Are world models the unlock for Physical AI? | Episode 19

Summary

Podcast Summary: "Are world models the unlock for Physical AI?" — This Week in AI, Episode 19

Date: June 25, 2026
Host: Alex (Jason Calacanis)
Guests:

• Boris Softman (CEO, Bedrock Robotics)
• Jeff Hawke (CTO, Odyssey)
• Ethan Baragas (CEO, Icarus Robotics)

---

Overview of the Episode

This episode explores the accelerating intersection of AI and robotics, especially the role of "world models" in bringing artificial intelligence from the digital to the physical realm—what the guests call "physical AI." The discussion spans the state of the market, technical challenges, economics of deploying AI in physical industries, adoption barriers, and the promise vs. reality of world model research, with insights drawn from construction, autonomous vehicles, and even space robotics.

---

Main Discussion Points & Insights

1. Vibes and Market Sentiment in Physical AI

• The AI and robotics market remains bullish among builders, even as broader public sentiment grows more cautious.
• Boris Softman: "You could do things with physical AI that were just infeasible five years ago... 80% of the world's GDP still grounded in physical industries." (02:39)
• Market energy is high but accompanied by challenges: talent scarcity, expensive compute, and complex system integration.

Timestamps:

• Market excitement: (02:39–03:34)
• Geographic/industry differences: (03:34–05:01)
• Gen Z/user adoption gap: (05:01–06:36)

---

2. What are World Models? Why Do They Matter?

• Jeff: "World models are causal multimodal systems that learn and predict and interact with the world over long horizons." (12:00)
• In simple terms, world models are like advanced simulators that can anticipate future states not just in language (like LLMs) but in physical reality—starting with vision, expanding to sound, touch, and more.
• They help bridge the gap between human and machine perception: seeing, hearing, and acting.

Notable Quote:

• "Think of this as like the 18 years you spend learning how the world works before you do 30 hours of driver training." — Jeff (09:28)

Timestamps:

• World models defined: (12:00–13:57)
• "Video game" analogy: (13:00–13:57)

---

3. Limits and Challenges: Data and Generalization

• Physical AI is harder than digital—unlike text, there is no single, open corpus for "the world's physical experiences."
• Scaling is limited by the cost and diversity of real-world data, especially for edge cases ("toddler behind a car" or "elephant in the road").
• Adversarial data and synthetic data are vital to upsample rare but critical failure states.
• Boris: "You can get a really great representation of common situations, but actually still be pretty exposed when it comes to the rare cases." (16:16)
• Zero-G, space, and other domains require even more creative data collection and world model training.

Timestamps:

• Data scarcity and adversarial scenarios: (14:07–18:15)
• Multimodal models & recent research: (19:24–22:21)

---

4. Practical Applications: Construction, Space, Simulation

• Bedrock Robotics: Automating excavation and machinery using large, transformer-based ML models.
  • Uses world models for simulation and training, but deploys transformers for real-world operation.
  • Rich data from operating retrofitted heavy machines. 1 TB/machine/day. (43:22)
• Icarus Robotics: Autonomous and teleoperated robots for space stations.
  • Rely on human demonstrations ('teleoperation') for initial deployment due to high cost and risks in orbit; full autonomy "is earned, not given." (30:24)
  • Extreme communication limitations (S-band vs. laser comms) and unique physics (zero-G) frame both engineering and data challenges.
• Odyssey: Building core world models as a platform for robotics and games, including recent "Starchild" (multimodal) and "Prowl" (regret-driven) research.

Notable Quote:

• "Full autonomy is this thing that's earned, not given." — Ethan (30:24)

Timestamps:

• Bedrock’s stack: (23:12–27:35)
• Icarus’s sequential, data-driven autonomy: (29:25–33:33)

---

5. Scalability, Modularity & Autonomy Hierarchies

• Both space and Earth-bound robotics benefit from hierarchical models: high-level planners and local control policies.
• Generalist models for reasoning; specialist models for safety-critical, domain-specific tasks.
• Boris: "You simulate the reasoning model more aggressively, and optimize the physically complex components locally." (33:33)

Timestamps:

• Autonomy orchestration: (33:33–36:47)
• Adaptive control in zero-G: (36:47–37:17)

---

6. Public Perception, Policy & the AI Hype/Backlash Cycle

• Public “approval” for AI is below 20% (46:48), endangering societal support for physical AI—echoes of industrial revolution Luddite panic.
• Positive sum game: The potential for expansion, not just displacement.
• Science communication failure: Space and AI must connect their revolutionary impact (“Keytruda cancer drug as a space spinoff” — Ethan, 52:24) to day-to-day benefits.
• Caution against dismissiveness and polarization: "The only 100% losing strategy when it comes to selling AI to the public is to be dismissive and rude." — Alex (56:54)
• Controversies: Should AI labs be nationalized? Consensus: That would stagnate innovation.

Timestamps:

• Industrial revolution analogy: (46:48–49:08)
• Science outreach and examples: (51:16–53:23)
• Nationalization debate: (54:27–56:54)

---

7. Competition and Industry Alignment

• Nvidia’s entry (Cosmos 3, Halos) signals validation for world models in physical AI; but companies must vertically integrate for safety, performance, and interface control. No one-size-fits-all ("can't Mr. Potato Head it").
• Varying architectures, sampling rates, priorities; the field is still fanning out.

Timestamps:

• Nvidia and "encroachment": (58:02–60:24)
• Industry harmonization and divergence: (61:05–62:44)

---

8. Precision, Robotics vs. Humanoids, and Specialization

• Construction: Must compensate for machine variation, failure, and maintain safety margins (Waymo's aligned fleet vs. 72+ attachment types in diggers).
• Space: Constraints of human-defined environments, small scale, need for millimeter-level precision.
• Humanoids: Great for economies of scale on Earth, but purpose-built robots prevail in unique domains (underwater, zero-G, etc.).

Notable Quote:

• "On the financial side, you'll never beat a purpose built machine or robot for some task with a humanoid...the point is I can make a hundred thousand and employ them everywhere and democratize labor." — Ethan (67:34)

Timestamps:

• Real-world precision: (62:44–66:43)
• Humanoid robots versus domain-specific form factors: (66:43–69:19)

---

9. Looking Forward: Are We at the GPT-3 Moment for Physical AI?

Consensus:

• Not yet—the field is still maturing, highly vertical, and awaits explosive demand and generalization.
• Jeff: "We're on the trajectory but not there; GPT-3 was marked by a spike of explosive rollout and demand.” (69:45)
• Boris: "It'll be a while longer before you can genuinely go into brand new environments and be super capable like this." (70:41)
• Ethan: "We're so early...once you see the demand signal spike and capability spike, that’s when it will explode... two, three years time." (71:50)

---

Memorable Quotes by Timestamp

• "This is the year of physical AI." — Alex (00:00)
• "World models are causal multimodal systems that learn and predict and interact with the world over long horizons." — Jeff (12:00)
• "Full autonomy is this thing that's earned, not given." — Ethan (30:24)
• "There's not enough the communication around it has certainly not been great, but there's so many successes..." — Boris (46:48)
• "If you break down the environment and first principles, a human has been evolved to work on Earth very well, but not space... the robot itself is still a general purpose platform." — Ethan (67:34)
• "We're on the trajectory (to a GPT-3 moment for physical AI), but we're not there." — Jeff (70:05)
• "You will for longer see more verticalized applications... it'll be a while before something can genuinely go into brand new environments and be super capable." — Boris (70:41)

---

Key Takeaways

• Physical AI is just beginning its exponential curve. World models could be the unlock, but are bottlenecked by data, safety, and highly customized applications.
• Scaling in the real world is much harder than in the digital one, especially when human lives or expensive hardware are at risk.
• Simulation and synthetic data are critical, but the physical world's edge cases are hard to fully simulate.
• Collaboration between specialist startups, established giants (like Nvidia), and open platforms will be vital to accelerate progress—but integration and safety mean verticalization still rules.
• Public perception and communication remain the industry's Achilles heel; making the benefits real, tangible, and positive is a top priority for avoiding backlash and stagnation.

---

Suggested Segments for Listening

• World Models Explained: 12:00–14:07
• Practical Applications & Data Pipelines: 23:12–27:35, 29:25–33:33
• Industry Challenges & Horizons: 46:48–53:23
• Debate on Nationalization/Policy: 54:27–56:54
• Closing GPT-3 Moment Discussion: 69:19–72:39

---

Tone & Style

The episode maintains a balance of technical depth, optimism about progress, and frankness about obstacles—frequently mixing nerdy enthusiasm (“video game” analogies, technical acronyms) with playful banter and occasional irreverence about funding rounds, regulatory bottlenecks, or industry PR woes.

---

Summary prepared for listeners who want actionable insight into where the ‘physical AI’ revolution stands as of mid-2026, what’s working (and what’s not), and where the first seismic shifts are likely to happen.

Transcript

[Alex]

It's the year of physical AI.

[Boris Softman]

80% of the world's GDP still grounded in physical industries.

