Latest / Automated with Brian Heater / Russ Tedrake on Robotics, Physical AI, and the Future of Work

Transcript

Russ Tedrake on Robotics, Physical AI, and the Future of Work · Automated with Brian Heater

• 0:00
  SPEAKER_02: Machine learning success empirically has gotten far ahead of our ability to understand it theoretically.
• 0:07
  SPEAKER_02: I feel like we've had to change from being engineers where we design everything on first principles to we're becoming more scientists, almost like behavioral scientists, where we're building things we don't fully understand and now we have to go probe them to figure out what the heck just happened.
• 0:19
  SPEAKER_01: Things are changing in such fundamental ways that it must be both an exciting and very challenging time to be running a physical AI startup.
• 0:27
  SPEAKER_01: Totally.
• 0:28
  SPEAKER_01: But one has to be focused enough to not chase after every shiny bauble.
• 0:34
  SPEAKER_01: How do you determine when something is a shiny object and when something is like actually a real game changer?
• 0:41
  SPEAKER_02: That's a good question.
• 0:43
  SPEAKER_02: So I'm spending a lot of time these days talking to labor economists and with people that I think could be potentially impacted and just trying to build my empathy muscles and build my understanding.
• 0:54
  SPEAKER_02: I do think if we're successful, the field is successful, then it's going to change the very nature of work for people.
• 1:00
  SPEAKER_02: And it's going to have a hopefully hugely positive impact on the world.
• 1:04
  SPEAKER_01: I'm sure that there have been plenty of points in your career where you've been approached to do something like this, and I'm sure that you've thought about it in the past.
• 1:18
  SPEAKER_00: First of all, I think Hello and welcome to another episode of Automated.
• 1:33
  SPEAKER_00: My name is Brian Heater.
• 1:34
  SPEAKER_00: I am the managing editor at the Association for Advancing Automation.
• 1:38
  SPEAKER_00: Here's a uh great conversation from our marathon of Boston shows from a few weeks back.
• 1:44
  SPEAKER_00: Russ Tedrick is somebody that we've been looking to get on the show for a while.
• 1:48
  SPEAKER_00: I figured that we would have to wait until uh he officially had something official to announce.
• 1:54
  SPEAKER_00: But it turns out that showing up at his doorstep also worked.
• 1:58
  SPEAKER_00: Uh more info on that new project soon, but here is some insight into his work in the space to hold you over.
• 2:04
  SPEAKER_00: If you're enjoying the show, don't forget to like and subscribe, and please check out the newsletter over at automated.fm.
• 2:11
  SPEAKER_00: And with that, please enjoy this chat with Russ Tedrick.
• 2:16
  SPEAKER_01: You grew up in Detroit.
• 2:17
  SPEAKER_01: I grew up just outside of Detroit.
• 2:19
  SPEAKER_01: Your father was an automotive?
• 2:21
  SPEAKER_01: That's true.
• 2:22
  SPEAKER_01: General Motors.
• 2:23
  SPEAKER_01: So you were like, in a sense, you were kind of in and around automotive manufacturing automation.
• 2:28
  SPEAKER_01: Yeah.
• 2:29
  SPEAKER_02: My dad actually uh collected classic cars.
• 2:32
  SPEAKER_02: We had a drive-out basement and we had a 21-packard touring car and a 48 Cadillac, which was the very first of the tail fit.
• 2:39
  SPEAKER_02: But uh but I learned a lot about cars back in the day, yeah.
• 2:41
  SPEAKER_02: Tell me a little bit about the uh the internship that you had at the uh Ford plant.
• 2:46
  SPEAKER_02: So when I was in high school, so a lot of the um the big car makers had a lot of small contract shops uh there that were around town.
• 2:57
  SPEAKER_02: There was one that my neighbor ran uh that did paint shop improvement.
• 3:01
  SPEAKER_02: So they went into paint shops in the local plants and they would optimize the airflow and other things about the paint process.
