← Design Better
Design Better
Leonardo Giusti: Archetype AI's co-founder on physical AI and the limits of the chatbot
00:28:44
We speak with Leonardo about what it actually means to treat emerging technology as a design material, why the chatbot is a primitive interface for the physical world, and why he believes augmenting human intelligence might be the most important design challenge of our time.
Loading summary
Transcript22 lines
- [00:01]
A
Design can play a critical role in transforming R and D technology into products. R and D often is driven by the curiosity, and it's a beautiful thing. Sometimes we forget this. But these scientists, these engineers, are driven by their own passion, their own curiosity. As a designer, our role is to understand it and try to materialize it into something tangible and concrete.
- [00:30]
B
Leonardo Giusti has spent his career in the spaces between disciplines, between art and science, between research and product, between the physical world and the digital one. It's not a conventional design path, but it's one that's led him to work that most designers never get near. He's the co founder and chief design officer of Archetype AI, a company building foundation models trained not on text or images, but but on the continuous stream of sensor data flowing from the real world, like factories, power grids and city intersections.
- [01:00]
C
Before that, he spent nearly seven years at Google's Advanced Technology and Projects group, where he led design on Project Soli, a miniature radar chip that taught devices to understand human gesture and presence, and Project Jacquard, which wove interactivity into everyday objects like Levi's jackets and Yves Saint Laurent bags. He holds a PhD in Human Computer interaction from the University of Florence and spent years as a postdoctoral researcher at MIT's design lab. He's also filed more than 30 patents. What makes Leonardo's thinking distinctive is his insistence that the metaphors we use to describe AI shape everything. How we build it, how we regulate it, and who it ends up serving. He's skeptical of the dominant vision of AI as an autonomous agent that does things for us and is pushing towards something different, AI as a tool we think with.
- [02:02]
B
In this conversation, we get into his unusual path to design through cognitive science and robotics, what it actually means to treat emerging technology as a design material, why the chatbot is a primitive interface for the physical world, and why he believes augmenting human intelligence might be the most important design challenge of our time. This is DesignBetter, where we explore creativity at the intersection of design and technology. I'm Eli Woolery.
- [02:25]
C
And I'm Aaron Walter. If you're hearing this, you're not currently on our Premium subscriber feed. DesignBetter Premium subscribers enjoy weekly episodes. That's four episodes per month rather than just two, and all of them are ad free. Plus you'll get an invitation to our monthly AMAs with the smartest folks in design and tech. And if you subscribe at the annual level, you'll also get our Toolkit, a collection of our favorite design and productivity tools like Perplexity, Miro, Read AI and more. You'll hear a preview of this episode, but if you'd like to hear the full conversation, please consider becoming a Premium Subscriber@designbetterpodcast.com subscribe the podcast is available to everyone through our scholarship program, so if you can't afford a subscription, just shoot us an email@subscriptionsecuriositydepartment.com we'll help you out. We'll return to the conversation after this quick break. Design Better is brought to you by WIX Studio, the platform built for all web creators to design, develop and manage exceptional web projects at scale. Learn more@wix.com studio and now back to the show.
- [03:57]
B
Leonardo Giusti, welcome to the Design Better podcast.
- [04:00]
A
Thank you for having me.
- [04:01]
B
We're excited to have you. You have a very interesting background, worked at Google for a number of years as a head of design and now you're working at what looks like a pretty unique company, working at this intersection of physical and AI products. So we've got a lot of stuff to talk to you about, but maybe first tell us about that sort of transition you made from working at Google on some really interesting cool products to starting your own thing here with Archetype.