[Ethan Baragas]

There's people that are using Claude code and Codex and cowork on their own, developing things that we've never seen before and make 100,000 and deploy them everywhere and democratize labor. And when you look at the Earth and you say, okay, half of the Earth's GDP is people doing things. It's human labor. There's a very compelling argument there.

[Alex]

Are we still approaching the GPT3 moment for physical AI, or have we already gone past it?

[Ethan Baragas]

I think we're so early. I think it's a far, far time. Thanks to our friends at PayPal, the exclusive sponsor for this Week in AI. Try the payment and growth platform that's trusted by millions of customers worldwide. PayPal open start growing [e-mail].

[Alex]

hey, everybody. Welcome back to this Week in AI. My name is Alex. Now it's the year of physical AI. It seems that AI driven robotics are all the rage. People are worried that hardware will become the only mode possible. No one has enough compute. And companies ranging from digging in the dirt to flying in orbit are pursuing different approaches to understanding and interacting with the world. So on today's episode, we're sitting down with the CTO of a world model startup, the CEO of a company automating construction equipment, and the CEO of a company that wants to take autonomous robots up into orbit. I really want to understand how quickly the physical AI market is developing, which type of AI model is the best suited to bringing AI into the real world, and what the economics of AI in the world, or perhaps AI of the world, look like. So please join me in welcoming to the show, it's Boris Softman, the co founder and CEO over at Bedrock Robotics. Boris, good to have you here.

[Boris Softman]

Great to be here, Alex. Thanks for having me on.

[Alex]

And Bedrock just raised a $270 million Series B. So shout out. We also have Jeff Hawke, the co founder and CTO of Odyssey. Jeff, how you doing?

[Jeff Hawke]

Doing well, thank you. Thanks for having me.

[Alex]

$310 million Series B just pipping what Bedrock put together. And then we have Ethan Baragas, the co founder and CEO over at Icarus Robotics. Ethan, welcome to the show. And where the hell is your series Beat?

[Ethan Baragas]

We're working on it, Alex. We're working on it. But thanks so much for having me.

[Alex]

No, I'm so glad you guys. So glad you guys are here. I want to start with something a bit broader though, than going super deep, because when I look out around the tech market today, it feels like in the last couple of weeks the vibes have changed. And I think it's fair to say that we're all AI bulls here on the show. Don't mind a data center when I see one. But it seems that public sentiment has shifted. It seems that the markets are choking a little bit. And so I'm just curious if this vibe shift that we're seeing outside of the industry is also taking part inside of the world of AI and robotics, or if this is just stuff that's more on the media and public opinion side of things. And Boris, I thought we'd start with you.

[Boris Softman]

You look at the last couple of years, it's felt like it's just been a non stop march upwards in terms of energy and excitement. There is something foundationally true about the fact that you could do things with physical AI that just were just infeasible five years ago. And so it's very exciting because 80% of the world's GDP still grounded in physical industries. You have tremendous opportunities in transportation and industrials and manufacturing and all these areas. There'll be ups and downs. I think the general sentiment feels still full of really deep excitement. There's just also kind of a noisiness and kind of like a mass to the messaging. And so we haven't sensed too much of a reprieve in it. It's still incredibly hard to find talent. It's really hard to put all the pieces together. Compute still expensive, all these things are still present. But I fully expect there to be some ups and downs in the near future.

[Alex]

Okay, that makes sense. And that tracks what I'm hearing in the States. But Jeff, you're over in London. What's it like over there? What's the temperature? What's the level of excitement versus fear, if you will?

[Jeff Hawke]

Literal temperature is sweltering.

[Alex]

And I'll say the same thing.

[Jeff Hawke]

It's. We sort of exist in sort of the, what I probably call the sort of AI research epicenter where there's a lot of focus more on sort of the core fundamental AI research. And I would say the general sentiment is very strong and quite bullish. There's sort of two macro themes that a lot of Neolabs and the large labs are also focusing on, firstly around this idea of recursive development of AI and then also in world models. And both have a pretty major presence. And at least from what I've seen, the sentiment is, is very bullish on both. So I think it's early Days and I, in as much as we've seen a march up progress over the past year, I don't think the. I don't think we're done is maybe the way to say it.

[Alex]

Glad to hear that. I think there's a lot more stuff that we need to build. Speaking of which, Ethan, we're going to, we're going to talk about your company in a couple of minutes. But I hate to say this as the designated young person on the show today. What's the vibe like out there amongst people that can't rent cars?

[Ethan Baragas]

Yeah, still get that 500 bucks extra every time.

[Alex]

I know it's not, it is literally ageism, but I mean discrimination.

[Boris Softman]

Yeah.

[Alex]

I don't mean to put you on the spot, but I am curious because I feel like Gen Z being super digital is probably imbibing a bit more of the discourse here, if you will.

[Ethan Baragas]

So it's quite funny actually because like myself in the position I am in the tech sphere that I live in, I'm so far in this bubble. And then when you interact with other people outside of this tech sphere, it's way more polarizing than you would actually think. Think. And you know, even with my friends from my hometowns, whenever I get make it back and get to talk to them, you know, they're not implementing it in their daily lives. And this is actually something that's crazy. We're still so early. You know, we sit on two ends of the spectrum. One end of the spectrum is we kind of push the frontier in this very novel domain, which is zero G. You don't necessarily see world models built for it. You don't see lots of data of teleoperation collected for it and expert examples. You just don't see that. And when we try to talk to even some of the folks at NASA, it's a crazy conversation. Safety, it's not Cartesian control, monopoly control. It's this unknown value still that they've just never seen. But then when you talk to some of the people on the implementation side, even terrestrial, some of the stuff that we're doing still in the tech sphere, it's still super exciting. You know, this is the thing that people jump for and it's more of the mainstream, what you come to expect people talking about and the same excitement that you expect from the mainstream. But it's very different on these two other sides. And I think when you look at, you know, our demographic, it's even more split than ever right now of people that are saying, oh my God, AI like you can use it for what? And physical AI. What does this even mean? I've had to have that conversation so many times of what is physical AI that's not an LLM. And then on the other end of, you know, there's people that are using Claude code and Codex and cowork on their own, developing things that we've never seen before.

[Alex]

I'm a huge codecs guy and my spouse recently for the first time used AI to do a thing. And I was like, oh, we live in two different worlds inside of our house. And she was like, yes, honey, that's your world. And I was like, okay, fair enough. But it does surprise me, Boris, on the point about people being more excited about physical AI perhaps than purely digital AI. Are you seeing that same split that Ethan just discussed? Yeah, well, there's a feeling that there's

[Boris Softman]

a digital AI has just matured a lot. It moves fast. You have the exact same feature space that everything got to use. You had infinite data on the Internet. And so you started to see this snowball that was shocking, impactful, very, very exciting. On the physical side, it's like it's genuinely harder. You don't have the same sort of access to data. You have uniqueness in the hardware and the sensors and everything, but the opportunities are tremendous. And you see it in just how much Waymo started to progress and the value it's created and the way it's starting to scale. You start to see the opportunities in assembly and the sort of things we're doing. So I think there's an excitement that it won't necessarily have the exact same shape as Anthropic or OpenAI did in these kind of horizontals with things wrapped around it. But you can actually go and tackle these problems that have astronomical value to them. And what's interesting is that it's incredibly expansionary. It's not like a replacement of labor in a lot of cases, it's that you can actually start to do things that were massively supply constrained or just weren't possible before. I think we're not even scratching the surface of how transformative that can be because oftentimes AI is thought of as a cost cutting measure. In reality, it can be incredibly expansionary if it's actually applied in the right way. And you see that more in the physical world. In some ways, although you've seen that in coding and other aspects as well. On the digital side, it's also famously

[Alex]

expansionary when it comes to your budget line items that you have to then report to the cfo.

[Boris Softman]

Yeah, good to be AWS in this world.

[Alex]

It's great to be anthropic in this world. They put out a new model, they double the price and everyone goes, can we please have some more? I mean it's just the best place in the world to be. Jeff, Boris just brought up Waymo and I was going to get to this later on but let's do it now. I've been a little bit surprised to see the issues with self driving cars as they've scaled recently and I've talked to a number of people in the self driving world, Wabi and Wave and so forth and I've talked to them about world models and kind of where they use their technology and so forth. But, but I thought no matter what underlying technology, self driving cars would use that as they scaled actually in the physical world getting off simulation, getting out of the world model, they would get such rich data that we would see improvements. But instead if you look in the last month or two the headlines about self driving progress, I mean, you know, Waymo doing the major recall about construction zones and I think there was an issue with school zones and then Tesla is in trouble for several different things. Is there a problem here or is this just the standard kind of like you know, I meets the real world. The there's going to be some teething problems.

