Beyond Replication: Building Non-Human Intelligence Through Physical AI w/ Jaime Lien & Rashi Agrawal #238

We're doing a special in-episode feature with our friends and sponsor, Span. 2025 has been the year of AI adoption, and AI transformation is going to be the reality in 2026 in the years to come. Stay tuned for later in the episode, Steven Paletto, field CTO at Span. Deconstructs the tactics behind how top companies are navigating the AI transformation, navigating new bottlenecks and accelerating work. Hello and welcome to the Engineering Leadership Podcast brought to you by ELC, the Engineering Leadership Community. I'm Jerry Lee, founder of ELC, and I'm Patrick Gallagher, and we're your host. Our show shares the most critical perspectives, habits and examples of great software engineering leaders to help evolve leadership in the tech industry. Today, we're sharing a special conversation recorded live at ELC annual 2025. If you missed it, the entire conference was dedicated to explore how AI is fundamentally transforming engineering leadership. And so we looked at a lot of different things, you know, from the way that we build products and structure our teams and organizations, to the actual day-to-day ways that we work and build together. But what stood out to me most about this specific session with Jamie Lien and Rashi Agrawal is that it moves beyond the height. It takes a hard look at how what is possible from a product and technology perspective is shifting, and examines the implications of that on the product experiences, the problems that we can solve, and the tools that are now possible for us to create. Here is some background context on this session. The true promise of AI isn't in replicating human intelligence. It's in developing new forms of non-human intelligence that perceive and understand the world in fundamentally different ways. This means building machines that can directly sense the physical world, form their own interpretations, free from our biases, and translate that understanding back to us. This conversation explores the frontier of physical AI with Jamie Lien and Rashi Agrawal. Jamie is the co-founder and chief scientist at archetype AI. She previously led radar sensing development for Google A-taps, Project Soleil, and contributed wireless communication and localization expertise at NASA's Jet Propulsion Laboratory. Rashi is leading AI at Hinge Health. She formerly was head of AI engineering at Goodleap, leading enterprise-wide AI initiatives. Previously, she scaled engineering teams at Yahoo, advancing its multi-billion dollar advertising business. Enjoyed this conversation from ELC annual 2025 with Jamie Lien and Rashi Agrawal. Today as we know, everybody is talking about AI, but I'm especially excited to talk to you about the world of physical AI and what you're building at archetype AI. So let's start with this grand vision or what archetype AI is really about. archetype's mission is to build a foundation model for physical reality. Paint a picture for this audience five years from now, what becomes possible with this technology that feels like science fiction today? Thanks, Rashi. So fundamentally our vision is that artificial intelligence will be able to directly observe the real world in real time. And what this means is that you can have a wide array of different sensor modalities and have that data fed into a model that can synthesize all of these, translate them into some sort of human understandable interpretation or meaning. And not only are we talking about the more familiar sensor modalities like images or sound, but modalities beyond what we as humans can directly perceive. And when we think about the wealth of phenomena that is existing in the world around us, electromagnetic waves, radar, infrared, ultrasound, chemical processes, a wide array of phenomena that are at tiny scales or astronomical scales that we as humans fundamentally just don't have the capacity to directly perceive and interpret. You can see the potential for AI to really expose us to some new understanding if it's able to directly observe these and learn from them in a self-supervised way. And so when we talk about things that become possible, I think the fundamental thing here is discovery, that AI that has access to all of these observations of the world and the ability to learn from them, rooted in direct observation and self-supervised learning, could potentially help us to discover entirely new phenomena, accelerating complex sciences like astronomical physics, and in more familiar arena, helping us to do things like, you know, very early health diagnostics because it's able to detect pre-symptomatic signs of illnesses or different conditions. That's a fascinating vision. When I first learned about archetype AI, my mind immediately went to the Westworld reality, the Westworld, you know, world that was created. But what you're really talking about is not an alternate reality, but a true partnership. How physical AI can truly augment the way humans perceive the world around us and what it can unlock for us, right? Exactly. Is that a fair way of thinking about it? Yes, exactly. Our goal here is not to replace humans or even to automate what humans can do, but to expose humans to new ways to perceive reality and perceive the physical world and the ability to synthesize this into human understandable insights. And so there's really, I think, a component of this innate, rooted intelligence, the machine being able to form its own interpretations of the data and its own latent representations of the data, and then a second component which is being able to translate that understanding to us. And so fundamentally, we see this ability and the potential for AI to learn directly from observation and