# The $15B Physical AI Company: Simulation, Autonomy OS, Neural Sim, & 1K Engineers—Applied Intuition Latent Space · 2026-04-27 Applied Intuition co-founders Peter Ludwig and Qasar Younis argue that the real bottleneck in physical AI is deploying intelligence onto constrained hardware, not model intelligence itself. Their $15B company builds simulation, operating systems, and AI models for autonomous trucks, mining equipment, and defense systems. Starting as YC-era tooling for robotaxis, they now offer 30+ products across simulation & RL infrastructure, vehicle operating systems, and autonomy models. They compare fragmented vehicle software to pre-Android phones, and their OS enables reliable updates and L4 driverless operations (trucks running in Japan today). Verification uses statistical nines of reliability, and they internally adopt coding agents like Cursor and Claude Code. They hire 1,000 engineers at the hardware-software boundary. ## Questions this episode answers ### Why does Applied Intuition compare the physical machine industry to the pre-Android phone market? Peter Ludwig explains that before Android and iOS, phone makers used around 50 different operating systems, making it nearly impossible for Google to run its apps across all devices. Similarly, today’s vehicles and industrial machines have highly fragmented operating systems. Applied Intuition is building a consolidated OS to solve this fragmentation and enable reliable AI application deployment, much like Android did for phones. [0:00](https://addtry.com/fe59c3ec-a7c6-4efd-8cd9-c274905636b3?t=0) ### What is the biggest constraint in deploying AI onto physical machines? Qasar Younis states the primary bottleneck is not model intelligence, but deploying AI onto constrained hardware. Physical AI systems must run on embedded chips with limited compute, power, and strict latency requirements. Peter Ludwig adds that models need to be extremely efficient and small, as every millisecond matters for real-time control. Safety-critical demands make this much harder than screen-based AI. [47:02](https://addtry.com/fe59c3ec-a7c6-4efd-8cd9-c274905636b3?t=2822000) ### How does Applied Intuition close the sim-to-real gap for autonomous systems? Peter Ludwig details an iterative validation flow: they feed real-world data back into simulation parameters until the simulator accurately represents reality. This process is essential—skipping it leads to unreliable results. Qasar Younis notes they balance cost and compute, using world models where effective but never abandoning real-world testing. An example is modeling actuator overheating in humanoid robots to improve reinforcement learning policies. [37:33](https://addtry.com/fe59c3ec-a7c6-4efd-8cd9-c274905636b3?t=2253000) ## Key moments - **[0:00] Opening** - [0:00] Physical machines today have as many operating systems as phones before Android, causing fragmentation — Peter Ludwig - **[2:56] Applied Intuition** - [2:56] Applied Intuition's mission is to 'build physical AI for a safer, more prosperous world' across cars, trucks, construction, mining, agriculture, and defense - [4:24] Applied Intuition runs driverless L4 trucks in Japan right now, Qasar Younis reveals - [5:33] Starting a tooling company in 2016–2017 was out of vogue: VCs saw tooling as 'just workflows' and not interesting, Qasar Younis recalls - **[6:52] Eng Culture** - [7:46] Applied's technology stack completely changes roughly every two years; they've done four full evolutions since founding — Peter Ludwig - [8:43] 83% of the company is engineering, employing over 1,000 engineers and 40‑plus founders — Qasar Younis - **[9:40] Tech Stack** - [10:02] Applied Intuition's tech falls into three buckets: simulation & RL infrastructure, vehicle operating systems, and fundamental AI models for autonomy - [10:57] Peter Ludwig says they built their own OS because existing options weren't good enough; now it's a real business enabling reliable updates - [11:46] 'In order to write and run great AI, you need a great operating system,' states Peter Ludwig - [12:24] Human‑machine interaction is shifting from physical buttons to voice and cabin awareness, with machines teaming with humans — Qasar Younis - [14:36] LiDAR is 'hands down a useful sensor' for AV R&D, providing per‑pixel depth for model training; Tesla still uses LiDAR on its R&D vehicles today — Peter Ludwig - [16:32] A vehicle OS goes deep: real‑time motor control, sensor streaming, latency, memory management, and fail‑safes for cosmic ray bit flips — Qasar Younis - **[19:28] AI Tooling** - [20:01] Peter Ludwig draws the Android analogy: a consolidated OS is prerequisite for modern AI apps on vehicles, just as Android unified phones for Google services - [22:36] Q: Does Applied Intuition lock customers into its full stack? A: No, it's an open platform; customers can mix and match OS, autonomy, and tools - [23:13] Low‑level vehicle software uses C, C++, and occasionally assembly, with Rust emerging as a new option — Peter Ludwig - [23:41] Internally, Cursor was the hottest coding tool, but 'Claude Code has taken the reign'; Applied runs an internal leaderboard to encourage adoption — Peter Ludwig - [24:35] Applied's Sensor Studio now lets AI agents configure sensor suites via text, likely reaching better results than the traditional GUI — Peter Ludwig - [25:44] AI is creating a bimodal distribution of engineers: those embracing AI tools have an enormous productivity gap over those who haven't — Peter Ludwig - **[26:08] Safety & Sim** - [27:40] Using Claude for GPU shaders a year ago would be underwhelming; now the latest model is 'blown away' effective, even in embedded domains — Peter Ludwig - [28:13] 'You're not gonna trust your life to AI‑written software that's not been very carefully checked by humans,' asserts Peter Ludwig on safety‑critical code - [28:28] Verifiability is the hardest problem right now: as models get better, finding their faults gets harder, but proper evaluation remains crucial — Peter Ludwig - [29:41] End‑to‑end autonomy demands 'Neural Sim' — a hybrid of Gaussian splatting and diffusion to generate sensor data for RL training — Peter Ludwig - [31:06] Validation has shifted from binary Euro NCAP tests to statistical measures: how many nines of reliability can the system achieve? — Peter Ludwig - [32:51] Applied educates governments on validation: 'We're more teaching the government,' says Peter Ludwig, because they're not AI labs - **[43:47] Embedded AI** - [46:49] In physical AI, 'we're not constrained by model intelligence — it's deploying them on constrained hardware,' says Qasar Younis - [48:57] Google's Gemma 2B can run on an embedded system, but it needs heavy customization for autonomy tasks — Peter Ludwig - [50:27] Old‑world mining autonomy used hand‑coded RTK GPS; now end‑to‑end systems bring perception and dynamic decisions to off‑road environments — Qasar Younis - [52:01] Applied uses a diversified bet strategy across autonomy approaches because no single method guarantees success across all industries — Peter Ludwig - [53:57] 'Everything can be boiled down to a next token prediction problem,' says Peter Ludwig, on applying autoregressive models to physical AI - **[55:13] Production** - [55:28] China is organizing a humanoid robot marathon to push reliability, akin to the DARPA Grand Challenge — Qasar Younis - [56:26] Peter Ludwig predicts humanoid robots will run a full marathon 'any day now' - [57:25] After a decade, Peter Ludwig can look at any company's demo and list 'exactly the next 20 problems they're gonna hit' - **[58:16] Startup Advice** - [58:57] Qasar Younis's startup advice: impose a commercial constraint early to focus and survive long enough for compounding technology to pay off - [1:01:16] Qasar warns: don't apply mature‑company strategies like full vertical integration to a nascent startup — Apple in 1982 was very different from Apple in 2007 - [1:04:14] Qasar Younis, former YC COO, says 'YC advice from 2014 just would not apply in 2026' because the market and AI ecosystem have fundamentally changed - [1:06:11] Peter Ludwig's wishlist: efficient model distillation for embedded hardware, and robust model evaluation for safety‑critical systems - **[1:07:27] Hiring** - [1:07:27] Applied hires engineers who appreciate the hardware‑software boundary and understand low‑level systems, not just superficial technology — Qasar Younis - [1:09:25] General Motors Institute was founded a century ago to solve an engineer shortage; Applied upskills internally with extensive training — Qasar Younis - [1:12:17] The essential trait is an 'engineering mindset' — wanting to understand lower levels, even down to 'what is light? what is a radio wave?' — Qasar Younis ## Speakers - **Alessio** (host) - **Swyx** (host) - **Peter Ludwig** (guest) - **Qasar Younis** (guest) ## Topics Robotics, Reinforcement Learning ## Mentioned Applied Intuition (company), Cruise (company), Google (company), NVIDIA (company), OpenAI (company), Scale AI (company), Tesla (company), Waymo (company), YC (company), Android (product), Claude Code (product), Cursor (product), Gemma (product), Sensor Studio (product) ## Transcript ### Opening **Peter Ludwig** [0:00] Physical machines today are more akin to the state of the phone market before Android and iOS existed. Part of the, the reason that, that Larry at Google decided to get into Android was they, they wanted to run Google products on a bunch of phones, and they, they bought all of these phones from the industry, and it turned out they had, like, 50 different operating systems on these phones. And it was virtually impossible- -for, for Google to, to make their app run on all 50 devices equally well. And so the solution was, well, actually what if, what if they created a, a really great operating system and, and made it attractive to all of these phone makers? And that was sort of the genesis for what Android was and, and why Android existed. It was a way for Google to get their products onto really wide diversity of devices. The state of the, of the physical, uh, industry right now, it's a little bit like that. So many different operating systems, it's so fragmented, and to actually get a modern AI application to run on these vehicles, y-you actually-- you first have to consolidate the operating system, and so that's, that's why we've done that. **Swyx** [1:03] Before we get into today's episode, I just have a small message for listeners. Thank you. We would not be able to bring you the AI engineering, science, and entertainment content that you so clearly want if you didn't choose to also click in and tune into our content. We've been approached by sponsors on an almost daily basis, but fortunately, enough of you actually subscribe to us to keep all this sustainable without ads, and we wanna keep it that way. But I just have one favor to ask all of you. The single most powerful, completely free thing you can do is to click that subscribe button. It's the only thing I'll ever ask of you, and it means absolutely everything to me and my team that works so hard to bring Latent Space to you each and every week. If you do it, I promise you we'll never stop working to make this show even better. Now let's get into it. **Alessio** [1:52] Hey everyone, welcome to the Latent Space Podcast. This is Alessio, founder of Kernel Labs, and I'm joined by Swyx, editor of Latent Space. **Swyx** [1:58] We're very honored to have, uh, the founders of Applied Intuition, uh, Qasar and Peter. Welcome. **Qasar Younis** [2:03] You guys really know how to turn it on to podcast mode. That was, that was, uh... You guys are real, real pros at this. They were just joking around right before this, and then they flipped it pretty quick. **Alessio** [2:15] Oh, yeah, it's good to have you guys. Maybe you just wanna introduce yourself so people know the voice on the mic and they'll know who they're hearing. **Peter Ludwig** [2:20] Oh, sure. Yeah, I'm, I'm Peter Ludwig. I'm the co-founder and CTO of Applied Intuition. **Qasar Younis** [2:24] And, uh, my name is Qasar Younis. I am the, uh, CEO and co-founder with Peter. **Alessio** [2:28] Nice. Can you guys give the high-level overview of what Applied Intuition is? And I, I was reading through some of the Congress files, uh, when you went out there, Peter, and 18 of the top 20 global non-Chinese automakers, you two guys, you have customers in agriculture, defense, construction. I think most people have heard of Applied Intuition tied to YC when it was first started, and then you were kinda in stealth for a long time. So maybe just give people the high-level overview of what it is today, and then we'll dive into the different pieces. ### Applied Intuition **Peter Ludwig** [2:56] Yeah. So at Applied Intuition, our mission is to build physical AI for a safer, more prosperous world. And so we work on physical AI for all different types of moving systems, everything from cars to trucks to construction and mining equipment, uh, to defense technologies. And, uh, and we're a true technology company, so we, we build and sell the technology, and we sell it to the companies that make the machines. We, we sell it to, to the government, uh, really anyone that wants to, to buy a technology to make machines smart. **Qasar Younis** [3:24] Yeah, and I think, uh, in the broader AI landscape, a lot of the focus, uh, rightfully so in the last, uh, three years, has been on large language models and so everything that fits in a screen. You know, like, uh, whether it's code complete products or, or, or