· S0, Qasar Younis
[00:00:00] S0: How much of the drive to build Dana was building where the puck is going, and how much of it was these are customer pain points or friction points and maybe we can enumerate some of them and then we need to go ahead and build the solution for them. And this is like the next evolution of the stack that we're building. Tools, OS, autonomy stack. And now we have Dana, how much of it was outside in versus inside out? [00:00:00] S1: Both. I'd say both — we use our own tools to develop autonomy as well. So we're our own customer and those are different parts of the company. And I would say our most aggressive customer feedback comes from internally where there's very little patience for anything that doesn't work quickly and fast and on first try. We're an enterprise company and for almost a decade we've been deploying tools to customers and we get feedback from many thousands of engineers who depend on us. They're also using things like Claude. So then it's very natural if you're using Cursor and Claude and say, hey, where is the Claude cursor thing but in physical AI? And we're the company to develop that. This is, frankly speaking, our bread and butter. I think if you're in the space, it's very obvious it should be that easy, like it is to use one of the major coding platforms.
· Qasar Younis, S0
[00:00:00] S1: Both. I'd say both — we use our own tools to develop autonomy as well. So we're our own customer and those are different parts of the company. And I would say our most aggressive customer feedback comes from internally where there's very little patience for anything that doesn't work quickly and fast and on first try. We're an enterprise company and for almost a decade we've been deploying tools to customers and we get feedback from many thousands of engineers who depend on us. They're also using things like Claude. So then it's very natural if you're using Cursor and Claude and say, hey, where is the Claude cursor thing but in physical AI? And we're the company to develop that. This is, frankly speaking, our bread and butter. I think if you're in the space, it's very obvious it should be that easy, like it is to use one of the major coding platforms. [00:00:00] S0: Help us understand why Applied Intuition is uniquely positioned to go ahead and build this rather than another company, or as you said, there are other generic AI assistants or coding agents that could potentially do some of this. What's so different about Dana and why Applied Intuition specifically?
· Peter Ludwig
[00:00:00] S2: So it's important to understand a bit more about the technology itself. So the general-purpose models that come from companies like Anthropic or OpenAI, those are great and they're very useful for very general-purpose tasks. But when you're dealing with things that have a very deep safety-critical component, things where lives are literally on the line based on what is being developed, and they require a very complex development toolchain. Just the model is not enough. That's like 1% of the full solution. The Dana platform itself, this is building on top of everything that we've built over the past ten years. Everything has an API, and has been re-architected to work in a model with AI agents at the forefront. These very, very complex workflows to actually build physical AI, they maybe would require switching between 20 different tools in the past for different tasks. And deeply understanding this precise flow of information between all of those things to accomplish your end goal. So to actually make a physical AI system work, there's many, many layers to that technology stack. For example, it's like, why couldn't you use Claude to build the Linux kernel? It's like, well, because the Linux kernel is actually a very, very complex piece of technology that's been built out over many years and using many, many other different tools. The same thing in physical AI. Fundamentally, these technologies are very, very complex. And now with Dana, all of those complex workflows can be very seamlessly orchestrated from an agentic interface where you're able to write things in plain English and receive answers in plain English and do very detail-oriented things that are very deep in data science and production deployment of this type of AI.
· Qasar Younis
[00:00:00] S1: Use an example. Let's use an autonomous lawnmower. What are all the things that you need? First, you need some sensors, you need some compute, then you need to create a software package that understands this is the yard I'm going to work in, the physical space I'm going to work in and don't go in other places. So now the sensors have to understand that physical space. Then you want to create scenarios that lawnmower has to successfully pass in simulation. So you need a simulation framework. And so how do you simulate your backyard? And so that also is complex. Simulation itself is a massive industry. There are companies that are worth tens of billions of dollars that only do simulation. And that's where we started. Our bread and butter was in simulation, in the world model universe. So now that you have a simulated backyard, you have scenarios in the simulated backyard that you can run again and again, and you have to do that in the cloud. That is a whole orchestration that needs to happen. And then ultimately, once you're performing at a certain level of efficiency and fidelity, then you're going to deploy that first version onto the physical lawnmower. And then a bunch of things are not going to work and you have to figure out, well, why didn't it work? Why didn't the actuation happen as you thought it would? Why are the control systems maybe not behaving? And so then there's a whole loop of feedback. You can't do that all in an LLM. It's not made for that. LLMs are really made for a different environment. So we keep hammering things like the hardware interface or the safety criticality, but it's, frankly speaking, underemphasizing just how different it is to build a web app versus an AI product in the physical world.
· Peter Ludwig, S0
[00:00:00] S2: Also in that process, you're doing model training and evaluation, you're doing data collection and post-processing of that data. These are very complex systems, but that exact analogy, you can apply that really to any kind of machine or any kind of robot, and those same things apply. [00:00:00] S0: Your team, while we were prepping, actually described this very beautiful loop where you can take information and data that you're getting from a tractor or an underground mine, and then that's relevant to, let's say, a drone or autonomous vehicle that's also relevant to an autonomous sea vehicle, and they all feed into this broader platform and inform how the platform gets made. I would love for you to describe that loop and how that helps build a general-purpose, broader platform.
· Qasar Younis
[00:00:00] S1: An AI system really is always two big components. One is the actual platform that you develop the intelligence, and the other is the actual intelligence. So it's almost like the world model and then the actual intelligence that you'll deploy on the machine. On that second half, on the intelligence that you're deploying on the machine, the way to think about it is this data engine. It's a feedback loop. So as the car is in the real world, consumes data, as in it consumes the world around it and creates data packages, it also is not successfully navigating specific scenarios. So you can almost mark, hey, the car had difficulty in doing this. So then how do you help the brain on the car navigate that scenario that it wasn't able to navigate last time more successfully next time? Well, you can do it a couple of ways. You can expose it to lots of scenarios that humans maybe have already literally human-driven. So it imitates how humans handle that scenario. You can create a synthetic environment where you show it, this is how you would navigate this type. And there's a bunch of techniques, but the macro point is there's a data loop. It's just feedback. So the machine interacts with the scenario, it can and cannot navigate that. Now step back, don't just make that a car, make that any type of machine. It can be a drone, it can be a mining dirt mover, it can be a combine. And the same thing happens. The interesting thing we've learned in our development of intelligence and models is as we take scenarios from, let's say, a drone, or we run autonomous trucks right now, L4 trucks in Japan, we take data from those trucks, it actually makes the performance of those models in fairly different environments better. So what's really happening is the model is getting a sense of physics in the real world. This should elicit some corollaries in the chatbot universe. Chatbots used to be very, very specific. Transformers happen, and general chatbots now actually can perform really, really well. The same thing has happened in self-driving. We benefit a lot from that because we see this diversity of data in all these different use cases. And so it feeds that data loop and the data engine.