Dream Machines: Physical AI, Robotics, Being Human
01:04:48 · Evan Helda, Charles Wong
[01:04:48] S1: Mhmm. [01:04:48] S2: Right? And for humans, when we humans learn, we just call this, like, either... You know, it's called either doing the thing or imagining it. Right? Like, the difference between the way robotics and human... The robot and humans learn. Right? Like, if you're a human... If you're a robot... Like, let's let's start with robots. Right? Let's say you're a robot. You make a mistake in in production, and you have to go back, and then your, you know, your your engineer has to figure out where did it fail, how do we improve this, how do we collect more real data, you know, improve the dataset, retrain, redeploy, yada yada. Right? For humans, let's say we go to work, we make a mistake, what happens? We go home, we reflect, we ruminate, we play out several scenarios in our mind of how this could have been better, and then tomorrow we come back to work, we do a better job. There was no retraining, right? There was no, you know, figuring out, you know, collecting more real data and practicing until we get it right. You could just... For many permitting tasks, right, you can just sit and reflect and imagine. Right? So I like to think of imagination as a proxy for this, where, like, eventually robots will get to such a point where they are able to autonomously imagine and self learn and self correct on the job, where this paradigm of, like, taking it back and retraining whether with real or synthetic data, that will subside or, like, reduce significantly. Yeah. Yeah.
01:06:08 · Evan Helda, Aravind Kandiah, Charles Wong
[01:06:08] S1: What's the role of your of your business in that future? I mean, that could come pretty quickly. Like, I just had a meeting with a customer that's building a, not a VLA, but a video action model where this is what it does. It it it imagines the split second future. It's about a good experience, and then it knows how to go do the thing. And then you have, like, world models that are starting to accelerate, in quite a way. And and I don't mean the world labs world models. I mean, like, the world models understand physics and cause and effect in lots of settings. To me, this sounds like this is a huge risk to Bifrost, and Bifrost doesn't really need to exist in that world. Tell me where I'm right or wrong. [01:06:54] S0: Yeah. So there's there's two parts of that. Right? There's two futures. One future is imagine intelligence is solved. Right? You you have a brain that can work anywhere. Same in physical intelligence and skill, they release the the final API. Right? Yeah. The final API. They can they can do anything and everything. That's a [01:07:15] S1: cool idea. I'm writing that I'm writing that down. Final API. [01:07:17] S0: That's a that's a [01:07:18] S1: catchy title for some kind of essay. [01:07:20] S0: All these credits, man. I gotta I gotta look [01:07:22] S1: at these credits. Is what I'm not getting credit for. I might just steal it, but that's okay. [01:07:28] S2: I'm I'm a professional world, everyone. Groupies, so, you know, do it that way you will.
01:07:28 · Charles Wong, Aravind Kandiah, Evan Helda
[01:07:28] S2: I'm I'm a professional world, everyone. Groupies, so, you know, do it that way you will. [01:07:31] S0: Indeed. Indeed. So the... So you mentioned you have the final API. Right? And you are working in a in a science lab. Right? And you're like, okay. You know, I I can't hire any lab assistants. Let's go buy Unitree G1, and then we're gonna load up the physical AI API, and I want it to do all these very domain specific tasks for me. Are you gonna, a, just load it up and hope it works in the lab? Or are you going to load it up and run it through a whole bunch of specific scenario tests and make sure you can do the task first before loading it into the real robot? So our thesis is actually that you won't need training at all in the future, but you will need very thorough rigorous evaluations for domain specific things. [01:08:19] S1: Mhmm. [01:08:19] S2: Right? So it's
01:08:19 · Charles Wong, Aravind Kandiah, Evan Helda
[01:08:19] S2: Right? So it's [01:08:19] S0: like, you wanna do it for domain specific testing, will need to evaluate it for your domain specific things, and you can't just evaluate it in real life across the thousand scenarios that you wanted to do. And every time a new update comes or a new, like, know, you just fine tune on the edge and you update the model, you still want to CICD your model. Right? So there will always be a need for testing simply because we're just deploying the systems into, like, the real world where real infrastructure, real lives are all at risk. Right? So for us... And we see that happening today as well. Right? Like, most of our customers actually use us for eval more than they use us for training. Right? So we're seeing that shift slowly [01:09:00] S1: That's fine.
01:09:00 · Evan Helda, Aravind Kandiah
[01:09:00] S1: That's fine. [01:09:00] S0: Where it's more eval centric. But to Charles' point, you know, everyone's like, okay. You know, if we can load up our world model, a world action model, and we do test time inference. Right? So we we roll up 10 different policies. And, you know, like, this is, approach one x takes as well. It'll be highlighted recently as well. But I think there's a there's a few problems there. One is taking a world model and running it at test time inference is crazy because you need to do 40 different rollouts in faster than real time on an edge device. What you will need is you will need an abstracted low resource, low compute, way of rolling out these scenarios. And if you have a three d representation that is low resource, low compute, you could actually roll out things much, much faster. Right? And we think that's a... It could be another thing, but that's, like, much, much further down the road. For us, like, the main thing is, like, focusing on evals. But if we could become that imagination engine, and that's like that that moonshot thing. Right? Imagination engine that can actually run on edge compute. That's a that's a thing that we're trying to work. And it's actually a tractable problem. It's just like, okay. When is the right time to get into that market?
01:10:20 · Charles Wong, Evan Helda, Aravind Kandiah
[01:10:20] S2: Yep. We believe it's a compute problem, and that compute problem will be solved eventually. Right? We are going to build that evaluation framework for robots to imagine whether they're, you know, plugged into an Ethernet cable and and, you know, interfacing or something else, or they're doing it on their own brains, the difference is just a a compute problem. Right? So we build a framework now. We build a great evaluation framework now. We, you know, we integrate all the latest models, you know, that all these labs are coming up with. Right? We create that easy test bit for any developer to, you know, quickly activate self learning on their own robots. And it's only a matter of time until robots get smart enough where you can actually just put in their own brains. So [01:10:58] S1: Very cool. I'm excited to see this this eval platform and product. So there's... I'm I'm seeing a lot of these eval products pop up. Before we wrap, I'm curious, how are you guys? And you... We've kind of danced around a little bit here. You've gotten the signal... Some sense of that. But what's gonna make your eval product different? And and then, you know, I'm I'm thinking about like the like the light wheels of the world. There's a few other companies like this. How do you think about really being differentiated as an eval platform? You [01:11:31] S0: will see in a month. [01:11:34] S1: It hasn't launched yet. Sorry. It's just launching in a month. Got it. [01:11:38] S0: Yeah. We're going live in a month.