- [04:28]
A
As you said, I'm now one of the co founder of Archetype AI, which is a physical AI startup where we are building these AI foundation model that is designed from the ground up to live in the physical world, helping people run machines, factories, infrastructure, cities, et cetera. So it's really like a deep tech startup where it has this very strong R and D component. And so as a designer I always found myself working very close with R and D. I'm always being more comfortable, you know, to work with engineer and scientists a bit outside to what you may consider to be the mainstream of the design community. I believe this is a little bit the result of my upbringings as a designer. I didn't really take a straight path, but many different detours that ultimately brought me where I am today. So I have a very mixed background as a designer. As you can probably say from my accent, I'm Italian, I studied there and when I was in college I was starting to be very interested in cognitive science and artificial intelligence and human machine interaction and I stumbled upon design later in my career. I was doing my PhD at the time. I was working on robotics, try to design them in a way that they can interact socially with people and we did this project with Domus Academy. It's a design institution in Italy that captured the outcomes of the radical design movements in Italy in the 60s and the 70s. It's a design school, it's a design institution. They also do research. We started working together and I was exposed to this idea of design design though as a critical practice. It was this idea that design can be seen as an inquiry method, a way to question the status quo. And so this really clicked with me, this idea that design is not just about solving problems, but really it's about discover the problems that are worth solving, you know, to outline new narratives and vision for the future. In this context, I learned to see prototyp, for example, not just a way to approximate the final product, more like provocations to think critically about what you are building, about the implication in society and so on. All these things click for me and I found my passion and I took all this bag of experience and I moved to us in 2011 when I was a researcher at the MIT Design Lab in Boston. I spent a couple of years and then worked a little bit at Samsung and finally I landed at Google atap. Google ATAP is advanced technology and project. So I was actually working in an R and D lab that really helped me refine my design practice as a designer in this highly technical environment. When I was there, we were working on some very interesting projects. One of them was Project Soli, which was this tiny radar that we were able to embed inside consumer electronics and it was able to understand what was happening around them. When we started, really we didn't really know about what do we do with this. You know, it really started from the curiosity of some scientists and some researchers and a designer. When you are there, it's kind of interesting because you can't really think already about the products or product features or why it's useful. But really the first thing that you need to do if you want to sit at the table with these people is to work with them and try to understand why they are curious about these things and what are the properties of these new materials, basically that you're working on. The core idea of Soli was that this tiny radar was embedded in consumer electronics that now become aware of their surroundings and aware of people's behavior. And this radar was able to understand different aspects of nonverbal communication. For example, it was able to understand when you were moving toward a device or away from a device, they were able to understand different type of gestures so that you can start to establish a more seamless interaction with these devices. What brought me to archetype the team that was building solely, we realized very quickly that in order to make it work to understand all these different behaviors, you have to use very advanced machine learning. And machine learning was very expensive at the time. In order to recognize, for example, even one single gesture that you do on top of a nest device or a pixel phone devices. It requires a huge amount of data collection. For example, we spend like years to collect data of a simple swipe gesture. And we collect 5 million samples of this gesture all around the world to take into consideration different cultural difference, demographic difference, and so on. And it was very expensive. Every time you want to do a new gesture or a new behavior, you have to redo this training once again. And so very quickly we realized that this really doesn't scale much. And it's probably one of the key problems that IoT had in the past. In parallel, there were all these interesting new things happening in AI, these foundation models. And so we thought, what if we actually built a foundation model for sensor? So every time we want to create new use cases, we don't have to retrain from scratch, but we can actually just fine tuning with a smaller set of data. And this is a little bit how Archetype started with this idea to build a foundation model for sensor data so that we can develop very quickly new use cases, a sort of like horizontal platform for the physical world based on the continuous understanding of sensor data. So we are able to translate the sensor data into meaningful information on the project solely.
- [10:00]
B
I'm curious, did you bring any of your cultural background to this project? I'm just thinking that Italians like to speak a lot with their hands, which I love. So there's naturally, like, I'm sure maybe an interest in like, what could we do with sort of a gestural communication, but perhaps it doesn't translate across different cultures. Yeah. Just curious what part of your perspectives you brought to that.
