1000 Designs a Day: Neural Concept's Thomas von Tschammer on AI-Native Engineering

Watch Episode Here

Listen to Episode Here

Show Notes

Research Notes: AI for Engineering with Thomas von Tschammer (Neural Concept)

Engineering physical products has gone through exactly two revolutions in the last forty years — and Thomas von Tschammer, co-founder and Managing Director US of Neural Concept, argues we are living through the third. First came the move from physical prototypes (build a car, crash it into a wall, rebuild) to computer-assisted design, where finite element analysis software could simulate a crash instead of staging one. That took the industry from five or ten prototypes a year to fifty or a hundred simulated designs. But those numerical solvers remained the bottleneck: a single crash simulation can take days on a large compute cluster. The third revolution — physics-aware AI models that learn from simulation and test data — collapses days into minutes, and fifty designs a year into thousands. Neural Concept, spun out of EPFL in 2019 with model architectures that ingest 3D geometries directly and predict aerodynamics, deformation, and temperature, sits at the center of that shift.

The headline proof point comes from Jaguar Land Rover, which presented its work with Neural Concept at NVIDIA GTC (session video). JLR's external aerodynamics workflow — already highly parallelized on traditional solvers — topped out at about fifty design evaluations per day, mediating the perennial trade-off between the studio team's aesthetics and the aerodynamicists' drag targets. With AI in production, JLR now evaluates 1,500 designs every single day, a 30x jump. Suppliers designing battery cool plates have cut development cycles by 80% — and because exploring more designs means finding better ones, they are also shipping cool plates that cool 20% better on batteries 15% lighter. Crucially, Thomas is clear that AI is not replacing numerical simulation any more than CAD replaced prototypes: simulation moves later in the process and gets used more surgically, while AI handles the wide early exploration.

Architecturally, Thomas's bet matches what he calls Jensen Huang's "five-layer cake" of AI: general-purpose frontier reasoning models at the core, with companies like Neural Concept building the application layer on top — the tools, skills, and context that let an LLM-class reasoner work with 3D geometries, injection-molding constraints, and CAD systems. A plain LLM, he notes, "has no very accurate 3D reasoning" and is "nowhere close to solving actual fluid dynamics equations." The winning paradigm is capable general-purpose reasoners equipped with the right tools — including the ability to call Neural Concept's specialized physics-aware models, send new geometries to CAD, and dispatch validation runs to high-fidelity solvers automatically. Today there is no off-the-shelf foundation model accurate enough for every OEM's aerodynamics, so models are trained and fine-tuned per customer on their proprietary simulation and wind-tunnel data — which doubles as a knowledge-retention flywheel: every data point an engineering org generates gets captured by the model and compounds into faster development cycles. Foundation models for aerodynamics are coming fast, Thomas says — it's the low-hanging fruit — and Neural Concept's strategy is to be the domain-specific layer that makes them deployable inside hundred-thousand-person organizations regardless of who trains them first.

The competitive stakes are stark. A Western European or US OEM takes 48 to 60 months to develop a new car from launch decision to manufacturing; in China it's 18 to 24 months. Thomas attributes the Chinese advantage more to manufacturing agility — highly automated plants that Western engineering executives now tour for lessons — and to freedom from legacy processes than to faster design per se: Chinese OEMs "can just pick what's best out there today, not what was best yesterday." He predicts massive disruption and exponential gaps between companies that adopt AI-driven engineering workflows and those that don't, with digital-native hardware companies (consumer electronics firms building hardware for only a decade) adopting far faster than hundred-year-old OEMs. His prescription for a legacy OEM: in year one, make every iteration in flagship disciplines — crash, aero, powertrain — AI-led, for 20-40% speedups; in year two, break the silos so agents orchestrate across crash, aero, thermal, and manufacturing constraints simultaneously, where gains compound to 50-60% reductions in development cycles.

The Formula 1 material is a highlight. Neural Concept serves multiple F1 teams, which operate under an explicit regulatory cap on the CPU hours they may spend on aerodynamic simulation — and the cap is a sliding scale: the better you finished last season, the less compute you get, because regulators can tie simulation compute directly to on-track performance. (Nathan flags this as a suggestive precedent for AI governance.) F1 engineers are also the picture of the "token-maxing" engineer: they translate the next race's profile into aerodynamic requirements, kick off AI-driven workflows that generate and evaluate tens of thousands of design configurations overnight, and arrive in the morning to an interactive trade-off dashboard from which they pick the design for Saturday's race. Because F1 teams iterate the car between every race, Neural Concept uses them to stress-test the workflows every OEM is tending toward.

Those overnight explorations produce what Nathan frames as Move 37 moments: engineers report designs they "would have never thought would be good" — designs they'd have scrapped on sight — that outperform anything they could come up with, sending them back to the dashboard to reverse-engineer the AI's choice and rebuild their own intuition. Thomas calibrates the surprise carefully: the models remain grounded by physics ("you will not reinvent the physics"), but they routinely escape the bounds of human intuition. On value, the math is blunt: a supplier winning one more battery cool-plate program is worth tens of millions; an OEM compressing a billion-dollar car development program by 20% saves hundreds of millions; and a part that's 10% cheaper because engineers spent their reclaimed three months tuning the manufacturing process wins programs worth hundreds of millions more. Pricing, he argues, will accordingly migrate from seats to value as agents rather than individuals do the work.

The conversation closes on the long arc. Thomas sees no fundamental capability gap blocking an end state where RL-style loops span the whole product cycle — spec ingestion, design generation, simulation, even virtual market testing — with the binding constraints being infrastructure, governance, and data flows rather than model capability. Engineers, he insists, keep the final trade-off decisions: a car is a multidisciplinary optimization with effectively infinite coupled constraints, and the engineering IP a company encodes into its AI workflows is precisely what will differentiate one OEM from another — otherwise "a GM car would be similar to a Ford would be similar to a BYD." On why cars all look alike, he points to autonomy: most robotaxis are drivable cars retrofitted for autonomy, like Waymo's Jaguar I-PACE fleet or even the Cybercab, while autonomous-first designs like Zoox ("it's not a car") show what unconstrained form factors look like. Humanoid robots in plants? Real, but second-order — general intelligence applied to design and machine tools comes first. His parting message: AI-first product design is not a forecast; it is already happening in automotive, aerospace and defense, and consumer electronics, and the next breakthroughs in those industries will carry AI-driven workflows inside them.

Topics covered

(00:03) How cars get designed: 40 years from physical prototypes to CAD and finite element analysis
(00:08) The physics domains of a car: aerodynamics, crash safety, thermal management, electromagnetism, structural dynamics
(00:13) Hybrid training: combining numerical simulation with wind-tunnel and test-lab measurements
(00:16) Why models are trained per-customer — engineering know-how as data, and the knowledge-retention flywheel
(00:18) JLR's production results: 50 → 1,500 design evaluations per day; battery cool plates 80% faster, 20% better cooling
(00:22) Assistant, not black box: AI proposing design spaces while domain experts make the trade-offs
(00:25) Frontier reasoners + tools + specialized physics models; Jensen Huang's five-layer AI cake; Neural Concept's 2019 origins in 3D deep learning
(00:29) Specs in engineering vs. software: RFQs, human interpretation, and coupled constraints
(00:32) Could frontier models design next year's car? Where engineers hold the final trade-offs
(00:35) Pricing in the agent era: from seats to value
(00:36) Disruption math: 48–60 month Western development cycles vs. 18–24 months in China
(00:41) Designing for manufacturing: the additive-manufacturing lesson; embedding manufacturing constraints in the model
(00:44) Decomposing China's advantage: plant agility and freedom from legacy processes
(00:46) Foundation models for aerodynamics: the low-hanging fruit and Neural Concept's strategy
(00:48) Formula 1: sliding-scale CPU-hour caps on aero simulation as performance handicapping — an AI governance precedent?
(00:57) The token-maxing engineer: tens of thousands of designs overnight, trade-off dashboards by morning
(00:59) Move 37 moments: AI designs engineers would have scrapped that outperform — grounded by physics, beyond intuition
(01:03) Quantifying value: programs won, billion-dollar development budgets compressed, manufacturing time reclaimed
(01:08) A two-year roadmap for OEMs: AI-led iteration in year one, cross-discipline orchestration in year two
(01:11) Engineers as artists: adoption resistance, ML teams as early adopters, what flips the skeptics
(01:15) Why cars all look alike — autonomy, commoditization, and new form factors (Waymo, Zoox, Cybercab)
(01:20) The end state: RL loops across the whole product cycle; speaking products into existence
(01:22) Humanoid robots in plants: necessary breakthrough or second-order problem?
(01:25) Closing: AI-first design is already shipping in automotive, aerospace and defense, and consumer electronics