• 3:09
  SPEAKER_02: I took an internship there over the summer.
• 3:11
  SPEAKER_02: So I got to write some automation that tried to control the airflow through the plants.
• 3:16
  SPEAKER_02: Uh, I was the junior guy on the totem pole, so when somebody had to climb into the sludge pit with the sensor, that was my job the first time.
• 3:23
  SPEAKER_02: Uh, but I did work up to you know writing some of the initial codes that would sort of automate the systems.
• 3:29
  SPEAKER_02: The most defining part of that experience for me was the day that my, I thought, clever code, which when there was an exception, like somebody tripping over an Ethernet cable, uh, I thought the safe thing to do would be to turn off the air fans.
• 3:45
  SPEAKER_02: Velocity, you know, airspeed go to zero.
• 3:47
  SPEAKER_02: Uh, but that was the wrong thing to do because the temperature in the booth went from, I don't know what it was, eighty one to eighty-two degrees, which was the unionized um threshold when people walked off the line because they didn't have to work at 82 degrees, and I got yelled at.
• 4:01
  SPEAKER_02: Like I really learned a lesson, a hard lesson that day about um what it means to stop the line in an automotive plant.
• 4:08
  SPEAKER_02: But it was yeah, I grew up a lot that summer.
• 4:11
  SPEAKER_01: Yeah, I was so yeah.
• 4:13
  SPEAKER_01: I mean, I I I guess also the lesson is just like really kind of like knowing the product that you're creating and knowing all the competitions.
• 4:20
  SPEAKER_02: But um, but it was yeah, it was awesome to to see the plant in action.
• 4:24
  SPEAKER_02: That was in the Ford Wayne Assembly plant.
• 4:26
  SPEAKER_01: I mean, it's wild that they entrusted you with something like that.
• 4:28
  SPEAKER_01: There were people watching me, but I but I did screw that one up myself.
• 4:31
  SPEAKER_02: Yeah.
• 4:32
  SPEAKER_01: Yeah.
• 4:33
  SPEAKER_01: Um, and then uh you were also uh studying video games at a certain point.
• 4:37
  SPEAKER_02: That's true.
• 4:38
  SPEAKER_02: So I um when I was at University of Michigan, Go Blue, they just won the national championship last night.
• 4:44
  SPEAKER_02: Uh robotics program was kind of in flux there.
• 4:46
  SPEAKER_02: Dan Kodichek had actually just uh been leaving right around then.
• 4:50
  SPEAKER_02: Uh John Laird was a was studying AI there.
• 4:54
  SPEAKER_02: I worked with him, and the closest thing to robotics I could get into at the time was doing AI for video games.
• 4:59
  SPEAKER_02: So I got to spend time with John Laird programming AI for video games.
• 5:04
  SPEAKER_02: I spent a summer at Microsoft uh research doing uh AI for video games at Microsoft Research.
• 5:12
  SPEAKER_02: Yeah, that was my entryway.
• 5:14
  SPEAKER_01: I mean, it it sounds in a sense like in hindsight like a pretty good entryway, especially looking at the way kind of NVIDIA and everything played out as far as simulation and and physics and robotics.
• 5:25
  SPEAKER_01: For sure.
• 5:25
  SPEAKER_02: Yeah, no, it it worked out.
• 5:27
  SPEAKER_02: I learned a lot.
• 5:28
  SPEAKER_02: Uh also just being at Microsoft at that time was really a fantastic experience.
• 5:33
  SPEAKER_02: I think the company has has grown and changed a lot since then, but I really enjoyed those summers.
• 5:38
  SPEAKER_02: Yeah.
• 5:39
  SPEAKER_02: How how did you get into bipeds?
• 5:41
  SPEAKER_02: So when I first came to MIT, uh so Gil was actually just Gil Pratt was just uh um he was running the leg lab, but he was leaving the leg lab, so he didn't take on new students at the time.
• 5:54
  SPEAKER_02: But I got to hang out with Jerry Pratt and Dan Pelluska and a few of the Pratt's.