- [10:17]
A
Definitely when you look at the way we interact with each other as humans, it's very intuitive, right? If I see you walk behind me, I can probably keep a door open for you even without saying a single word. I immediately understand that you want to come through the door as soon as I keep open to you, if I see that you are touching your glass and maybe pour water for you, that there is the intuitive understanding that we have between people that is based on this understanding of subtle cues, this little nonverbal communication cue. But this sort of understanding doesn't really exist with technology. Technology is, if you want, very rude. It interrupts us at the Wrong time. It doesn't have the ability to understand any of these social cues that make our interaction with other people very smooth, easy and ongoing. And so really the idea there was, can we bring some of this social intelligence into consumer electronics? Can we design this device in a way that can participate in our everyday life in a more polite way, if you want? It's not that we need to make them smarter, but maybe they can be a little bit more polite and participate in a more harmonious way in our day to day life. And so we took inspiration from the way people interact socially to build a design language, an interaction design language for consumer electronics. And this was, in a nutshell, what Project Soli was in the end.
- [11:39]
C
So Project Soli, it shipped in Pixel 4 and also in the Nest hub. So it went from being this abstract idea to something that was like a commercial feature in existing Google products. You also worked on Project Jacquard as well, which was building interactive textiles that were deployed commercially with Levi's jackets and Yves Saint Lauret bags as well. What did you learn from that project? Project Jacquard, I mean, it feels like it is related to Soli, but different in many ways.
- [12:15]
A
Project Jacquard, as you said, was this conductive fabric that we could embed into different products. The core idea was to bring interaction capabilities to everyday objects. There was this idea that we can bring computational capability into everyday object, but often it results in the design of new things that after six months end up into a drawer and no one use them anymore. We have seen in the past this explosion of Internet of things, right, that they've tried to do all sorts of things that are cool for a little bit, but then disappear. What we really tried to do with Jacquard was to really try to understand how to augment the things we already use, like a jacket, a backpack, a pair of shoes. For example, a project we did with Adidas, it wasn't to create something new, but to really try to leverage what we already use and exist and make them better. And in order to do that, we have to really study how things were made. So we worked with this company to understand their manufacturing process, for example, to be sure that the technology we were building could be part of the way they do things, so they didn't have to change, for example, how you make shoes or how you build a jacket. So this was to me, a very important understanding coming from Jacquard. When we want to bring technology and intelligence and connectivity to the physical world, we should start from what we already have and try to make it better. In order to do that, we need to deeply understand as designers how things are made. Because if we create friction to the manufacturing process, it will never happen, right? I mean, the costs start to raise and so on.
- [13:57]
B
So a few of your projects have made it from this R and D phase into actual commercial applications. And that's not often true for R and D projects. And we've talked to the Google X folks a few months ago, and they're sort of in this similar arena of developing very speculative in some ways, products and experiences, and some number of them make it on what are your insights about bringing something from a more experimental phase into something that's more commercial?
- [14:24]
A
I believe that design can play a critical role in transforming R and D technology into products. I do believe that design is sort of like this cog that sits in between R and D and productization. And very often, one of the mistakes that I have done, I've seen other people doing is going to work with the people that do R and D with a product design mindset. And I'm trying to explain what if you go there and you start immediately talking about product features, end users, you create constraints to the R and D process that at that time are not really necessary. R and D often is driven by the curiosity, and it's a beautiful thing. Sometimes we forget this, but these scientists, these engineers, are driven by their own passion, their own curiosity. As a designer, our role is to understand it and try to materialize it into something tangible and concrete. And of course, this is, to me, is always the first step when I start a new R and D process. I try to work with these people, try to run what I call material study. I try to understand any type of technology as a new material. And I try to study it first, try to understand what are the characteristics of this new material. For example, we saw it was radar, it works on the electromagnetic spectrum. But you really didn't understand from an interaction point of view what can actually do. So the first study we did was to understand, okay, how does it react to the human body? Can you understand large movements like, I'm here, I'm not here. But surprisingly, as we were running this study, we discovered that it can also detect tiny, tiny movement of your fingers. And so in this first phase, really, as a designer, we need to be very humble and we need to work with these people, try to work with them to take what they're building and try to materialize into prototypes, experiments that allow you to understand the affordances and the properties of this new technology. Because then this can give you the ideas. When we realize that Soli can understand large body movement and tiny mommy movement, then we. Oh, interesting. We can now understand different aspects of nonverbal communication, like we do as human like presence, large body movement, gestures. And then from there we build a coherent interaction design language. And of course, this is the first phase and this allow us to have a very material understanding of this technology, of the properties that they have, which is quite common in other disciplines outside interaction design, like in architecture, the importance of material, for example. Right. It's critical for the development then of the final output. This is the first step. And then you need to understand how to translate that. What you do by doing this, basically you build almost a map, a legible map of the technology so that someone with a more product mindset can look at this map and understand what journey they want to take into the product development. And so I think at the very early stage, the role of design is to build these legible maps of the technology to inform further on, like the development of new products. The only example that I've seen, like, I mean there are a lot of people, the Bauhaus was big in these. They had like these basic courses, like when they were teaching the student how to explore the properties of materials, for example. I think another interesting reference is what they were doing at Pixar. And there is this famous quote, like, technology inspires art, arts challenge technology. And I see at the very early stages a very similar approach where technology really inspire design, but design has the opportunity to push the technology from the very beginning. If you're willing to sit at the same table with these people so that you can accelerate and try to find a path over time that makes sense from a product point of view.