things like that. Um, and the-- what's different about us is we're deploying intelligence onto a lot of things that don't have screens. You know, they're physical machines. There are sometimes screens within the cabin or for, for example, of a car or a truck or something like that, but, uh, most of the value we provide is putting intelligence that is in safety critical environments. So that... The, the, those two words are really important because learned systems can make mistakes if you're asking for, like, you know, some, you know, so-something like, "Tell me about these podcast hosts- **Peter Ludwig** [4:14] Yeah **Qasar Younis** [4:14] ... that I'm about to go meet." But, uh, you can't do that obviously when you're... You know, we run, like, as an example, we run driverless trucks in Japan right now, like, m-m-as we speak. You can't have errors. That, those are L4 trucks. Yeah. **Alessio** [4:26] Yeah. Was that always the mission? I remember initially, I think people put you and Scale AI very similarly for some things about being kinda, like, on the data infrastructure side of things. What was the evolution of the company? **Peter Ludwig** [4:38] Well, from the very beginning, we, we always wanted to, uh, really be a technology company that, uh, that helped generally push forward the industrial sector. And so we started off working in autonomy. Our, our very first customers were robotaxi companies. And, and we started off doing a lot of work in simulation and, and data infrastructure. And then over the years, we've expanded our portfolios. Now we have, uh, over 30 products, and it's a, a pretty broad technology play within the, the landscape of physical AI. **Qasar Younis** [5:05] Yeah, I think the, the Scale reason is because we're all YC universe companies, you know, and so, uh... But it was a very, very different company. You know, Scale, uh, was, is more of a services company, data labeling company fundamentally. We started, and still are, uh, you know, do a lot of tooling. So, like, you think, you know, developer tooling is now in vogue again, thanks to, thanks to, you know, thanks to the AI boom. But honestly, 10 years ago it was out of vogue. It w- w- w- like, doing a tooling company in 2016, 2017 was not, like, the thing to do because I don't know if you remember, the, the VCs generally, their views was, are the toolings are, they're just, they're just workflows, and workflows ultimately are not really interesting. Uh, and we've gone and come, you know, full, full circle with that. But when we started the company, our, our kind of l- you know, it's kinda like in the periphery of what the company wants to be. It was like, uh, from our earliest days, like, we wanna deploy software on Physical machines, like on cars and on trucks and things like that. And obviously, we didn't know that the transformer boom was gonna happen. We didn't know that autonomy systems would become end-to-end. Those things we didn't know. And why that's important with autonomy systems becoming end-to-end, it is just now you can-- those models can be generalized to, you know, multiple form factors. And so back nine, 10 years ago, tooling was a great way, and still is a great way to, you know, build the technology and sell technology to our end customers, a lot of them who wanna build the stuff themselves. And so we just offer, like, a spectrum of solutions from you can just use, like, one part of, of a development suite of tools all the way to buying the full thing. The way to think about the company, or at least the way we think about the company, is, as Peter said, a technology provider. It's kinda like, uh, you know, what NVIDIA does or what an AMD, but we just don't do chips. ### Eng Culture **Peter Ludwig** [6:52] Mm-hmm. **Qasar Younis** [6:52] We don't do silicon. But we're a technology provider fundamentally. And I think even, you know, we used to joke when we started the company, like, you know, we're not the guys to build, like, Instagram. Like, that was just-- or this is not our-- or that's just not us in a mo-most fundamental way. I mean, I, uh- **Peter Ludwig** [7:06] You have thoughts. **Qasar Younis** [7:07] Yeah, yeah. Well, it's, it's-- I mean, I think it's just like what... And, and, and I mean, we worked on Maps and stuff, Google Maps. Consumer products are extremely difficult for a lot of different reasons. It just, I think, doesn't scratch the itch. I think we're like Michigan guys who are kind of more of that traditional engineering kind of a realm, uh, or lineage. Uh, we used to joke- **Peter Ludwig** [7:27] I, I, I gotta say, though, what, what was clear 10 years ago was that there was so much more that was possible with software and AI in vehicles. **Qasar Younis** [7:34] Mm-hmm. **Peter Ludwig** [7:34] And that was generally the space that we started in 10 years ago. **Qasar Younis** [7:37] Yeah. **Peter Ludwig** [7:37] And the precise path that we've taken over the years, I think we've been strategic, and we, we've adjusted to make sure that we're actually building stuff that's valuable to the market. And, like, the technology has changed so much. Like, our, our own technology stack has, has completely changed, I would say, roughly every two years. And so now we've probably done, let's say, four complete evolutions of our own technology stack, and I sort of see that cadence roughly keeping up. **Qasar Younis** [7:59] Mm-hmm. **Peter Ludwig** [7:59] And, and so the way even we think about engineering is almost on this two-year horizon, we're preparing ourselves that, hey, like, we wanna invest the appropriate amount, but then also be very dynamic as the research gets published and as our research team figures out new advancements and adapting to that. **Qasar Younis** [8:13] Yeah. One thing that has been consistent is the type of people we've, we've recruited, frankly speaking. It's engineers who are-- fall into the, you know, sometimes very traditional, like, you know, Google- **Peter Ludwig** [8:24] Mm-hmm **Qasar Younis** [8:24] ... Gen suite, um, but way different from, you know, other companies. We are hiring folks who really know the intersection of hardware and software, who know really low-level systems. Obviously, traditional ML re-researchers and folks who've, uh, actually, you know, put ML systems into production. That's been pretty consistent. I, I think that, like, you look at the mix of our engineering, 83% of the company is engineering, so it's, like, a giant lease. **Peter Ludwig** [8:51] Mm-hmm. **Qasar Younis** [8:51] A lot of engineers. **Peter Ludwig** [8:52] Which, by the way, 1,000 engineers- **Qasar Younis** [8:54] Yeah, we're 1,000 engineers **Peter Ludwig** [8:54] ... at least. I mean, that's on your website, so I imagine it's all a date. **Qasar Younis** [8:57] Yeah. It is, it is up to date. Yes, yes. **Peter Ludwig** [8:59] Uh, okay. And then 40-plus founders. **Qasar Younis** [9:01] Yeah. We would tend also, uh, this was more luck than, than strategy. Um, we, we've recruited a lot of ex-founders. It's been a great place for founders, YC and non, 'cause obviously I know a lot of the YC folks. It's kinda like we recruit a lot of Google people- **Peter Ludwig** [9:19] Mm **Qasar Younis** [9:19] ... for, for them to exercise both their technical and non-technical skills because, you know, we're, we're, we're on the applied side. We have a research team that we do fundamental research, we publish, and we've, we've had great traction there. But fundamentally, the business wants to take this intelligence and deploy it into production and, and there's, like, a certain type of person that's more interested in that. ### Tech Stack **Peter Ludwig** [9:40] Mm-hmm. Yeah. You mentioned the tech stack, Peter. Uh, so I just wanted to give you some reign to just go into it. I'm interested in where Applied Intuition, uh, starts and ends. In, in, in, in some sense, what won't you do? What, uh, do you do that's common among all the verticals that you cover? There's a few buckets of, of work that we do, and, and we've been at this for almost 10 years now, so the technology's pretty broad. But, uh, we got started- **Qasar Younis** [10:03] Yeah, with 1,000 engineers, like, you could work on lots of stuff. **Peter Ludwig** [10:05] There's lots of stuff, yeah, espe-especially with AI tools now. **Qasar Younis** [10:08] Yeah. **Peter Ludwig** [10:08] So we got our start in, in simulation and, uh, simulation tooling and infrastructure. And so generally, if you're trying to build a very complex software system that involves moving machines, you need to test that, and the best way to test it is it's a combination of virtual developments, a simulation, and then also obviously real-world testing. **Qasar Younis** [10:25] Mm-hmm. **Peter Ludwig** [10:25] And then there's a very careful process of that correlation between the simulation results and the real-world results and ensuring that the simulator is in fact accurate to that. Simulation's a, a very deep topic. **Qasar Younis** [10:35] Mm-hmm. **Peter Ludwig** [10:35] We have a whole, whole suite of products in that, and we can talk for many, many hours about that specifically. Um, but that, that is one part of what we do as a company. Reinforcement learning as a sub-part of that is also super critical. I think a lot of the, a lot of the best advancements happening in, in a lot of these AI systems right now in some way relate to reinforcement learning, and with now we have lots of compute, and you can do tons of interesting things for reinforcement learning. The second bucket of work that we do is on operating systems technology, a true operating systems. Like, uh, uh, think, think about, uh, schedulers and, and memory management and, uh, and middleware and message passing and, uh, highly reliable networking and data links. Like, the reality is, if you wanna deploy AI onto vehicles, you need a really good operating system. And when, when we were getting deeper into that space, there wasn't really anything that we were happy with. **Qasar Younis** [11:25] Mm-hmm. **Peter Ludwig** [11:26] Like, things existed, absolutely, and we were using what was available in the market. And a-as an engineering organization, we roughly realized these things aren't great. We think we can do this better, and so let's, let's build something. And that was then the, that was the, the moment of inspiration that started our operating systems business, which is now a very real business for us. And in order to write and run great AI, you need a great operating system, and so that-that's what got us into that. And then the third bucket that we work on, it's, it's true fundamental AI technology. Models, we do a lot of work in, Qasar mentioned, the foundational research, but then the-- also the, uh, the world models and the actual autonomy models that are running on, uh, on these physical machines, as that's across cars, trucks, mining, construction, agriculture, and, and defense. And, uh, and so that's both land, uh, air, and, and sea. **Qasar Younis** [12:17] And also, uh, a, a smaller subsector of that third bucket is the interaction of humans with those machines. **Alessio** [12:24] Mm-hmm. **Qasar Younis** [12:24] So, uh, that's a multimodal, uh, experience. Historically, if you're moving a dirt mover or any of these machines, there are like, you know, buttons you press- **Alessio** [12:34] Mm-hmm **Qasar Younis** [12:34] ... whether they're actual physical tactile buttons or, or something like a touch screen. That's just, that fundamentally is changing to where you're just talking to the machine and the machine... And you're teaming with the machine. **Alessio** [12:44] Voice? **Qasar Younis** [12:45] Yeah, voice, absolutely, yeah. **Alessio** [12:47] Oh. **Qasar Younis** [12:47] And al- also the machine just being aware of who's in the cabin, what their state is. Um, you can think from a safety systems perspective, the most simple version of this is, like, the driver is tired, right? **Alessio** [12:58] Mm-hmm. **Qasar Younis** [12:58] And they're, they're... You know, if you get those alerts when you're driving your car- **Alessio** [13:01] Yeah **Qasar Younis** [13:01] ... it says, "Maybe take a coffee break." That, take that times, you know, a couple of order of magnitudes up. Uh, but this concept of teaming man and machine is important. When you think about running agents or just running, you know, different instances of, uh, you know, Claude and doing work for you in the background, you, you can take that analogy out, almost copy and paste and put it into, like, a farm, where you have a farmer who's running a number of machines. So where they interact with the machine is where there's maybe a critical decision or a disengagement or something like that. But generally speaking, the agent on the physical machine is running and making decisions on the behalf of the farmer until there's something maybe, you know, critical, and that, um, that's also what we work on. So that, that's not pure autonomy. It's a little bit of a mix, but it falls under, uh, autonomy. In the automotive sense, that's typically defined in SAE levels as an L2++ system- **Alessio** [13:51] Mm-hmm **Qasar Younis** [13:52] ... with a human in the loop, but just take that idea, you know, to, to other verticals. **Alessio** [13:55] Yeah. You've not mentioned hardware at all, like sensors or, you know, obviously we, you mentioned you don't do chips. I think even in AV there's, like, a big, uh, you know, cameras versus LiDARs. Like, what, what are, like, in your