Resources

Neural Concept

Neural Concept — the AI-first engineering platform (platform overview)
AI Design Copilot launch (CES 2026) — physics- and geometry-aware copilot
$100M Series C led by Goldman Sachs (Dec 2025)
Thomas von Tschammer on LinkedIn
EPFL — where Neural Concept spun out in 2019

Quotes worth pulling

"Thanks to AI, you don't get results in days, but you get them in minutes. And if you can get results in minutes... you don't explore 50 designs a year, maybe 100 designs a year, but now thousands of different options." — Thomas von Tschammer

"They went from fifty designs evaluated per day to one thousand five hundred every single day in production." — Thomas von Tschammer, on JLR

"Today a Western European or US OEM takes between forty-eight to sixty months for a new car development... In China, it's eighteen to twenty-four months. So everything those companies in the US and Europe care about is: how do we get from sixty to twenty-four months?" — Thomas von Tschammer

"The AI model came up with a design that I would have never thought would be good. If you had showed me this design, I would have said, 'Hey, scrap this, this is not going to work.' But actually, those designs are better than anything we could come up with." — Thomas von Tschammer, relaying an engineer's Move 37 moment

"You remain grounded by physics — you will not reinvent the physics, that's for sure... But we've seen scenarios where engineers thought it was a mistake, thought it was a blunder, similar to Move 37 — but it actually was not." — Thomas von Tschammer

"How will the differentiation happen tomorrow between OEM one and OEM two? It's going to be the ones that are able to deeply ingrain their engineering IP into these AI workflows." — Thomas von Tschammer

"Competing with Japanese companies in the '70s and '80s is potentially going to be easy mode compared to competing with AI-native companies if things go a certain way." — Nathan Labenz

Mercury: Command is Mercury’s new conversational interface, giving you natural-language access to your finances and helping you take actions within your existing permissions and approval policies. Visit https://mercury.com to learn more and apply online in minutes.

Sponsor:

Claude:

Claude by Anthropic is an AI collaborator that understands your workflow and helps you tackle research, writing, coding, and organization with deep context. Get started with Claude and explore Claude Pro at https://claude.ai/tcr

CHAPTERS:

(00:00) About the Episode

(03:52) Special Sponsor

(05:40) AI design revolutions

(12:00) Physics models and data (Part 1)

(18:40) Sponsor: Claude

(20:32) Physics models and data (Part 2)

(21:56) Copilots and workflows

(33:22) Automation versus engineers

(40:39) Industry speed gaps

(48:26) Foundation models and racing

(58:03) Surprising AI designs

(01:06:15) Adoption and differentiation

(01:17:02) Robotics and abundance

(01:24:36) Episode Outro

(01:28:10) Outro

PRODUCED BY:

https://aipodcast.ing

SOCIAL LINKS:

Website: https://www.cognitiverevolution.ai

Twitter (Podcast): https://x.com/cogrev_podcast

Twitter (Nathan): https://x.com/labenz

LinkedIn: https://linkedin.com/in/nathanlabenz/

Youtube: https://youtube.com/@CognitiveRevolutionPodcast

Apple: https://podcasts.apple.com/de/podcast/the-cognitive-revolution-ai-builders-researchers-and/id1669813431

Spotify: https://open.spotify.com/show/6yHyok3M3BjqzR0VB5MSyk

Transcript

This transcript is automatically generated; we strive for accuracy, but errors in wording or speaker identification may occur. Please verify key details when needed.

Introduction

[00:00] Hello, and welcome back to the Cognitive Revolution!

Today my guest is Thomas von Tschammer, co-founder and US Managing Director of Neural Concept, a Swiss company that uses specialist models for domains such as aerodynamics, heat dissipation, and collision safety to help automotive manufacturers, and other clients, accelerate their product design and engineering processes.

As a Detroit, Michigan native, this topic is of particular interest, because my father actually started his career at General Motors in the drafting department, back when designs and assembly instructions were hand-drawn on paper, and I have vivid memories of watching him use early Computer-Aided Design platforms on take-your-kid-to-work day when I was a young boy.

The work, at that time, was still highly manual and often quite intuitive, but as in so many fields, it's become far more computerized over time.

By the time my dad retired, designs were routinely tested via physics-based digital simulations before the physical manufacturing process began, and this increased iteration velocity by an order of magnitude, but still… as we've seen in biological structure and binding prediction to materials science to robotics controls, the compute required to run these simulations often became a bottleneck unto itself.

Today, as you'll hear, Neural Concept’s models can deliver similar results to expensive physics-based solvers in minutes and they now also offer an Engineering Copilot product which can call the domain-specific prediction models as tools, and actually use the core CAD platforms to make design changes as required.

This tick-tock combination of agentic optimization and domain-specific validation is the perfect recipe for Reinforcement Learning, and already, it allows manufacturers like Jaguar Land Rover to conduct aerodynamic testing on more than 1000 designs per day; frees human engineers to explore much larger regions of design space, and to focus their attention on navigating higher-level trade-offs with other parts of the organization; and occasionally produces surprising, Move-37-like designs that outperform human designs and actually alert human engineers to new possibilities.

Neural Concept has even found a niche in Formula 1 racing, which, I was surprised to learn, limits the amount of compute that teams can use for aerodynamic optimization from one week to the next.

The bottom line is that we can add engineering to a long list of domains where essentially the same pattern of development is working over and over again. What once could only be done manually in the physical world was first digitized, and then dramatically accelerated with specialist models.

Today, agentic workflows are accelerating things further and Neural Concept is beginning to evolve from training models on a per-customer to a future of more general purpose foundation models for engineering.

All of which makes it pretty easy for me to imagine a future engineering superintelligence that combines the general-purpose design skills with super-human intuition, all in the same set of weights. As we reach that point, and probably even before, we can expect faster and faster product cycles and an explosion of new form factors, all with higher quality and better resource efficiency than we've ever experienced before.

If you've ever felt that promises of AI abundance were a bit too hand-wavy or detached from physical reality, I think this episode should serve to inspire you, and so, I hope you enjoy this preview of the AI-powered future of engineering, with Thomas von Tschammer of Neural Concept.

[03:52] The cognitive revolution is brought to you by Mercury, the fintech that more than 300,000 ambitious companies and individuals trust to run their finances. I've wired AI into nearly every corner of my life. My e-mail, my messages, my calendar. I even gave Mercury virtual cards to my agents with low limits and category and merchant restrictions for their autonomous use. But still, my AI's access to my financial data has remained limited. With a normal bank, I might export a bunch of statements and have my assistant process them for me. But for real-time, up-to-date information, and certainly for taking any action, trying to get your agent to use the bank via the browser is just too hard, too slow, and too error-prone to be worth it. And that's why Mercury's new conversational interface, command, is such a big deal. It's built directly into Mercury, which means you get natural language access to your finances without exposing anything outside of your bank account. No exports, no spreadsheets, no pasting your transactions into third-party tools. I really think a lot of people are going to prefer it this way. And it can already help you take actions too, with everything bound by the permissions and approval policies that you've already set up in your account. I am genuinely impressed to see this level of AI integration in banking in 2026. And so I invite you to join me in the future. Visit mercury.com to learn more and apply online in minutes. Mercury is a FinTech company, not an FDIC insured bank. Banking services provided through Choice Financial Group and Column NA, members FDIC. Thank you to Mercury for supporting the cognitive revolution and now on with the show.

Main Episode

[05:41] Nathan Labenz: Thomas von Schammer, co-founder and managing director of the United States at Neural Concept, welcome to the Cognitive Revolution.

[05:48] Thomas von Tschammer: Thanks very much. Thanks for asking me.

[05:51] Nathan Labenz: I'm excited for this. We have not done much on AI for engineering on this feed, and with 350 episodes under our belts, probably a bit of a miss. Especially because I'm sitting here in Detroit, Michigan, where I know you have some customers and where there is a long tradition of engineering physical products for the physical world. My dad actually, fun fact, worked at GM. He started back when the drafting was still done on pencil, paper, big tables with slide rules and stuff like that. And then he moved into the CAD era and now he's retired. And so he's not going to be working through the AI-assisted, increasingly AI-automated era. But I still am very excited to kind of pick up where I left off the thread with him on Take Your Kid to Work Day years ago and fast forward to where we are now. With that in mind, folks who listen to this feed are like very into AI. And there's certain concepts that you don't need to introduce. But I think we probably can't assume common knowledge in terms of what does the life of an engineer, what does the life of a neural concept user look like if you go watch over their shoulder and see them at work for a little representative sampling of their working life? Maybe you could give us a bit of a sense of, in brief, obviously, how do things get designed? Who's doing that? What are the key skills? What are the key iteration loops look like? Before AI, and then we'll obviously add that AI layer to our understanding.

