Next, we're joined by Insight Partners Managing Director Devin Perk to discuss where AI goes next, from software to hardware. He is joined by tech pioneer and investor Tony Fadell. Tony has helped create some of the most recognizable devices in the world, including the iPod, iPhone, and Nest.
So, Tony, look, you've helped change how we listen to music and how we change our temperature, and now you're investing in deep tech. You're writing books. Just tell us what you're doing today.
What am I doing today? These days, right now, I'm still investing and taking care of the portfolio at Build Collective. So, we're a deep tech investment firm, and we're in around 150 companies, something like that. We've been investing for about 15 years. Author, right? So, Build, now in 31 languages around...
You're my agent. And then I just last week got appointed to be a designer in residence at MIT in the Media Lab and Architecture and Engineering. So I'm going to go up there, and they just said, 'Make trouble.' So I said, 'Okay, I'm going to come and make trouble and figure out what's going on.' So that's what I'm mostly doing right now. And I'm a shadow CEO at four companies.
That means companies are not doing that well.
I don't want to be doing this. Let's put it that way.
All right. So, let's talk about the topic of the day, AI. One of the things you were quoted as saying is that AI are like know-it-alls. And you know, I don't necessarily think of know-it-alls as a positive statement. So, as you think about why do you think that, and then how has that influenced what you're investing in?
Well, so you know, I think we all grew up with someone who was a know-it-all, right? You know, and we were always like, 'Okay, here we...' or you may, you would never hire one who was a know-it-all because it was like you could talk to them and you're like, they say everything the way it's supposed to be said, but then when you actually go behind the scenes, you go, 'Wait a second, that doesn't make sense, that's not true.' And so we have a lot of great examples of AI where it works and it works brilliantly, but we don't, we don't have all the stats of where it doesn't work. And I've been asking the AI community multiple times, have we done a bunch of error, like do we have error tracking on all this? We just show the percentage of success, we show the failure, but how failed are they and all the other failures they have. And so that's okay for consumer, lots of consumer apps, it's okay if you don't get a picture that looks the way you want it, whatever, if you're on the creative side. But when you're dealing with, you know, a financial, you know, a financial bot, or you're dealing with an AI medical scribe, or you're, you know, you can, a legal, you know, a legal assistant. When you're going to put your brand and your company on the line, you need to know how these things work, right? You need to know you can trust it, or you know that the person that's using it has tools to be able to evaluate the output, right? Um, you know, luckily with AI coders, you can see if the code's working or not, and then you can use more tools and everything else to debug that, but we don't have those things for all these other industries. So, when you're going to put your brand, when you're going to put your, your, taking a person and either augmenting or replacing them, you better know what's going to work. And we have a lot of, we have a long way to go for that. And so I invest in companies that really understand the data and their training models very, very well and have all kinds of tools to then assess the accuracy of those things. We have MRI, or excuse me, medical imaging ones that are FDA approved. We have medical scribes, these kinds of things, but they're trained very, very differently than just using general purpose LLM.
But you are still building them on existing models. Are you...
No, we're building, we're building them on, you know, open-source ones with different training data.
They're specific to that.
Yeah, they're very specific. Yeah.
So, look, you said you talked about having a big portfolio. You're effectively a coach to your portfolio. Maybe a shadow CEO. And...
That's why I wrote the book. I was like, I can't coach all of them. I'm going to write the book to give it to the, to the...
I think all of us who have portfolios right now are dealing with the reality that we have some companies that are AI native day one, but a lot of companies that are not. Like a lot of these portfolios got built in the last five years, seven years, longer.
So as you talk to your entrepreneurs, what do you think are the, what, what are the best entrepreneurs doing with respect to reinventing their business for an AI native world, and what do you think the most common mistakes people are making?
Well, you know, a few of them, you know, they're like, 'Oh, we're going to go and replace customer support, right?' And if any of you have been on any of these AI-driven customers, they're absolutely horrible. They say they work. They might work, but they really don't when you're on the other side and you're having a problem. And so that's like, I don't want to put things up in, and I tell the companies that we're invested in, don't put things in front of your customer that you can't absolutely know that your brand is going to be, it has to be properly represented. And so I'm always about, okay, what's in the back of house? How can we do other kinds of automation and not necessarily human replacement? If we're doing human replacement, we're doing it with things that do not touch directly the customer. It has to go filter through a couple of chains of humans before it gets there. And so I'm trying to do much more on the operations and the interior side than on the exterior side.
So kind of moving to kind of just the way we interact with AI. I mean, if you think about the success you've had, so much of it early in your career was around beautiful design. You know, like I still have kept my first iPod. Like, I mean, it'll go Smithsonian one day, right?
