, also in case of emergencies, for example, is, you know, a lot more critical in a professional space than a consumer space.

Jake Gilbert (35:55.356)
That makes sense. guess, like you said, from a security point of view, you have probably those two different ownership challenges there, right? Of someone maybe choosing to implement a system and ultimately put those features in and live under it versus someone walking into a building and have that operating without necessarily someone like yourself down having put in 200 devices in place and knowing what they’re all gonna do at the same time. interesting.

With that in mind, we talk about data already and the accuracy, the quality of it needed. What else aside from that, because we know that now, that’s quite a clear thing. What else technically or operationally needs to improve or get up to scratch in order for this to be more widely realized?

Dan Levine (36:47.408)
I think the reality is a lot of these systems are, if they’re connected, they’re connected in a way that is to specific clouds and they have their own data models and they’re not necessarily interoperable with each other. So I think the key enabler is going to be the abstraction of data across these systems.

Jake Gilbert (37:05.241)
Mm.

Dan Levine (37:14.506)
and then having agents leverage that abstracted data to deliver outcomes that are reusable and scalable because many of the systems in both buildings and homes are siloed, as you will, to how they’re operating. And I think that that’s going to be the key enabler for this.

If I may add one more dimension to that is for me it’s also the energy efficiency surrounding it. Just to give you an idea, know, mean lighting is now 130, 140 years old. The first light bulb had an efficiency of less than 1%. 99 % of heat, right?

Today we have 70, 75 % efficiency. However, now we add a network to it, we add a sensor to it, et cetera. And now we add AI agents to it. so now we basically all the energy we save, we waste again in data. Not waste, but we burn it in data centers, you could argue, right? So Alex, you learned from you and your team, I thought it was fascinating. How can you become very frugal and super efficient in programming an AI agent? So I think today we’re in the very beginning, right?

Jake Gilbert (38:12.259)
Yeah. Yeah.

Dan Levine (38:27.13)
we basically burn tokens by the second, so to speak. But in the end, eventually, the efficiency of performing an agentic action, I think, will become super critical.

And I don’t know, Dan, but I think you study a lot on this, right? But I think we are also at the very beginning of it, right? We are. But a lot of the methods that have been used in cloud and consumption-based technology really apply. And the FinOps practice is really, really important when it comes to building scalable AI solutions. And as these models are expanding and growing bigger, doesn’t always mean better, especially for cost.

You don’t need intense reasoning if you want to keep a room bright enough and cool enough for somebody to be in. having those levers to be able to adjust along the actual pipelines and workflows is monumental.

Jake Gilbert (39:35.379)
And to maybe put that back to a bit of wishful thinking then. So if you could wave a magic wand maybe more in signifiers context and get over a hurdle that there’s in the way at the moment for that, what would it be?

Dan Levine (39:58.82)
Well, think at least my spontaneous reaction would be, I think we’re not yet in a position to answer the question because we haven’t hit it. Because, you know, yesterday we had a review on this topic. It really just at the very beginning. So I wish we were already in the position to answer your question, but I think we aren’t yet.

Jake Gilbert (40:05.307)
Good answer.

Jake Gilbert (40:20.623)
That’s probably a more revealing answer than maybe thought interesting. And I know there’s an element to this which we certainly wouldn’t want you guys to start releasing the secret sauce recipe live over a podcast. But what are the kinds of things maybe that we should be looking out for in this kind of topic when it comes to signifying the products there, the engineering?

Dan Levine (40:52.026)
Well, think, think Signify as it’s evolving is really starting to work together as a, as one. And historically there’s been so many incredible inventions that have come out of the businesses within Signify. And now with, with our new leadership and a focused effort on

how do we work together and how do we leverage technology and especially data across our ecosystem to delight customers? And I think that you’re going to see a lot of focus around customers and how do we make their lives better as the core focus for us. And that’s where we’re going to be applying AI. data is absolutely critical for that.

Let me add that one last comment there. What signifies the only truly global player?

Jake Gilbert (41:49.237)
and so on.

Dan Levine (41:56.608)
So if you talk about the widespread application of AI in all the different geographies, and of course there’s a lot of, let’s say, debate around this today, so we have to tie into all of these debates as a global player, which gives, I think, meaning on all of these regional implementations, let alone on a global scale, but we have to play on all of those.

Jake Gilbert (42:21.719)
And taking that maybe a bit into the future then. So we are where we are at the moment, as you said, still maybe finding feet in terms of what that’s gonna manifest. But I will ask maybe a bit of a crystal ball question around two, three years ahead, not necessarily from a signified perspective, but in terms of where this is going, what are the kinds of states that you think we’ll see with these systems?

maybe not so time specific and I won’t ask you to put money on it, but what do you think of the kind of aspirations for this?

Dan Levine (42:57.998)
I think AI is here to stay, and I think it’s going to change the world, and we’re going to get to that Iron Man Jarvis state. But I also, being in technology my whole life, I’ve seen how these transformational technologies come, and everyone gets very, very excited about it, and a lot of money is put into it. And the reality is it takes time.

for it to develop. IoT for 20 years plus has been top of people’s minds, but it didn’t really become as strong and prolific until the pandemic. That’s when it really, really started to ramp up. Look at the internet in the late 90s. Everybody was…

Jake Gilbert (43:41.808)
Yeah.