[07:27] Thomas von Tschammer: I can start with the example of the automotive industry, right? Julian is a good example, Nathan. So we've been designing cars essentially for the same way for the past four years, right? Four years ago, before CAN, which is computer-assisted design, when you wanted to design a new car, you were essentially building a prototype, right? And you were crushing that prototype against the wall, and ensure that are looking if the pedestrian or the passengers were safe inside the car. And then if not, you'd have to change the design and then rebuild the new prototype. As you can imagine, this was a very lengthy process, which means that ultimately you could explore 5 and 10 prototypes a year if you were looking before to go to production. Then about 30, 40 years ago, let's say 40 years ago now, we moved into computer-assisted design. Right? So which means that now, four years ago, we could build new designs on the computer directly. And then instead of crashing it in real life against the wall, we could simulate through what's called numerical simulations, the effect of that car crashing against the wall, but directly on the computer. Right? We use for that software that we call FEA software, Finite Element Analysis software, that would predict, simulate the effect on the crash. Since you now, you don't have to build as many prototypes. You could go to at least 50, 100 of different designs of car a year, which is substantial, right? substantial to the app. However, these tools, they remain very complex to use and very expensive, right? A single crash simulation can take days to be run because we are solving the questions of physics on the computer. You need very large clusters, typically. and then you need to wait as engineers several days, maybe one or two days, to get that result out, which today is still the main bottleneck when you want to iterate on your design, right? And there's been improvements, of course, in the algorithms, compute that we have so that we could spin that up. But essentially, for the past 30 years, we've been using the same CAD tool and we've been using the same SEA numerical simulation source. And that's true for crash, but as I say, aerodynamics, for thermal management, that's for every single physics that you need when you build the car. And now using AI, we're seeing that thermal revolution, right? So from product to CAND and not theoretical solvers to AI, where thanks to AI, you don't get results in days, but you get them in minutes. And if you can get results in minutes, it means that you don't explore 50 designs a year, maybe 100 designs a year, but now thousands of different options. And that drastically accelerates your development cycles. That also means that you as an engineer can innovate much further because you have more options that you can explore thanks to these animals.

[10:33] Nathan Labenz: So that is a real echo of a pattern that I see across all kinds of different spaces right now where there's this ability for models to learn a sort of intuitive physics, I sometimes call it, maybe most famously in protein folding, right? We've had a similarly hard time in the past, either doing crystallography to eventually get to a protein structure or doing really compute intensive simulation to get there. And now somehow with enough data and the magic of learning, we can take a couple orders of magnitude out of the compute that's required. And that just changes the game in terms of how many designs we can explore. So that pattern, I think, is fairly familiar. What I realize I don't have a great intuition for is like, what are the different flavors of intuitive physics that models that we need to get models to learn in order to accelerate what different subdomains of engineering. You alluded to one a little bit with aerodynamics. And so I know that'll be prominent on the list. But how many different things are there like this? And what are the unlocks? What are the sort of fields that to which they apply? What are the unlocks associated with them? And what has neural concepts role been in building these models?

[12:00] Thomas von Tschammer: Yeah. Great question. So there are many different domains, as you can imagine. I mean, think about the complexity of a car, right? Now, today in the industry, a GM or another OEM is simulating the entire car when developing it, which means that every single component or sub-assembly within the car is being simulated, is being evaluated, and is being integrated on, right? So that means that we are, as an engineer, we are evaluating many different physics on the car. Aerodynamics is 1, specifically critical for EVs, right, on the range. You want to improve the range on your next EV. Crash safety for pedestrians and passengers is another big one. Then we also have thermal management, when we want to cool the batteries, hold the engine, but also for ventilation systems inside the car, right? Electromagnetism, when we want to build an acceleration of electric motors, right, there is a big electromagnetism aspect to it. And then structural dynamics, generally speaking, for the car, your ability to the chassis or different Rd. profiles, and so on and so forth. And just this thing, the main categories, of course, as you can imagine, then this is being divided into component thermal assemblies. Ultimately, with a company like GM, you have thousands and thousands of engineers that domain experts on this physics, on those components, so that they can iterate and improve each of these specific assemblies. So those are the big domains, essentially.

[13:33] Nathan Labenz: Are all those different domains that you laid out now powered by a domain specialist model that has learned, for example, the intuitive physics of heat dissipation through a ventilation system? And how, if we just take that one example, if the old version was like actually having a simulation down to the level of I don't know if it was all this detailed, but going all the way down to molecules of air blowing through a space and how they bounce off of each other and what ultimately happens. What level of abstraction or sort of intuitive physics are we now able to get? How do we say to this like specialized model, here's a new design for a ventilation system. You like tell it, predict what's going to happen. What kind of inputs and outputs look like to those models? Are they trained on simulation data also? That's another thing that I've noticed is a real pattern.

[14:29] Thomas von Tschammer: Yeah. So to the last question, these models, these A models can be trained both on simulation data, but also on test data, external data, right? Because today, even today, there are some phenomenon that we're not able to simulate very accurately with traditional solvers. In that case, what we can do is that if we cannot simulate them, we can measure them in wind tunnel or in test. loops, right? And we can gather the data and then train the corresponding AMLs. So this is also a path of data. This is what we call hybrid training, essentially, where you combine numerical simulation that can be no more fidelity because we're not capturing the physics very well and measurements. Now to your question before, how far have we gone into that space? To be very clear, today we are not fully replacing numerical simulation. The same way that we never fully replaced prototypes, we're still doing prototypes today in the industry, right? But we're doing much less prototypes and much later in the process, right? It is going to be the same and it's the same for numerical simulation. We're going to fully replace numerical simulations, but we're going to make a much smarter usage of it for the most mature stages of development. And it's going to be exactly the same with AI. AI, earlier on in the development process, we enable you as an engineer to explore a much richer space, to explore the right candidates, and then narrow down the one that you actually want to simulate to validate before going to prototypes, right, what we're saying. Today, LGRM is such a specific, accurate field that there is not one foundational model that can be used to solve the aerodynamics. on every car, right? There is research definitely going in that direction and we are sort of the forefront of it at our concepts. However, it is not yet able to capture the level of affinity that you would need, that every car OEN would need to be able to deploy it on the shelf. What does that mean? That means that we are retraining and we are training the models, the company's specific data and numerical simulations.

[16:37] Nathan Labenz: So a the way that you interact with customers, if I understand that correctly, is models are typically, it sounds like maybe trained from scratch on a per customer basis because they are sitting on top of a bunch of the simulation data and some real world test data. And they're like, man, it would be great to take a couple orders of magnitude out of this as we explore the sort of optimization space around the core decisions that we've already made. So we can't jump from like a sedan to a Cybertruck, perhaps, with models that we have available. But once we're in the zone, we kind of know where we're going to end up. We can refine dramatically faster because we're able to train these models on all this existing data and then feel like this is sort of the epicycle development. It's like we're really dialing things in at the end.

[17:38] Thomas von Tschammer: Okay, exactly. So think about data as knowledge, right, know-how. So essentially, you can, and we also provide for some specific applications, pre-trained models, right, that we're training from existing data elsewhere. But today, we need to fine-tune these models with the company-specific data, because they have their own know-how, their best practices, and you want the model to match exactly their requirements, right? Now, what is very interesting as well is that these models are always evolving all the time in the sense that every time you fill them your data, they are being retrained and improved so that they can cover a broader and broader space and become more and more accurate. It's also a way for the component to retain knowledge and know-how, right? Because now every single data point that they generate, this is knowledge that is being captured by the model and reused such that the next development cycle can be even faster in there.

[18:40]Claude: Claude by Anthropic is an AI collaborator that understands your workflow and helps you tackle research, writing, coding, and organization with deep context. Get started with Claude and explore Claude Pro at https://claude.ai/tcr

Main Episode

[20:32] Nathan Labenz: So now the other sort of complementary AI that's starting to get introduced at the same time is I think everything we've said so far is more around the validation side. You could imagine a human sitting there and maybe you can tell me like some of the shortcuts again that already exist before we get to the sort of AI copilot or engineer. But you can imagine a person sitting there, I have this vision of my dad doing this from years ago, drawing these little shapes in three-dimensional space and kind of manipulating points in the point cloud. And it was that human intelligence that would say, okay, here's the test result I just got. Here's the spot that the way in which we're having some aerodynamic problems. Let me go back in there, tweak the design a little bit, sand that rough edge down a little bit. Now, of course, I've got these other constraints too that I got to keep in mind. So I've developed a certain intuition for how I can make these changes without breaking other constraints that are really binding on me as I develop this system. Now I'll do those changes and how manual those are, they were quite manual when I watched my dad do it. Again, you can tell me some of the shortcuts, but then they go back into this solver, right? New big thing, of course, is that the AIs are also coming to full iteration cycles. I want to hear too about how with the AI copilot, like how is that experience changing for engineers? How automated is it starting to get? How automated is it likely to get in the massive distant future?