space, maybe some of those design decisions that you made, and are they driven by the OEM's ability to put things on the machinery? Like, how much influence do you guys have on co-designing those? **Qasar Younis** [14:18] Yeah. So, so we don't make sensors. Like, we're, we're not a manufacturer. Obviously, we use a lot of sensors in our autonomy products. Um, in terms of what actually goes on the vehicles, we have a preferred set of sensors that, that we, uh, let's say, fully support, and then our customers, they can sort of choose from those. And obviously if, if there's a, a very strong opinion on supporting something else, we will add that to the platform as well. And the, the LiDAR question is at this point sort of the age-old, uh- **Alessio** [14:45] Right **Qasar Younis** [14:45] ... topic in, in autonomy, and the, the state of the industry right now is LiDAR is hands down a useful sensor, uh, specifically for data collection and the R&D phase of autonomy development. Um, uh, if you see, for example, a, a Tesla R&D vehicle, it actually has LiDAR on it- **Alessio** [15:03] Mm-hmm **Qasar Younis** [15:04] ... uh, to this day, right? In, in the Bay Area, we see these, uh, you'll see, like, uh, Model Ys or Cybercab that have- **Alessio** [15:09] Mm-hmm **Qasar Younis** [15:09] ... uh, have LiDARs on them just driving around. So it's, it's useful because it gives you per pixel depth information, so if you can pair a LiDAR with a camera, and you can say that, well, this camera's looking this direction, this LiDAR is looking this direction, and now for each, each pixel of the camera, I can see how far away is that pixel. Um, you can actually then use that as a part of your model training, and then the, that depth information then becomes- **Alessio** [15:31] Mm-hmm **Qasar Younis** [15:31] ... a learned, uh, a learned state of the camera data. And then when, when you're doing the production system, you can now remove the LiDAR- **Alessio** [15:38] Right **Qasar Younis** [15:38] ... and, and now you can actually get depth with just the camera. And so that, that difference between, like, a highly sensored R&D vehicle and then the down-costed production vehicle, we use that across our whole portfolio of products. Um, and of course, the end goal is you want super low cost and super reliable. **Alessio** [15:55] Right. **Qasar Younis** [15:55] And, uh, and then in certain use cases, you have some more, uh, bespoke things, like in, in defense as an example, you do things at night oftentimes, and so you care about sensors like infrared, uh, more so than... And you don't, you don't wanna be putting energy out, so you don't wanna use LiDAR or radar. **Alessio** [16:09] Mm-hmm. **Qasar Younis** [16:09] But you still need to be able to see at nighttime. So yeah, we, we work with the whole gamut. **Alessio** [16:13] Cool. So that's kinda, like, on the hardware level. Then on the OS level, how does that look like? What is, like, unique? I mean, my drive-- I drive a Tesla. Whenever I drive some other car that has a screen, it always sucks. It's some, like, cheap Android tablet. It's like, it's laggy and all of that. What does the OS of, like, the autonomy future look like? **Qasar Younis** [16:32] When most people, uh... It's really like what you just described. When you think about operating system in a vehicle, you're thinking about the HMI, right? The human machine interface, and absolutely that's a, an important part of it, but that's actually only one thin layer on top. Um, so when we talk about operating systems for AI in vehicles, there's many, many layers that go deep into the safety critical realm and embedded systems, and you're talking about the, the real-time control of- **Alessio** [16:59] Mm-hmm **Qasar Younis** [16:59] ... let's say the electric motors or the, the engine and the actuators, and you have different redundancies for, uh, for different, let's say the steering actuation in, in the vehicle. And all of these things, uh, need very core support in the, in the operating system. And then, of course, for autonomy, you have real-time sensor data that's streaming in, and the latencies there are really important, right? If you try to... Imagine trying to run Microsoft Windows- **Alessio** [17:21] Right **Qasar Younis** [17:21] ... uh, like streaming your sensor data in or controlling the vehicle. Like the latencies are gonna be absurd. Like you can never do that. And so what's special about what we do is we really have this system level thinking, right? So we're looking at, we care about every performance characteristics of the entire system, and then we also, because we're doing a lot of the software or all, all of that software, we can fine-tune and control all of those things. So we can very, very carefully tune in the latencies for every aspect of the system. We can carefully tune in the memory management. We can have the right, uh, fail-safes and fallbacks, uh, for, for different things. 'Cause you have to account for what if, what if there is a critical failure? What if there's a cosmic ray that flips- **Alessio** [18:00] Right **Qasar Younis** [18:00] ... a bit in the middle of the processor that causes some, uh, malfunction? Uh, and you have to have a fail-safe to all of that. And so the, the core operating system is a part of that. And then the, um, the one last thing, which is a lot less exciting, but is, uh, actually a very big topic, is reliability of updates. **Alessio** [18:16] Right. **Qasar Younis** [18:16] Uh, so the... I have a, a Tesla and, uh, you get updates fairly frequently, right? **Alessio** [18:22] Mm-hmm. **Qasar Younis** [18:22] Once a month. **Peter Ludwig** [18:23] Most companies that are making vehicles- **Swyx** [18:26] Right **Peter Ludwig** [18:26] ... are basically never doing updates. And they're-- And even if they are doing updates, they're usually only updating maybe one module. Maybe they're updating the, the HMI module. But they're not able to update, let's say, the, the safety-critical parts of the system. **Swyx** [18:37] Mm-hmm. **Peter Ludwig** [18:37] You have to go into the dealer for that. And so with our operating system now, we can actually enable highly reliable updates of any system in the vehicle, and that's way easier said than done. Like, there, there's lots of technical-technically deep stuff, uh, in the tech stack to, to do that in a way that you're not going to accidentally brick a vehicle. **Swyx** [18:54] Mm-hmm. **Peter Ludwig** [18:55] And right, if, imagine you're- **Swyx** [18:56] That would be bad. **Peter Ludwig** [18:57] Yeah. But bricking a car is a very expensive- **Swyx** [18:59] Yeah **Peter Ludwig** [18:59] ... uh, and, and ho-honestly, it, it, across the industry, maybe one of the most just pure impactful things that we've done is we've just-- we're, we're now enabling the industry to actually do software updates. **Swyx** [19:08] Just to clarify as well, who is the customer for this? Like, uh, I assume a lot of hardware manufacturers have their own firmware, and I'm sure some of them would just have you write it for them because you're experts, uh, and others would have their own. Like, who pays for this? Who, who invites you into the, the house? Is it, is it the end user, or is it, is it the manufacturer? **Peter Ludwig** [19:28] Yeah, yeah. So let me make an analogy firstly on the, on the fragmentation of software. So physical machines today are more akin to the state of the phone market before Android and iOS existed, right? So, uh, I, I worked on Android at Google, by the way, m-many, many years ago, and, uh, and part of the, the reason that, that Larry at Google decided to get into Android was they, they wanted to run Google products on a bunch of phones, and they, they bought all of these phones from the industry, and it turned out they had, like, 50 different operating systems on these phones. ### AI Tooling **Peter Ludwig** [20:01] And it was virtually impossible- ... for, for Google to, to make their app run on all 50 devices equally well. And so the solution was, well, actually, what if, what if they created- **Swyx** [20:12] Mm **Peter Ludwig** [20:12] ... a, a really great operating system and, and made it attractive to all of these phone makers? And that was sort of the genesis for what Android was and, and why Android existed. It was a way for Google to get their products onto really wide diversity of devices. The state of the, of the physical, uh, industry right now, it's a little bit like that. Like, there's y-yes, these companies have firmware, but they, they have so many different operating systems. It's so fragmented. And to actually get a modern AI application to run on these vehicles, you actually, you first have to consolidate the operating system, and so that's, that's why we've done that. And then, uh, your specific question was, who are our customers? It's, it's, generally it's the companies that are making these machines. And yeah, we're, we're, we're selling our technology to them to really simplify the architecture and then enable these AI applications to run on them. **Swyx** [20:59] How much is reusable a-across? Like, do you have, like, one OS that is just configured for everything, or is there some com- more customization that nee- it's needed? **Peter Ludwig** [21:08] Yeah, highly reusable. So the, um, the, the fundamental technology is, is quite universal, right? So things that we do have to think about, though, are like chipset support. And so, um, i-if you're, if you're coding, uh, let's say, uh, an LLM, and you have a, start with an assumption that, "Hey, oh, I'm gonna, I'm gonna use CUDA, and I'm gonna run this, uh, on an NVIDIA chip," then you don't really have to think about the hardware in that sense. **Swyx** [21:31] Mm-hmm. **Peter Ludwig** [21:31] Like, you, you're just, "Okay, I'm just, I'm in the CUDA/NVIDIA ecosystem, and I'm, I'm going to, to use that." But the, the hardware, especially in safety-critical systems, it, it's a lot more diverse. There, there's not one or, one or two players. There's a, a bunch of different chipsets that we have to support, and so our operating system doesn't just run on, like, the equivalent of x86. It has to, it has to run on a, a number of different architectures from chips from a, a bunch of different companies. Uh, but again, we've been working on this for a long time now, so we have, we have support for all of those, those chipsets. And then when you wa-want to then run the AI applications, we can then do that reliably across now a variety of providers. **Qasar Younis** [22:05] And I think that is, like, heavily inspired by Android, right? Android has a huge suite of testing, and, uh, and it's a reliable operating system that runs on thousands of devices. And, you know, we think we can, we can do the same in, in all these physical moving machines, with the difference that we're really in a safety-critical realm. Android isn't. **Swyx** [22:26] So on Android, I don't need to use Gmail. I can use Superhuman. Like, what about your machinery? Like, can people bring somebody else's automation to it, or is it kind of like all-in-one? **Qasar Younis** [22:36] You have to use us. No. Yeah. I mean, we're-- If, uh-- Yeah. Yeah, it's totally open. Yeah. **Peter Ludwig** [22:43] Yeah, yeah. Our, our, our philosophy is that we are a technology company, and so we, we license our technology to customers to use how they want. And so if a customer wants to-- If they want to license our autonomy tech and our operating system, then great, we'll license those. If they just want to license the operating system and then use different autonomy tech, that's fine also, and we have great documentation- **Qasar Younis** [23:04] Or if they just wanna use developer tooling. **Peter Ludwig** [23:05] Yeah, exactly. Yeah. **Swyx** [23:06] Yeah. It's like a better together if, uh, obviously, uh, if you, if they work together. Is it all C++, I assume? Is there different compile targets? **Peter Ludwig** [23:13] We use a lot of C++. **Swyx** [23:14] Yeah. **Peter Ludwig** [23:15] Uh, I mean, uh, Rust is, is sort of a hot, the new hot kit on the block- **Swyx** [23:18] Yeah **Peter Ludwig** [23:18] ... for a, for a bunch of things as well. Uh, but yeah, when, when you-- The, the lower level you get, especially when you, when you get to real-time constraints, you hit C, C++ at some point, and at some point, maybe you work your, work your way into assembly when needed. **Swyx** [23:30] Oh, damn. I'm curious about the, um, coding agent adoption, just, like, since you're mentioning more esoteric languages. Like, what's the adoption internally? What have you learned? **Peter Ludwig** [23:41] Yeah, we, we use everything. Um, I mean, so Cursor was, I think, the, the hottest tool in the company for a good while. Now Claude Code, I think, has, has taken the, the reign on that. We have a internal leader, leaderboard that we use just to sort of encourage adoption- **Swyx** [23:55] Mm-hmm **Peter Ludwig** [23:55] ... uh, with-within the company. And, uh, yeah, it's, they're phenomenally useful. I, I mean, it's, uh... Honestly, we, we take inspiration from, from some of those tools also in how we're adapting some of that mindset of thinking to the physical realm. Like, oh, if it's so easy to, to build an app for this or that thing that lives just on a screen, we can-- We, we're taking now a lot of the same ideas and, and applying that to, okay, well, if you wanted a physical machine to do something, how easy can we make that, uh, using our own tooling and platform