also understand, in parallel, our human mental models of the world, well enough, that it can fluently understand and translate those interpretations back to us. Yeah, that makes a lot of sense. And I think everybody in the audience, they can very well relate to the explosion of LLMs, right? LLMs have had their transformer moment, which literally kicked everything off. But for physical AI, can you help us understand what was that key breakthrough? The missing ingredient, if you may, that makes what you're talking about possible now. Yeah, I think in general, there are two buckets of technological breakthroughs that make this the right time for physical AI. One is something that has applied to LLMs as well, and this is basically the advances that make general purpose intelligence feasible. So things like the advent of transformer based models where you can paralyze the training and the inference of massive, large-scale models, basically that have the capacity to take in a lot of data and build in a large general understanding of the underlying structure and patterns within that data. The second one is advances in self-supervised learning. So no longer are you constrained by this need to have massive amounts of labeled data, but you can actually learn from the data itself and thereby bootstrap massive amounts of unlabeled data. And then the third, of course, is the compute. So the ability to actually run all of these advances on feasible machinery. So this is one bucket. I think the other bucket, which is very specific to physical AI, is what we've seen in modern days with the commoditization of hardware, and specifically with sensor hardware. And what this means is that today, the ability to produce sensor hardware smaller, cheaper, faster than ever before means that we can scale-ably develop and deploy sensors at scale into our environment, at a scale on a rate that we've never seen. And so when we think about things like the advances in semiconductor manufacturing with MEMS, with SOC integration, all of this means that complex sensor modalities that previously were extremely expensive, extremely time-consuming to build and deploy and we're very brittle, we're very subject to different environmental conditions. Today, it's a sensor on a chip and you can imbue it into environments, you can put it into consumer devices. And what that means is that we have these perceptive nodes that have the ability to do these kinds of measurements and observations of the world directly as we live in it. So it almost sounds like it's the perfect storm, right? Like finally, all the software innovations from the LLM world are meeting this explosion of cheap scalable sensors. In fact, right. It sounds like, you know, the hardware has finally caught up to software's ambitions, right? We're taking a quick break for a special feature, taking a look at how AI adoption is actually transforming the Devcycle with our friends and sponsor Span. Steven Paletto, field CTO at Span, shares how top companies use metrics for strategic inquiry, effectively drive performance improvement and make smarter business decisions. Across the board, a fairly common pattern that we're seeing is that the adoption of these AI tools does speed up and yield more output in aggregate. So teams are shipping more code more frequently. We're also seeing a slowdown in code review and more of a burden being placed on senior engineers to review the code that the Gen AI tools are producing. We've been highlighting this trend and identifying opportunities for training and coaching more junior engineers to review the code before it goes up into the pull request process. And also think about when AI is going to genuinely speed you up because it's good at producing a first draft this largely going to pass the code review process versus lots and lots of iteration and back and forth where it may actually be faster to write the code by hand. In that adoption chapter, I would say we've seen tops down approaches, but there's also been more bottoms up, more grassroots style adoption. And in both cases, I think leaders are looking for some level of visibility into where it's being adopted, where it's not being adopted. There's organizations that have decided to go all in on a singular tool, like maybe they're all in on cursor or all in on cloud code or what have you. There are also many organizations that are experimenting with multiple tools. And so having a single pane of glass visibility that cuts across multiple tools has proven difficult. And so one of the challenges is just getting all the data into one centralized place. Span connects to all the places that your engineering team does work from source control to ticketing, calendars, gen ai tools, and provides a single pane of glass for engineering leaders to understand what's happening in their software development life cycle. As AI adoption is transforming that process, we provide a lens on which metrics are improving, which are degrading. We can provide engineering leaders with a map of where their engineering time is going. These are the major projects and initiatives. This is the rough allocation of time that this project has received from the team. And then cut it by AI usage. So you can understand for the code that's being entirely produced by human generation versus those that is being produced by these tools and provide insight for engineering leaders to figure out where to focus their efforts and improve. Span is the AI native developer intelligence platform bringing clarity to engineering organizations with a holistic human centered approach to developer productivity. So if you want to get a complete picture of your engineering impact in health, drive high performance and make smarter business decisions go to