[21:57] Thomas von Tschammer: Yeah, very good points. On the metrics, I can give you a few examples that are public out there. The first one is Jaguar Land Rover, JL Arc, which is OEM out there based in the UK. They've published their work with us. at the latest NVIDIA GTC conference that was back in March this year. And they're using AI for all their external aerodynamic workflows, right? As I was mentioning a couple of years ago, they were using crypto numerical solvers. And even in your design, they were going through this very expensive solver, which time consuming. And they had already highly parallelized, highly optimized this stream. And they got to about 50 designs evaluated every day. So 50 designs and iterations between the studio teams, which are responsible for the aesthetic, the design of the car, and the aerodynamic system. And at the end of the day, it is a trade-off between the best looking car and the most aerodynamic, because you want to improve range, right? With AI, and that was what the publicly announced a couple of months ago, they went from 50 designs evaluated per day to 1,500 every single day in production. So you can imagine the level of spin up that it brought to them. We have other suppliers as well that are designing matrical place to cool the battery that were able to reduce by 80% their development cycles, becoming 80% faster to develop a new battery cold place. And on top of this, more design cycles, they could get better performance. right? Because now, if you can explore many more designs, it also means that you can innovate better. You can find new options that you could not think of before as an engineer, because you just rely on intuition, right? And we see examples where the company is cooling 20% better on the battery, it's 15% liner. If you think about aerodynamics, we also help engineers to find designs that are 2, 3, 5% more aerodynamic, which is can change of them.

[24:05] Nathan Labenz: Yes, let's dig into that last point. How is that happening? Again, to map this onto a space I've studied in a little bit more depth. There's the protein generation models now, right, which allow us to go beyond a biologist's ability to say, oh, there's a couple other proteins in the protein bank that are similar to this that I can pull in, or even I have a little intuition of my own. And now we've got models just throwing out new designs, which can then be evaluated. So where are we on this sort of language assistant with like tool calling paradigm to intuition of design. You can imagine a language freeway. You could imagine a version that's just, here's your constraints, here's the point cloud, here's the feedback that we got, generate a new point cloud. And you could just go point cloud to point cloud. Where are we on that in those relative paradigms? Do you, how are they starting to converge?

[25:05] Thomas von Tschammer: Yes, today what we're seeing is that this AI-driven engine workflows are not in that box, right? They are today an assistant to the engineer. so that the engineer can take the right data-informed design decisions, right? We are not in a world where we're just sending a spec sheet to the AI and expect the AI to get back to us as a black box with the best optimized car, right? We are rather in workflows where the AI ingest the spec sheets, understand the requirements, will set up the base model, but this will be done in interaction with the engineer that will validate steps along the way and will also guide the model. Why is that? It's because engineering is a very complex space where you never have one single answer. It's always trade-offs between disciplines, between costs, between design, between performance, right? So you want to have these domain experts in the loop that then ultimately contain the right trade-off decisions. But essentially, we're moving from a world where, as you were saying with your data, where in the past you were relying on intuition, hey, my data doesn't work, but I think if I tweak it that way manually, it should work out, right? To a world where now the AI is providing a space of solutions of options to the engineer. And then the engineer will explore that space and decide with which design they want to move forward. And that means that here in that process, what is very interesting is that the AI is able to interact with the different tools, right? With CAD, we talked about CAD. So now the AI is able to interact with CAD. to send your geometries, is able to interact with numerical simulation solvers, right? Hey, I want to validate this design I came up with, and sends that directly to the high-finity simulation solvers, and all of that becomes automated.

[26:56] Nathan Labenz: So what models are you serving as the co-pilots for engineers today? We're talking on Fable Plus One, and we're seeing all these incredible projects that people are just vibe coding their way to where 3D landscapes that are really becoming like extremely elaborate. The models are just kicking out. So it sure seems like there's been some important thresholds crossed in terms of the model's ability to reason physically and use physics, coding engines and things like that. I would expect on some level that Mythos class models are like pretty good at just sitting down, so to speak, and using a CAD product. And that sort of performance is probably like pretty hard to find aside from maybe one other provider. But then I also think, boy, the loop that you want to create is probably one where you can start to train based on the feedback of these validators, which have been the kind of crown jewels of the system so far, which as you pointed out, embody the company's know-how. What does that look like? How much are we going to be relying on a few frontier models to assist our engineers? How much do you think you're going to start to see this loop get closed and create other like specialist models that play that role instead?

[28:26] Thomas von Tschammer: Yeah, great question. So if you think about what's happening, yeah, Jensen mentioned that later. is seeing today AI as a five-layer cake from the foundations to the top. And the last layer at the very top is the application layer, which he explains himself as being the most important one. And today in the industry, there are many, many companies that are specializing in building that application layer for some very specific application, right? So it's actually taking the generic models, models, and then bringing the right domain-specific skills capabilities, such that these models can have value in very complex environments. This is exactly what we're doing in engineering, essentially, right? So we leverage the standard JRIP models, state-of-the-art that you can think of, right? But we are giving them the right set of tools, and they can work with 3D geometries, they can generate image geometries, specific set of skills, so they have the context about what a mill means to generate a geometry. What are injection molding design requirements? If I want to do a mix, what do I need to optimize for the right complex? But then they can have an input. Today, if you just take a plane, you will not be able to go very far because they have no very accurate 3D resorting. Yes, they are becoming more and more realistic for video rendering, for example, but they are nowhere close to solving actual fluid dynamics equations for extra aerodynamic, right? So there is a big gap to bridge here. And that's what we're doing, the company, not accept.

[30:08] Nathan Labenz: But it sounds like you're expecting reasoning to be mostly provided by frontier models for the foreseeable future. The paradigm you think is winning is Increasingly capable general purpose reasoners equipped with the right tools.

[30:29] Thomas von Tschammer: Exactly. Equipped with the right tools, I think, is the important part where these models are able, again, to interact with CAD, but with also other class of models, right? If we take a step back and we look at our history, how did we start? We started in 2019 by building the first AI model architecture. based originally on computer vision that could directly ingest really geometries and learn from physics. This is how we started. It was not LLMs back then, it was a different type of architecture, but these models could learn and directly take as inputs a CAD geometry and then predicts aerodynamics, deformation, temperature, and so on and so forth, right? But these are also models that you want as part of your workflow, right? And you want the agents to be able to call the specialized physics-aware models so that you can actually speed up overall design process.

[31:27] Nathan Labenz: So it would be helpful to do a little kind of comparison to software. I think audience is super diverse, but one common profile is the sort of AI engineer, software engineer who's now doing increasingly everything with AI, both in terms of writing the code, but also the products that they're building are increasingly AI-ified. And it seems like we're hitting this point where if you can specify what you want in a clear and accurate way for a really astonishingly large set of things that people might want, the AIs can just deliver that for you now. And they're also getting obviously pretty good at even flagging the areas where you were ambiguous and they might need your help to make a decision. So it's Putting the premium, of course, on the spec and the clear thinking about what it is that we actually want. I guess I don't really know. Intuition in engineering would be that these specs are like better typically than they are in software. I would imagine that there's a more disciplined process culture around saying exactly what we really need, because we know that there first of all are hard realities around things like heat dissipation and strength that we just literally have to have. Whereas in software, we kind of figure we'll patch that on the fly later if it's not scaling the way that it needs to. Something like literally breaking, we have the ability to kind of reach in and fix that, even if we're in production. Obviously not so with a car. So am I right that there is much better specs? And does that put models in a really just position, or are there ways in which specs are actually still not so well specified as the naive person might think, and we're relying on a sort of human fuzziness to unpack those such that it remains difficult for AIs in some ways?

[33:23] Thomas von Tschammer: Good question again. Unfortunately, if you are an automotive super OEN today, it's more than matter, right? Specific RSQ, as they call it, so request for quotation and specification, and they is still not fully streamlined, not fully automated, not fully defined. There are standards that OEMs are trying to impose and sets, right? But there is always human interpretation, especially when, for example, we think about a car, which is such a complex problem with an infinite number of dimensions and constraints, right? Imagine that if you change the thickness of a single component somewhere under the hood, this might impact the overall engine block, right? And you have a lot of constraints that are tied together, which means that ultimately it's not as deterministic as one might think, which is why those problems are also extremely complex to solve, right? Yes, starting from a set of specs that the model can read and translate that into 3D, this is exactly what's happening today, right, already. However, why it is not yet as black boxy as it can be for software engineer, it's because the dimension of the problem is much, much broader, much richer, right? You have many more trade-offs you need to make. And there's never one way to get to a solution, right? Which why it's not going to replace engineers anytime soon, but it's going to empower them to be faster, actually. It's going to remove or eliminate the low added value tasks That's for sure, that's already happening, right? Where the engineer needs to manually set up a new simulation, needs to manually go on the CAD and roll a new design. This is going to, this is being eliminated as we speak, right? But it's never going to replace the engineer taking those design decisions, even the demonstrator.