as well? **Alessio** [24:26] Are you changing any of like the OS architecture, kinda like the way you expose services to like be more AI friendly or- **Peter Ludwig** [24:35] Yeah, absolutely. The, uh, in the early days of our tools infrastructure work, it was a lot about, um... You had engineers that were experts in certain topics, but the things that you're dealing with, they're oftentimes more mathematical or more abstract, where actually gooey tools are very, very useful for certain things. Like, a- as an example, we have a, a product we call Sensor Studio, which is, uh, it helps you design the sensor suite for, uh, for your autonomous vehicle. Whether, again, it could be a car, it could be a, a drone, could be a mining, mining equipment, could be a robot. And, and you, you place sensors in different places. You- there's different-- uh, there's a library. You can understand what are the trade-offs that you're making in, in the design of, of that system. And that was like a very, a very gooey, intensive, uh, thing 'cause it's a little more like a CAD tool in that sense- **Alessio** [25:23] Mm-hmm. Yep **Peter Ludwig** [25:24] ... if you've seen CAD tools. Nowadays, though, uh, right, we, we expose all of the underlying APIs for that and, and now using, uh, AI agents, you can actually configure a sensor suite with just text and, and likely reach a better result than you could have through the gooey in the past, and we're taking that thinking now through the whole product portfolio. **Swyx** [25:44] Another thing I was thinking about is just in terms of like AI, uh, adoption, does it change your hiring a-a-at least a little bit, or how do you, how do you sort of manage engineers, um, differently? **Peter Ludwig** [25:54] Yeah, a-a-absolutely, it does. Um, we, I think like every company in, in the Valley right now, are evolving our, our hiring practices- **Swyx** [26:02] Yeah **Peter Ludwig** [26:02] ... um, because the, the skills required to be effective are changing so fast, right? I mean, you, you used to really select for just rote implementation ability and, and now it, it is more the AI engineer skill set, right? ### Safety & Sim **Peter Ludwig** [26:16] Where it's like, yeah, you know how to implement, but actually just banging out code is, is no longer the core job, right? It's, it's actually knowing what questions to ask, knowing how to tie, how to tie together these different AI tools. And so the, the interviews that we give now, I think, are way harder than they've ever been. But, but we also allow, right, selective use of AI tools to solve the problems. And I think in that, you, you start to see more of a bimodal distribution of engineers, right? You, you start to see like, wow, there's, there's this subset of, of people that they, they really get it. Like they're, they're all in, and they've, they've clearly invested the, the hours needed to learn these tools and, and how to be effective. **Swyx** [26:55] Mm-hmm. **Peter Ludwig** [26:55] And then there's sort of the, the group of people that haven't done that, and that the productivity gap is just enormous, and so we're, we're trying to obviously select for the people that are, are really, really into this. **Swyx** [27:06] I first wrote the, my AI engineer piece three years ago, and when I first wrote about it, I was like, "Actually, not everyone should be an AI engineer," 'cause I think there's a, there's an extremist stance where, well, every software is an AI... engineer is an AI engineer. And my actual example of people who should not be a adopting AI was embedded systems and operating systems and database people. Are they adopting AI? **Peter Ludwig** [27:28] I think it's the classic bitter lesson, uh ... uh, topic, which is the, um, uh... Six months ago, I would have said the same thing, but it's, it's becoming super useful for every domain. **Swyx** [27:40] I'm sure. **Peter Ludwig** [27:40] Right? Like, uh- ... th-there was, uh, I think six months ago or maybe, maybe a year ago, if you tried to use, let's say, the latest Claude model for writing shaders, uh, GPU shaders, the results were probably underwhelming. And if you use the latest model now to do that kind of task, y-you're a little bit blown away, like, "Wow, that actually worked." That's amazing. And we see the same thing in, i-in the embedded realm. The, um, no question though, especially when you get into safety-critical systems, the, the human validation is, is 100% key. Um, like I, I... you're not gonna trust your life to, uh, like AI-written software that's, that's not been very carefully, uh, checked by humans. And so I think now the, um, really the challenge is about that appropriate level of, of human validation for, for these safety-critical systems. **Alessio** [28:28] How do you think about, yeah, touching on the simulation side, I think verifiable reward and reinforcement learning is like the hardest thing. What have you done internally to build around that, and like what, what gives you-- what makes you sleep at night? Like, if somebody's like, you know, just vibe coding something or like- ... wants to try something new, you have like a good enough system. Because I think the opposite is also true, is like if it's super easy to write anything- **Peter Ludwig** [28:50] Mm-hmm **Alessio** [28:50] ... then it puts a lot of work on like the verifiable- **Peter Ludwig** [28:53] Yeah **Alessio** [28:53] ... side of it. Like, what does that look like for people? **Peter Ludwig** [28:56] Yeah, yeah. So, so verifiability, a, a broader bucket of like evaluations, right? Like how, how do you evaluate the, the results that you're, you're getting? I think this, this is probably the hardest problem right now because the... as the models get better, it can be harder and harder to find the faults in the system. And so like the, the problem of doing proper eval to find those faults, uh, like that problem also keeps getting harder as, as the models get better. But it's no less important than it's ever been, right? If you still-- There, there are still going to be edge cases that are not met and, and whatnot. And so it's, it's a big area of, of investment for us. The, um, on the reinforcement learning topic, I mean, the, the key thing is there's all these new requirements that, that come to be in, in the latest generation of, of these technologies. So for example, end-to-end is the big thing right now in, in autonomy and physical AI, which is you can now train these models that can effectively take sensor data in and then put control signals out and get really good results out, out of that. But the way that you train and improve those models is really different from, uh, from the previous generations. And so to do reinforcement learning on an end-to-end model, you now need to actually simulate all the sensor data, right? So then this becomes, uh, we, we call our, our, uh, work in this neural simulation, but it's- **Alessio** [30:12] Mm-hmm **Peter Ludwig** [30:12] ... think of it like a, uh, a hybrid of Gaussian, uh, splatting and, and diffusion methods and, uh, where you really care about performance. Like performance is, is everything. If you can't do enough simulation fast enough and cheap enough, you actually can't get results that are, are worthwhile, uh, in the end. It also gets, gets to a lot of our work in embedded systems, which is like- ... performance critical work, and that, that performance optimization, performance criticality, it carries over to a lot of the, the, the model training work, um, because like, the only way to make it affordable is it has to be really fast. **Qasar Younis** [30:44] I think it's worth a, a few minutes talking about our own, uh, evolving thoughts on verification and validation within- **Peter Ludwig** [30:51] Mm **Qasar Younis** [30:51] ... kind of, uh, traditional simulators, which are, you know, uh, you can think of like vehicle dynamics or something like that, which you're just taking textbooks and taking those formulas- **Peter Ludwig** [30:59] Mm **Qasar Younis** [30:59] ... and putting them into software, to like now this neural sim/world model universe. Uh, I think that's an interesting topic. **Peter Ludwig** [31:06] Yeah. Yeah. So, so in, um, in more traditional development, right, you, you oftentimes would have, ah, more black-and-white answers to questions. **Qasar Younis** [31:14] Mm-hmm. **Peter Ludwig** [31:15] And so the, uh, in, in Europe as an example, there's a, a regulatory, uh, system, it's called Euro NCAP. It's the European New Car Assessment Program, and as part of that, the vehicles have to pass a bunch of tests, and those, those tests actually, uh, in- include, um, CP systems. So automatic emergency braking for a child that runs in front of a car- **Qasar Younis** [31:38] Mm **Peter Ludwig** [31:38] ... uh, or let's say an occluded child that runs out- **Qasar Younis** [31:41] Yeah **Peter Ludwig** [31:41] ... and, and you hit it. And, and so you, you have-- you end up with sort of these binary answers of like, well, did, did the car under test pass this specific test? And there's a very, very well-known set of, of test cases- **Qasar Younis** [31:51] Mm **Peter Ludwig** [31:51] ... that the vehicle has to pass. And that was how the industry worked, let's say, uh, un- until ten-ish years ago. But what's changed now is with these models, everything is stati- statistics, right? Like you, you no longer have a black-and-white answer, but it's like, well, how many orders of magnitude or how many nines of reliability can, can I get in the system, and how can I, how can I prove that to be true? Um, and, uh, and, and the big unlock, honestly, for physical AI as, as an industry, is that these models are just becoming much more reliable, right? Things like, things actually work a lot better. It's like the number of nines you can get out of these systems are, are now good enough that it actually becomes cost-effective to, to really deploy these things. And so the, the big shift in, in sort of verification and validation has been from a little bit more of a, again, in the past it was strictly requirements and are you meeting or not, and, and now it's more of a statistical, uh, verification and validation case where it's all about how many nines of reliability and, and meantime between failures, that sort of thing. **Swyx** [32:51] And, uh, is the target audience regulators or even the customers or cou- Yeah, if you... I imagine the customers are bought in, and it's mostly regulators that need to be satisfied. **Peter Ludwig** [33:01] We do work with the US government. We do work, of course, with the E- European governments and, and the government of Japan. And, um, the, the government is not like an AI lab by any means. **Swyx** [33:11] Mm. **Peter Ludwig** [33:11] So, uh- **Swyx** [33:12] They just care about the outcome. **Peter Ludwig** [33:13] They care about the outcome. **Swyx** [33:14] Right. **Peter Ludwig** [33:14] And, and so we, we do education, uh, in, i- in that regard and, like, sort, sort of teaching about, hey, this is how we think validation should be done, and this is an approach that, that we think is, is reasonable and how to think about, like, when is a, a driverless system actually safe enough to, to, to go on the roads and that, that sort of thing. Um, but, uh, uh, I wouldn't say that the government is asking for it. It's like we're more teaching the government in that, in that sense. It's honestly, it's more so for our own, our own comfort, right? Like, we want to build very safe systems, and then of course our customers care deeply about that as well. But i- i- in that context, we're also typically educating our customers. **Qasar Younis** [33:47] Yeah. Our first, I mean, uh, uh, our first core value is on road safety. So, uh, uh, I think we can't underline enough that, uh, us also verifying and validating that the systems that we're deploying are safe to, to us is probably as important as, like, some regulator or a customer saying, you know- **Swyx** [34:05] Of course. Yeah, you have to satisfy yourselves. **Qasar Younis** [34:08] Yeah. **Peter Ludwig** [34:08] As I say, as, as a whole, across the world, regulation oftentimes it's like a almost lowest common denominator. But, like, you really have to substantially exceed what the regulators are expecting to make good products. **Swyx** [34:19] Yeah. One thing I often talk, talk about, I think, and, and I try to make this relatable to the audience also, is Cruise, where they had an accident that basically ended the company. I wonder if people overreact to single incidents, because incidents are going to happen regardless, right? Because it, it's a statistical thing. But as long... I don't know if regulators understand that, uh, you cannot extrapolate from a single incident. But we do, because that's all we have to go on. And your sample sizes are necessarily gonna be lower than, I don't know- **Qasar Younis** [34:46] Yeah **Swyx** [34:46] ... consumer driving. **Qasar Younis** [34:48] Yeah, I think the con- the, the Cruise example wasn't a technology failure. Um, there was, the, the real, uh, compounding issue there was just how did the company talk to the regulators and, and what was their kind of behavior, and I think that became more of the issue. If you look, you know- **Peter Ludwig** [35:06] It isn't, it definitely was a technology failure, but it was made much worse by the- **Qasar Younis** [35:09] The car backed onto a woman. **Peter Ludwig** [35:10] Yeah. **Qasar Younis** [35:11] Yeah. Yeah, yeah. And le- let