[Jeff Hawke]

I think it speaks to the sort of core problem that all embodied AI faces which is really one of solving generality and solving sort of an open ended world problem is really, really, really difficult. In many respects that was one of the core theses of what led to Odyssey. This idea that you have to Boris, you mentioned for example needing to scale data and it's just hard to get a lot of data. The same is true regardless whether it's construction vehicles across different types in different environments. It's an immense challenge. So many respects or the bet we're making is that war models are that unlock for generality for general purpose robotics. The good thing is it's not just an embodied AI problem obviously as applications there, but in many respects I think it's one of the best ways we have of using the sort of the digital AI phrase used to really help fund and help that development of solving this broader foundational intelligence. So certainly I'll bet the feature would bring world models to the table as maybe sort of analogy. Think of this as like the 18 years you spend learning how the world works before you then go and do 30 hours of driver training.

[Boris Softman]

It also has a disco element to it where you drive 200 million miles, you're kind of poking statistics so many times that you start to really, really feel the long tail. And all these kind of interesting challenges that just wouldn't have come up when you're doing kind of testing and development and small scale deployments. And so it's the real world, like you start to really get exposed to it.

[Jeff Hawke]

Scale is big.

[Alex]

Does that make world models? And we're going to get into defining those in just a second. But does that make point you make there, Boris? Make world models more or less useful when it comes to bringing AI out of our screens and into the world?

[Boris Softman]

I think it's incredibly useful. I mean, in fact, we kind of go through similar thought process where we're creating general capabilities for specialized heavy machinery. So these are manipulation machines that manipulate whether it's like earth or aggregate material, lumber, farmland. Then these are all kind of specialized manipulation machines. And you can get off the ground by using more traditional simulation approaches where you're modeling kind of Earth physics and sensor models and so forth. At some point, it is unequivocally the right long term answer to go to world models where you like your own data, kind of teach these systems in order to represent the right physics. It becomes more scalable, it becomes easier to start doing things like reinforcement learning where you need to computational efficiency. So it feels like it's still, you know, there's a lot of complexity to it and you do have to have the right data for particular domains. But it feels like that's the future of simulation. When you fast forward to a number of years.

[Alex]

Let's talk about world models. The way that I think about this, Jeff, is that LLMs model language and world models model reality. Your company Odyssey says that world models are causal multimodal systems that learn and predict and interact with the world over long horizons. Can you translate that into English for everyone out there who does not speak nerd?

[Jeff Hawke]

Think of this as like a learning to simulate what happens so you know some information about what you've seen. And what you're learning is potential futures. So you think about like you would a sort of computational simulation rolling forward in time at its simplest. That's kind of the best way to think about it. Now the question is, what are you simulating? The easiest form of data is visual state. And that's where we're seeing most progress and certainly in the sort of physical AI category is most world models have an inherently visual element. We don't see ourselves stopping there. Audio is a very Useful extension, the way I like to think about it is it's almost like mimicking human senses of sight and sound. Right. It's how we perceive and interact with the world. And I think that's a good place to start in terms of state.

[Alex]

I think a lot of people consider world models from the perspective of self driving cars, which make it appear to be kind of like a video game style interface. And is that the way we should think about visualizing a world model or is it something a little bit more esoteric than that? And I'm thinking about the Candy Cane version versus the Big Boy edition in one level.

[Jeff Hawke]

You can think of it as a video game. I mean, to be honest, the industries we focus on primarily are games and robotics. In many respects it gives us the sort of maximal coverage of a couple of industries that are big in their own right. But from a capability perspective allows us to think about most of the things that matter to world models. So yeah, I honestly do think the same model will be used for general purpose robotics in terms of autonomous driving and will also sit underneath future video games. And I think that's kind of the key crux of this, where if you want to really, really solve this foundational technology, you have to solve generality. And as a result that means that learning from games benefits, for example, construction diggers.

[Alex]

But does it help, Ethan? Because generality I think works quite a lot when you're in our gravity well at 1G. When you're up in 0G there probably breaks down. Ethan, am I wrong here?

[Ethan Baragas]

No, yeah, it's something that we actually face day to day right now. I mean we have our photorealistic sim, most people use ND Isaac sim and things that look like this. We use it internally and it just, it doesn't get you there right now. And actually it's kind of a question I have for Boris as well where, you know, you know, we're on that tail end of embodied AI using robot learning from expert examples, we can teleoperate using laser based comms from Earth to space. And this is amazing. Within low Earth orbit, once you move out farther and farther away, you're bound by the laws of physics. You have communication delays, you can't do it. And there's this ongoing debate right now within the field of how many expert examples, how much data can you actually throw at these models before it's diminishing returns? And it ties into what we're talking about a second ago about Waymo, about Tesla with autonomous driving and how much actual data can you throw at these current know models and heuristics and training pipelines that we have before you need to move into something new. And I think that's why it's so key that in the future you have to have things like world models to collect some of this data and augment what you have in the real world. But we're still missing that key point. And so for certain industries, you know, things like construction, where you literally cannot move as much earth as a digger can, right. It it's so important to have those capabilities or in space when it costs, you know, $130,000 an hour to keep an astronaut alive to do something you actual capability. Right. And so I think these are some of the key questions that the industry on the research side has to figure out moving forward.

[Boris Softman]

The challenge, it's an interesting challenge because like scaling w, it's been surprising in the space how far they they still go where you get improved performance if you have the right capacity. We saw this to Waymo and autonomous driving. You see this in research with you know, grasping and kind of humanoids, you see with lms. And so it's kind of, you know, you think more and more of it should be clever architecture and so forth, but there is just like an element of, you know, scaling laws. Oftentimes it's getting the right data. And so you end up having to create adversarial scenarios that are very rare and kind of like push your learning rate faster because.

[Alex]

Boris, can you explain what adversarial means in that context? Yeah, I think a lot of folks listening don't fully get it.

[Boris Softman]

Of course, of course, yeah, sorry. This is where you are basically up sampling really rare scenarios that matter a ton, but you don't show up too often. And so in the case of autonomous driving, maybe that's like a toddler jumping out from behind a coated car. And you're not going to drive around hundreds of millions of miles waiting for that to happen. You're going to try to create that using synthetic data or close course testing. We use mannequins or splicing in real data into a previous log. And so you're effectively manipulating your distribution so that you're creating examples that really matter and give you the most learning value out of it. And so, and this is where you start to see hybrids of simulation in real world. But you have to get creative because what's challenging when you get into these like safety critical applications is that you're really optimizing for the 1 in 5 million mile case in the case of, you know, driving or a rare interaction with a dangerous heavy machine, or you know, these sort of like, you know, rare scenarios that, you know, Ethan might be dealing with in space. And so this is where you actually have a bit more sensitivity. And the challenge of the world models has to overcome is that by definition you're modeling the distribution that you've learned. And you have a challenge now of how do you prove that that's actually representative of what you're going to encounter. And so this is one of the interesting jumps where you go from an LLM where the consequence of being wrong in most cases is not that significant. And maybe you start to get into tougher long tail challenges with medicine and law, but you go to something like autonomous driving or these much more like very, very consequential physical robotics. You're actually quite sensitive and if you're not careful you can get a really great representation of your very common situations, but actually still be pretty exposed when it comes to the rare cases. And so this ends up being one of the deepest challenges of a lot of robotics applications. We have these dangerous physical complex systems that need to be incredibly safe in many respects.

[Jeff Hawke]

One of the big challenges that we've seen is getting that, for example, you mentioned what's a 1 in 5 million mile example? How much driving data is there with, I don't know, an elephant in the road. As big as WAMU is probably not many. And as a consequence the question is, where do you get that information? I think this problem compounds as well with humanoids or anywhere where you are sort of needing some very specific embodiment, specific environment. And then actually Ethan, you point out it gets worse as soon as you hit space. There's nothing inherent in the architecture to these models that is inherently limiting in how they're used. I think the challenge is really understanding how to build them. There's still a lot of algorithmic work going on coupled with where do you bring in data and how do you sort of bring them all together? That sort of is the recipe that I think is necessary. So to your question of whether can this work in zero G? Yeah, I think it absolutely can work in zero G. It's just a question of are we bringing in that right data? I think it'll come right. If we think back in terms of language model evolution, the early versions of GPT were kind of rubbish at biochemistry. Now they're pretty good. And there's a whole lot of things that in terms of maybe academic fields or very specialized domain Knowledge which will be brought into world models, even if they aren't.

[Alex]

Today I want to go a little bit more on world models because we're talking about them in a very specific context. But Odyssey does quite a lot of research. You guys recently put together something called Starchild and also this is the best, I presume, acronym I've ever heard. But if you take the prioritized regret driven optimization for a world learning model, it becomes prowl. So tell us a little bit about the state of the art for world models and then we're going to apply that to our two other guests use cases. But what's the state of the art? How fast is it progressing and do we have enough compute?

