This Week in Startups
00:36:06 · Oliver Cameron, Jason Calacanis, Lon Harris
[00:36:06] S2: Simply that if you want the fly's path to be the best possible performance, that's where you focus your GPU time. And so everything outside of that, unessential to be computed. And so I think we... You know, some of our work is around how do we scale up agent learning to massive scale. Right now, we can support something like, 20 agents inside a world model session simultaneously. [00:36:38] S0: Got it. [00:36:38] S2: And so we can generate 20 viewpoints, unique viewpoints, simultaneously. Now how do you scale that to a billion? How do you scale that to 8,000,000,000 to Yeah. [00:36:49] S0: If you were actually simulating eight eight billion people's reality, half of them or a third of them would be asleep at any point in time and dreaming or not dreaming. [00:36:58] S1: But, I mean, if we really are in a simulation, simulation, how do you know there really are... Like, you only know there's 8,000,000,000 people on earth because you've read it, because you've been told, like, we don't... Maybe all of us who are being simulated are in Texas and nothing outside of Texas exists. And you know, [00:37:13] S0: like Yes. When I go to Japan, it pops up Japan as a concept and a [00:37:17] S1: it because now one of the people with actual perspectives are there. There could be just 10 of us in here.
00:37:17 · Lon Harris, Jason Calacanis, Oliver Cameron
[00:37:17] S1: it because now one of the people with actual perspectives are there. There could be just 10 of us in here. [00:37:22] S0: And and that's a limitation of the tech. Now if we were to look, Lon, at what we experience as humans in evidence, we try to understand things like dark matter. We try to understand the universe. We try to get off this planet. Pretty hard to get off the planet. It's pretty hard to get to the core of the planet. It's pretty hard to get to the Mariana Trench, but we have these little interludes where we can get to the moon or we can send a probe out to the end of our solar system. We can do our solar system, but we don't get much further behind that. Then whatever else we get is part of the big bang or not the big bang, this crazy story we tell each other that we're kind of guessing on, but it's incredibly low fidelity. That is evidence that you're in a simulation and that what Oliver just described to us, like why would I do the compute of more than 20 people in the simulation? I don't have that kind of compute money. Nobody does. And we took all the compute on the planet, Oliver, we took all the computer on the planet, we gave it to you for simulation. How many people would you have in the simulation? A million? 2,000,000? Yeah. Have you done that calculation? [00:38:31] S2: Yeah. But we we haven't. No. But I I do think... So LOGO is the double slit experiment, or it's inspired by the the double slit experiment. [00:38:39] S0: Explain.
00:38:39 · Jason Calacanis, Oliver Cameron
[00:38:39] S0: Explain. [00:38:40] S2: Yeah. So the double slit experiment is this famous experiment that basically says when you pass light through two slits and you observe it going through those two slits, do you get a pattern that is, conducive to light being a wave or to a particle? And, that is something that we very much believe is indicative of us understanding, how the universe works. When you see something, it observes this pattern. When you don't observe something with an instrument, it has this different pattern. And that might be indicative of some sort of, I don't know, performance characteristic of the universe. [00:39:30] S0: In the double slit experiment, you you you have, like, a light going through two slits in a piece of wood, a piece of paper, whatever it is, and then you have it projected. And so here's your light source. It goes through a metal sheet, and then it creates light, and then the light hits a screen. Now, when you're observing it as a human, you see a different pattern than when you're not observing. In other words, the observer changes reality. Am I correct in describing that as a layperson there, Oliver? [00:40:03] S2: To an extent. I don't think we really understand yet what observation really means. To us, yes, we're observing it vis a vis. Is it that simply articles are interacting? Is that observation? And there's really limited understanding of this.
00:40:03 · Oliver Cameron, Jason Calacanis, Lon Harris
[00:40:03] S2: To an extent. I don't think we really understand yet what observation really means. To us, yes, we're observing it vis a vis. Is it that simply articles are interacting? Is that observation? And there's really limited understanding of this. [00:40:18] S0: And then popular science fiction, Lon, you have this moment in time when Kryptonians come to Earth because they have supernatural perception. When they take their mask off and they get hit by the sun, they hear everything on the planet. They see everything on the planet, and general Zod drops to his knees because he's disoriented, and Superman's been on the planet for thirty years. [00:40:40] S1: Earth's yellow sun. Yeah. [00:40:42] S0: Yeah. But Superman's learned how to tune things out and fine tune it. So what is the business here, Oliver, now that we've gone down a science fiction rabbit hole for fifteen, twenty minutes, which makes it super interesting. You know, obviously, you're not building a simulation to, have, you know, philosophical discussions like this and pop culture discussions. You're doing it for a business. What is the business here, and how close are you to that business?
00:41:05 · Oliver Cameron
[00:41:05] S2: So the business is world models for robots and world models for agents. And before starting Odyssey, I spent eight years building driverless cars, and it turns out we were building that technology all wrong, I believe. And the reason we were doing it all wrong is because what intelligence should learn from is a very general set of concepts before it tries to accomplish a task. Now what we did in Driverless Cause is just show the same task over and over again. We showed those models, here's how you drive a car, here's how you drive a car, here's how you drive a car. And so then we're a little bit surprised when it can't handle crazy edge cases or it gets stuck and does something dumb. Well, the reason it does that is because all it's seen is the world through the lens of a a car. And so, really, what a world model proposes to be is a much more general foundation for robots, big and small, like cars to build upon. And so the paradigm now with world models is that you take this very general foundation model, a world model, and then you show it just a few hours of experience. So for example, we as humans, we learn for seventeen years, then we go learn how to drive, and we're exposed to about twenty, thirty hours of experience of driving a car. And so the equivalent of that for world models is that you show it twenty to thirty hours of humans driving cars. And then what we see is that the model's able to take those very general concepts and the twenty to thirty hours of experience and then become a really good driver. And this same concept applies to robots that you can show at twenty, thirty hours of manipulating, some task that you want the robot to do, and then it can perform that task with lots of robustness to the world. And so this this whole... That's our business, to build that foundation model to then work
00:43:01 · Jason Calacanis, Oliver Cameron
[00:43:01] S0: on customers. Are robotics companies, factory... Companies that own factories, or own depots? [00:43:08] S2: We we sell to physical AI companies. And so that's robot labs, humanoid labs, self driving car companies, drone companies. They take our base model, build upon it, and then they have more reliable [00:43:18] S0: Have have they started, actually paying for it, and they don't see that as their core business? They think it's better to just buy yours because it's, you're working on that problem, and it's like they got enough problems on their own to get the hands working on these things. [00:43:30] S2: Yeah. Yeah. We generate revenue today. And I I think something changed when language models, really landed, which was... I think everyone went from in deep tech, we have to build everything ourselves to be more open to doing post training as the value add. And so the robot labs we work with, they primarily post train Odyssey-three, our latest foundation world model, and then they deploy that model into their robot, then they deploy that in the field. And so that's definitely been a shift, in the in the space.