• 5:59
  SPEAKER_02: Yes.
• 5:59
  SPEAKER_02: Um uh I got to hang out in the bottom in the basement of NE43 where the leg lab was.
• 6:04
  SPEAKER_02: Pete Dilworth was a was a guy there at the time.
• 6:07
  SPEAKER_02: Finger across yeah, Wobbleworks.
• 6:10
  SPEAKER_02: Yes, yeah.
• 6:10
  SPEAKER_02: Um, but at the time he built Trudy the dinosaur robot, which to my young eyes was just like that's what I want to do with my life.
• 6:18
  SPEAKER_02: Uh so I hung out with Pete and with Dan, and I worked on M2, and I worked on Trudy a little bit, and I fell in love with that.
• 6:26
  SPEAKER_02: So um I had to find my own path uh through that.
• 6:30
  SPEAKER_02: And the the path I took was through passive dynamic walkers.
• 6:33
  SPEAKER_02: So walkers that were uh inspired by simple toys that could waddle down a ramp, but showed that the dynamics of walking was beautiful.
• 6:41
  SPEAKER_02: And so that was my my thesis was a little bit more on the dynamics of walking.
• 6:45
  SPEAKER_01: Showed that the dynamics of walking was beautiful, but um I don't know, uh something that could actually be integrated into robotics at a certain point.
• 6:55
  SPEAKER_02: If you think about what robotics was at the time, there we had uh a lot of robots that were walking very conservatively.
• 7:01
  SPEAKER_02: So like a stiff astronaut, you know, kind of yeah, I think Asimo kind of kind of.
• 7:05
  SPEAKER_02: Asimo was there.
• 7:06
  SPEAKER_02: Asimo was actually beautiful, but even before that, that we had a lot of really stiff walkers, really slow.
• 7:10
  SPEAKER_02: We had Mark Raebert had built some hopping robots that were throwing themselves through the air, which was extremely inspiring.
• 7:16
  SPEAKER_02: And then there was this kind of outlier, a guy, Tad McGear, and then Andy Ruina at Cornell also followed up wonderfully on this work that built these robots that you put on the top of a ramp, you give them a little push.
• 7:28
  SPEAKER_02: And some of them had like incredibly human-like gates, way more human-like than what we were seeing from the actuated robots.
• 7:35
  SPEAKER_02: But they only walked down a ramp.
• 7:36
  SPEAKER_02: So it seemed a natural milestone would be to be able to figure out how to add a little bit of actuation, let physics do most of the work, but um, you know, add a little just a little kiss of of power and control to make it walk around.
• 7:49
  SPEAKER_02: And so that was the inspiration.
• 7:51
  SPEAKER_02: That was my little thesis robot.
• 7:53
  SPEAKER_02: I used machine learning, I used reinforcement learning back in the day, before it was cool, um, to make that work.
• 7:59
  SPEAKER_02: So I had a little robot that learned how to walk in 20 minutes.
• 8:01
  SPEAKER_02: That was the basically the thesis work.
• 8:04
  SPEAKER_01: So so the idea was if we can get it to uh you know move down a ramp, then you know, the learnings uh potentially we can use those to get get it to walk, you know, in other situations.
• 8:16
  SPEAKER_02: You got exactly right.
• 8:16
  SPEAKER_02: So the you know, learning to walk from scratch was considered very hard at the time.
• 8:21
  SPEAKER_02: And so we basically kept it, it was already walking if the with the motors were commanding only zero, and then you just slowly lowered the ramp and it basically kept learning, kept learning, and then just walked right off the end of the ramp.
• 8:32
  SPEAKER_01: So as as somebody who was, you know, who was sort of steeped in in I mean who was steeped in in bipeds at the time, um are are you surprised at all how things have sort of really come around and come around this quickly to bipeds and humanoids?
• 8:47
  SPEAKER_02: It was always the dream.