[Jeff Hawke]

How fast is it progressing? Very quickly. Do we have enough compute? Never to such a state of the art. So starshild was the first multimodal world model. So this is basically not just visual state, but also bringing in jointly generated audio. So this is a bit different from what you might have seen in VEO or for example, avatar models with speech. This basically means that as you're predicting frames forward, you're getting both pixels coupled with an audio wave at the same time. Which turns out was a more difficult brief problem than we expected, which is probably why people hadn't done it. But nevertheless, I think, why was it harder? Maybe the easiest way to describe it is if you think about predicting one frame of video or a very short snippet. The actual audio information you're dealing with is very short, so it might be half a phoneme. So phoneme being sort of a phonetic sort of unit of sound, it makes it quite difficult to get sort of stable, coherent generation, particularly as soon as you get speech. Now this might seem like it matters more for sort of non embodied AI, but I think it's quite an important signal in many respects. Models that can generate and model the world well are also very useful for understanding the world. And audio is a big part of what happens, right? Be it robot grasping, if something cracks or sort of crumples, that's a very useful feedback sound. If something happens sort of out of sight, but you hear an audible sound, that matters. So hence we see this as a core fundamental capability that has to be baked into world models. So that was Starchild you mentioned Prowl. That was a piece of work we're quite proud of. Historically, world models have been an intrinsic part of reinforcement learning, where you could think of the world model as a learned transition model or a learned simulator regulator coupled with an RL agent. And these are Kind of two sides of the same coin. Now what was interesting historically is most people are focused on improving the agent in terms of decision making. Papers like SEMA from DeepMind is a good example. And that's really where a lot of the research is focused. We looked at us and said, well actually you kind of have garbage in, garbage out problem, so why don't you actually improve the world model? So what we did is basically stand up a simulation environment using Minecraft and effectively trained an RL agent to provoke failure in the world model as it was exploring the world. Maybe it was failure to adhere to actions, maybe the geometry failed, maybe the physics failed, lots of different reasons. And essentially it's rewarded to find failure modes. So as a result you get a really nice framework that allows you to essentially boost the level of your performance in your world model. Given that simulation environment in an automated way. So you crack the flops, world model goes up.

[Alex]

How do you determine what is a failure state? Because if I'm thinking about construction machines, a failure state is if the bucket hits someone in the face place and kills them. Right? Bad, very bad failure state. In Minecraft there's not quite the same level of impact, which is slightly funny given what I just said. But it seems like a difficult environment to define failure state.

[Jeff Hawke]

Typically you've seen games are often the starting ground of where a lot of AI development happens. This is a very long history through a lot of the early work in DeepMind's a canonical example also OpenAI. So I think it's the right sort of environment to explore these kind of concepts. There's no reason that you are limited to Minecraft and we'll have some more work coming on that soon. So for example, any ground truth simulation environment can be used in this way to juice the performance of a fully learned what model. So that could be for example a space simulator, an earth moving simulator. There's no reason that can't be used in a close fit manner in the same way.

[Alex]

Okay, so Boris, given all of that, I'm curious about how you guys have picked what AI technology to use inside of your machines. Because you guys said in your public materials that you put large scale machine learning models. Is that a one to one to world models or is there a variation in what Bedrock uses to power its machines?

[Boris Softman]

No, world models will be used for your kind of simulation and training purposes. The actual ML model is effectively a large transformer based end to end architecture. It's taking as an input sensor, data, machine signals goal of what you're Trying to achieve a variety of other factors that give you the pose. And then it's processing all of this with a particular structure and outputting a behavior for the machine, as well as potentially other behaviors like honking the signal that you're ready. And so that behavior can actually govern a complicated machine like an excavator that has six to nine degrees of freedom. But the exact same architecture could power a wheel loader, a bulldozer, a dump truck, and other aspects. And so you're effectively learning from a variety of inputs. The core of it can be human demonstration where you have a lot of examples of huge amounts of hours of actual real work. This is actually similar in spirit to how the foundation for Waymo's training was. The big shift that really caused it to break through was learning from hundreds of thousands and millions of miles of driving that captured the nuances of how you interact with downtown San Francisco. The same subtleties exist in how do you interact with complex, different materials and, you know, get like inch precision on, you know, cuts and grades, do demolition, deal with different tools, different machine sizes, different types of machines. And so we're learning from all of this and effectively enabling these systems to be able to behave like an operator and work towards achieving a goal. And so that goal could be a large kind of cut to build a data center. It could be a trench to, you know, do kind of piping, a foundation. Sure. And then, you know, eventually, like, this becomes a. The property that you start to see is that the foundations of what you learn in these machines, they actually carry over and give you a huge shortcut towards the next capability or the next machine. And we saw this at Waymo, where going from San Francisco to Los Angeles, Phoenix, Austin, Atlanta, Miami, you start to get a higher and higher subsidy and less and less new data required until you're almost more of an operational and qualification problem. We saw the same thing going from car to truck, from surface, street to freeway. And so if you structure things in the right way, you start to learn more and more of a foundation of what is it that enables large physical machines to manipulate the world around them. And that does more heavy lifting. And you need less and less incremental data that's specific to the new application. And it starts to get a similar property to how LLMs have generalized across language and other applications like grasping and so forth. And so it's quite exciting. And then the world models end up being a really powerful potential tool to simulate and do offline, kind of like iterations, training, get more training Data that doesn't just have to be physical, because in all these applications, physical testing is so expensive that what you really want to be doing is using the physical world to train your simulation. And then the simulator is actually where you develop and qualify your systems and deploy, because that scales much better. And so you oftentimes will see one to 100 to one to a thousand kind of price differences in a physical hour versus a simulated hour. So you want to push it into that direction as much as you can.

[Alex]

So you start by collecting real world data, put that into your simulator to create a strong world model that you can do more testing, then take the results of that and then put them into the machine again and send it off on its own.

[Boris Softman]

Yeah, and the war model becomes like a tool to do all this. We've kind of experimented with the mix of elements, but yes, you do a lot of training offline, where you're effectively training a large scale model to emulate the best behaviors of a human. And this is where it's interesting, because you're not just optimizing just for safety. You actually can have all types of metrics that you actually care about. And the same way that a waymo is 10 times safer than the human, you can actually become superhuman by upsampling the best behaviors and down sampling the worst. And this is where you become the superhuman operator that absorbs more and more scope over time.

[Alex]

And just explain for me in idiot terms why you don't want to use a world model on a machine in the world and you want to use this transformer based approach you mentioned earlier. Is it just a lower compute footprint? Is it just a better fit for the work?

[Boris Softman]

I would think of a world model as an environment in which you can realistically emulate the real world and simulate when you're on a machine, you are in the physical world, so you're actually kind of like now acting on it. World model becomes an incredible tool for development in these sort of autonomy applications, where you can climb the quality of your safety, your behavior, create scenarios you never encountered, test new versions of your software in scenarios that you would have to physically recreate. When you deploy physically, your world is your world, your world is your simulator in some sense. And so you basically just apply your final output in the final model you've trained. Now, as Jeff was mentioning, in the case of a video game, your world model could actually be the output because your end product is a virtual experience. In the case of robotics, your end product is a physical behavior. And so you use world Models more as a development tool.

[Jeff Hawke]

Maybe it's worth adding. There's a fairly recent sort of approach that people have been using, basically using the world model as a backbone. If you're familiar with the concept of a vla, a vision language action model, you can kind of think of this as deleting the language model and putting a world model in its place as a backbone and then effectively using that. At least the current research in literature suggests that effectively, that is a more useful prior or more useful information to learning it. But as Baro says, you still need to train that backbone and actually train it for the sort of the behavior that you're looking to solve with your robot.

[Ethan Baragas]

I mean, if I can jump in here, like this is something that we use internally on the VLA side of things. And like Boris, I have a couple questions for you actually. Like, how long horizon can your tasks actually get? Are they high scale primitives right now? Are you going end to end very long horizon tasks across an entire construction site of multiple different things? Because this is something that we deal with, number one. We just don't have the quantity of data of any other domain right now. The amount of data that's been collected of high scale, high fidelity manipulation in zero G, dealing with the dynamic coupling problems of picking up something that is unknown mass and then trying to manipulate that and move it, is probably very similar to the physics of picking up unknown masses of dirt and earth and moving this treasure. In fact, you know, some of the people on the team wrote controls for Caterpillar and some of the heavy machinery there because of how similar. We call it the dump truck problem, actually, if you pick something up of sufficiently heavy mass and you try to move it, the controls of this entirely

[Boris Softman]

changing your entire controls.