Dan Levine (43:52.346)
coming out with these products that were going to change the world. But the reality is that did happen, but it took years behind the actual hype. So I would say that I would be surprised in the next few years, some of the companies, like the biggest companies now, like the Open AIs, may not be as important as they are today. So.

The real question is, who is going to survive and really be the ones that change the world?

Jake Gilbert (44:28.123)
Well, gets on to when we look at the kind of economics of that and certainly in the US and you look at the top 25, top 50 businesses by market capitalization in the US at the moment, the picture changes so much of a relatively short time period. You’ve got companies like Nvidia, OpenAI, others at that stage who just didn’t even…

exists or certainly weren’t anywhere near up there not that long ago. And the pace of that kind of replacement of new businesses, think the US is a bit of a different context because, you know, in the UK, top, the Fertsy 100 wasn’t looking too dissimilar not so long ago in terms of companies, but the US you see that pace of change. And I think you’re right in, it’s really all to play for, certainly in the next two, three years. And the use cases that will come out in the wash, as they say, will be ones that we

couldn’t have even sat here thinking about.

Dan Levine (45:24.854)
I would make a of an alert there because you’ve always, you know, magnificent seven, basically, you know, the large seven in the US, the players, today it’s a bit like driven by, you know, the bigger better is the better, right? So if you take OpenAI, you know how many trillion of parameters can you crunch?

Jake Gilbert (45:32.207)
Yeah.

Dan Levine (45:44.313)
If you look at China, obviously also limited by access, let’s say, to Nvidia infrastructure, they took a different approach and they basically say, let’s see what we can do with what we have.

And I have to say, you know, we are a close observer of that. It’s phenomenal what you can do with even one generation back, so to speak. So I would not be surprised, you know, if eventually over time, you know, with all the tools we have available already today, there will be new players emerging that know how to most economically apply those for certain use cases. And as I we’ve seen some of these first examples, DeepSeek obviously, and China is a good example, right? But there will be more.

Jake Gilbert (46:28.091)
So I think that’s quite a nice note on which maybe to bring it from that future point back to today then. So I did ask this question before, so hopefully I haven’t come out of the blue with it. But for both of you maybe, what do you think of this conversation, this topic at the moment, signify, context or otherwise, is the main thing that think people should take away from this conversation? Maybe just to stew on and have a bit of a think about.

Dan Levine (46:55.92)
Want to go first? So for me, I thought about this obviously when we prepped it up.

Let me put a bit in a historic context. As I said, I just come from a conference in New York and that’s also where Thomas Alva Edison in 1878 switched on the first light. It was magic, right? It’s never happened before. So he built a little power plant right in the middle of New York City to power the first street lights. And it was magic to people then. And then, know, in 1891, Phillips started in Holland doing exactly the, know, in a way, exactly the same thing. And so the last hundred or so years, you know, light

light involved into what we just described in the past half hour or so, into this extremely complex infrastructure. My dream would be that with the help of AI and AI agents, we actually go back to where we started. That lighting is as simple as switching on one switch. And making that magic happen again, without taking away all the complexity underneath.

Of needs to happen today in all these buildings and so on. But the simplicity of, you know, OK, I turn on the light and it’s simply light out there. And that’s what people like and experience. I hope that we can eventually, after this journey of 150 years, come back to a bit of this magic again.

Jake Gilbert (48:16.653)
one switch to power them all.

Dan Levine (48:18.832)
In a way, but you know what I mean with the equivalent of that, right? But to experience that simplicity, go away from all this complexity and let the agent in the background do the work.

Jake Gilbert (48:23.899)
Yes.

Jake Gilbert (48:30.779)
do everything else. Yeah, I like that. Dan, am I putting you on the spot by asking the same thing?

Dan Levine (48:37.776)
So I think the unique position that Signify has being probably the largest sensor network in commercial real estate and the largest IoT connected deployments in the consumer space. I think that there’s a real opportunity to connect

various systems and capabilities together and us help enabling and being that bridge. And I think AI is really helping accelerate that for us being that we’re a company that was founded in innovation and we’re going to leverage these innovation technologies to transform.

the experiences we deliver to customers and how we operate as a company. So I’m very encouraged in being a lifelong IoT person and it’s a great privilege to work for Signify and be able to have this type of impact.

Jake Gilbert (49:49.275)
I think that’s a great final point guys. It’s been a great conversation. Really appreciated having both your time here to talk about this. There’s a lot in it. I’m sure we could probably go on for another hour or two in a different life. I think we’ve covered a solid cross-section of what we’re talking about with the Gentic AI, what it means in practice, how it can start to change those systems, and hopefully what we’re looking to see out of it in the future. I think one of the other interesting things particularly is that this isn’t happening.

in isolation, right? Having to be layered onto existing infrastructure for better or worse, you know, old buildings, new technology, and there’s a lot of work to do here. So great conversation, and it’ll be interesting to see how it evolves for sure. So for both of you, thanks very much again, coming on, sharing your perspectives. And for everyone listening, as always, if you found this useful, please feel free to share it with someone who might be working in this space or interested, and we’ll see you on the next one.

Dan Levine (50:48.538)
Thank you. Take care.

Jake Gilbert (50:50.502)
Thanks.