span.app to learn more that's span.app. So let's take a step back, right? The title of our session or our talk today is Beyond Replication building non-human intelligence. It is a powerful phrase. What does non-human intelligence actually mean to you? And why is that distinction so crucial for the future that you're building? So fundamentally we believe that the potential for AI is not just mimicking what we as humans already know. But to be able to help us to understand things that were beyond our previous realm of knowledge. And the way that we believe that this can be done is by giving AI direct observation capability of the physical world and the ability to learn from that in a self supervised way. What this means is we're trying to bring up this model without the injection of human labels and without the injection of human biases. This includes things like human-defined physical laws, human-defined concepts and semantics to describe what's going on. And if you think about it, this is actually fundamentally different from a lot of approaches that people have today. We see a lot of models where basically you're trying to build this with physics-informed architectures. You're pre-defining these physical laws that the model should follow in interpreting the sensor data. And so this idea that rather than imbuing these human-defined laws into the model, that the model itself can learn its own representations directly from the data is what we're trying to pursue. This idea that the model doesn't need to learn from us, but we potentially can learn from the model and its own findings of the data. That's a very critical and crucial distinction. Thank you for clarifying that for our audience. It sounds like what you're trying to say is the physical AI is trying to bypass the inherent biases or limitations of human perception. We are emitted by our languages, by our five senses. But to get to what a fundamental truth of how our world really works, that is where physical AI is truly unlocking the capability. That's exactly right. And I think there's a direct contrast here with LLMs today, where you see that these models are trained basically on a human interpretation of the world through language. And we all know that language is inherently lossy, it introduces biases, it's not necessarily a one-to-one match for what's exactly happening in the physical world. And so this idea that language is, it's neither necessary nor sufficient for true intelligence, for developing understanding as a prime hypothesis and a key foundation of what we're trying to build today. And what we think is that we can leverage language not to build this innate understanding, the latent representation of the data, the reasoning of what's going on, but language as a communication tool to translate from that machine understanding into human interpretable outputs. And so as I was saying, there was this element of the machine understanding, native understanding, and the element of translating that into language or other types of human interpretable forms. So let's take it to the real world, right? And coming from Goodleap, this is where it gets very real for me. At Goodleap, we support solar installers, contractors who work in messy, unpredictable, physical environments. How can physical AI help them in that critical last mile? Collaborating with their expertise to make better, safer decisions right there on the ground. Yeah, and thank you. I really like this question because one thing we found in our work is that these skilled traits, people, these technicians, they have a wealth of expertise and domain understanding. And so we believe that archetype AI, it's really important for these people to maintain agency over these sort of tasks. In addition, we have a lot of respect for the complexity of this problem. As you were saying in the last mile, you often encounter unpredictable conditions. You encounter failures that you didn't plan for. And so this gap between the planning and the execution is where this really comes into play. So I think where physical AI can help here is really in augmenting the sensory perception of the person on the ground. And what I mean by that is basically supplementing their own perceptive view of what's going on with additional sensory inputs and not just the raw data of those, but synthesize down into a form that they can very easily and quickly understand. We all know that there is a bandwidth on human mental capability. If you're focused on one task, it becomes very difficult to perceive what's going on outside of that. And so the ability of AI to supplement with, for example, chemical conditions or impending failures that you might not notice, whether conditions which might indicate that something dangerous or unstable might happen, I think comes into play. The other area that I think AI can really help here is in providing a more seamless interface to the machines itself. And what I mean by that is that because of AI and the ability to perceive how the person is behaving and the way that they're maintaining or interacting with these machines, you can potentially unlock nonverbal types of interfaces to your machines. And it makes it much easier for them to activate controls or perform their tasks without having to manipulate very bulky or unnatural interfaces. Yeah, so I think the key theme that you're trying to underline over here is AI, especially physical AI, is here to augment, you know, all of these skills, skilled experts. It's not here to replace them. Definitely not to replace their judgment. The way I see it is physical AI or no companies like archetype AI, they're trying to provide superpowers to the skilled experts on the ground. And as you said, right, like the ability to see the world in like an entire 