And remember, the Nest Learning Thermostat was one of, it was probably the first consumer product ever to have AI in it. We just couldn't talk about...
In 2011. In 2011, we had AI, but we couldn't call it AI. We had to because people would be scared.
You have artificial intelligence changing my temperature. You're like, no. So we had to call it the learning thermostat. But we had real AI in it on a little processor, and it was running battery operated. We've sold millions and millions, first, you know, million seller.
But you have, but you had a beautiful interface in both cases.
Well, you had to have that, too.
And, you know, so you got all, AI has gone from the cloud to the edge, but we're all sitting around still typing into chat boxes. Is that going to be the ultimate interface?
I think, you know, if we look at how computing started, it all started with typing.
So chatbots are where we're at right now with, with typing. But I, I remember when my daughter, before she could read, she started interacting with Siri, right? And so if you start thinking about the next generations, and I, and I already see people, you know, younger people already using, you know, like ChatGPT voice.
Yeah, and just interacting that way with it. And I see that more and more, much more than you ever saw Siri or any other things. So I think we're going to go through the natural progression of, well, it's text first, but then it's going into other modalities. And I think a lot of the listening, you know, a lot of the, you know, you're starting to see these devices. How many people in here use Granola? Anybody know what Granola is? Yeah. Like that's like a simple thing where you're just, it's in the background and it, it generates incredible, you know, notes.
And OpenAI seems to have invested billions of dollars in some sort of device that will come out sometime in the next couple years, right? What do, if you had to... I know you're not going to go back and work at OpenAI for Jony Ive, but if you were out thinking about what that next hardware device would be for AI, what would it be?
AI, AI is all about context. So when, when we have, when we're talking about, you know, chat sessions and you're saying what's the next generation AI, if AI only, AI hallucinates less, you have to ask it less if it has the best context. And what we were trying to do at Nest was actually create incredible context awareness inside your home. So when you used an AI agent, you could get what you wanted to be done when you say just four words because it knows you're home, it knows you're in a certain room, what have you. And so these devices and the things that are being worked on now that are beyond the smartphone, but the smartphone can do this as well, are things that give much more context, real-time context to chatbots or voice bots or whatever you want to call them, or whatever, however you use them, so that they can answer much more accurately when you ask it a much simpler question. Right? I can, right now you're asking me questions. We know this context. We know what, what's going on here. The chatbot needs to know who are you talking to, in what space, at what time of day, all these different things. And so what is going on, and because of Apple's restrictions and Android restrictions, a lot of the software apps cannot get sensor data, and they can't get access to all this stuff. So people are going to be creating these either wearable ones or put it on a desk, put it in your backpack, whatever, sensor blocks, and they get, and they'll pass all the sensor data over, you know, Bluetooth or what have you, the network to get into the bot to give the context because Apple, Google, whatever doesn't want to give out that data for privacy or other reasons. And so third, these companies, OpenAI is one of them, are making sensor boxes. So think cameras, multiple microphones, temperature sensors, location sensors, all kinds of different things in that so they can pipe all of that context into the chatbot, voice bot, whatever it is, so that it can give much quicker answers. Now what is also going to happen is those boxes, what should happen is because the problem is we're seeing with, you know, trillion dollars of data centers and everything is most of the time when you're doing consumer apps, not when you're doing deep research, these other things, you don't have to do a lot of inference work, right? Inference work for an AI to work. And so if you're seeing everybody's making small models, those models need to run right next to you because the power requirements of going over the network, everything is the latency of going over the network. You need everything right there. So you have the sensors, you're going to have the little compute because you can't use the compute on your iPhone. iPhones have tons of AI compute, but it's not being exploited, right? So they're going to put that in the box as well, right? It can do the voice right there. It can do simple, a lot of the, a lot of the basic things, not just simple, but basic things that you would want done with an AI in a consumer mode. You're not asking it to do a, you know, a chemistry experiment on the, you know, right there. Not everyone's asking that every day. If it needs to do it and doesn't understand it, it can go back, right? We're always seeing that with ChatGPT-5 right now. It has, you know, quick mode, thinking mode, and then it self-selects. Well, you're going to see all kinds of modes like that, and they're going to run local or they're going to run in the cloud or some hybrid version, those kinds of things. And so that's all going to be coming, and that's what you're going to start seeing in these sensor boxes or in glasses or in what have you, you know, even earphones. They're really great over-the-ear headphones too. So that, you know, you're not going to, your smartphone's not going to be cast aside. It's still going to be there. Why? Because visual information is best shown visually, right? It's just, and you're not going to carry yet another screen. And even these little screens in your, like Meta, don't, don't, don't cut it.
So, so you're going to still see a, these are all going to be peripherals to a smartphone.