[Ethan Baragas]

Yeah, exactly. And so this is why we have to rely so much on real world data. Because as much as you can have photorealistic sims and world models, it's just not enough fidelity, number one, for safety, but number two, for you to have operationally, whatever your primitives are done autonomously. And so for us, we're lucky where, you know, the business case closes. I can pay someone, you know, $200,000 a year to sit here and operate this robot. And because the cost of labor on orbit is so high, it closes for us. And this is another hot take I have. Teleoperation gets such a bad rap right now. Everyone's like, oh, it has to be fully autonomous. Full autonomy, full autonomy. That's the only way robotics work. And in certain industries this is just not True. People don't care who moves their cargo on orbit. They don't care if it's a magical rainbow unicorn, an astronaut or a robot. They just need that task done. And so for us, we found this one place where we can deploy in orbit. We can have a human in the loop, which NASA loves, because safety wise, deploying autonomous autonomous systems that are based off of simulated data and not collected data, as well as based off things like world models, it's just not there yet because you have humans floating around in a tin can up there. If you poke a hole in the wall, this is a very bad situation for everyone involved. And so for us, the way that we think about it is you deploy sequentially and full autonomy is this thing that's earned, not given. And so you deploy first the teleoperation with the you collect all of your sensor data of the environment around you. You understand the physics of just moving in six degrees of freedom, your base before you even pick something up. And then as you pick something up, then you move forward and you start to understand what does this look like of moving an object that's similar mass to our mass of our robot. And from there you can start to roll at high level primitives. So let's say I've collected, you know, with a vla, for example, like Jeff was talking about, let's say I've collected, you know, 150, 200 examples of myself picking up an ISS cargo bag and moving it to, you know, the Columbus node or Columbus module. I can then at a high level tell this robot, hey, go pick up cargo bag in slot 3L and move it to Columbus node 3 or whatever it might be. And we can do that high level primitive. And over time, as we collect more data of those edge cases and train and simulation of those edge cases and augment our actual model, we can then build out more and more longer horizon tasks that are not compound. Instead of moving a cargo bag from A to B, I can now move the cargo bag A to B, unzip the cargo bag and unpack it. And after I've done that, I can start to implement experimentation and plugging things in and plugging things out, all on one learned behavior, all on one actual task horizon. And over time, as we do this, we build that corpus and everything becomes more and more higher fidelity. And then you erg autonomy. But it's not necessarily something where you can just go in the computer, run thousands of hours and deploy. Simterial is still way too Ethan.

[Alex]

Ethan, pause. Let Boris respond to that then keep going, otherwise he's going to forget the first thing you asked him and then we're going to have to do it all over again. I love you, but let's, let's give it. That was like seven paragraphs of interesting things. We have to unpack it slowly.

[Boris Softman]

Super fascinating. I mean this is an interesting challenge in like every kind of autonomy application where you can actually take advantage of potentially a split where you think of the long horizon planning as almost like a reasoning model that lives above everything else that's happening. Our version of this is you might have autonomy for an excavator doing a task and then a bulldozer and a dump truck. And each of them internally has the model that kind of operates their behavior, but it's more local. Like they're doing this kind of strip of a cut, they're doing kind of loading tasks. The reasoning model becomes kind of the orchestration model becomes an opportunity to coordinate all of them. And you can take advantage of the fact that you have very different constraints in that model where the physics are a little less relevant and you can learn from much broader data and break up this problem in a way that isn't as vulnerable to the sort of physical challenges you mentioned. And we have the exact same ones where you could be carrying 2 tons of earth in one bucket, kind of stretched out 20ft. And that creates tip over risks, different vehicle dynamics. And so if you're able to do reasoning in a higher level model that then hands off to this much more physically complex kind of reasoning has to happen locally. You can potentially simulate that reasoning model a lot more aggressively and then optimize the components that are actually much more vulnerable to the physics of space or the specific controls of an arm or whatever the challenges might be. And so it becomes a decomposition of complexity where you kind of have a way to lean into the strengths of each approach and it makes it more simulatable versus kind of taking on the biggest pain points of every single one altogether.

[Alex]

I've been looking for this for the last two minutes. So here it is. If you're curious what the Columbus module on the ISS looks like. Not that world's best resolution image, but there she is right there. It is literally a tin can. It's like one of those eight ounce Cokes you get at the dentist's office. You don't really want very small little squat thing. All right, Ethan, back to you.

[Ethan Baragas]

No, I mean we think about it the same way. So you can have this higher level thing planning these individual tasks and then you have to get much less high fidelity data of that, you know, one to one human operator. And then you can start to expand this out across and then, you know, to kind of jump on. You know, one of the biggest tasks for us is, you know, every about 60 to 90 days, depending on resupply schedule, we send up three and a half tons of cargo. And we have some of the most highly trained people, highly trained scientists, just go move the bags in and out. It's kind of ridiculous that they do this, you know, out of a 16 day duty cycle. They'll spend 14 full days just moving things, things. And so that's why we look at things very similar to the construction industry of that dump truck problem of I've just moved this bucket of two tons of dirt. How do I actually control this? Because that changes your entire vehicle dynamics. And if you think about it in space, you move your arms forward, your body moves backwards, and so this problem becomes magnified tenfold. And the only ways that we found to solve this is to use things like adaptive sliding controllers and use humans in the loop, but then also learn from that with RL and VLAs.

[Alex]

Adaptive, what was the mechanism you used

[Ethan Baragas]

to move things around, like adaptive sliding controllers? So controllers that will actually then change themselves based off of what you're picking up, based off of estimates that you actually have and in different environments.

[Alex]

Got it. So highly adaptable based on whatever the task is. Got it. Okay, Jeff, you want to weigh in here on building world models for zero G, because I feel like everyone's talking a lot of smack about not having enough data and this is really tough. And I'm curious when we're going to get Moon one from you guys. It's going to solve the problem.

[Boris Softman]

Hold my beer.

[Jeff Hawke]

I was going to say give us six months. And to be honest, none of this is intractable. I mean, for example, the hierarchical planning that you were sort of both talking about, this is pretty common. I mean, almost every autonomy stack has got looks of some kind of flavor of this at every point in history in the last decade, frankly. So the way I kind of see this evolving is you'll always have some kind of system decisions which will be dependent on the application that might be safety related or something like that. There's probably going to be some kind of controller on the loop. And the question then is, where do you want to devote the horsepower of AI in terms of key decision making? And there can be different abstractions essentially, as the core foundation models get better to maybe use in a language parallel LLMs are amazing, provided you can fit things in context and essentially you hit limits as soon as things drop out of context. So that's getting better. Right? That's almost as like this rising tide that improves the sort of core reasoning and planning capabilities of these models. War models will be very, very similar. Right. It's a relatively early technology in comparison. Imagine sort of the algorithmic march over a number of months. So what I would expect is basically a lot of the complexities of that to sort of collapse over time, which is what we saw play out in autonomous driving saw play out LLMs, the same thing will happen. So the question I think is how do we make a space world model? I think it's a combination of two things. Either a closing loop through a simulation environment where you can kind of close the loop through some like physical interaction manipulation environment with gravity set to zero. Great. That's something that any world more company can work with. Similarly from a data perspective, any observational data of visual state and what happens and how this works from a visual perspective also great. So I think it's just simply a matter of that and bringing that into the data mix. But there's nothing inherent beyond that that is intrinsically limiting. So yeah, if you want to send some data our way.

[Alex]

Yeah, I was just, just about to say that, like Jeff, So let's say Ethan is going to go up to space to the ISS with Voyager Technologies, a public space company. They're going to take their. Ethan, correct me here. Is it Joy or Joyride your robot?

[Ethan Baragas]

So, yeah, the robot is named Joy and the mission is Joyride.

[Alex]

Okay, I got this mixed up in the end. Okay. By the way, good branding. So Joy goes up into, into the Columbus module or whatever and bops around doing its testing and so forth. If they took telemetry data, sensor data, video data, as much as they could from a reasonably long mission, and sent it over to Odyssey and said, can you guys try to take this and improve a zero G world model for us? How far do you think a mission's worth of data from Icarus would get? Odyssey in building that, how long is the mission?

[Jeff Hawke]

Like how many hours?

[Ethan Baragas]

Yeah, we're up there for a full year. So Q1 27 to Q128. So we'd have thousands of hours to ship to you.

[Jeff Hawke]

I mean, I was going to say like 1000 hours is probably sufficient in most cases. Certainly maybe it's sort of a parallel. So we have some of our world models in use and robot foundation model applications we're not a robot foundation model developer. Rather we basically provide this as a model to people who then build on it. In this case, there's often some domain misalignment in terms of data, but it's like order of magnitude sort of in the sort of hundreds of thousands to maybe a low millions of samples. Usually gets us a pretty long way, might be a little bit more in terms of zero G, but there's also quite a lot of data that exists in the public Internet of things moving in zero G. So we're not starting from scratch in that regard. So certainly very happy to see add some zero G to the mix.

[Ethan Baragas]

Yeah, we'll ship you an SD card or something.

[Alex]

I think it's going to be a little bigger than an SD card. I mean, dear God, if that's all the data you send over, I think Jeff would laugh, he'd be like, what is this data for ants? A pathetic.

[Jeff Hawke]

I don't think I've ever worked with less than a petabyte of data in my life.

[Ethan Baragas]

Yeah, exactly.

[Alex]

Okay, but okay, this actually goes back to what Ethan was talking about earlier on about, you know, space distance and temporal lagging communications. If you wanted to get a petabyte of data down. Ethan, is that even possible with current throughput between us and leo low Earth orbit?

[Ethan Baragas]

Yeah, so there's a bunch of different communication schemes and they all have different latencies and different bandwidths. So traditionally to talk to the iss we use a S band relay. You go from Earth to Geo Geo down to leo, Leo back, the Geo Geo back.