360 view or even more than that multi-dimensional, which they may not be able to see. And I could think of very simple examples like solar installer may not be aware of, you know, an electrical fault that may be going on in the house or the home. Or, you know, there might be structural weakness. Exactly. He may not be able to find right away. Right, that you can't even see visually. Exactly. Absolutely. So, you know, to tie this home, could you give an example to our audience? Like, could you share a real one example of where archetype is doing this today already? Sure. So one example that I like to bring up is in the city of Bellevue, we've actually deployed our model in order to interpret safety conditions for pedestrians. And so the ability to synthesize not just video data of what's happening in intersections, but also combine that with information about the traffic control signals means that we can actually activate traffic changes when the conditions are most safe for pedestrians. And so we've shown that this has the ability to prevent things like accidents and near misses where a car might turn into an intersection when the person is intending just to step out of the curb. And so the ability to do that not just for one intersection, but at scale means that now you can effectively produce a big increase in safety, not just in a small, vocalised area, but at scale. That's so true. And I remember you talked about another use case about injury prevention, a little bit that you alluded to, right? Like, that is such a big unlock. So it's not like, you know, you're replacing humans or, you know, taking away their task, but how do you make their day-to-day work safer? Exactly. Thank you for sharing that. Let's, you know, pivot to another point, you know, which our audience can, you know, take away from. So as we all know, building a foundational model is a more new mental task for all the engineering leaders in the audience. What's the biggest thing that you've had to unlearn? What's the, what's fundamentally different about leading a physical AI company? If I can sum it up in one word, it would be the messiness. The messiness of the data. I think in a lot of, you know, applications for SaaS companies and traditional AI, you kind of have your data structured in a nice database, you hook it up to your pipeline and suddenly things are working. And this is totally not the case when you're dealing with live sensors and data from the real world. Not only is the real world itself messy, but there are observations of that, you know, have imbued noise. They might have bias. They might be subject to changing sampling frequencies, missing data. And so all of these considerations, we believe, have to be built into the system from the start. It's not something that you can patch up afterwards once the system is built. I'm sure everybody over here can relate to the idea of embracing messiness in the real world. And it sounds like, you know, you have built a culture that treats noise and unpredictability as a core competency and not as an edge case. Exactly. You know, on a closing note too, for all the leaders out here whose products touch the physical world in logistics, in manufacturing, or even fintech like mine, what's the single most mindset shift that we need to make to prepare our teams for this new era? Personally, it would be embracing ambiguity and fast changing conditions. If we look at the state of the art today, there's basically a new advancement or a new technology every day, something new is coming out. And so I think the ability to adapt to changing conditions is key. I think not just in technology itself, but the whole attitude and mind shift of how we relate to AI. I think this idea that there's something to be learned from AI. It's not just going to be something that we teach, but there might be something fundamental in what the machine has learned from the physical world itself that might aid our own ability to understand and perceive the world. Yeah, that's a fantastic and a perfect note to end on. To keep an open mind, right, as we navigate this world of software AI, physical AI, and when the boundaries start transending. Thank you, Jimmy, for sharing your vision with us today. It's been an incredible conversation. I'm sure our audience has many more questions that they would like to ask. So I would love to open it up for some audience questions. Hi, my name is Konstantin. As from the studio said that don't teach AI physical laws. Let this cover based off could clarify this point because I think it will be not possible for Tesla to crash thousands of cars in hopes that FSD will learn how to drive. Yeah, so basically what we want the model to learn is some underlying kernels are underlying patterns that dictate how physical behaviors evolve. And so we believe that that fundamentally can be learned in a self-supervised manner, that you don't need to teach the model with human labels what is going to happen next. But through the data itself, you can design training objectives, which are similar to natural language based around the prediction of mass data or what is going to happen next. And what we found, especially in NLP, but what we're seeing with sensor data and quantitative data today is that that sort of training objective, despite the lack of explicit human labels, is sufficient for the model to learn some underlying structure. That basically the ability to predict or forecast what is going to happen based on previous data does actually encode some underlying information about the physical process itself. Yeah, kind of related to this. So in the LLM world, we have this checks and balances. I mean, you can test your outcomes with different models to basically validate the outcomes that your model