- [18:19]
C
History is filled with examples of consumer products that came from some sort of R and D scenario, but they just showed up at the wrong time. There's the Apple Newton, which was a brilliant device. Google Glass, the Segway. All of these things now are in our lives and like normal and not that magical though they were at the time when they were introduced. But they just weren't timed very well in that R and D to productization cycle. How do you know when something is ready for consumers or vice versa, when consumers are ready for that device?
- [18:56]
A
So the way I see the innovation cycle, right, very often we talk from zero to one, which is like from an innovation to a product. And as a designer, we have a lot of techniques, design thinking and so on. To really work this through, right? We take some existing technology, we reshuffle them, we play with them, we understand how they fit into society, use cases and we build beautiful product. But there is another face that it's often not considered, especially as a designer. It is the face that goes from the invention to the innovation and the invention of a new technology, like radar or AI for example, as well. This phase from minus one to one is where we translate some early stage ideas into a technology. It is an innovation itself. And then later on we translate into a product. I think there are different ways. And I've talked a little bit about how you can translate early stage invention into possible innovation, like a new technology with coherent, for example, interaction model for Soli and so on. That innovation, is it going to become a product? That's, I think when it gets quite hard and it really depends on timing, is society ready for that? Is the economic context and cultural context able to receive that technology? And the reality is a lot of people run studies and you can try to predict and figure it out, but it's typically really hard. Especially with Project Jacquard, we weren't really sure about the viability of Jacquard as a product that actually fit into people's everyday life. We have a vision, we have an idea, but what we were able to do was to start really, really small. And so not every company can afford this, of course, but the idea that you have a branch of your company that is dedicated to, to create these products which you build in small batches and you actually test them before you go a full scale. The idea that you can try different version of a product, it's typical in software. It doesn't really happen in hardware very often because of obvious reasons, but I think is what we're trying to do with Jakarta, we're seeding products into the market to understand and study how people react if it works and if it doesn't. So understanding the right timing for a technology and if there is the right social, cultural content, I believe it's really hard to do on paper. And I think there are ways in which we can these days, especially try to test it, even if it is an hardware product, the same way we do with software.
- [21:35]
B
Let's shift and talk a bit about what you're doing now with Archetype. And a lot of our audience is familiar with large language models and where they're trained on, let's say, all the text that exists on the Internet or all the imagery. And what you're doing is Slightly different. You're building models that are based from data from the physical world via sensors and other technology. Maybe you could give us a specific example of the way that your technology works in a project that you're working on.