[Alex]

Oh, you actually do the. Oh, I didn't know about the last loop there. Interesting.

[Ethan Baragas]

Yeah, and so this historically has very low bandwidth. Like for example, on that S band we'll get about a gig a day, like pushing it. And so this really sucks for us, but if you look at the new generation of infrastructure going up there, so I'll pick one of them out of a hat Voyager with STAR Lab or Vasp with Haven. These have laser based comps, so you get about 100 milliseconds latency and communication and you get about 25 gigabits per second. And so this is a massive increase from what currently is state of the art. And what's really exciting about, you know, what we're doing is while we won't have the full bandwidth while we're up there on the space station because it's like a 30 year old Toyota Corolla at this point, what we do get is we get all the data actually back down because we're getting the robot back. So it's coming down on a Dragon capsule and so actually be able to physically bring it down and actually process the data afterwards.

[Alex]

What's the backup if it doesn't?

[Ethan Baragas]

Yeah, yeah. And then we have the, you know, the one gig a day over the course.

[Alex]

That's the backup.

[Ethan Baragas]

Okay, yeah, that's really the backup. So you're going to lose a lot of data, but, you know, let's just hope that that Dragon capsule comes down and, you know, the astronauts don't die as well, so.

[Alex]

Oh, well, now I feel like an asshole. I guess there would be humans in there too. I didn't. I was just thinking about your one pathetic little SD card getting too hot and burning up. Boris, we're talking a lot about, you know, where the rubber meets the road here, or I guess where the robot meets the zero G. But you're definitely out there in the real world. I was really impressed that your company recently moved 65,000 cubic yards of dirt on one project alone. So are you now just like in data abundance as a company? Because I feel like if you've done that much actual earth moving, after all the training you've talked about, all the data, all the sampling, all the modeling, you must just have a ton of fresh and useful data that it's accumulating.

[Boris Softman]

Everything's relative. Where I think there's always, particularly given how many types of machines and types of work are out there, there's always a frontier of new diversity, new variety and so forth. It's funny, Zetan was talking, I'm thinking the data problem, everybody has a version of it where we generate a terabyte per machine per day, and so it's 1,000 gigabytes. And so it's like, yeah, not quite doable with Starlink. So we're still kind of shuffling physical data on and then there's kind of like working away, trying to hope that you compress your way into something that could be purely through satellite. Yeah. So for us, we have a really helpful property that we are operating on existing heavy machinery that we can retrofit. And so we do not need to create a brand new machine that is obviously expensive, complex, and also not in existence, which makes data harder. The fact that we can go and take existing $600,000 Caterpillar machine and turn it into autonomously capable, whether it's for data collection or for autonomy testing, that opens up a lot more flexibility in how we get the data. And so we are able to have really rich partnerships with general contractors and subcontractors both, you know, for data, but also to go. And in our actual deployments with them, like what you mentioned where we were helping build a factory in Arkansas, in this case, we've been on a number of data centers that a manufacturing facility, a couple warehouses, and we're not far off from going. Those were supervised autonomy tests where we were doing autonomy, but with a safety operator. And later this year, we'll be going full driver. So that means that these will be operating on real sites, doing work all day long with nobody in the cab. And in the initial use cases, it'll be mass excavation and then moving on to more and more types of work. But what's exciting is that some of these projects, you know, we're now doing projects that will be many hundreds of thousands of cubic yards of earth, and these are subsets of projects that get into the many millions of cubic yards of earth that become the opportunity. And so when you think of some of these data centers, for example, they're getting built, they'll be getting built on many hundreds, if not thousands of acres and can be, you know, 2, 5, even 10 million cubic yards of earth. And so these are the scale of projects that are actually happening.

[Alex]

That's almost as much dirt as Peter Thiel dug up to build his bunker in New Zealand. No jokes aside, 65,000. Sorry, it's not fair to make you guys enjoy my humor. 65,000 cubic meters is exactly 26 Olympic size swimming pools. So that's how much dirt you guys moved, in case you wanted it a more prosaic term. All right, so, so listen, all this is great. I'm excited about everything from world models to, you know, saving time, construction sites, building things faster, going up, learning. It's great. It's so great. It's optimistic, it's very human, positive. Why? Why does this not form more of the conversation we have at the public level? And what do we need to do as people who are really bullish on having more intelligent machines in the world and better models behind them? Change the conversation. I'm so bummed to read the news and read what people are saying, Read the polling about data centers and everyone being very concerned and then coming here, having a blast and just there's an

[Boris Softman]

optimism, right, like on what's possible with a lot of this work. Yeah, maybe I'll kick off. It does worry me because when you see approval rating for AI, like below 20%, that puts it in danger of being a bipartisan issue that actually impacts the space In a way that is detrimental not just the space, but to overall society. And so I think there's not enough the communication around it has certainly not been great, but there's so many successes in, you know, the breakthroughs you can start having in medicine, the way this democratizes education, like the way that, you know, today in construction, there's like a astronomical shortage because it's been systematically under invested in. And now when the demand is skyrocketing because of, you know, infrastructure and data centers on training, manufacturing, you start seeing housing prices go up and so forth, and you have a world that is like fundamentally can be expansionary and everybody does have a legitimate fear of local job disruption. But almost always it's much easier to see the local direct impacts where something is lost than all of the astronomic positives that have happened in every single technological transformation that are actually harder to predict. Where you see the mobile economies that started to form after mobile phones, what the Internet created, even if it was that hurt retail stores locally in some places. And the biggest example that I can remember that kind of hits on the physical transformations that we're talking about is the Industrial revolution, where it has that sort of a feel today where when it happened, everybody was petrified that all the jobs would be lost, there'd be complete unemployment, everything would be absolutely horrible and just a few people make money off of it. And what ended up happening is that even though there was locally some really painful kind of transitions that might be geographically or in profession fixated, by the time that dust settled, the productivity skyrocketed by orders of magnitude, plus the number of jobs actually increased, not decreased and the average salary doubled and the rate of poverty significantly decreased in the entire country. And so you have these really, really non zero sum impacts that I think we're not doing a good job of conveying external.

[Alex]

Jeff, there's the big sovereign AI pushover in the uk. I know there's talk about building kind of a UK specific AI model. I know you're going through some Prime Minister turnover, but it did seem that the Keir Starmer government was at least generally in favor of building out AI. So are y' all doing a better job with the picture, the optimism, the forward looking element of this, the we are going to have better medicines and better technologies and cheaper goods.

[Jeff Hawke]

Yeah, maybe the context, we're partly transatlantic company with actually half the team here in London and half the team in Palo Alto. I think AI is ultimately a global problem and a global question. And the reality is companies who are Pursuing this, particularly the sort of foundational scale, you have to think about winning this globally and I think it's a good thing. As Boris said, it's not a zero sum game. In my mind, it's more about focusing on how do we make this a positive sum game. And I think there are so many things where like many technologies in many respects, I think all of us here are building something that is not possible today and the world will be better at the outcome. And I think in many respects that's really why certainly what motivates me and I think certainly what motivates most folks in the AI field. So as to whether we'll be doing a better job, I'm not entirely convinced. I think every society is grappling with this in terms of what it means, how the world will work, how relationships between countries will work. It's a bit of an unknown, but I don't think the unknown is to be feared. I think in many respects it's a thrill of discovery and that motivates people to do ridiculous things like go into space. I think something to be celebrated.

[Alex]

I wonder, Ethan, if we just got everyone to read 10 times as much science fiction, we could solve this branding problem. Because if you pick up a copy of Red Mars or just even something that's older but just aspirational about going out into space or rendezvous with Rama, just pick another classic. To me, those got me so excited about the possibilities of going faster and further. Your company being a good, I would say, step towards our, our, our front porch as a species before we go off even further. How do we, how do we get folks to be a little bit more open minded and optimistic, especially amongst the younger set that are currently dealing with the butt end of AI today, which is a slightly softer job market for recent college graduates.

[Ethan Baragas]

You know, it's two things for us. Number one, we play on the end of physical AI in the AI field, but then also we work in the space industry and there's nothing more expensive looking than the space industry. When a rocket goes up, it looks expensive. It looks like someone took a pile of cash and lit it on fire and people are like, what does this do for me?

[Alex]

If you're New Glenn, that's true.

[Ethan Baragas]

You know, and like the thing about like programs like New Glenn is there's so much of the positive end where NASA and space have very bad PR problem. They don't like to take credit and they should take a credit for a lot more things than we have day to day. And we really have to ground what we do on how it impacts people's day to day lives. And I think that's the thing that every time I've sat down and had a conversation of why it's so important to do what we do, then this is what lands with people, you know, like for example, you know, your cell phone. The reason why we have miniaturized computers was because the Apollo program, we needed them to put on the rocket to take us to the moon. The reason why you have memory foam, that's NASA. It created Invisalign. It created from a NASA program Invisalign.

[Alex]

The things of your teeth, Invisalign.