is actually providing. In the physical world, do we have a parallel? How do we? Because here, we are not advocating having a human in the middle. How do we do the testing and validation of these models? Are they performing it the right way? Yeah, I think that's a really good question. And maybe one of the differentiators from LLM's is that the physical world has a defined ground truth of what actually happens. And so there is actually a basis for evaluating the accuracy of the model. It's not like a subjective language generation where is your poem good or not? It's kind of up to interpretation. But the fact that there is quantitative data and quantitative observations so that tasks like prediction, you can actually validate and measure the performance of the model is key here. And what we see is one big differentiator from LLM's today. I think the other thing that I would point out here is that I think humans innately have the ability to question and probe whether something is true when presented with the hypothesis. And so I think this natural human curiosity to actually understand something that's presented to them in combination with new tools for interpretability of models together provides some means just through our natural, you know, personalities and human nature that will drive a feedback loop of validating what the model is actually telling us. Hey Jimmy, thanks a lot. It's a very interesting topic. Training LLM's at least in the current, like, state of art is very compute and data-intensive task, right? When you move to learning like reinforcement learning or like for physical intelligence, I assume that's kind of fundamental. While the sensors can feed the data, is there any other unlock that need to happen before learning in real time can happen? Or is that not a requirement for systems such as this? Do they need to be a lot more compute at the end point for these kind of systems to work, right? Sensor data is coming in. I think that there's a lot of improvements there, but do we need the kind of compute that is required and say that is currently needed by LLM's for these kind of systems to actually work? For some architectures that might be true. What we're exploring actually is ways that we can enable are provide incremental steps toward that even with the compute we have today. And so one of the paradigms that we have implemented is disability to do basically prompting our n-shot examples, not only with natural language, but with data itself. And what that means is that as new data comes in, you can actually feed that back into the model without having to actually retrain the weights. You can actually steer the model through multimodal prompting and examples in real time and in a dynamic feedback loop without having to actually go and train and provide data for modifying the model weights. In the long term, we would like to see the disability exactly as you're talking about, maybe at the edge, to do this sort of active learning and adaptation to new data. And I think with the trends that we see today and compute becoming less expensive, you know, more commoditized, shrinking down into form factors that become feasible for deployment, like you're describing, that becomes feasible in the near future. Hi, hi, Jebi. Thanks for the talk. So I work for a building security tech company. So we make cameras, we make access control systems, different kinds of sensors. So, you know, we live and breed sensors and we have our own, you know, AI models, not surprisingly. I mean, I'm curious, your tech, I'm trying to understand exactly what it is, but like, does your model or models, does it do things like object detection, face recognition, some of the more traditional CV type things? And is it, is it, is it kind of meant to be a best in class at those things? There's a meant to more add this sort of AI capability in places where it didn't exist before and just let someone who normally, like, whenever have the means to build an AI model, like deploy it in their system? Yeah, I would say the biggest value proposition we provide here is as a general purpose model. And what that means is that you can have it do a wide variety of tasks, you know, out of the multitude that become possible very quickly and deploy a solution very rapidly because you start from a pre-trained model. And so things like, you know, object detection or traditional CV tasks are also possible, but the fact is that you don't need to start from scratch when training these models. You have something that out of the box can already perform, you know, 80, 90 percent decently and potentially with additional fine tuning make up that extra performance gap that has needed for your use case. And so the ability to basically bring up models for any number of use cases is really the amok here that you can design solutions customized for your use case and your application in a matter of days at most rather than, you know, the months that might be traditionally required. Got it. Thank you. All right. I think that concludes our Q&A. So let's give a round of applause to our speakers. If you're listening to this and you're wondering, how can I connect with other engineering leaders in my city? Pull up your phone right now and go to elc.community. Click our chapters page. You can see that on the menu on the left. Find your local chapter and click join. We're hosting virtual and in-person events all the time. And this is the best way to help you get involved. Expand your network in your city and support your leadership and career growth. So pull up your phone head to elc.community. Join your local chapter and get involved. A huge thank you to all of our local leaders who make community happen. And thank you for listening to the engineering leadership podcast.