Sources with Alex Heath
00:51:34 · Alex Heath, Tony Zhao, Ellis Hamburger
[00:51:34] S2: I needed that tonight, Chris. [00:51:37] S0: I I think the idea there is, like, when when the robot is doing certain things, right, it's it's actually delivering the hand to where it wants to be. But the rest of the body, let's say, where elbow is, doesn't have to be at the exact elbow position where it is right now. Right? Let's say you're you're... You have a kids and they bump into the robot, the robot shouldn't be there, like, rigid, but it should kinda tucked in and maybe, like, yield when someone hit it. So this is what we mean by compliance that it is not like a industrial robot, that all it does is to follow the rigid instructions going from point a to point b, and no matter what you put in front of it, it will it will destroy it. Right? That's not how we think home robot should be. I think to achieve that is, again, a mix of, like, very very deep into the system side that we need to choose the right hardware, and we need to do... Build the right electrical system so that we can enable the hardware to be compliant. And then our model also need to consider that. So it's it's always like a complicated answer that involves all the parts of the system that we need to, like, put in, like, right at the beginning. [00:52:42] S1: I think I need to upgrade my hardware because I break glasses and plates constantly, expensive ones. And it just, I mean, it just makes you wonder like, are we gonna see silverware, forks, knives, plates, bowls specifically designed for robot? [00:53:02] S0: Robot friendly? Yeah.
00:53:02 · Tony Zhao, Ellis Hamburger
[00:53:02] S0: Robot friendly? Yeah. [00:53:04] S1: Yeah. Robot friendly, or do we not have to worry about that because Memo can handle your grandma's China just fine? [00:53:12] S0: I think that's a that's a very good question of whether there there exists like robot friendly, like utensils or [00:53:22] S1: OXO Good Grips, the most goaded brand in all of kitchens worldwide with those little rubber grips on them. I love that brand.
00:53:30 · Tony Zhao
[00:53:30] S0: Yeah. And I think maybe the broader question here is how can we... It's like when we deploy a robot into people's homes, I think the adaptation... Of course, the robot will adapt to how you live your lives. Right? But I think human will also adapt their behavior to the robot. For example, like, let's say you have a butter knife that is very thin. And for some reason, after you eat, you just really want to put it on a table. And it's just essentially flat with the table, and the robot really struggles to pick it up. Right? And what you you could do is, like, you know, keep doing this and the robot keeps struggling. Maybe it take, like, five minutes to pick it But what you could also do is put it on top of a plate, and that is infinitely easier for the robot to pick up without you actually intentionally trying to do things that are so different. So I think there will be things like this happening in the future that we will also kind of evolve our lives a little bit in exchange for more convenience that a robot can bring us. For example, when you're choosing a new coffee machine and you realize there's one one of them, the robot knows how to use, but the other one doesn't, and they maybe produce, a similar amount of, like, quality of coffee. And I think people will go for the ones that they can have the robot to do. Right? So I think it's a exciting feature too for us as well is that we don't need to handle, like, a million types of coffee machines, but we own... I would, like, only do handle the ones that are more popular, and we can kind of change users' behavior a little bit by steering them into getting, like, the ones that works well.
00:55:12 · Alex Heath, Tony Zhao
[00:55:12] S2: How long before we get to what you're talking about? I mean, I've asked before, like, when are these gonna be in homes? But are you... Is this like a ten year vision? Is this a five year vision? Like, how far out are we from this actually in your mind? [00:55:25] S0: Yeah. It's it's definitely less than five years. And Really? Yep. Wow. Because, like, if if you think about where are the fundamental bottleneck is, it's data. Like, I think the... I'll put another hot take here is I don't think robotics, like, training an AI to to, like, do households, right, is fundamentally harder than, like, math Olympiads or, like, all the crazy reasoning models that we see right there. In terms of the fundamental AI architecture, I think the rest of the fields, large language models or vision, are actually quite a hit. And the only reason that the robots cannot do the choice right now is the lack of data. That is almost like the singular biggest bottleneck that is hurting progress. Because without the data, you cannot put in... Like, do very meaningful research to iterate on figuring out what are the right methods and figuring out all the science questions. We're just, like, blocked there. But once we're able to unblock the data part, I think I think the capability will be more like an exponential growth than, like, we're kinda stuck somewhere. So that's, like, kind of how we think about it. It's all about the data, and we want to get as much data, as diverse data as possible.
00:56:42 · Alex Heath
[00:56:42] S2: And you think the best way to get the data is the hand versus just like training on videos of humans interacting with the real world or all the... You know, how Elon talks about FSD is the way that you're gonna learn for optimists. Like, everyone has a theory of how to get the training data right now. Right? And yours is a very unique one. I got... I have to give it to you. Like, I haven't seen anyone else approach it that way. And you obviously have conviction that that is the way. And I'm curious, like, you probably evaluated all the other ones. Like, are all the other ones just dead ends?
00:57:14 · Tony Zhao
[00:57:14] S0: I don't think that that's how I'll put it. Is I think it's true that right now, I think robotics is still at a at a fun part that, people are debating. Right? And these days in in large language models, people don't debate about, like, whether you need training. Right? They're like, this is... Like, give me money. I'll give you the the corresponding results. Right? There's sort a fun part that people disagree that there are many fun ideas that are... If you are, like, kind... Trying to, like, combat each other with narratives, right, essentially. But given the amount of information and the concrete evidences that we saw, we think dealing with gloves, like, collect the data, transfer it into the robot is the best way. But this is not... It's a strong opinion that is, I would say, like, weekly health. As in if there is another breakthrough that we see that changes some of the assumptions, we're more than willing to, like, update our prior and figure out how to best leverage that. Right? And that can come in a lot of different flavors. Let's say, world models that people talk a lot about these days. It could be simulation. It could be, like, anything. But the piece that will never change in my opinion is, like, real high quality data. And that is going to be the concrete thing that we'll have either way that would differentiate us no matter what's the downstream algorithm is. Let it be like training the robot directly with, like, reinforcement learning, with imitation learning, or more models. I think we're largely open minded about how to use it.