[Ethan Baragas]

You know, that was a spin off of technology that was made by the NASA program and NASA funding things like, you know, your credit card, every time you swipe it, that timing signal that comes from a satellite, your GPS when you came into work, you would not have that today without a satellite, without the space industry. And I think the one for us is we have the most advanced lab in the world floating around us with people working on it every single day for the last 25 years. And they created this cancer therapeutic called Keytruda. It's the number one cancer therapeutic in the world right now. And so we did the research and development of this drug where you can do protein crystallization on orbit. And we found out the way that proteins crystallize in space, totally different than Earth, way more effective. We changed the way we manufacture this drug right now on Earth with a random positioning machine to try to replicate zero G, because we can't. And that one drug made $29 billion last year in revenue for Merck, the pharmaceutical company, and has saved hundreds of millions of lives.

[Alex]

There you go, that's the one you want. I think we need to drop the financial one, even though we're all capitalists here and just go straight to the petted kittens and planted trees. Because I think that there's so much money going around right now that every time we bring it up, it doesn't land well. Like the joke that I made at the start about how Icarus hasn't raised a nine figure series B yet is a joke amongst us because we know that the markets and problems are so big that it's worthy of investment. But to an average person they're like, my school fell down. Why are you giving Jeff $310 million? He's got nice hair, but does he really need all of that? Couldn't he get by with 35 million? So maybe you should see our GPU

[Jeff Hawke]

bill,

[Alex]

maybe we need to, like, you know, definancialize things a little bit. I don't know. I just. I don't want to live in a world where we have the shot at such rapid progress and a brighter future and we end up talking ourselves down or dumbing ourselves down to the point in which we slow down. And I don't know if we're going to lose to China, per se. I think it was Jeff talking about competition, but I don't even want that to be a possibility. As a fan of democracy,

[Boris Softman]

do we

[Alex]

need a new leader in technology? Have we just failed to find the right person to carry this torch?

[Jeff Hawke]

I think in many respects, revenge of the nodes is a good thing in minarespecs. I think a lot of what built the technology industry is people being like, oh my God, I can build something ridiculous in my garage and it's really, really cool. It enables amazing new things. And I think it's often been sort of couched in large business interests, which are true in many respects. It's a good thing. It brings the capital behind the problem. But at the same time, I think. I don't know, I think that soul is valuable and I think sometimes being sort of a bit more open or human with the public in that regard can be a useful thing.

[Alex]

Yeah. Boris, do you think we should have partial AI lab nationalization? We're seeing the horseshoe theory in practice here. We got Bernie Sanders and Donald Trump both coming to the same conclusion. I don't think it's a particularly good idea, but I'm happy to be wrong. What's your take?

[Boris Softman]

That kind of scares me as well. I'm with you. There's not a huge amount of examples that's actually worked out well in other industries. At the very, very least, it massively slowed down innovation. If you now have that gate across now, doesn't mean that they're given the power of some of these models. You don't have some checks on how and when, the sort of ways that they can be applied. There's export controls on GPUs, there's strategic reasons on why that might make sense. Right. But it kind of scares me. It's almost like saying we're going to government control software. Right. It's such a math in this case. Right? Math. It's like, okay, well, there was proposals to control models in general. It's like, well, at what point does it become a large model? What if you're one parameter under. It's just kind of silly in the big picture. And it's trying to kind of apply a hammer to something that is just operating on a very different dimension. So it worries me. But at the same time, I think it's probably not doing the industry. Is doing the industry a disservice to bulldoze forward and put your head in the sand and not actually try to communicate and kind of create a bridge to government, to the rest of society, where people do genuinely feel like they're not benefiting equally from it, even if a lot of their benefits are actually going to come in five years when all these technologies mature.

[Alex]

I think the only 100% losing strategy when it comes to selling AI to the public is to be dismissive and rude to people who have concerns, because I think what that does is just poisons the conversation. I'm not going to play it because I want to save a question for. For actually for you again, Boris, but I saw a video of Theo Vaughn, the podcaster who became very well known during the last election, and he was talking about data centers, and he was talking, what I would say is a standard tick tock set of points about data centers. And everyone was just dunking on him instead of trying to, like, you know, politely educate him, help him out, be supportive. They were just like calling him an idiot, calling him various slurs. And I was like, that's not at all how we're going to build the coalition here to keep this from becoming a bipartisan issue. It's going to get crushed. So I'm hoping for some more positivity. All right, I want to do one more thing before we go and talk about the return of Fable for fun. As a final question, Nvidia recently announced Cosmos 3, which is a, quote, open frontier model for physical AI and the world's first omni, modal, unifying vision, reasoning, world action, end generation. So, Jeff, are they encroaching on your core domain here, or is this something that's more like a test bed for other folks to learn from?

[Jeff Hawke]

Overall, I think it's positive for us. In many respects, they're another voice sort of advocating for this as a category. And Nvidia historically has done this where they use their sort of research arm as a way to basically proof markets and validate markets. And essentially as soon as the market is mature, they typically go and find the research tiers another way. And this has been played out quite a number of times. So, yeah, in many respects there are things that are comparable to our models. There are things that are different in terms of our models. There's some great researchers on it, but I'm quietly confident in our models.

[Alex]

Yeah, well, I was just thinking that Nematron, the models from Nvidia are a good example of open weight AI from the United States, which we could see a lot more of, and I think we wouldn't be any poorer. Now, Boris. There was also something from Nvidia called Halos for Robotics, which is full stack open robotics safety system. Is this something that you can then bring to your robots, or is this something you've already figured out and they're just helping other people avoid the, not the dump tech problem, but the swinging bucket problem?

[Boris Softman]

We're constantly kind of like surveying and kind of trying out these tools. When you get into these incredibly complex systems, we've at least found that you have to more deeply, vertically integrate and really control the full stack because you have to deeply understand the interface for the, the latencies, the physics and controls, the cycle counts, the performance of every model. You have to have an architecture that is guaranteed to fail safe. And so we are constantly looking at these types of tools as a support for various building blocks. But I think given that there's still so much novelty to these large kind of physical systems and things are just not standardized, and how does a car work versus an excavator versus a humanoid machine? There's no magic bullet. You kind of tend to have to still have a pretty deep control all the way through the stack.

[Jeff Hawke]

I think it's an understatement. Every single robot, furniture, model, company, including autonomous driving that I've spoken to has a uniquely different view and sample rate, frame rate in terms of the action rate, in terms of where they parameterize trajectories, including resolution, the number of cameras. Do people like diffusion? Do people not like diffusion?

[Boris Softman]

It's like, what's safety critical? What's not safety critical? What's your safety case? Yeah, it's almost like you can't Mr. Potato head it. Like, it almost. It just never really works.

[Alex]

Like, when you talk to them, they're all so confident that they're correct and everyone else is wrong. And I'm always like, guys, you can't, you can't all be. But maybe you can all be right, but you can't all be wrong. I, I don't know these very smart people.

[Jeff Hawke]

This is natural. Like, I mean, you tend to see this, right, where everyone kind of fans out and you kind of imagine you're kind of like throwing darts at a big dartboard and you're trying to get points and everyone's sort of exploring and over time basically things converge. I think autonomous driving is sort of a little bit ahead of the curve compared to some of the other fields just because it's a bit more of a mature industry, a little bit more defined problem. At some level the objectives will converge over time. It's just a question of enough people getting enough information. You end up with cross pollination between companies. This will start to consolidate in methodology. I mean.

[Alex]

Okay, and so do you think once we get the methodology kind of aligned or harmonized, do you think we'll be able to go faster or will that actually slow us down down because we'll be exploring fewer odds and ends and nooks and crannies in the research game.

[Jeff Hawke]

It'll make my life easier. I will get such a wide range of feature requests.

[Alex]

Jeff, I'm not actually optimizing for your personal retirement. I'm not going to lie.

[Jeff Hawke]

Typically people consolidate around things which work better. And typically something that works better in sort of adjacent domains tends to work well in another. So maybe things which work well in autonomous driving will probably have in many cases will work well in maybe construction, although maybe that's a little bit different given you also have got manipulation of your environment at the same time. So there are adjacencies and typically where there are adjacencies you accelerate.

[Boris Softman]

I'll take one slightly different lens on this one, is that this focuses on the autonomy model and kind of like the core approach of development. What's deceptive is how much of the challenge of solving, for example something like autonomous driving is actually not at all in the autonomy model. But, but how do you evaluate how are you able to improve your system when you're in a Super long tail 100 millions of mile case without regressing accidentally in 15 other places. How do you actually qualify it? And so there's actually so much surface area that in a lot of ways people have gotten quite good at climbing quality of models. When you have the right data and so forth. And the hardest problem eventually becomes how do you get the right signal that you actually can drive 500 developers to optimize over or how do you actually test it and make sure make sure that it's qualified and ready for driverless. When you know you you have safety critical applications, it's actually harder in your

[Alex]

case with bedrock because if you think about like modern equipment, it's often hydraulically driven. Right? If you're doing a big, a big digger and you know what's not super precise hydraulic driven Large buckets full of dirt. And so earlier in the conversation, you mentioned something like precision and like getting within an inch or something. I've seen these machines, they're not always very well maintained. So you probably have to also take into account, like, you might tell it to do something and you're going to get a slightly different outcome and have to handle that as well in the real world.