[35:17] Nathan Labenz: Are you sure? What does a wake-off look like there when it comes to sitting down thinking what would be a good product in the market, even if it's something like high level as that. Models are getting pretty good. And then there's all these steps from this sort of initial ideation to breaking things out into subsystems and thinking about the constraints that each of those has to have, and then throwing these things into actual 3D representations in these systems, setting up all the simulations. I'm increasingly struggling to find the place where I'm like, in that whole sequence of events, like here's the ones that the AIs can't do. And I think there's some we may not want them to do, or we may want to hold on to final judgment, final calls, all that kind of stuff. But leaving aside like what we want to hold on to, it doesn't feel like it's so far off that you could have a little society of fables pick up where the humans left off and literally like design the next model year of car. If that's crazy, why do you think that's crazy? What is the part that we're so far off on?

[36:33] Thomas von Tschammer: I don't know. So you have a good point. All of these tasks today, I strongly believe that with the models and capabilities that we have, can be automated instrument. These models have capabilities to translate any specification into a design. and then simulate that design automatically using solvers. All of that can already today be automated through agents, and that's what we're doing as a company. Where I believe there is another level of complexity is the dimension of the problem, right? Yes, this is being done today for specific products, right? Battery core plates, aerodynamics, crash, right? Where I don't believe this will become a blank box, is you add all the dimensions you want to have for a card, right? Because then that's where you really want engineers to hold on to that. I believe that we always want to have engineers hold on to the final trade-off and the final decisions, right? Because that's where you need the domain is deeply ingrained, right? And that's also where the differentiators are happening between you and the competitor, right? How will the differentiation will happen tomorrow between OEM1 and OEM2, it's going to be the ones that are able to deeply ingrain their engine and key into these AI workflows. That's going to be key. That's what many good companies are realizing.

[38:01] Nathan Labenz: Yeah, that's fascinating. What do you think the odds are that we get, like, I guess I'll only start with one about your alignment of your business model, the role of the human. I assume that over time you've probably had some sort of seat-based pricing and potentially also a sort of compute-based pricing for simulation execution. I'd just to know where you've been on that historically. But then as we go forward, do you think you actually can sustain a seat-based model or is the model going to have to move in another direction?

[38:37] Thomas von Tschammer: I think essentially we have to be based value, right? What's the value we're providing, right? And that's the way you price. essentially, the price for a value ratio, right? The question is, how do you monitor value? And I think it depends on the industry and the applications. But ultimately, we're going to move to a world where pricing is going to be based pure value you want to deliver for companies. And I think that's actually very healthy, right? That's what you want to have. But yeah, because the value is not going to be tied to an individual anymore to assist, but now potentially to agents, right? Pricing will evolve accordingly. And I think it's going to be the same for every company.

[39:13] Nathan Labenz: How much disruption do you expect to see in these like 100 year old industries like auto where obviously companies have changed a lot, but it's largely companies that were formed 80, 100 years ago that have consolidated and there's certainly been only the strong survive dynamic, but not too many new entrants recently, right? Like a couple. But when you think about the what matters most being like figuring out a distinctive way to encode your know-how into an AI flywheel process, especially one that like might even enter into its own kind of recursive self-improvement loop. I just talked to some OpenAI forward deployed engineers last week who are doing this for tax and pace with which they are able to convert feedback from a tax professional on something the AI did wrong. into an improved scaffold that prevents that error from happening next time. The case of progress is so fast right now that they're really rapidly climbing the hill, in their case, of accuracy of tax prep documents. But do you think that there's something fundamentally different about manufacturing that would make those hills really hard to climb? Or if you're good at climbing those hills, is it a moment where you actually think we could see new entrants into the market come up and rival the incumbents.

[40:39] Thomas von Tschammer: For sure. I think it's going to be a massive disruption in the markets. And I think it's going to create exponential gaps between the companies that are able to adopt this AI-driven issue and the ones that are not, right? And the gap is going to widen essentially over the past the next few years for sure, right? So there's going to be a lot of disruption. What is complicated for companies if you think about traditional OEM? is that they've been building cars essentially the same way for the past 20 years, right? Same teams, same tools, same know how, right? So you're asking engineers that have worked the same way in a long time. Now to change their thinking and even the governance of their own company, right? Seeing as some are embracing it faster than others, right? And our role is to make sure that we can support them in that journey, show them how others are doing it, what are the best practices out there, and move them through the process. But I also believe that, yes, there is also an opportunity today for a company that is building hardware to come into come fast, right? We talk a lot about these digital native companies, right? And we work with many of those outside of automotive, right, that are building electronic products, consumer electronic products, right? So those companies have been building hardware for the past 10 years, max, maximum, right? And you feel that they're able to pick up these new workflows much faster. Right. What do I mean by that? I mean that in a year, they've had massive impact, massive speed up of their development workforce across many applications, right? And within the slower pace in the larger, older companies, engineering companies out there.

[42:22] Nathan Labenz: Yeah, that's been the story of Detroit for quite a while. It's been companies that they came up, they got so big and so powerful and had such market dominance. that they forgot that they might need to evolve. And they've come a long way since then. I'm talking, that's like a 50 years ago phenomenon. They've come a long way, but they're now going to be faced, I think, with their most profound challenge ever. And competing with like Japanese companies in the 70s and 80s is potentially going to be easy mode compared to competing with AI native companies if things go a certain way. So it's going to be really interesting to watch the organizational dynamics and challenges of that.

[43:06] Thomas von Tschammer: When you think about it, some numbers today, Western Europe or US OEM, it takes them between 48 to 60 months for a new car developed. From the moment they want to launch it to the actual moment it hits the plant and it's being manufactured. In China, it's 18 to 24 months. Everything those companies in the US and Europe care about is how do we get from 60 to 24 months, right? Because that's what's going to create the next competitive advantage finally.

[43:39] Nathan Labenz: Yeah, that's a sobering stat. I mean, the number of iterations in a given time is a pretty hard deficit to overcome long-term. Quite sure how to phrase this question, but one thing I wonder about is, in a way, if we really super optimize the design process. It seems like at least in a naive way, we might end up making things really hard on manufacturing. The sort of, going back to like my dad's, the start of his career, there was literal like pencil on paper and like annotation of it should be this much, right? And the tolerances that that existed on the machining side were just a lot more generous than I think they are today. And I think maybe even you could imagine, again, if the design gets so optimized and we're really satisfying these constraints down to the absolute maximum in our designs, it sounds like that would create really super tight tolerances and really super difficult manufacturing challenges. And so how do you think about I guess maybe one answer is this is the role for the human engineer, but don't satisfy, you know, let's not satisfy ourselves with that answer. How should we be thinking about like just how far we want to push design, optimization, and when we need to meet manufacturing a little bit more in the middle?

[45:04] Thomas von Tschammer: It's a great question. So when we're saying that we want to optimize designs, we also want to optimize them for manufacturing, right? What do you mean by that? means that when I'm saying that building a car is a multi-disciplinary and multi-optic optimization. I also mean that because you want to incorporate design rules, manufacturing constraints as early as possible into your design duration, right? Because that was the promise of additive manufacturing 10 years ago, right? That you could design freely, because then you could print anything, right? But then we quickly realized that this would not work out because it's expensive, it doesn't scale in production. So now let's say you have a manufacturing plant and you have your stamping process. You know what are the manufacturing constraints. You know what you can and cannot do with stamping. But what you want to make sure of is that those design rules, those manufacturing rules, are embedded and are available for the AI model to consider as good as possible. So that whenever the model evaluates evaluate, explores in the design, it explores it, no way that it can be manufacturing, right? And that's a lot of the work that we are doing as well. We are embedding within the model this know-how to make sure that every single design being explored and generated is valid for manufacturing at reasonable costs. So starting the physics of some manufacturing and costs along the way that you want to bring earlier into the dream process.

[46:35] Nathan Labenz: If you had to break down the dynamics or realities that give the Chinese companies such an advantage in iteration time, how much of it would you say is on the design side and how much of it is on the manufacturing speed side? Obviously, those things can't be fully decoupled. Hopefully, you kind of get a sense of what I'm getting at. You know, if I show up with, are they doing their design a lot faster in China or are they getting from the point where they have a design that they like to, you're actually rolling a line? dramatically faster. My sense is it's a little bit more of the latter probably, but I'm not really sure how to think about what is contributing to that huge advantage.

