00:19:18 · Adam Miller
[00:19:18] S0: So I'm sitting here in Chiperdal right next to our office, and Sachin and I just came back from New York. We're on a long documentary tour there, and now we're just heads down building our Stealth startup. And the thing I can't really get out of my mind is this idea of trust. Can customers really trust that what we're building is gonna help them create great content and marketing that helps reach the buyers that they need to reach. That's the thing I think about all the time. And that's one version of trust, but there's so many different types of trust you need to build with customers. Does your product work? Is it gonna help them get value? Is it gonna work in six months? Do they trust your business model? Do they trust your team? There's so many different things. It takes a really long time to build that trust. And that's data security and compliance. Because with a company like Vanta, they can help you automate your compliance in weeks instead of months. We all know those acronyms, SOC two, HIPAA, the complex acronyms that enterprise buyers will throw at you when a deal starts to get serious. Vanta helps you take care of all of that. Using Vanta is really a no brainer solution. Those aren't things that I wanna focus on. I want Vanta to do that for me. Because, you know, the worst story that you'll hear about in startups is when you've found great customers, you've built a great product, you're talking to these big enterprise customers, and you're, you know, you're going over a deal, you're negotiating, you're getting to the final dotting the p's and the i's in the contract, and then they realize that you don't have all the compliance that they need and the deal falls through. No startup founder wants that to happen. We want companies like Vanta that can take care of that for us so then we can just focus on our product and on our customers. So if you're interested in this, head to vanta.com/sachinandadam and you'll get a thousand dollars off.
00:21:04 · Anastasis Germanidis, Sachin, Adam Miller
[00:21:04] S2: Yeah. So this is, essentially, this is an example of, our world model being used as a simulator for robotics. So what you're gonna see is, our robotics model that's basically, rolling out a policy, which is basically performing an action, performing a task inside the imagination of, GWM one, which is our world model. So [00:21:36] S1: what we're looking at is instead of a robot going into this desk room or whatever and and actually just collecting training data, you're collecting training data through simulating it through the model [00:21:46] S2: exactly and you're essentially simulating how a robotics model will perform and you can see you can simulate things like you know like folding clothes which is traditionally very difficult things that to simulate with traditional methods [00:22:01] S1: and what are the inputs of these simulations actually look like? [00:22:05] S2: So the inputs is essentially you're providing you're controlling this robot arm in a similar way that you could control the robot arm in the real world so the input you're providing is the movement of how how that robot arm moves in space. [00:22:21] S0: Yeah. Are we looking at AI generated video or are looking at a robot that has been trained on it? [00:22:27] S2: So this is all generated footage. So our... This is our world model GWM one that's generating two different views of a scene and a robot performing this task. [00:22:40] S0: And with your recent models, you're able to generate a hundred milliseconds per frame, believe. How is that even possible?
00:22:40 · Adam Miller, Anastasis Germanidis, Sachin
[00:22:40] S0: And with your recent models, you're able to generate a hundred milliseconds per frame, believe. How is that even possible? [00:22:48] S2: It's a difficult research and engineering challenge to essentially start from a base video model that's traditionally very slow to getting to distilling it so that it can actually, you can run with inference at real time, frame rate. So, it's it's necessary in order to provide interactive control, and it's something we've been really focused over the past past year of how to distill our base video model to be real time. [00:23:14] S1: If I think of some of the kind of companies we profiled, physical intelligence, we went down to SciQuantum in the South Bay a couple of days ago. I I imagine there's been so many use cases of technology like this for companies like that. In terms of becoming like physical intelligence, how do you think they could use technology like this?
00:23:14 · Sachin, Anastasis Germanidis, Adam Miller
[00:23:14] S1: If I think of some of the kind of companies we profiled, physical intelligence, we went down to SciQuantum in the South Bay a couple of days ago. I I imagine there's been so many use cases of technology like this for companies like that. In terms of becoming like physical intelligence, how do you think they could use technology like this? [00:23:32] S2: Yeah. So a company like physical intelligence and other robotics labs and, you know, we're we're in collaboration and conversations with a lot of them. They spend a lot of time running evals, basically understanding how the robotics models that they've trained can perform, in in the real world, and it's very difficult in order to differentiate between, you know, you've trained a few models, you wanna see which models are good at what tasks. That's extremely time consuming to do. Very often it can break the robots, like it's... And so being able to minimize the number of real world evals that you have to do, by doing the evals in simulation and scaling them, you know, as much as you want is is very, very useful and kind of significantly accelerate kind of the the research, kind of progress of them. [00:24:22] S0: And is there a big risk there? Because this data would need to be pretty perfect. Right? Because if you're using synthetic data to train other models then that would just degrade quality over time right unless these sort of generations are near perfect?
00:24:22 · Adam Miller, Anastasis Germanidis, Sachin
[00:24:22] S0: And is there a big risk there? Because this data would need to be pretty perfect. Right? Because if you're using synthetic data to train other models then that would just degrade quality over time right unless these sort of generations are near perfect? [00:24:38] S2: So they don't need to be perfect as long as the performance inside the world model correlates with real world performance, I think it doesn't need to be that, you get this perfect kind of, recreation of how a robotics model will perform in simulation versus performs in real world as long as two robotics models, if one wins over the other in simulation it also transfers in the real world. But yeah there's definitely a lot of work that we're doing to further improve the diversity of the data we're trying on to make sure that, we improve the realism even further, but it's still quite useful to be able to, to basically create scenarios so quickly and increase and that makes the models a lot more robust as well because now they can be trained on the same the different variations of the setting different variations of the task of the objects and so they're much more robust in different variations in the real world as this is is also an interesting one to show, so this is a generated video of, an accident in an AV setting. [00:25:51] S1: Wow, so that could be used for, self driving cars and?
00:25:51 · Sachin, Anastasis Germanidis, Adam Miller
[00:25:51] S1: Wow, so that could be used for, self driving cars and? [00:25:56] S2: Yeah, so, so self driving and AV research specifically, it's very difficult to collate data of, edge cases of crashes of accidents it will be very expensive to generate to create data sets of accidents but it's very easy to generate those scenarios with a world model and so you can basically create you know, adverse weather conditions, different scenarios of pedestrians moving, moving in the streets, accidents, and that really helps in terms of, like, improving the robustness of the self driving models that essentially control the car. [00:26:39] S1: And when would this technology actually be available and where do you think is the first use case where it's gonna go really mainstream? [00:26:47] S2: So this is already... These kinds of models are actively being used today, to accelerate robotics research. So we're, we're working with a lot of robotics labs to essentially, incorporate GWM one into into their work. So this is already available and already proving to be useful. Of course, there is a lot of room to further improve those models, but GWM one is already kind of essentially in the private beta at the moment and will be generally available in the coming months. [00:27:21] S0: Wow. Sounds exciting.