me put it another way. There is a version where Cruise still exists. **Swyx** [35:15] Right. Right. **Qasar Younis** [35:16] Right? It- **Swyx** [35:16] It was like the last straw. **Qasar Younis** [35:17] It- **Swyx** [35:18] Like a long chain of- **Qasar Younis** [35:19] Yeah **Swyx** [35:19] ... like, issues of- **Qasar Younis** [35:20] So even if like ATG had that horrific accident of someone actually dying, uh, because, you know, uh, that was a homeless person crossing the street. So yeah, I think, I think, uh, we can't understate enough that ultimately, like statistical validation of something, that's one part of it, but it's not the only part of it. Like consumer and, let's say, mainstream adoption of these technologies is also gonna be part of that conversation. I think companies like Waymo are doing a lot of s- you know, service positively to the industry in the sense of they're, they're setting a high benchmark, and they're showing, you know, kind of in a very responsible way how to, how to deal with these. There have been Waymo incidences as well. They've just not been as significant as, as the Cruise one that you mentioned. But yeah, so I think, I think you'll just continue to see that. I think pr- probably the long-term question is really gonna be, again, around, like it is very clear humans are way worse drivers statistically. **Swyx** [36:15] Yeah. **Qasar Younis** [36:15] Like there's no, there's no debate. And so at what point... But we're emotional animals. **Swyx** [36:21] Yeah. So my thing is like we have to get to a point as a society where we accept horrific accidents that would never happen by a human, because statistically we understand that it is safer overall. In the same way that planes, they're safer, uh, than, uh, I think they're the safest mode of transport that we have. Um- **Qasar Younis** [36:36] Yeah, I mean, it's more dangerous to drive to the airport than it is to get on a flight. **Swyx** [36:40] So, so if you're ever- **Qasar Younis** [36:40] Yeah **Swyx** [36:41] ... if you're ever getting nervous about getting on a plane **Peter Ludwig** [36:43] Just think, "I just gotta get to the airport." **Qasar Younis** [36:45] Yes, you're flying. **Peter Ludwig** [36:45] If I get to the airport, I'll be good. **Qasar Younis** [36:47] But then it's, planes also concentrate the tail risk if, if planes just- **Peter Ludwig** [36:50] Yeah. And I was, I, I, I don't think we honestly have to worry about there ever being, uh, accidents from these systems that are mu- like much worse than what humans would cause, 'cause humans do, do terrible things. **Qasar Younis** [37:00] Yeah. **Peter Ludwig** [37:00] People fall asleep at the wheel all the time. **Qasar Younis** [37:03] I have. **Peter Ludwig** [37:03] Yeah. **Qasar Younis** [37:04] I, I'll ca- I- I've been a drowsy driver. **Peter Ludwig** [37:05] Kind of dri- drunk drivers, and that's- **Qasar Younis** [37:06] Yeah **Peter Ludwig** [37:07] ... that's the extreme end of the example. But, um, I mean, these AI systems, you have redundancies, you have fallbacks. Like, there's m- many, many things have to go wrong for there to actually be something catastrophic because there's, there's so many, uh, fallbacks that these systems have. **Qasar Younis** [37:21] Yeah. **Alessio** [37:22] I mean, your simulation is, like, so vast because there's so many use cases. What are, like, maybe things that worked in a simulation and then you put it out and it's like, "Fuck, this is- ... this just did not work at all"? **Peter Ludwig** [37:33] Yes. That's m- maybe a bit of a misconception, uh, about simulation there. So let me go a little bit more, more technical on this. So, um, at first go, no simulation is, is going to represent the real world. There, there's always a process of this, uh, sim to real matching- **Alessio** [37:48] Mm-hmm **Peter Ludwig** [37:48] ... where you actually, you need the real world feedback to basically feed into the parameters that are being used in the simulator, and you have to do that, uh, it's like this validation flow, uh, a number of times until you can get some confidence that, like, I, I think the simulator is now accurately representing- **Alessio** [38:05] Mm-hmm **Peter Ludwig** [38:05] ... what's gonna happen in the real world. Now, if, if you have a situation where you've done that full validation and, and you thought that it was accurate and then there's something different, those are much trickier cases. And that's, that absolutely can happen, but really the, the validation process is a really important part. You can never skip the simulation validation process, like, where you're actually ensuring that, "Hey, actually, my sim to real gap here is, is small enough that I can trust these, uh, these simulation results." And, and there's, there's so many fun things that you can do when you get into it. Like, I, I'll give one, one fun example that came up recently is, like, in these, uh, these humanoid robotics, uh, systems, overheating actuators is a real problem, right? So, uh, obviously phenomenal demos. I, I- **Qasar Younis** [38:47] Right **Peter Ludwig** [38:47] ... the most amazing- **Alessio** [38:48] Words can't- **Peter Ludwig** [38:49] The most amazing eye candy. I, I love, I love watching robots do acrobatics like everybody, but the, these systems actually overheat, right? If, if, like, uh... And one of the ways you can use s- simulation though is you can actually have that, the temperature of those actuators be one of the parameters that's represented in the simulation. And, um, and if you're doing reinforcement learning over a certain task, then the robot can actually adjust its motions in the simulation to account for the fact that, oh, it knows that as it's moving, it's actually beginning to overheat this, this motor. Um, but if you didn't have that parameter of, let's say, the heat of that motor represented in the simulation initially, then your RL policy might, it will disregard that. And now you run that on the robot, and the robot will overheat and, and fail. **Alessio** [39:30] I guess the question is, like, how do you have all of these parameters taken care of while also understanding the deployment environment? Like, temperature is, like, a great example, right? Well- **Peter Ludwig** [39:39] Yeah **Alessio** [39:39] ... why did you make my robot worse when it runs in, like, a freezer? So it actually shouldn't worry about that. You know, it's like, uh, yeah, how do you design these simulations? **Peter Ludwig** [39:49] This is honestly the, the-- this is what makes simulation so hard, right? Uh, it's because you... Simulation is, is fundamentally about you're trying to optimize the development of a system, right? Like how, how can I build this system faster and better and cheaper, and, and what are all the levers that I have to actually accomplish that? And, and because simulation's just a software program, you can, you can change it a lot more easily than you can hardware systems. And then what's particularly awesome about the, let's say, world models and using that as a part of simulation is now the simulation doesn't just scale with, let's say, adding new math equations in- **Alessio** [40:23] Mm-hmm **Peter Ludwig** [40:23] ... but we can actually scale the simulation environment now with, uh, with additional real world data. And, and that, that also unlocks, as I would say, a whole new field of robotics. **Qasar Younis** [40:32] There is a meniscus line where you cross where still doing real world testing is better. There, there's a, in this sim to real gap, you can reproduce reality at exceedingly expensive costs. And so nothing is free. So really, you, you have to... you're finding that line where you're getting great performance, you're getting great feedback, whether it's on the training side or on the eval side, but it's way cheaper than doing it in, in the real world. At some point, it, that doesn't make sense. And so even, you know, from our earliest days in autonomy, our view was you're still gonna do real world testing. You know, there's, there's not, there's not this, you know, magical land where you're not gonna do that. And, uh, maybe even, like, a more nuanced version of this in, like, traditional software development is, you know, most of your testing for software in a vehicle, 95% of that can be, like, traditional CI/CD kind of, uh, flows that you'd have in traditional web development. But once you have, now let's say you have a truck. Well, you can do, like, 4% of those in, like, a rig which has all the components, the electrical and electronics of a truck but doesn't have, it doesn't have the, the tires and it doesn't have the phys- And then you have the 1%, which is actually the vehicle. There's something sim- There, there's a similar analogy in terms of using simulation for intelligent systems. You can do a lot in a simulator, but, um, in, in using world models, uh, but ultimately it's, it's physical AI. So you're gonna deploy it on physical machines and- **Alessio** [42:03] Yeah **Qasar Younis** [42:03] ... the freezer example kinda comes to, comes to light. **Alessio** [42:06] The world model thing has been, to me, the hardest thing to- **Qasar Younis** [42:09] Yeah **Alessio** [42:09] ... wrap my head around. Like, we have Faith Lee on, on the podcast. **Qasar Younis** [42:11] We've been doing a small series with, like, a- another Intuition company, General Intuition as well. **Alessio** [42:16] Yeah. **Qasar Younis** [42:16] Uh, yeah, and I mean, lots of, lots of coverage on NeRFs and- **Alessio** [42:19] Yeah. It, it feels like, like, uh, we talk with, about, um, the heliocentric system, right? It's like in a world model, if you just feed visual data, the model might learn that the sun spins around the Earth. It makes sense, right? And it's like, well, not really. And I think what are, like, some of these other things that, like, like hydroplaning is one thing I think about, is, like, can a world model understand hydroplaning and, like- **Qasar Younis** [42:43] What amount of water, like, causes it to happen? And it's like, yeah, to me it's like I don't understand how you guys do it, I guess is like the, the real thing- ... is like when you're doing both cars and the highway in Japan versus the, you know, excavator in a mine in, uh- **Peter Ludwig** [42:59] Arizona **Qasar Younis** [42:59] ... wherever you're... A- Arizona, wherever you're deploying them. **Peter Ludwig** [43:01] Yeah. **Qasar Younis** [43:02] How much of it are you relying on the world models to, like, generate the simulations for you and then try and close the gap after, versus, like, giving the world models as a tool to your engineers to, like, curate the simulations, if, if that makes sense? **Peter Ludwig** [43:15] Yeah, totally. So, so, um, yeah, I can say at a pure engineering level, I think if, if you're hoping to do real-world deploys and you're purely relying on a world model approach, you probably won't get to something that works, uh, before you go bankrupt. So there, there is just a, a very practical mindset of like, uh, world models are, are amazing and they're extremely useful for a lot of use cases, but there are a lot of other things that you need to do to actually get something started and something deployed and working. Uh, most fundamentally, world models are all about... It's understanding the world, but also understanding what's going to happen. It's like the cause-effect relationship, right? ### Embedded AI **Qasar Younis** [43:48] Yeah. **Peter Ludwig** [43:48] And so, like, it, right, if you have a, take some sort of construction tool, uh, and that construction tool's gonna be doing some work on the Earth in, in some way. It's gonna be move- moving earth. Um, the world model needs to understand that cause-effect relationship. Like, okay, when I, when I take this material from here and put it over there, now I have things that are over here and, and not over there anymore, and, and that, that cause-effect, uh, relationship. Um, data obviously is a, is a big problem. The hydroplaning- **Qasar Younis** [44:12] Yeah **Peter Ludwig** [44:13] ... one is, is actually a really great example because it's actually quite non-obvious sometimes. Right? It's like, well, it's, it's raining and, uh, and well, this, this road, uh, has, uh, let's say the, the appropriate curvature to it- **Qasar Younis** [44:25] Mm-hmm **Peter Ludwig** [44:25] ... so the w- the water is running off the road and cars are driving faster here. And then you approach a road that's very flat and water is now puddling on that road, and all of a sudden cars are driving slower because when they were driving faster, they were starting to, to lose control. Um, and, uh, there are a lot of visual nuance- uh, very nuanced visual cues in the scene, and so I do think in, in, in the world model concept, there's a good chance that the model actually would learn that you should just drive slower when these visual cues exist, and that's obviously the beauti- the beauty of, uh, these kinds of models, right? They just- **Qasar Younis** [44:59] Mm **Peter Ludwig** [44:59] ... they learn these non-obvious things. **Qasar Younis** [45:00] It doesn't need to know about hydroplaning to know that it needs to drive slower. **Peter Ludwig** [45:03] Exactly. **Qasar Younis** [45:04] I guess it's... Yeah, yeah. **Swyx** [45:05] I wanna ask questions about, uh, also de- deploying models. I presume, like, you, you use a lot of these world models for training data and simulation, but what about