[Boris Softman]

So that's exactly right, because you now have a variance where you do not. Like, you know, Waymo has thousands of jaguars on the road. They're tuned to be like, as similar as you possibly can be. You know, you have like 72 different types of attachments to an excavator. You have tons of different sizes from, you know, like in terms of tonnage and, you know, advances as well. And so what you have to do is create safety cases that are actually resilient to this and work at everything from subsystem to kind of overall system level. And so, for example, you can have controls, systems, impose systems that estimate and all these models, and you can empirically kind of like measure your errors and your variances and collect data incredibly quickly over a wide range of machines and have very precise tolerances and be confident that you have a particular type of margin of error that's going to be your some giant percentage of time you'll be within this sort of tolerance. And then part of this is knowing that you will have exceptions where, for example, let's say a sensor broke or a machine is just like fundamentally different and you fall outside their range. You have to have fail safes to where it's always preferable to fail safely than to try to do something that's outside of your space. You've qualified.

[Alex]

Ethan. It's funny how much you're in the opposite problem to this. Because you're building one robot, you're controlling all the sensors, all the tools, all the pieces you don't have to go out in the field with. You know, someone didn't maintain this digger, so it's a complete mess. How precise do you need to be when it comes to your machine interacting with the physical world? Are there still tolerances you can play with or is it like millimeter precise requirements?

[Ethan Baragas]

No, you still have to be within those centimeters. Especially because the environment that we're in is working hand in hand with humans. And so where you can have very clear workspace with something like large hardware that's on a construction site and make sure not many people are around. So even if you do go into whatever failure state you're in and you can be safe for us. If we go into the wrong failure state next to somebody and you hit them, it's not like there's a doctor up there. Everyone's trained in medical, but that can be a very bad situation. So when we talk about the systems that we're interacting with, from the switches to the science experiments, to even the cargo bags, these are made for human level fidelity. You're very good at grabbing very small things. You have tabs and zippers that are sub centimeter sized that we need to grab ourselves. And if you take robotic grip and you kind of break it down, you get two things, you get pinch and you get power. And so the way that we've made our end effectors are actually to replicate that. So it's actually just two small fingers because of how precise we need to be where we're flipping switchers on certain experiments to turn them on and plugging in mil spec connectors. And for us, the two things that make it hard is, number one, every single action in space, the reaction is just so exaggerated. On Earth, you have one G dampening every single movement. For us, we just don't have that. You move your arms forward to the side, your entire body starts rolling, your cameras start rolling. And then when you pick up things, sometimes we're picking up things that are more massive than our vehicle itself. And so that shifts outside of your control matrix of your thrusters, that shifts your center of mass way outside of this. So traditionally, if you have your center of mass inside of this and you try to strafe left on the X axis and just move, you'll start pivoting in a circle. If you've sufficiently moved your center of mass far forward enough. And so these are the types of things that we have to really care about to get our precision fine enough that you can do the tasks that matter.

[Alex]

You know, I find it really funny that the reason why a lot of people want to build humanoid robots and they've, they've told me this is because you can take a robot then and put it into a human environment and it can perform well. And they often do this in places where you have lots of room and could actually change the definition of work and create more of a purpose built robot. You're going into a literally human defined space that's very small, very complicated, and you're not going humanoid. So I wonder if everyone else kind of has it backwards because you're making, it looks like a robot. If I imagine like, you know, what are the dimensions on Joy? Like a pizza box by like four pizza boxes?

[Ethan Baragas]

Yeah, that's actually like perfect. I've never thought of explaining it that way, but that's probably a great visual.

[Alex]

You're welcome. But it's that size and it's supposed to interact with human things quite well. So, Ethan, are people just is figure at all making the wrong bet when it comes to what they're actually building to do the work?

[Ethan Baragas]

So I think it really depends on the environment and what you're trying to do. Like humanoids, the reason why these are compelling is economies of scale. If you can get something that does every single task on Earth made for a human at 80% and you can just print these, it makes sense. On the financial side of things, you'll never beat a purpose built machine or robot for some task. With a humanoid, it just won't happen. But that's not the point of humanoids. The point is I can make a hundred thousand and employ them everywhere and democratize labor. And when you look at the Earth and you say, okay, half of the Earth's GDP is people doing things, it's human labor. There's a very compelling argument there. When we look at space, we say, oh my God, I'm going to go look at these astronauts. I'm going to go talk to these astronauts. What do they actually do in this environment? Turns out, legs suck. They don't help at all. They mess around your center of mass. The only things you ever use your legs for is hooking into footholds so you can react against the vehicle itself so you're not floating and spinning in a circle. And so when you break down the environment and first principles, a human has been evolved to work on Earth very, very well, but not space. We sent them there. It kills them. You know, your heart degrades, your eyes go bad, your bones. Exactly. And so when we broke it down to first principles of what was most important, you know, it takes shape as something very, very different. But the machine itself and the robot itself is still a general purpose platform. And so I think there's a very big difference between humanoids and general purpose platforms. Depending on the environments, you don't want a humanoid to go swim underwater. You want something that's an underwater ROV or similar class. The same way that you don't want a human to fly in the air, but you might want a human to walk on Earth. So I think depending on your domain, you get very different general purpose machines.

[Alex]

All right, listen, I could keep asking you guys Questions literally all day long. But it is getting a little bit late in London so we're going to, we're going to wrap towards. Don't let me for this, wrap towards a final quick round of questions here. So for each one of you, one, when does Fable come back? I'm curious. And then two, are we still approaching the GPT3 moment for physical AI or have we already gone past it? And Jeff, we've heard the least from you lately, so we'll start with you.

[Jeff Hawke]

Oh, when we're getting Fable back Soon I hope would be my thought. I have no unique insight into when they might come back. I think that's probably out of my hands.

[Alex]

You have to put a date on it. You can't, you can't waffle. You can't waffle here. No one pretend you're from Texas.

[Ethan Baragas]

Two weeks.

[Alex]

Okay, good.

[Jeff Hawke]

In terms of GPT3 moment for physical AI, I don't think we're there. I think in my mind the GPT3 moment is when you see this spike of sort of explosive rollout and demand. I think we're seeing promising signs, but I don't think we're there. I think we're on the trajectory.

[Alex]

All right, Boris, same questions for you.

[Boris Softman]

Oh, let's see. When does it come back? Well, and not in a form where they kind of like Nerf the capability, but for real, I'd say it sounds like it's a little bit longer because they got a little bit of a mess with like kind of, you know, government now they got to, you know, degavit. So I don't know, I would say eight weeks is my guess. A couple months.

[Jeff Hawke]

Yeah, I prefer two weeks.

[Boris Softman]

Yeah, me too. I don't know, I'm a little less optimistic. Sounds, sounds messy like Jeff, I actually think we're not there. Maybe a little bit more pessimistic on the GPT3 moment anytime soon. Just because that benefited so much from almost infinite data in a common feature space of language. There's so much nuance and variability to the hardware, the sensors, the environments, the safety challenges, everything that I think you will for longer see these more verticalized applications that really focus and, you know, take things end to end. Even though the industries can be double digit percentages of gdp, I think it'll be a while longer before you just have something that can genuinely go into, you know, brand new environments and be super capable like this. But likewise, I don't think it's impossible. I think we're on the way. It's just Going to take longer. Okay.

[Alex]

No, I really appreciate the context there, Ethan. You get the last couple of words here. So when does fable come back and the GPT3 moment for robotics, future or past?

[Ethan Baragas]

Yeah, so when it comes to, you know, like mythos and fable, I don't think we're getting anytime soon. I'd put it, you know, two and a half, three months.

[Alex]

This is getting worse.

[Boris Softman]

Guess you'd be like pessimistics, like.

[Ethan Baragas]

No, I mean, the government moves at the speed of government. I think once you classify something as a weapon and you put itar controls on it, this just makes everything so much more messy. You know, we've had our dealings with them. Just getting documents signed takes weeks. So I can imagine that this will take a very long time. And then like as far as GPT moment for physical AI, I think we're so early. I think it's a far, far time away. And I think like Boris said, you're going to find very vertical niches in very, very constrained use sets right now. And then once you see that demand signal spike and once you see the capability spike is when it will explode. And so I think that's what makes it exciting for me is because we are on the greenfield, bleeding edge of it right now and I think like two, three years time we're going to see that massive explosion.

[Alex]

Well, it's really optimistic because I hate doing work and I hate driving and I like to write. So if anyone can take care of all the physical stuff, I'll do all the typing and life will be good. All right, guys, we got to leave it there. This has been this week in AI My name is Alex and today we are joined by my new besties, Boris from Bedrock, Jeff from Odyssey, and Ethan from Icarus Robotics. Thank you all much. We'll have you back and we'll see everyone next week. Goodbye.