[47:15] Thomas von Tschammer: Yeah. So I agree with you. I tend to think it's more than latter, right? That much more agile when it's about rolling out a new plant. Also the way the plant is being operated, right? It is highly, highly automated there. I mean, now you have engineering executives from Europe, from US. traveling to China to understand how they are setting up their plans, right, taking lessons from it, and then going back and applying those best practices in their country. This is happening already today, right? So the fact for sure. On the design side, I think the benefit that they have is that they can take more risks because, again, they don't have that legacy to work with, right? Legacy of processes, of tools, and they can just pick what's best out there today and not what was best yesterday. That's a big differentiator, right? So, I do believe they work with best-in-class tools, for sure. They're able to take more risks, again, because they are in the history of developments, and they have less processes that are deeply great, right? If you don't have processes that were built in the 2000s, it's much easier to work in a much more agile way today. Then we're going back to the digital native type of companies that are able to pick up new standards.

[48:27] Nathan Labenz: We, the quote unquote West, have our work cut out for us, it sounds like. How big of a deal is it going to be for Neural Concept to move from the one model per company based on their data paradigm to a more foundation model type paradigm? Foundation models obviously could be like with varying breadth, right? But simply going from one aerodynamic model per company to a general purpose aerodynamic model that would allow you to be like, what if we did a Cybertruck and I'd move from your historical product line to something quite different? Seems like it would be a huge value on lock. Then you could also imagine going even more modalities, right, and trying to bring, I say this all the time, so I apologize to listeners, but you haven't heard me say, integration that we see on language and pixels coming out of the nano banana omni sort of line does really show to me that integrating quite different data modalities into the same model is a super powerful thing to do. And then there's another version of the foundation model, which is all these constraints generate me the design in the 1st place, you know, as opposed to the validation side, which I was speaking about before. Do you guys have kind of a How would you describe your strategy on that? How big of a deal is that? How soon do you think we will get there? I assume it's got to be inevitable on some level. What's your sort of strategy to bootstrap into those things?

[50:04] Thomas von Tschammer: Yeah, So you're right, it's going to happen, right? It's moving extremely fast already. So we are going to have foundational model for aerodynamics. I think it's going to start with aerodynamics, right? This is the low hangings fruit today, because even the physics are complex, it is relatively similar. across components. It can be easily replicated. So it is going to happen very quickly, right? And of course, it's a very big difference, right? So we are doing actually research in that direction. However, we may not be the first ones to have that foundational model, right? But then, as we, and I'm going back to Jensen's final cache of AI, our role and also our focus is to make sure that whatever the foundational model is, We provide the right domain-specific capabilities that it can be fully integrated into a complex engineering environment in 100,000 people organization, right? So that you can visualize your designs, you can tweak the geometry as well, leveraging those conditional models and so on and so forth. So this is also our role and what we are already building towards so that when the foundational capabilities arrive, everyone will be ready for it and to leverage it.

[51:21] Nathan Labenz: One thing I learned in researching neural concept I thought was super interesting is that you guys are serving, in addition to a bunch of enterprise customers, a number of Formula One teams. This is, I've honestly never really been super into motor sports or engineering sports. I don't know a ton about it. But it does strike me that in the run-up to the countries of geniuses in a data center, we really do stand to learn a lot from highly competitive and performance-oriented organizations like Formula One teams that are under just this incredible pressure to turn things around quickly, right? So how does that look like right now? And another interesting detail was that apparently Formula One teams have a explicit, it's like one of the big rules that they work under is that they can only spend so much compute aerodynamic simulation. That was a real surprise to learn. Is that just to prevent an arms race? There might be something for like our AI governance listeners to learn from Formula One as well. But what do you think we should be learning from the, what are we seeing? What should we be learning from the Formula One users?

[52:38] Thomas von Tschammer: Yeah, I mean, that's super interesting, right? So indeed, today, Formula One teams are capped in, to be most specific, the CPU hours that they can run. So CPU hours is the computes. to run excellent aerodynamic simulations, right? And what's even more interesting is that depending on your RAM key from the previous year, you don't get the same numbers for the next season. The idea is that you want to try and make it more equal across teams, right? And you don't want to make it a race as to, okay, the biggest budget, the biggest compute, so I just win. Essentially, so they're trying to equalize that in some way.

[53:12] Nathan Labenz: So they handicap. Essentially, if you win, you get less. You get less confused. The NFL does, if you're first place, you have to play a first place schedule. The NBA does if you're like, they're now changing it because people have been tanking to try to get better draft picks. But I'd never heard of this level of actually changing not just the talent acquisition process or who the opponents are going to be, but actually changing the fundamental rules of the game itself in terms of how you are allowed to prepare from one race to the next. That is really interesting.

[53:48] Thomas von Tschammer: Now they can directly tie compute simulations that are being run to performance, right? Because if you can run more simulations, as an engineer, you can explore more designs and you can improve your CAF for them, right? Which is... in Formula One, right? You want to get the best car out there for the next race. So that's a way to balance the formula between different teams, actually. That's super interesting.

[54:12] Nathan Labenz: So what are we seeing in terms of their cultures, their practices that you think will diffuse into broader engineering, ultimately manufacturing culture?

[54:28] Thomas von Tschammer: If you think about it, Formula One engineers are the state of the art of engineering, the most agile teams you can think of. Design of the car is changing between every single race, right? From one week to another. You don't see it because it's very fine details, but the car is actually different, right? They are improving it week over week. So that means that they have to reach an extreme level of automation of design iteration processes and speed, right? We see in Formula One teams, we believe what every single OEM is trying to tend towards, right? Trying to aim for the way they work, the way they iterate, the way they take design decisions, right? It's a good way for us, essentially, to proof tests and to stress test our models, our workflows, our platform, right? Because if it works for an F1 team, we can build it for an F1 team, we can believe that then it can work for the more traditional OEMs out there, essentially. So that's really the way we're saying it, and we're asking those teams to really push the limits of the models and the workflows. to break phase, essentially, because that's when the break phase that we see where we have to focus and what we have to do.

[55:38] Nathan Labenz: So maybe again, we can make a little bit of analogy to software where we have the token maxers who are trying to really push the limits on what their agents can do for them and increasingly not writing code anymore. And then, of course, there's a lot of places where we haven't quite caught up. And so we're maybe still writing called the old-fashioned way, or we're like doing a little autocomplete or whatever that's useful and giving a little speed up, but it's still an assist in the old paradigm versus a genuinely new higher level of abstraction as the base place where a human spends their time operates. Could you paint a little bit of a picture for analog in engineering? What is the F1 person do when they are token maxing? What kind of, what are these moments of like key decision or sort of judgment on particular trade-offs that actually rise to their level? What does that look like? And I think we kind of know what the old school one looks like. So yeah, what is the token maxing F1 engineer's life look like today?

[56:44] Thomas von Tschammer: So they will do typically is that they would look at the next race profile, right? That has more terms than the previous one. And then they would associate that to a list of requirements on which they need to improve the car. Hey, we need to make the car better in straight lines for next realistic. We have many more straight lines and we're not going to have to overtake much more, right? So that's the baseline. Then they transcend that today into engine requirements in terms of how the car can improve and the aerodynamics of the car. And then they are running these AI-driven workflows so that are taking those aerodynamic requirements that have information, awareness about the geometry, about the 3D. And then overnight, the AI-driven workflow will generate 100, thousands of configurations of design options. We'll evaluate them, we'll evaluate the corresponding aerodynamic performances. And on the morning after, the engineer, the aerodynamicist, will have a dashboard, a report, interactive. He sees the thousands of points, data points on a dashboard, and he can look at the different trade-offs, look at the corresponding design, and pick the one they want to move forward for the next race, which is next Saturday, right? So, token maxing is essentially thousands, 10s of thousands of designs being explored overnight, fully automated by these AI models.

[58:04] Nathan Labenz: Have we seen any surprises come out of that process? Of course, the legendary move 37, it seems like increasingly with the Mythos class Fable models, we're starting to see Move 37s might be strong, but they're definitely stepping outside. I had a really interesting experience today where I, or yesterday, where I ran a skill that's a very familiar workflow. And Fable stepped outside of the workflow and proactively asked me a bunch of questions, actually presented a webpage to me to collect information. Not instructed to do that, never been part of the process before. Opus never did anything like that, but it took it upon itself. to say, I've got all these inputs, but I think I could use some more inputs to really do a great job. And here's what I need from you, the human, to really knock this out of the park. That was like a little mini Move 37 in that it was, I've probably done this workflow 50 times over the last few months and nothing like that had ever happened. What's the sort of most Move 37-like thing that we're seeing in engineering?