deploying it, uh, onto the systems in, in production? Uh, presumably you have mo- you have, like, GPUs on device- **Peter Ludwig** [45:22] Mm-hmm **Swyx** [45:22] ... but they're s- I keep saying on device. What, what's the, what's the right term for that? **Peter Ludwig** [45:26] On machine. **Qasar Younis** [45:27] On machine. **Peter Ludwig** [45:27] Or embedded, yeah. **Swyx** [45:28] Yeah, yeah. What, what is the embedded world like? Uh, because for, for, for people who are not used to that world, um, this is very alien. **Peter Ludwig** [45:36] Yeah, yeah. So it's actually... We, we call it onboard and offboard. **Swyx** [45:38] Okay. **Peter Ludwig** [45:38] So, like, um, onboard software and offboard software. **Swyx** [45:41] Yeah. **Peter Ludwig** [45:41] And the, the great thing about offboard software is you don't have to care about time, uh, and you can run really large models, right? So you can, you can say, "Well, uh, this, this model, I don't care if it takes one second for it to give me a result or 10 seconds for it to give me a result, uh, because we have time." And the models can be really big and they can run, uh, in the data center or on a, on a huge GPU and, and you can obviously have distributed compute, et cetera. But onboard you don't have, have any of those benefits. You're like, "Well, I need-- I have this many milliseconds where I need an answer from this model." And so a lot more of the energy then is about, uh, think of it more like distillation and, and it's, like, truly efficiency and, like, literally every, every fraction of a millisecond counts. And, and, uh, and you, you, you can't have a situation where the model takes too long because then the vehicle can't actually function. **Swyx** [46:27] Yeah. **Peter Ludwig** [46:28] And so you can, you can still use a lot of the same techniques, uh, and, and the, the, the models themselves you can think of as like a derivative of, of larger models that you can run offline, and then you're, you're trying to just get a, a model that is still performs really well but it's, it's a, it's smaller, small enough version that you can then run on this embedded system where you care about latency and power. **Swyx** [46:49] Yeah. **Qasar Younis** [46:49] Yeah. And I think, like, the, the broader point I think, uh, which, which, uh, maybe is not obvious but it's worth saying is in physical AI world, we're not really constrained right now by, like, the, the intelligence of the models. It's actually what Peter's talking about, it's actually deploying them i- in- **Swyx** [47:06] The hardware they give you. **Qasar Younis** [47:07] Yeah, on the hardware they give you. **Swyx** [47:08] Right? **Qasar Younis** [47:08] And, and so... And, and there's just a reality is of safety critical systems. So those end up being the, your, your limiting factors- **Swyx** [47:15] Mm-hmm **Qasar Younis** [47:15] ... rather than let's say a limiting factor for, you know, uh, a, a foundation model company- **Swyx** [47:20] Yeah **Qasar Younis** [47:20] ... is gonna be just capital maybe, you know, or, or, or researchers. **Swyx** [47:24] Yeah. **Qasar Younis** [47:24] So we're, we're in that way dealing with, you know, for us as people who kind of come in that realm, like a very interesting, those constraints force creativity. **Swyx** [47:33] And I imagine, you know, nobody was deploying or giving you the hardware for transformers back in 2018, whatever, but now they are. What's the evolution like? Uh, just peel back the curtains a little bit. **Peter Ludwig** [47:45] Yeah. Transformers first off, I think the paper was originally published in 2017. **Swyx** [47:49] 2017. **Peter Ludwig** [47:49] Yeah. **Swyx** [47:49] So there's no time. **Peter Ludwig** [47:50] And, uh, I, I think it's- **Qasar Younis** [47:51] But I'm just saying-- I guess I'm saying, like, you know, embedded s- uh, ML systems usually, like, a lot less parameters, a lot less compute, and now, like, orders of magnitude more. **Peter Ludwig** [48:00] Yeah, yeah. T- A- A- Absolutely. Um, what I was gonna say though was I think in the, uh, in the original paper in 2017, uh, may- maybe it's in the last paragraph, s- s- somewhere in the paper they talk about, it's like, "Oh, by the way, this, this, this technique might, might be useful for, like, images and videos as well." **Qasar Younis** [48:16] Yeah. **Peter Ludwig** [48:16] These last sort of things. **Qasar Younis** [48:17] Yeah. **Peter Ludwig** [48:18] And, uh, it took a few years for that impact to really hit. Uh, but, like, now, I mean, we're seeing, uh, transformers are everywhere. **Qasar Younis** [48:25] Yeah, vision transformers. **Swyx** [48:26] Yeah, yeah. Cool. **Peter Ludwig** [48:27] And, uh, the, um, yeah, the, and then the compute just keeps getting better and better. Um, but you, you, you do have this fundamental trade-off, right? It's like you- Power, you have cost, uh, and, and performance and, and, like, uh, g-getting the right, getting the right mix of those things in an embedded package that can also be, like, shaken and, and baked in all the- **Swyx** [48:46] Right **Peter Ludwig** [48:46] ... conditions that these things have to h- have to operate in. But yeah, I think that they're only going to keep getting better. And so we also try to plan our strategy understanding that, uh, we know the rate of improvements of these systems. **Swyx** [48:57] Yeah. So, so, like, Google just released the Gemma 2B model. **Peter Ludwig** [49:01] Mm-hmm. **Swyx** [49:01] That effective 2B model. Is that useful to you guys, or is that too big? **Peter Ludwig** [49:04] You can run that model on an embedded system, definitely. **Swyx** [49:06] Yeah. **Peter Ludwig** [49:07] Um, the, um... So, so yes, it's, it's useful in, in that regard. The bigger question is, like, what do you use it for in an embedded system? Like, you actually need to customize it quite a bit to make it useful for, for something. But, uh, but yeah, you, you could run a two bl- billion parameter model definitely. **Swyx** [49:21] It also interesting, like, what percent is a custom ML model that only does that thing versus a generalist LLM- **Peter Ludwig** [49:28] Yeah **Swyx** [49:28] ... uh, which probably is not that useful actually for your context. Like **Peter Ludwig** [49:33] You, you, like, you can imagine d- different use cases, right? **Swyx** [49:34] Yeah. **Peter Ludwig** [49:34] So the- **Swyx** [49:35] The voice stuff, yes. **Peter Ludwig** [49:36] Yeah, the voice stuff, totally, yes. **Swyx** [49:37] Yeah. **Peter Ludwig** [49:37] So for the, um, the actual, uh, autonomy elements, I mean, that's 100% in-house. Like, uh, we, we do every bit of that, the, the data simulation, the model, everything. Uh, but when you get into the more generic use cases like voice or a voice assistant kind of thing, that's where these, these more generalist models like, uh, Gemma actually can be quite- **Swyx** [49:55] Mm **Peter Ludwig** [49:55] ... can be quite useful. **Swyx** [49:55] Yeah. And then there's also obviously a trade-off between, like, what percent must you do on machine, uh, versus just call home. **Peter Ludwig** [50:03] Yeah. It's all about latency. **Swyx** [50:04] Well- **Peter Ludwig** [50:04] It's all about latency. Yeah. **Swyx** [50:04] Yeah. Well, like, you know, I think actually in a lot of contexts, especially in the US, you can just have a connection to the, to the web. **Qasar Younis** [50:12] Yeah. I think, uh, uh, I think though most of our universe is everything has to be fairly, uh, uh, you know, embedded and local because just the nature of... Even in the US, there's a lot of pl- like mines- **Swyx** [50:26] Patches **Qasar Younis** [50:27] ... don't have- **Swyx** [50:27] Yeah **Qasar Younis** [50:27] ... have coverage, right? And if you look at, like, uh, the old world of autonomy within mining, which is, like, l- long before transformers and, and, and kind of, uh, neural networks, uh, in the, like CNN and, uh, uh, uh, a kind of a universe, they were really just hand-coded, you know, systems. They were just like, "This machine is gonna run to that place with this-" **Peter Ludwig** [50:49] That was RTK, is like very accurate GPS. **Qasar Younis** [50:51] Yeah. And so that worked, and that works for 20 years. So why would we actually need to use transformers or kind of mo- more modern end-to-end systems? Mainly because you can only really run a path and run backwards. That provided a lot of value, but n- not as much as you get when the machine is actually intelligent. It's, it's seeing, it's perceiving, it's acting in a dynamic world. **Swyx** [51:13] Mm-hmm. I looked up RTK, real-time kinematic, uh, one to two centimeter accuracy. **Qasar Younis** [51:18] Yeah. **Peter Ludwig** [51:19] Yeah. **Qasar Younis** [51:19] Fantastic. But, you know, and fantastic in faraway lands where there's not gonna be cell phone coverage. **Peter Ludwig** [51:26] Yeah. So it's widely used on the legacy mining and agricultural autonomy systems today. So like, uh, for example, a combine that can be precise within one or two centimeters as it's driving down the, the field, uh, they use RTK. **Swyx** [51:39] Yeah. **Peter Ludwig** [51:39] But it's, it's expensive. **Qasar Younis** [51:40] Yeah. And it's, it's, it's autonomy, but it's not intelligent in the way that I think all of us- **Swyx** [51:44] Mm-hmm **Qasar Younis** [51:44] ... if in 2026 we'd be talking about intelligence. **Alessio** [51:47] One of your blog posts, you mentioned research on lar- large-scale transformers that are similar to those doing modern generative AI. What are, like, the big differences other than, "You're absolutely right. I should steer the car, so, um, you probably wanna remove that"? **Peter Ludwig** [52:01] We have a diversified bet strategy internally, and, and the reason we, we've done that is because we operate in now a bunch of industries, a bunch of geographies, and, and each of the approaches has, uh, honestly a different risk to them. **Alessio** [52:13] Mm-hmm. **Peter Ludwig** [52:13] And so, uh, like, we're not going to put all of our eggs in, in a single, in a single basket for a s- a single approach because that approach may not work out. **Alessio** [52:22] Mm-hmm. **Peter Ludwig** [52:22] Uh, and so that's, that's one of the bets that we have, and it has certain advantages in, in certain scenarios. And then, but the way that these things play out in practice is it has certain benefits, it also has certain drawbacks. And, and then, and then the research team tries to then work on, uh, the, the situations where that's actually worse than, than these other approaches and, uh, to ultimately arrive at, at a, a really great solution for all these things. **Alessio** [52:43] Is there a plan mode for physical autonomy? The idea of a planning step and then a action step or... **Peter Ludwig** [52:50] So s- short answer is yes, right? So just like you can use, uh, a Claude code to, to plan out some complex coding task and, and you get some almost specification written out, those similar, similar approaches absolutely can be applied to physical systems. Because imagine you're trying to accomplish some task. The easiest to think about w- is, is robotaxi, but I think- **Alessio** [53:09] Mm-hmm **Peter Ludwig** [53:09] ... things get more interesting, let's say, in, in the defense context or in, uh, in the, in the mining context. You actually do have to think about many steps in advance. **Alessio** [53:18] Mm-hmm. **Peter Ludwig** [53:18] It's, it's not just this one thing, but to accomplish the goal, there's 100 steps. And then the, this concept of the plan mode, it's, yeah, very applicable, uh, in this. **Alessio** [53:26] Yeah. I, I was gonna say, to me, driving feels like a great next token prediction thing because you're kind of like on a path and, like, it doesn't really matter what you've done before. You know, you can always turn around. **Qasar Younis** [53:35] Planning. Yeah, yeah. **Alessio** [53:36] Yeah. Versus like mining, it's like, "Oh man, I took a, you know, I took a scoop out of this thing." It's like, "Now we can't really- ... you know, I can't really go there anymore." You know, it's like, is there like a huge difference? Like, how, how would you... I guess, like, do you have like a taxonomy of, like, these different types? So there's kind of like driving- **Peter Ludwig** [53:53] Yeah **Alessio** [53:53] ... excavating, like flying. How, how do you- **Peter Ludwig** [53:57] So the interesting thing is, uh, I think probably everything in the world can actually be boiled down to, like, a next token prediction- **Alessio** [54:03] Right. **Peter Ludwig** [54:03] ... problem. Um, and, uh, in any workflow, anything, uh, can be thought of almost as like there's this, this sequence of steps or this sequence of trajectories or what- whatever you wanna call it, and it can be boiled down actually to, to that sort of thing. And in the mining case, you can imagine, like, t- taking that scoop. Okay, that was that set of, of tokens, and, and now that's, the, the model is now understanding that, okay, that the state space is different, and now the next time I do token predictions, it's going to, going to be modified by that. But yeah, these... The remarkable thing about these techniques is just how, how universally applicable they are, right? I mean, it's, it's truly is incredible. **Alessio** [54:40] What else is underrated about what you guys are building on the physical side? I think there-- I mean, we were talking about it before