- [22:02]
A
Our foundation model is designed from the ground up to continuously analyze sensor data and translate them into an interpretation that is useful for certain tasks. The model is able to receive as input data, such as accelerometer cameras, any type of time series data and is able to translate them into a description and is able to understand hidden patterns in the data, like anomalies and so on. The core idea of the model is that we are able to map this sensor data into the same mathematical space that language. And so by aligning them together, we can both control the model using natural language and actually have a description of what's happening in the world in a language that we humans understand. Sometimes we refer it to like Arosetta stone, that is able to map different types of input to language and make them intelligible. We are working with many different customers, mainly in the area of industrial technology. We have several partnerships with manufacturing companies, for example, where they have a lot of data. There was a time 15 years ago where factories have embedded a lot of sensors into machines and they are sitting now there with a huge amount of data and they don't know what to do with that. And now finally we have AI that can actually able to understand this data. And so you have entire factory that have been already sensorized that we can actually provide additional value there. And in use cases, like, for example, one of the most typical one is predictive maintenance. We know, for example, when a machine is about to break, we can spot, for example, signal in the machines before it happens that it's going to break. And that's very important because these machines sometimes are very expensive and even few hours of downtime is worth a lot of money. We are able to identify anomaly, for example, in certain data in a certain machine. That may prompt an operator in the field to go there and check what's going on if there is something is broken or not. Other use cases that we have outside the manufacturing and industrial setting is, for example, we are working with different cities as well. In this case, we are using camera as input. And one of the projects that we have is with the city of Bellevue in Seattle, where they have a lot of cameras at the intersections and they are interested in optimizing and increase the safety of this intersection. Our model is able to identify what they call near misses when an accident is about to happen, but that didn't really happen. And by analyzing this data across all the different intersections they can build a sort of like heat map of the city, understanding what intersections for example are more problematic than other and then potentially intervene. A follow up project that we're also doing with them is our model can also recognize if someone with a disability for example, or a mother with kids is crossing the street and if it taken longer then the lights allow. They can automatically extend the lights so that they can finish and crossing the street. So these are completely different use cases, one in factories, one in cities. But the idea is like wherever there are sensor our model is this general purpose intelligence that can help to extract insight from this data from the real world and provide the people working there with additional intelligence if you want. We are trying to augment them to make the right decisions.
- [25:32]
C
How do you think about privacy and ethics in this space? There's tons of very valuable scenarios where this could be helpful. You described some that are very business to business kind of manufacturing efficiencies and operations in the hands of a government. It could be helpful at an airport to detect potential threats. You know, is this person walking differently? Is there something different about this person's luggage? So could TSA for example, screen more effectively or get some clues about where to look for threats? I can also think of government scenarios where if you've got an authoritarian government and they've got this sensor data, how might they use this to change behavior or control masses? There's a lot of different ways that this could be used potentially to dangerous outcomes. How do you think about those sorts of ethical implications?
- [26:27]
A
I think we have a problem as a society that is beyond what archetype is doing, what you can do with sensor, because these sensors are already there. Some of these capabilities are already there and I do believe it.
- [26:41]
C
If you'd like to continue listening to this conversation, you'll need to subscribe@designbetterpodcast.com subscribe. Once you do, you'll get access to every full length episode, all ad free monthly AMAs with inspiring people in design and tech and recordings of all our past AMAs. The podcast is available to everyone through our scholarship program. If you can't afford a subscription, just email us@subscriptions the curiosity department.com and we'll help you out. Your support makes design Better possible. Invest in yourself and the design community by subscribing at Design Better Podcast. Design Better is supported by masterclass. I've been fascinated by the research from Cornell University showing that we get far more lasting satisfaction from experiences than from material things. They call it the experiential advantage. The idea is that while a new gadget eventually becomes just another thing, the wisdom and skills you gain through learning and actually become part of your identity. That's why I've made Masterclass such a consistent part of my routine. It's an investment in an experience that pays dividends for years. And I'm not alone in that. 88% of surveyed members feel that Masterclass has made a positive impact on their lives. I use the platform to stay curious across the board. I've been using James Clear's class to refine my daily habits, and I'm looking forward to Michael Lewis's class to find better ways to tell the stories that we share right here on Design Better. Whether I'm in the car, using audio mode or at my desk, it's a way to ensure I'm actively becoming a better version of myself. With Masterclass, you can learn from the best to become your best. With plans starting at just $10 a month billed annually, you get unlimited access to over 200 classes taught by the world's top minds in design, business and leadership. Plus, there's no risk. Every new membership comes with a 30 day money back guarantee. Right now, our listeners get an additional 15% off any annual membership at MasterClass. Com DesignBetter. That's 15% off@masterclass.com DesignBetter MasterClass.com DesignBetter.