[59:12] Thomas von Tschammer: Yeah. We have very similar analogies, and I think to be fair, it's one of the favorite part of the job field. That's where it becomes very exciting. We have engineers that are using these workflows, right? So using AI to explore and asking the AI to explore these hundreds, thousands of configurations overnight. And when they come the morning after, they're looking at the results, and then they're getting back to us and telling us, hey, very impressive. The AI model came up with a design. that I would have never thought would be good. If you had shown me in this design like this, I would have said, hey, scrap this, is not going to work. But actually, those designs are better than anything we could come up with. And now I need to get back to the dashboard to understand why it isn't so much better, right? So I need to rethink my intuition because I didn't think it could be that good as a design, right? So you learn as well from these models. Again, because they explore this much richer space, they go out of bounds. They go beyond their intuition, similar to what you were saying, I think. So that's why it becomes super interesting because then something really clicks with the engineers, they become very excited because they understand that they can also learn from the model. I didn't get there. How can I learn from it? Because explore new physics or new phenomenal that I was not aware of when I was only working with intuition, essentially. So is it possible without even trying to reverse engineer the design themselves?

[1:00:38] Nathan Labenz: learning from the AI's advances and its occasional leapfrogs over us is definitely a really exciting, thrilling, slightly scary part of this new future. Could you give us a little bit more intuition for like how, like just how radical these moments are? It may be a little bit hard for somebody not in the domain to really grok it, but I'd love to try, to get a little bit better sense on kind of, are they really good optimizations or are they really like stepping out and exploring different regions of the design space that people, because I think what made Move 37 qualitatively so compelling was like, no human would have made that move. Like initially, I think the live stream commentators like thought it was a blunder, right? When we see these surprises in engineering, like how big of a surprise are they? Are we seeing like, oh, that's kind of interesting. I wonder if that could work. Or is it like, that looks kind of crazy, but in defiance of all my intuitions, it actually does work. Just try to help me calibrate on how big those surprises are.

[1:01:55] Thomas von Tschammer: Yeah, so you remain drawn by physics, right? So you will not reinvent the physics, that's for sure. So you will not get. completely insane design is breaking the physics because everything is corrupted by physics. However, we've seen scenarios where the engineer is telling us, hey, there's no way in the world that I would have done that design because I don't think, didn't think it could work, right? So now that I know it works, I need to go back to the dashboard and understand why it does, right? We have the physics as a baseline, we cannot break the physics, it's going to be there. But we've seen occasion scenarios where the engineers I thought it was a mistake. I thought it was a blender, right? So you had to move 37, but it actually was not. And it let them rethink the way they were approaching the problem and their intuition around the problem.

[1:02:44] Nathan Labenz: Maybe another way to think about this is how much is it worth? You talked about like value pricing earlier. And there's so much discourse right now around Are people going to be willing to pay for Mythos class models? Notably, the price originally previewed with Mythos has already come down a lot with Fable being significantly less than that original Mythos preview price. I'm not sure how often it is the case that these insights are like readily quantifiable in terms of money. So because it could be a little bit more efficient, that moves the needle on my range. I can say now my car gets this many miles on a charge where it used to be only this many. How much is that worth in the market? Obviously a whole other question. So how are people assessing the value of the, especially the, you speak to all aspects of it, including the smaller optimizations, but really interested in these sort of move the small scale or move 37 light type moments. How much value do people perceive in it? Are they able to measure it? Are they willing to pay? Are we going to see people continue to run the neural concept co-pilot, the engineering co-pilot with anything less than a fable model? Or is it going to be like, nah, you got to pay for the best because that's where all the insights come from and the bill is going through the roof, but like we have no choice. But to do that, to stay competitive, where do you think we're going to land in the short term on this willingness to pay question?

[1:04:24] Thomas von Tschammer: It's a very good question, again. Considering the value is key, and I mean, it's a decision we always have also with the companies we work with, right? Typically, when you think about design breakthrough, right, so much better performance than what you could get before, we have discussion with suppliers, right, that are selling to the big OEMs, the big OEMs down there. are similar aspects. If you can get a better design, probably means that you are much more competitive on the market. Probably means that you will win more projects, more programs with wins and research more controls. If today, let's assume you're building battery cool plates to cool batteries, right? If you can win one more program per year, that's millions, 10s of millions of dollars, right? Just one more program on the 350 you're winning every year. Right? So that's very tangible. But the other way to see it is, if I can get much faster to a good design, yesterday it took me six months, now it takes me 3 months. I can spend the other three months optimizing my manufacturing process to reduce the cost as much as possible and I become even more competitive to my clients. Right? And this has indirect dollar value as well. If your part is 10% cheaper in competition because you spent three months, you could allow yourself to spend 3 months working on the manufacturing process, tune in the details, then you win this one, two, three more programs that are 10 hundreds of millions of stocks. And then if you go on the OEM side, if you can shrink down your development times from 48 to 24 months, that's also millions of developments you're selling, you're selling for every car, right? A car is typically a billion, right, to develop or from scratch. But if you can even 20% of it, the amount is pretty quick.

[1:06:16] Nathan Labenz: Yeah, a lot of opportunity for savings in there. What do you think American car companies, to take one very salient example, or you can broaden it as well, what do you think they should set as their kind of critical milestones, like must-hit accomplishments with AI over the next Let's even just say one to two years. We know on that time scale that OpenAI is planning to have a very large chunk taken out of ML research itself in terms of automation. We know that we're already iterating at half the speed of the Chinese companies. We see some hard manufacturing places where the product cycle has accelerated. Look no further than NVIDIA for probably the most dramatic example of this. Forget about like what would get you initially like a, coffee spit take and laughed out of the room if you said it. What do you think is like the actual achievable speed up that you would, heart of hearts tell the CEO of GM, like this is what you really need to be able to do in terms of speed up if you want to be meaningfully, you know, durably competitive in the AI era.

[1:07:34] Thomas von Tschammer: So I think there are different scales, of course, to this. In year one, You'd be looking at some core departments, core areas, crash safety, aerodynamics, and powertrain, let's say. And you'd want to make sure that every single iteration is being AI-led. AI-led, I mean, there is an AI workflow that can orchestrate different tools. That's the base, right? And that can already lead you to 20%, 30%, 40% speed up on those flagship disciplines, right? That is for year one. Then year 2 and year. What you'd be looking at is that orchestration across disciplines, right? So you break the silos between crash, between arrow, between thermo, right? So that the agent can not only orchestrate the aerodynamic optimization, but can orchestrate it while taking into consideration the safety aspects, the manufacturing aspects, as we talked before, right? Automatic design constraints, right? And break those silos between teams and disciplines. Once you do that, then that's where the gains really compound, the benefits really compound. And that's where you can really break down and reduce the cycles from 50 to 50 to 60 percent, and that's what we are seeing already, right? Today, there is not an OEM that was done at scale for right, but we are seeing this multidisciplinary automated AI-driven workflows happening already for some specific disciplines, and we are seeing 50%, 60%, that's massive.

[1:09:03] Nathan Labenz: Yeah, interesting. How do people respond to it? my sense is that you probably won't have a hard time convincing CEOs of these companies that this is really important and they can look again at the China iteration speed. They can look at Tesla's ability to update its manufacturing processes much more dynamically than they're accustomed to doing. And I think increasingly they're going to feel the heat. Now, Another question is translating that through a legacy organization where probably a lot of people have a lot of different feelings about this, how they want it to go or if they want it to happen at all, many of which very understandable feelings, by the way. I don't mean to dismiss those feelings, but they're an obstacle in many cases in the way of the company actually transforming in the way it probably needs to be competitive. How would you characterize maybe your kinds of roles? I'm thinking for whatever reason, the ML researchers are like most keen to automate their own labor. And then we see a lot of artists are very hostile to the technology, certainly not all, but that's like a pretty common point of view, especially if it's like, I love doing this. You know, why do we want to automate something that I love doing? Where are engineers in that space?

[1:10:23] Thomas von Tschammer: Yeah, it's a good question. I would, I think we do see both end of the spectrum, right? I do see that engineers are artists in a way, right? They made their own intuition on how to build a car and they really enjoy that aspect, right, of manual iterations, leveraging intuition, thinking about the physics of the problem, right? And there are many engineers that's not easy, right? They've been doing it the same way for the past 30 years and they enjoy that. It's totally understandable that when they see these type of new workflows, they are sometimes a bit resistant as to, okay, what would it actually bring, right? And they also believe, right, we saw that they need to be in the loop because ultimately they are the domain experts and they are the ones that are the brand of the company, right? Otherwise, a GM car would be similar to a Ford would be similar to a VYD, right? So you need to have this human aspect as well. But we also see all the way, the end of the spectrum, ML researcher, There are also methods team and machine learning teams within those organizations at the forefront in ZooPro. And those ones are the first ones to adopt it. And those are the first ones that want to explore, want to try new models, frontier models, and to benchmark them and experiment. So we really have, in the same organization, those two extremes, right? How do you bring them together? That's also part of the complexity of those very large organizations. But the thing we've seen a lot, though, is Once you manage to break that final, that barrier, and actually have those engineers hands-on and working with the AI models, you see a top-event response. Because then they understand how powerful it can be and how much it can actually empowers them to make their job even better, even funnier, right? Because then they don't have to spend time, which is low added value, again, setting up simulation, waiting for the simulation for it to load, for it to compute. but they can actually interactively query the AI model, get results, try different options, try much more what-if scenarios. And that's the fun part when you're an engineer. You want to try this phase out, to test scenarios, test engines, right? And that's what AI enables you to do today.