the episode. There's a lot of humanoid companies that do these great demos, um, and then I can't buy it, so obviously it can't all be there. In your case, you're, like, in production on real streets with, like, a lot of customers. What, what are, like, the things people are underestimating? The same way the Waymo demos seven years ago were great and then took seven years to actually get them on the street. Can you share about maybe, like, the last one percent that was really hard to, to get done technically? **Peter Ludwig** [55:13] Yes. So certainly, productionizing stuff is really challenging no matter what. Um, so I, I maybe would-- I would split the answer maybe into research and then also in, in production. First, on the production side, there's just so many problems that you find when you actually get the stuff to, to go in the real world. ### Production **Peter Ludwig** [55:28] And so, I mean, the classic problem in, in humanoids right now is these systems are actually pretty brittle. **Alessio** [55:34] Mm-hmm. **Peter Ludwig** [55:34] Uh, and so, um, uh, I'm not talking about any one company, but just as an industry, these systems are, are pretty brittle. I mean, interestingly, I saw this thing, uh, the other day that, uh, I think China is doing a, a marathon with humanoids. **Qasar Younis** [55:47] What? **Peter Ludwig** [55:47] Yeah, yeah. So, uh, in, in government, and not China specifically, but in any government, there is a, um, there's a concept called, uh, prize policy, which is so that there's, there's different ways of, of influencing an industry to go a certain direction. Like, you can, you can regulate it, right? You can do mandates, or you can actually just do these competitions. So the US version of this was the DARPA Grand Challenge. Uh, that- **Alessio** [56:06] That worked. **Peter Ludwig** [56:07] But it really worked. It- **Alessio** [56:08] That really worked **Peter Ludwig** [56:09] ... the industry. But, uh, I think China is literally doing this marathon because they know that reliability, uh, of, of these humanoids is a problem. And so what cooler way to solve that than to have a competition where humanoids need to run 26 miles, right? **Alessio** [56:24] Are we there? Can, can, can robots run a marathon? **Peter Ludwig** [56:26] I think it's happening any day now. **Qasar Younis** [56:28] Yeah. Yeah. **Peter Ludwig** [56:29] So it's- **Alessio** [56:29] So we're there. **Qasar Younis** [56:30] By the way, also, you know, automotive, there's a version of this which is, like, 24 Hours Le Mans, right? **Peter Ludwig** [56:34] Yeah. **Qasar Younis** [56:34] It's like Porsche wins 24 Hours Le Mans- **Alessio** [56:36] Three hundred miles an hour **Qasar Younis** [56:37] ... and then literally puts those, uh, the, you know, products into production. I would actually break it down. You, you know, talk about research and you talk about production. There's actually a step in the middle which is, like, advanced engineering, and I think a lot of the industry is moving into advanced engineering where it's like it's not fundamental research. Like, we're coming up with novel techniques. It really is advanced engineering for production. So what are the, the subcomponents that are gonna limit to getting into production? Once you're in production, you're dealing with another set of problems, which is, like, the deployment, maintenance, uh, of those machines that, that exist. So I would say, at least in our field, we're mostly in advanced engineering in the, like, automotive parlance. **Peter Ludwig** [57:15] Uh, honestly, every, every step is hard though. **Qasar Younis** [57:19] Yeah. Yeah. **Alessio** [57:19] Well, that's why you're worth $15 billion, so. **Qasar Younis** [57:21] You, you bleed every step. **Peter Ludwig** [57:25] Yeah. And I think- **Qasar Younis** [57:25] And it's fun. I mean, I think-- I mean, it's like, I don't know. I, I find it really enjoyable. **Peter Ludwig** [57:29] Yeah, but what-- it was also fun is like, so we've, we've been doing this now for almost 10 years, and, like, we, we've just seen... We've seen so much bad tire. And so right now, we, we can look at any company in this space and, like, get a demo and, like, I can, I can write down a list of I know exactly the next 20 problems they're gonna hit. **Alessio** [57:45] Yeah. **Peter Ludwig** [57:45] And, like, I, I can guess also what they're going to try to solve each of those, and I can guess which one's gonna actually work. **Qasar Younis** [57:50] Yeah, it's not because we're, like, particularly, like, geniuses. **Peter Ludwig** [57:53] We've just seen this stuff now. **Qasar Younis** [57:54] Yeah, we've seen enough of this stuff. We lived enough of this stuff. We, you know, our own kind of mental models of the world as, as leads in the company, we've tried so many things in the many... We're talking about the wins here. Like- **Alessio** [58:06] Right. **Qasar Younis** [58:07] There- **Peter Ludwig** [58:07] Plenty of losses. **Qasar Younis** [58:08] There's plenty of losses among that many people doing that many different things and, um... And so that kinda, like, get baked into your, like- **Alessio** [58:15] Yeah **Qasar Younis** [58:15] ... mental model of the world. ### Startup Advice **Alessio** [58:17] Yeah. **Peter Ludwig** [58:17] But I would say in, in general, like we're excited about robotics for sure, and like- **Qasar Younis** [58:21] Yeah **Peter Ludwig** [58:21] ... uh, the- **Qasar Younis** [58:22] Massive opportunity **Peter Ludwig** [58:23] ... massive opportunity and what's, what's happening now in the industry is like none of these concepts are, are new, right? What's new is, like, this stuff is actually working now. **Qasar Younis** [58:32] Yeah. **Peter Ludwig** [58:32] Right? The people have wanted to use, uh, neural nets for robotics for a long time, uh, but, but now, like, again, we now have the datasets, we have the simulation technologies where stuff is actually starting to really work, and yeah, we-we wanna be part... We, we're gonna be part of that for sure. **Alessio** [58:46] Do you have requests for startups or, like, advice against starting certain startups? There's a lot of, like, scale-up robotics, you know, companies. It's like, what, what do you think are things- **Qasar Younis** [58:57] A lot of, uh, a lot of Applied Intuitions for other things. **Alessio** [59:00] Right. **Qasar Younis** [59:00] I think you, you hit a, you hit a certain, uh, what is it, uh, you know, badge when YC- **Alessio** [59:07] X for Y. **Peter Ludwig** [59:08] X for- **Qasar Younis** [59:08] Right? You become like a, you know... Or, or literally the same similar names, like, you know? **Alessio** [59:12] Right. **Qasar Younis** [59:13] Um, I, I mean, I think my biggest advice, you know, in this, like, almost like commercialization of technology is I think often the, that constraint. So we talked about, like, hardware constraints, and we talked about, uh, there's also, like, on the commercial side, there's constraints, which is we're gonna only do things that fit in this box. That is, I think, very good for founders. The reason I think it's not often focused on is because you have plenty of access to capital, and the technical problems are so hard, you're like, "I already have a constraint," which is just getting this, you know, this technical problem solved. And I think the venture community, generally speaking, tends to be not very technical. For them, if you just say, "If we solve this thing, it's gonna be a lot of money," that's kind of enough for them. But you as a founder, I'm not giving you advice on how to pitch VCs. That'll work for VCs. You still gotta run a sustainable business. And I think we're really in, in, in that-- You know, question you asked earlier about kind of, you know, what's maybe not obvious about our company. It's like this is truly compounding technology. A lot of the work that we do just compounds. It- we don't throw it away. It gets better. The operating system work gets better. The dev, dev tooling gets better. The models get better. And so we're really gonna get a hu- I, I think you see it in Waymo as an example. Like, Waymo is a company that is, I would say, very interesting for a long time, but not worth a hundred and twenty-six billion dollars, right? So what happens, like, is that the human brain just doesn't emotionally understand the compounding effects. So that's gonna happen in our universe. So now if you're a founder, you're at the beginning of that, you know, that long, you know, walk, if you can put a little constraint on, on commercials that has a small ability for you to more likely see the other end of that, th-that walk. 'Cause if you can get to the other end, you will get the big return from compounding technology. Just a lot of people just don't make it. So yeah, s-s-summarize, like think a little bit about the equation of how you use money and where you use m-the limited resources and limited engineers that you have. I think sometimes when founders falsely kind of take very mature companies' strategies and then apply it to their, like, nascent, they're like, "Oh, well, Steve Jobs says be completely vertical." Well, yeah, in 2007, Apple is very different than 1978 and 1982. Those companies were different. They, they were literally just taking electronics from other manufacturers and just putting it in an enclosure. And so I think just be a bit more like, I don't know, be a bit more nuanced in your, in your commercial approach as it informs your technical approach. **Alessio** [1:01:49] Do you feel differently today? Like, I mean, you just joined X, right? **Qasar Younis** [1:01:53] Yeah. **Alessio** [1:01:53] You've been building this company- **Qasar Younis** [1:01:54] Yeah, yeah, yeah. **Alessio** [1:01:54] You've been building this company in stealth, and now you're like, "Well, I should probably be talking about what I'm doing." I think a lot of founders are in a similar way where they wanna raise a lot of money to signal they're strong, and you raised a lot of money without spending it. **Qasar Younis** [1:02:06] And to hire. And to hire, yeah. **Alessio** [1:02:07] You obviously like that. Do you think that's still possible to, like, have a very narrow approach of like, "Hey, we're kinda like building a compounding thing," without a grand vision right away versus- **Qasar Younis** [1:02:18] It's, it's very difficult to answer very general questions- **Alessio** [1:02:21] Well- **Qasar Younis** [1:02:21] ... like that. Uh, I, but I-- So maybe, like, maybe I reframe it as in, is it possible to build a product that has a small, let's say, problem space and hope that the problem space will grow? Maybe that's, like, a different way of asking the same question, but m-ma-more answerable. I think always yes. That is the old YC, like go really deep and then, you know, rather than very broad and shallow. **Alessio** [1:02:46] Mm. **Qasar Younis** [1:02:46] Very broad and shallow, unfortunately, there's just too many tech, especially in hard tech companies, there's just too many problems, and you can't-- you're gonna do all of them in a very mediocre way. And so the s- the, the, the, the full product is actually fairly mediocre. So yeah, I, I, I, I still in, I'm still in the camp of find a small problem space. The other question you're asking is a tangential is like, should you, like, build in, in stealth and anonymity? Well, yeah, if you're a YC COO - **Alessio** [1:03:14] Yeah. **Qasar Younis** [1:03:14] I mean, you can because- **Alessio** [1:03:15] Or Travis Kalanick. **Qasar Younis** [1:03:16] And we w-- Uh, yeah, we worked, we worked, you know, together at Google. We have a long history, and we don't-- and which mean, which is another way of saying we have big networks. I mean, our, our first of 400 people, majority were Googlers. Like, a majority of the company came from, you know, this giant company we worked at, and that's just very different. You're a founder who is, doesn't have that experience. You have to do these things. And I think it's k- that's k- So it's like, just don't take my version of the world or whatever other founder, Jensen's version of the world. They are different time and space. **Alessio** [1:03:48] Mm-hmm. **Qasar Younis** [1:03:48] And most importantly, their companies are in a different phase. **Alessio** [1:03:52] Yeah. **Qasar Younis** [1:03:52] And so then if you wanna take inspiration from other really young companies, that's also bad because most of them are gonna fail. **Alessio** [1:03:56] Right. **Qasar Younis** [1:03:57] So the only, the only solution you really have is use first principle thinking and say, "Based on my skills, my co-founder's skills, the skills of my early team members, and the, what I'm hearing from customers, what's the product space that I should, I should build?" And- **Alessio** [1:04:12] Yeah. **Qasar Younis** [1:04:12] Yeah. Does that make sense? **Alessio** [1:04:13] Yeah, it does. **Swyx** [1:04:14] Yeah. I, I, I mean, Sam, Sam Altman, he said he regrets a lot of the advice that he's given at YC. **Qasar Younis** [1:04:19] Yeah. **Swyx** [1:04:19] So I'm always curious to ask, you know, founders like you who nod then- **Qasar Younis** [1:04:23] So YC- **Swyx** [1:04:23] ... a long time ago **Qasar Younis** [1:04:24] ... everyone who leaves YC, like, does the opposite. **Swyx** [1:04:27] Yeah. It's, it's a- **Qasar Younis** [1:04:27] You know, we, you know, well, Sam was president, I was COO. **Swyx** [1:04:29] Yeah, yeah. **Qasar Younis** [1:04:30] So, and we'd have a CEO. So we worked together, you know, extremely closely would be an, an understatement- ... 