For, for the, for the time being until we make incredible agents that really know you where you can get rid of the smartphone.
And then you'll just have a small display that you can bring in.
One way I always get interesting insights about technology is talking to my kids. And so I, you know, I first heard about ChatGPT from my kid in college saying, 'Dad, you have to be on this right away.' But I was talking to them recently and I said, 'So, like, what are you using it for today?' He's like, 'Dad, have you asked it to describe you to yourself?' And I'm like, 'No, why would I do that?' He's like, 'Well, its memory is pretty good. Why don't you try that, Dad?' So I try it and I was like, 'Holy, it must be better than my wife.' And then I called him back. I said, 'That was kind of scary.' And I'm like, 'How?' He's like, 'Look, my friends are using this for therapy.' I mean, like it's... So my question to you is, how much, we talked earlier about accuracy. We talked about how these models are, you know, 80% accurate. It's not great for medical applications. How much does the fact that people are using it in that way or using it for relationships, how much does that scare you?
It's, it's, it's pretty scary to tell you the truth. It's pretty scary because these things start to become kind of information loops on you. And then I, because I've seen some people use them, and I'm like, that's the therapy help that it's giving you. It's basically just reflecting what you want to hear as opposed to helping you find another path for getting, getting help.
It tends to be very agreeable.
It tends to be very agreeable. So I actually invested in a company that literally was, that is a relationship coach, and it's fine-tuned on all kinds of real data that you say, 'Okay, I want to have a, I want to have a hard conversation with my partner, spouse, whatever it is.' And the AI therapist says, 'Well, you could do it these ways.' And then an AI version of your partner, whoever it is, you start talking to it. And then the therapist goes, 'Hey, you know, you might want to say it this way, and you might want to say it that way, and this is how it reacts.' And literally, this stuff is working today. So again, it's, it's got to be in context. And my, my worry is that these know-it-alls, we're going to get in a real weird, in a real weird situation when...
That's what I worry about happening with social media, you know, like you can see these things early and you either have to come up with a solution or, or not. But once it is in the wild and everybody's using it that way.
Yeah. Look at already, you know. Yeah. No, it's, it is, it's one of those things where it's, it's, it's...
It's scary, but it's also very, very interesting. And I'm also optimistic and hopeful. So I can't, I can't just put my head in the sand and go.
What's the, from your standpoint, what's the most overhyped part of the AI market? I don't mean necessarily company, it can be application. And what's the most underhyped where you think like there's going to be a real world impact and we're not really talking about it?
Well, I think I just, I think I just mentioned it. We're talking about incredible huge data centers, nuclear-powered data centers and everything. If you, if we start taking a lot of the applications that don't require that kind of level and you start pushing them to the edge and you start putting the compute on your devices that you have around you, we're going to, we're not going to have anywhere near the amount of compute that you need. You, we saw this with DeepSeek. Everyone was, there was the DeepSeek moment early this year and like stock market dropped and like, oh my god, we don't need all these data centers. And, and it's true, people just forgot about it. Those, those models exist. Even OpenAI just came out yesterday, I think it was, with their, the voice model that was 70, 70% much more, 70% smaller and can run on device.
Right. So it's like these things are happening and it's being, they're not being valued yet.
And all of this stuff has to move to the edge. A lot of this stuff has to move to the edge. It's going to be like supercomputers. We all don't have supercomputers of the 80s, 90s, and even today where we did run nuclear experiments and chemistry experiments and all these things. We still have supercomputing centers, but we're not all using them. And we're going to have, yes, we'll have supercomputing, quantum computing, all these other things for certain AI applications, but we don't need all of that stuff in our day-to-day lives. And so our data centers are getting incredibly massively built out for things that are going to naturally move to your, to your pocket or in your home or in your business or wherever because of the fact, energy, wireless communications, the cost, you're going to want to be able to afford these things. People are paying, you know, nowhere, like I heard that the stat is the GDP of Singapore is the same size as the investments in data centers this year.
And what is the revenue?
What is the revenue?
30 billion.
Somalia.
Okay, so like, wait a second, we have a mismatch here, and I don't care how much we build out, and what's going to happen is people are not going to be paying $200 or $1,000 a month individuals for these AI agents. They're going to have to run closer and closer to you.
So, and again, you don't have to name this in the context of companies, but you know where these companies are valued. Are you long or short there?
I'm not investing in them.
I'm investing in the picks and the shovels.
I'll ask the question a different way though. Like when you look at the capital intensity and you look at the announcements over the last three weeks of the amount of commitments that Nvidia has made to Oracle and Oracle's made back to AMD and AMD's made to Nvidia. You know, you add it all up and it's, you know, it's, it's a pretty big number.