[1:12:33] Nathan Labenz: Sometimes you've mentioned this idea of brands becoming the same and that they need to avoid that happening. This may be sacrilegious to say for long native to trader, but I feel like there's an awful lot of sameness out there in today's world, right? Manufactured products in general, certainly cars, right? I'm always kind of like, you know, really they look a lot alike. Let's be honest with ourselves, you know, there's been, there's been tremendous convergence. Do you think we're going to see a, or is this maybe even a way to think about success criteria for AI at a societal level, I feel like almost, a new trend toward more meaningful differentiation.

[1:13:23] Thomas von Tschammer: Yes, so I think what happened, and I think this trend is going to keep happening in the automotive industry, it's also lost due to autonomous driving, right? It's a bit pushed towards autonomous car, because essentially once you've sold autonomous driving, car becomes a commodity, right? So you don't need to own a car anymore, right? If you can just drive to your home, drops you anywhere you want. and you don't need drugs, right? Car becoming a community means that ultimately cars would tend to look more the same, right? By definition, by pure definition. However, before we get there, I do believe, and that's the difference I was saying before, that the winning companies are the ones that are going to be able to code the best practice brand within this AI workforce. And I'm convinced of that. And that can lead to again, widening the gap between the companies that did not unopt AI quickly enough and the one that did, right? And here we will see massively different differences.

[1:14:23] Nathan Labenz: People might be surprised by how, when you talk about taking the driver out of the car, so to speak, in a very, literal way that opens up like all kinds of new form factors, right? It could be the small delivery car that doesn't have any people at all. It could be sleepers that we get to overnight in on our way to grandmother's house, whatever. I think that has felt, even though people have talked about that, imagine that for a while when they think about the self-driving car world, it has felt like that is a long way off even once the technology works because We just haven't seen much change. Like why would we expect to see all these form factors pop up from a bunch of car companies that have given us like strikingly point for point similar product lines in recent decades. But this maybe could be very different in a world where a tremendous amount of stuff becomes automated. Is there like a, is there a bootstrap there that you think is interesting? One that does strike me is like the vehicle lists passengers in a, it wouldn't take much to perhaps create some slightly different regulations for those type of devices. And next thing you could really get a crazy flywheel going, perhaps, that then again, like leaks out into the rest of the broader industry. How, like, how fast do you think that kind of stuff could happen?

[1:15:55] Thomas von Tschammer: I think the main reason for the phenomenon you're mentioning is that today most of the autonomous driving cars are cars that you could drive, but they're not made autonomous, right? And because you need to drive them, then they look like the ones we used to. If you look at way more, right, this is what's happening, they take the chip hard, I-pace, and they make it autonomous. And if you give another cell, you need to be able to drive it by regulation if you have an issue. But there are a few companies If I take the counter example, that I've built autonomous first vehicles. If you think about Dukes, right, I guess you know about this company on the West Coast. I think the first tagline is that it's not a car, right? It's a robotaxi designed around you, if you recall correctly. Because they've built it fully autonomous minds and 1st, that's it. And it doesn't look like a car. If you think about it, it's a relation to anything we've seen. So I think the more that the technology becomes mature, the bigger the shift gonna be to new type of concepts as well.

[1:17:03] Nathan Labenz: How far does this go? Is there any limit to it? You know, I kind of imagine a, everything is an RL loop as kind of the end state. You know, you can imagine putting agents into every seat in the company. And again, we've, you know, keep in mind we're going to need some oversight for this, but in terms of, thinking kind of a first principles limit paradigm first, you could have product strategies, virtual market testing. There's increasingly models that in the same, general spirit as you have models that will validate the aerodynamics of your design, you can have models or scaffolds around foundation models that can sort of act as like virtual customer panels. So you could imagine, really from even the highest level, getting into a fast loop that's fed by RL, where you have at least a sufficiently good reward model to steer it in the right direction. And then that can kind of cascade all the way back, perhaps, right, to to the designs itself where you can imagine a not too distant future where I can sort of speak a perhaps rather complicated mechanical, electromechanical product into existence in a way that I can now like speak a video into existence. Do you see like any fundamental gaps in that vision? Like what, if anything, would prevent that from happening in five years' time?

[1:18:48] Thomas von Tschammer: No, we didn't. I think the capabilities are there. The main question for big companies to implement that is infrastructure, governance, mostly, right, and data, obviously, making the data flows and it's at the right locations. In terms of capabilities today, we have the right piece of the post. I strongly believe so, and it's going to go even faster. the latest frontier models that are being developed, right? I don't see any reasons for it not to happen. Again, we won't break the physics. The physics will be there. That's why we need to have those two combination of tools that are grounded in physics. We talked about primical simulation solvers, right? They will be there and they will remain there. They will be the fuel essentially for these models in engineering, right? But I don't see any fundamental reasons for it, but they're happening quickly and actually quickly.

[1:19:38] Nathan Labenz: Do you think humanoid robotics or robotics more generally is critical to this, especially on the unlocking the speed on the manufacturing side? I mean, I think there's a lot of different ways you could imagine unlocking more agility on the manufacturing side. How much do you think things like humanoids matter versus just kind of general intelligence that will also design the machine tools and, kind of bring all the same. You can kind of imagine the version that is like built on this. If humanoid robots are sort of the analogs of LLMs, you can imagine that's one kind of way that we get like crazy responsiveness from manufacturing. But maybe another is just that, again, the reasoning models bring all these same engineering paradigms to the machine tools themselves. And that, you know, is enough to kind of speed things up.

[1:20:38] Thomas von Tschammer: No, so I do believe that robots will be in our critical when we think about manufacturing, the manufacturing aspects and the plants, right? Now, do I believe those need to be humanoid? I think it's a different story, right? I think we as humans, we are optimized to do a lot of things. Are we optimized to be in the plants? I'm not sure we're the right form factor. But I think that's a different topic in itself. However, I think there are still some breakthroughs to be done for robots and humanity to be actively useful at scale within plants. The first one are the AI models themselves, right? I do believe that we need to much more advancement into the AI models so that those robots can actively interact with the physical world and be efficient, right? We are far from being there right now in the industry. And then there is another question on the hardware side, right? About the autonomy of these robots, how do you make sure they don't overheat, right? Building a hand, I don't know if you've done it, but there's a lot of research in universities on how do you actually build a hand that is as agile as what we have, that is so great, that is durable. You need a lot of breakthrough also on the hardware side. We want them to be efficient at scale. It's going to happen, but I think that there is some work to be done in that area for sure.

[1:22:04] Nathan Labenz: Yeah. But it sounds like you don't, to try to put a little bit finer point on it, do you think that we will be fundamentally bottlenecked if we don't have a highly generalizable physical intelligence, you know, that can kind of walk into a room and like troubleshoot some random thing and apply a wrench to something that, for a human mechanic, that wouldn't be so difficult. Obviously, robots really can't do that very well yet. Do we have to solve that part or can we, through sufficient intelligence, kind of take a more top-down route to highly efficient automation that doesn't require this like general purpose physical intelligence to be able to kind of patch things?

[1:22:50] Thomas von Tschammer: Yeah, I think I agree. This is definitely the somewhat second order. And thank you, I would say, I think with the first law principle, we can solve a lot of these bottlenecks through this more general automation and intelligence up front, right? The next frontier will be those robots in the plants, right? But I think to your point, this is definitely so.

[1:23:11] Nathan Labenz: The future of physical abundance, I think I'm feeling more than perhaps I ever have. I really appreciate you taking the time and walking me through all this and the contribution that you guys are making at Neural Concept is definitely a fascinating one. Anything else that we haven't talked about that you think, you know, I should have asked about or, you know, what blind spots would you detect in me that you could help me patch up before I let you get back to work today?

[1:23:40] Thomas von Tschammer: No, I think that's we covered the good top topics. Nathan, thanks so much. I think one thing I can add is that from the outside, I don't think we realize how close we are and how it's actually already happening in the engineering industry today, right? You have engineers, you have products that are being designed AI first, and it's true, it's already there, right? And the acceleration is just starting, right? So look out closely at the industry, automotive, aerospace and defense, consumer electronics, look at those industries very closely for the next few years and Most likely all the next breakthroughs in terms of designs, performance, products, they will be led or they will have an element of AI-driven workflows, iterations in them.

[1:24:30] Nathan Labenz: Thomas Von Shamer, co-founder of Neural Concept. Thank you for being part of the cognitive revolution.

Outro

[1:28:10] If you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts, which is now part of A16Z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI Podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at aipodcast.ing. And thank you to everyone who listens for being part of the Cognitive Revolution.