Going Direct Conversations
00:41:44 · Jordan Wolfe, Chris Paxton
[00:41:44] S2: and what's your goal of of um of doing the the weekly pod [00:41:50] S0: So this is actually Michael's idea, for what it's worth. So he was the genesis of it all. So I think he's the person to ask. But I think the idea is. Yeah, go ahead.
00:42:01 · Michael Cho
[00:42:01] S1: Well, I mean, there's a very selfish reason because I would love to talk to all these researchers and ask them questions that I don't understand. I still don't understand. The more I get into the podcast, the more I realize I don't understand. But I think it's a good thing, right? That's the only way to learn, like get really uncomfortable, but then just ask them. And I think it's just a great privilege for, you know, both of us to be able to talk to these researchers, ask them all these questions. And also that there's nothing like it, like, you know, I think, uh, There's really nothing like it even now. It's a very small audience because it's very technical. But those people who care about this technicality, they will listen. I think it's a decent couple. I don't know, Chris, how you think about it. The way I think about it is Chris obviously is a real expert. I'm like the guy who is trying to, who knows enough to ask, but then I'm like, I feel that I can be that bridge to maybe bring it down to earth a little bit, you know, and talk like a, because I'm really not a researcher. So I think that's a decent combo, but it's still quite technical to be honest. And we want to keep it that way. Because if you want to order like, you know, the big industry application, there's so many like VC podcast that I think that the market has plenty of that. But for something for this particular niche, I think there's literally no podcast like that, like long conversation with the technical expert doing very, very specific research and just get them you know, to talk about their work. And usually they're very willing to share and tell you very frankly, what is the capabilities and more importantly, what are the limitations? And sometimes you can't really just get that by reading the paper is you get a lot more to try them talking. Yeah.
00:43:41 · Jordan Wolfe, Chris Paxton, Michael Cho
[00:43:41] S2: Yeah, absolutely. Is there an area of research This is Chris for you. Is there an area of research that the more folks that you're speaking to at RoboPapers that's getting you really excited about maybe some kind of breakthroughs coming through around robotic training or anything else related to the field? I think the thing that has been most exciting lately has been reinforcement learning. [00:44:06] S0: There's a lot of interesting stuff going on. A lot of people are looking into how to do reasoning models for robotics and how to do long horizon reasoning for robots. Like I said, that's stuff that I've always found very fascinating. I don't know. There's so much exciting stuff happening, though. I think there's lots of separate little things which are really cool going on. [00:44:27] S1: Chris, I kind of feel that maybe you have changed your mind a little bit about like eco-centric video as we do more podcasts around research around that. I don't know why Chris has his doubts. [00:44:44] S0: Yeah, definitely. I definitely have my doubts. I still have plenty of doubts, but that's, yeah. [00:44:51] S2: I'm coming around about it. I'm definitely coming around about it. Why do you say that? First of all, you say ego or ecocentric? [00:44:58] S1: Ego. Ego. Ego-centric videos. [00:45:00] S2: Oh, ego. Ego, like first person. So tell me why you said that, because that's interesting to me. What's changed?
00:45:00 · Jordan Wolfe, Michael Cho, Chris Paxton
[00:45:00] S2: Oh, ego. Ego, like first person. So tell me why you said that, because that's interesting to me. What's changed? [00:45:06] S1: Well, because I know like Chris is pretty bearish about this stuff. But I mean, we still try to keep an open mind, you know, so we've talked to a bunch of researchers that have papers on it. And I kind of feel that Chris is, I mean, he has his doubts, you know. I think that's the thing with experts, right? Because of his years of experience, he must have biases, right, naturally. Whereas I'm like a blank piece of paper, so I was like, oh, okay, I have no idea. So as an observer, I'm just looking. And so even for Chris, I can see, okay, I mean, I should speak on behalf of you, Chris, but I think it's, yeah, it forced us to open up, yeah, a little bit, maybe previous biases. Yeah. [00:45:49] S0: Definitely. I think it's cool to see people, like, there are so many different opinions that people have, lots of cool research. Yeah, I feel like I learn a ton from this. I definitely feel like I'm, I mean, the fact that I've been doing this for so long makes me, like, by default, bearish about most things. Like, once you end up going into some stuff, like, I mean, mind changes. Yeah. It's changed a lot, honestly, over the last year. [00:46:16] S2: So talk to me about the egocentric data for you, either historically or why you're kind of more bearish on it. Just kind of explain why from a research standpoint.
00:46:16 · Jordan Wolfe, Chris Paxton
[00:46:16] S2: So talk to me about the egocentric data for you, either historically or why you're kind of more bearish on it. Just kind of explain why from a research standpoint. [00:46:28] S0: Historically, I actually talked about it earlier, is that it requires more engineering than it looks to make it work well. I think that's historically been the case, but these days, like Michael was saying, there's some stuff, the things that seem to change it are that you can now, with modern model architectures and training. It seems to be that you can do a little bit more co-training, do a little bit less hand engineering to pull the right frames out. With a lot of this, yeah. The problem with most robotic stuff, like you were mentioning earlier about the business side, and whether or not it was ready, the problem with most robotic stuff is that it generally takes a couple PhDs to make it work. So the more engineering you have to do, the less I buy it as something that will... scale, I guess. And that's been the source of my skepticism a lot of the time. It seems like wishful thinking that this will work. But then you start to see, oh, yeah, this is actually kind of working for people. All right, maybe my mind can be changed by that. Yeah. [00:47:44] S2: Cool. Yeah. Michael, I have to ask you. You mentioned incentive structure and to collect data and crypto is a way to use incentive structures. So that's super interesting to me for a variety of reasons. But break down a little bit more about what you're doing there.
00:48:09 · Michael Cho
[00:48:09] S1: Yeah, so to be honest, I was a crypto skeptic three, four years ago until I discovered this project called Helium. These days, they call these decentralized physical infrastructure networks. So, I mean, the way maybe I explain it is, I mean, if you look at, let's say, for example, everyone knows Bitcoin. Bitcoin is a project that has never raised any money. And yet it has amassed a huge amount of compute and every year it's spent, I read somewhere like 0.5% of global electricity bill is used to mine Bitcoin. So huge amount of resources gathered without raising money as a project. So there is something very unique about crypto that is very different as a way to do capital formation or basically gather resources that is different from the limited liability company kind of what we're used to for a few hundred years, right? And yeah, so I do think it maps quite well for some of these things we need to advance and body AI. A lot of this data, I think ultimately should be collected what they call in the wild. I think in theory, you can spin up some warehouses to collect teleoperation data full and close, but I tend to think that those models that are trained just using like robots in the warehouse will overfit to those control environment. Ultimately, these AR models are just compression of the underlying data. So actually, I have a sense that for research to, I mean, if you want to build very, very generalizable model, you want the underlying data to be collected in the world out there. And so it seems to me, it's something very hard to do centralized. And crypto is a great way to just, it's like a bit permissionless, right? You get people to just show up, hey, help me collect this data in your home, and then we give you a token in exchange, right? It's almost like, kind of like what, I mean, fundamental Bitcoin is doing, right? It's like, you help me do this, I give you Bitcoin, something like that. So, and I feel that if structured the right way, you can get a lot of resources at scale globally. Whether, when I say resources, it could be the teleoperator, the man hour, right? Someone needs to control the robot, someone needs to deploy the robots, someone needs to pay for the robots, right? And so, and even real estate, right? I mean, these robots need to be in somewhere, right? So paying for the real estate as well.