'cause the firm was also small. The- **Swyx** [1:04:36] Yep **Qasar Younis** [1:04:36] ... you know, YC wasn't, wasn't as big as like an OpenAI is. I, I directionally agree with that, but n- I would say that's not more of a YC function, it's more of the market- **Swyx** [1:04:48] Sure. Yeah, yeah. **Qasar Younis** [1:04:49] ... has changed. It is a different- **Swyx** [1:04:50] Right **Qasar Younis** [1:04:50] ... world. The AI industry is different, is, uh, the, the AI companies, I should say more specifically, and how they relate to the other YC companies and market, just so fundamentally different. The amount of money raised is different. The amount of investors, the sheer number of seed funds. One of our early investors is Floodgate. Um, and they did some analysis in the late, uh, 2000, like the double O's, where they were like, there's like single-digit number of funds that were like Floodgate, which were like writing sub $1 million checks, first checks, and they were not accelerating incubator. And Anne, who's, who's one of the co-founders there, uh, with, with Mike, they, they said that today they try to do, or like today as in like three, four years ago, they, they, they tried to do this analysis and they like lost count at like- ... 350 funds or something like that. So y- we're just in a different environment. So the YC advice from 2014- **Alessio** [1:05:42] Yeah **Qasar Younis** [1:05:42] ... just would not apply in 2026. But Sam is like way better at saying these things than me. **Alessio** [1:05:46] Right. **Qasar Younis** [1:05:46] Like, he somehow makes it like he, he says it in a shorter, most, more interesting and, than, than me. I, I can just give you like the... Like I, like if you ask me like, you know, "What is the purpose of a car?" Like open the owner's manual. I say, "Number one-" "... one, there's a steering wheel," and you know, instead of like, "It can change your life and will be there." **Alessio** [1:06:07] Yeah, yeah. Gives you autonomy and freedom. **Qasar Younis** [1:06:08] Yeah, exactly. Yeah, yeah. **Swyx** [1:06:11] Uh, and then for Peter, I was just kinda curious if there's any particular tech or research problem that you would call out as very meaningful for you guys if it was solved and unsolved, and if anyone is working on it, they should get in touch with you. **Peter Ludwig** [1:06:27] Yeah, I think, um, th-th-generally the, the making models very efficient, right? So it, because we have to run on actual vehicles, like p-phys-physical AI is literally, uh, it's taking like very large AI and now making it very small and very efficient. And, uh, and w- so we're constantly just at that boundary of these limitations of like, well- You have a great model, but now we need to make it faster and smaller and, and so that, that in general as a, as a field. And then I would say also, um, uh, folks that are just really passionate about, uh, like evaluating this technology, as in like mo- mo- model evals, um, is a- it's a hugely difficult topic, especially in safety-critical systems. And, um, uh, I mean, we, we have a, I think a really great engineering team that works on this now and, and, and researchers, but it's, it's a big area of investment. And so yeah, folks that are passionate about, um, yeah, performance, I guess, model performance, both in terms of capability and, and literally latency, and then, and then evaluation of models. **Swyx** [1:07:27] Awesome. Yes, any, um, specific engineering roles that you're hiring for? And especially like who are people that succeed at your company as engineers, I think that's always the most important thing. ### Hiring **Qasar Younis** [1:07:37] Yeah, I mean, uh, fly.co/careers, I think there's literally hundreds of roles. Uh, we're looking at all the topics we talked about from, you know, dev tooling and physical AI to operating systems to autonomy and, and, and, and AI, uh, within physical machines. The types of engineers, that's a great question. That's actually more interesting than than- ... the roles 'cause we're, you know, we're a larger company. We're roughly- **Swyx** [1:07:58] Hiring everything. **Qasar Younis** [1:07:58] Yeah. We hire everything. **Peter Ludwig** [1:07:59] Yeah. **Qasar Younis** [1:08:00] Yeah. I think, you know, we're a Sun- Sunnyvale company, and, um, I think just from this conversation and kind of our backgrounds, you can kinda predict a little bit of what that means. Uh, you know, we tend to hire fairly serious people, um, who are-- who understand low-level systems, not just the, just like a, as a superficial understanding of technology, uh, like engineers-engineers almost. We definitely hire folks who are, uh, like have some diverse skill sets. We hire tons of specialists as well, to be very, very clear, but they've seen production, and I think that-- 'cause that really informs how you, how you build technology. **Peter Ludwig** [1:08:39] Yeah. People that really appreciate the hardware-software boundary. **Qasar Younis** [1:08:42] Yeah, exactly. **Peter Ludwig** [1:08:43] Uh, like definitely in the vibe coding era, there, there are a, a crop of engineers that they don't think about hardware at all- **Qasar Younis** [1:08:51] Mm-hmm **Peter Ludwig** [1:08:51] ... and we don't have that luxury. And so people that are a little more passionate about going a little bit deeper. **Qasar Younis** [1:08:55] Yeah. If you're to contrast us versus like a, you know, a AI lab or something, that's where you're gonna get the biggest contrast, which is like we're just dealing with, with reality. I mean, what other things? All the other classic stuff. You know, you want, you want folks who work hard and who are, uh, who love the technology and like, like a podcast like this rather than like if you made it to this part of the podcast- ... you probably qualify for you're interested in this. **Peter Ludwig** [1:09:24] Yeah. **Swyx** [1:09:25] Uh, and Peter said that he, uh, likes the podcast as well, which is like- **Qasar Younis** [1:09:28] Yeah, yeah. **Peter Ludwig** [1:09:29] Yeah. I'm a fan. I'm a fan. **Qasar Younis** [1:09:30] Yeah. **Swyx** [1:09:31] Specifically on the hardware-software boundary part is, is something I think about of, of our education system, um, in the States, but also maybe just in generally, I feel like there is that retreat away from that classical computer science or EE education. **Qasar Younis** [1:09:45] Computer engineering or yeah. **Swyx** [1:09:46] Yeah. **Qasar Younis** [1:09:47] Yeah. **Swyx** [1:09:47] And like is there a point where you just do it yourself? Like, you know, 'cause y- y- at this point you guys are the world experts on this, and actually you shouldn't wait for some college system to spit them out for you. **Peter Ludwig** [1:09:58] Oh, you mean in terms of education and upskilling kind of thing? **Swyx** [1:10:00] Yeah. Yeah. Just, just grab like young- **Qasar Younis** [1:10:02] General Motors already did it. **Peter Ludwig** [1:10:03] Yeah. GMI. **Qasar Younis** [1:10:05] Literally. **Swyx** [1:10:06] The Harvard University. **Qasar Younis** [1:10:07] Yeah. That's where I went, uh, for undergrad. I went to the General Motors Institute. Uh- **Swyx** [1:10:10] Okay. **Qasar Younis** [1:10:11] Yeah. **Swyx** [1:10:11] That did not come out. I saw HBS. **Qasar Younis** [1:10:13] Yeah, yeah. **Swyx** [1:10:13] I didn't... **Qasar Younis** [1:10:14] Everyone sees HBS. The, the, the Harvard brand Lewis is, is, is high. **Swyx** [1:10:21] What's General Motors Institute like? **Qasar Younis** [1:10:22] Uh, it started a hundred years ago to answer this exact question, literally the question you just said, which is like- **Swyx** [1:10:27] Okay **Qasar Younis** [1:10:27] ... not enough engineers in Michigan. Uh, you know, you're talking about the early days of the modern corporation. **Swyx** [1:10:32] Yeah. **Qasar Younis** [1:10:32] General Motors being-- There's a great book, Alfred P. Sloan's, uh, My Years with General Motors, uh, that is highly recommended, which basically talks about what becomes a modern corporation. But a part of that is they're like, "We are-- We're basically buffering on engineers." So they started a school and, uh, actually even Google as, as, as most, as recent as probably ten years ago was thinking of starting a university internally. There was discussions on it. So yeah, it was abso-- We, we definitely up-- I mean, we definitely upscale folks as well. The amount of training we do internally is actually surprising. Yeah. But it's a luxury you have when you're at our size. **Swyx** [1:11:06] Yeah. **Qasar Younis** [1:11:06] When you're like twenty-five engineers- **Swyx** [1:11:08] No **Qasar Younis** [1:11:09] ... you just gotta survive. So again, take advice that's relevant for your company rather than like immediately start trying to take high schoolers and make them engineers. **Swyx** [1:11:17] But I, you know, like I, I did go to a class that w- that you taught 'cause like it sounds like you can teach a lot. **Peter Ludwig** [1:11:22] Yeah. Well, I think honestly, the-- one of the most amazing use cases of these large models now is education, right? **Qasar Younis** [1:11:29] Mm-hmm. **Peter Ludwig** [1:11:29] Like I, I've, I've taken, uh, an engineer who, uh, v- very good engineer, aerospace engineering background, and in a relatively short time- time span, like he's doing very confident front-end work, very confident back-end work, like with the help of these models. **Qasar Younis** [1:11:43] Yeah. **Peter Ludwig** [1:11:43] And like not only can you do the implementation with them, but you can also just learn, right? It's like you ask questions and you don't feel embarrassed 'cause the model's not gonna- **Qasar Younis** [1:11:51] Right **Peter Ludwig** [1:11:51] ... model's not gonna call you out on anything. **Qasar Younis** [1:11:53] Yeah. I think the, I think the thing you probably need more than an engineering degree, though engineering degrees are like very important. Like I, I don't know if there's a way to shortcut like fluid dynamics or heat transfer- **Peter Ludwig** [1:12:03] The fundamental stuff **Qasar Younis** [1:12:04] ... the fundamental stuff, uh, at least on the mechanical side, is you need an engineering mindset. And that sometimes is actually-- not everybody actually has that. Some people are emotionally drawn towards the arts or something else, and it is completely fine. There's no judgment there. But I think the engineering mindset maybe in a more usable way is like wanting to understand a lower level and the lower level and the lower... Like, like how do photons move? **Peter Ludwig** [1:12:29] And extreme curiosity. **Qasar Younis** [1:12:30] Extreme cu- curiosity, like what is light? What is a radio wave? Like these really fundamental questions. **Peter Ludwig** [1:12:36] Right. If and, uh, if you get curious enough about software, you ultimately end up in hardware, right? **Qasar Younis** [1:12:41] Yeah. **Swyx** [1:12:42] Yeah. **Qasar Younis** [1:12:42] And so- **Swyx** [1:12:42] That's the L and K quote, yeah. **Qasar Younis** [1:12:43] Yeah, exactly. **Swyx** [1:12:44] So I'm trying to make analogies and then do all these things like, you know, you're kind of a blend between new General Motors and Tesla autonomy division for everyone else. **Qasar Younis** [1:12:53] I mean, uh, you know, we do work in all these other fields. I think if you talk to our trucking customers, they wouldn't even think-- perceive, you know. They like some sense like, "Oh, you guys did some automotive stuff, but you're, you're really helping us." So- **Swyx** [1:13:04] Automotive is not trucking? **Qasar Younis** [1:13:06] No, no, I know. That's the, it's like- **Swyx** [1:13:07] It's a whole- **Qasar Younis** [1:13:07] Yeah. Yeah. **Swyx** [1:13:08] Okay. **Qasar Younis** [1:13:08] It's, it's, it's, it's separate. There's different problems. The masses-- And you have, you have the general categories of on-road and off-road. I think that's what you're thinking. So there's on-road and off-road, uh, but within on-road, there's all these sub- subclasses of the machines. **Swyx** [1:13:20] Ah. **Qasar Younis** [1:13:20] Especially when you talk about, you know, you know, you look at, uh, a delivery robot that doesn't have a human in it. That's actually very different because now you're not concerned with like the actual feeling that you have when you're in a self-driving system. You don't have to account for that. You can- **Swyx** [1:13:34] Just break. **Qasar Younis** [1:13:35] You can, you can brake hard, and you don't care about jerk and all of these metrics don't, or become, become insane. **Peter Ludwig** [1:13:40] The way to think about it, h- honestly, is a little bit like, uh, any system that you as an, as a human would need special training to operate, you can think of it a little bit differently. So like the license to operate a truck is different from the license to operate a car- **Qasar Younis** [1:13:50] Mm-hmm **Peter Ludwig** [1:13:51] ... which is different from the license to fly a plane. It's different from... You get it, right? **Swyx** [1:13:54] Awesome, guys. Thank you for taking the time. **Qasar Younis** [1:13:56] Yeah, thanks for having us. **Peter Ludwig** [1:13:57] Thanks for having us. **Qasar Younis** [1:13:57] Yeah. **Peter Ludwig** [1:13:58] Thank you. --- This library is powered by PodHood (https://podhood.com), the podcast website platform.