• 8:49
  SPEAKER_02: I actually I'm um am I surprised?
• 8:51
  SPEAKER_02: I'd I think uh it's amazing to me that it is as turnkey as it has become, that is awesome.
• 8:59
  SPEAKER_02: That is success.
• 9:02
  SPEAKER_02: The hardware is maybe more amazing than I might have dreamed in terms of the cost and the capabilities that we're seeing.
• 9:08
  SPEAKER_02: Uh but this, I mean, I'm I always believed that's why I was doing this.
• 9:12
  SPEAKER_02: Yeah.
• 9:12
  SPEAKER_01: But it just it it does seem like, I don't know, it it strikes me that maybe we skipped a couple steps in there, especially when it comes to building, you know, useful uh robots in the fields.
• 9:22
  SPEAKER_01: I mean, we were out there talking to uh Locusts yesterday, and obviously when we talk about really uh successful robots out there, uh we think about AMRs.
• 9:33
  SPEAKER_01: Um there there is kind of a uh a halfway point or or you know, like a missing link between an AMR and and a humanoid robot.
• 9:42
  SPEAKER_01: Um are we going after AGI?
• 9:46
  SPEAKER_01: Are we going after general purpose maybe a little too quickly right now?
• 9:51
  SPEAKER_02: Well, so um I don't think of walking the solutions that we have for walking, I don't think of as being directly on the path to AGI or or or needed for that.
• 10:00
  SPEAKER_02: Um I think there was inevit maybe inevitably, people who were living the dream of trying to understand, you know, biology makes this thing like bipedal walking look easy, and our robots are sort of failing at this.
• 10:13
  SPEAKER_02: So what is that's a that's a puzzle that needs to be solved.
• 10:17
  SPEAKER_02: And we've had uh you know great minds pushing on that for for decades.
• 10:22
  SPEAKER_02: So it's great that it's been solved as well as it has been solved.
• 10:25
  SPEAKER_02: You know, I don't know that there's I don't think it's completely solved, but I think it's it's pretty turnkey at this point.
• 10:30
  SPEAKER_02: Uh separately from that, I do think there's something extremely exciting about building more humanoid-like robots and exploring what AI can potentially do with a uh powerful general purpose body.
• 10:44
  SPEAKER_02: I think that's uh you know one of the most exciting things that's happening in the world right now.
• 10:48
  SPEAKER_01: Yeah, you were talking about being uh ahead of the curve when it came to machine learning.
• 10:52
  SPEAKER_01: Um what what was the state of machine learning at the time and and and you know, in what ways was it being deployed in robotics?
• 11:00
  SPEAKER_02: So people knew about reinforcement learning for a really long time.
• 11:02
  SPEAKER_02: Rich Sutton and others had been um vocal advocates for many, many years.
• 11:08
  SPEAKER_02: Uh I remember in 2004, that was when I did my thesis.
• 11:11
  SPEAKER_02: Peter Abil had his uh his helicopters that were doing a type of apprenticeship learning.
• 11:16
  SPEAKER_02: And I had my little toddler, we called it toddler, the learning robot, that was just using very simple reinforcement learning to learn how to walk.
• 11:24
  SPEAKER_02: You know, that was the first time in a handful of years that we had seen RL and learning working on real robots.
• 11:30
  SPEAKER_02: So that got people kind of excited, but there was still a lot of skepticism.
• 11:34
  SPEAKER_02: I mean, I used to teach in my class, I used to teach um a whole curriculum of robotics, and I had a section, I had like a third of the class was set up for reinforcement learning.
• 11:45
  SPEAKER_02: And you could tell people were just kind of like, yeah, I don't really care.
• 11:48
  SPEAKER_02: I don't think this is gonna go the distance.
• 11:49
  SPEAKER_02: You know, reinforcement learning's a cute idea, but it's not gonna really work.
• 11:53
  SPEAKER_02: And then I actually kind of slowly weeded it out of the curriculum uh because people just didn't seem that interested in it.
• 11:59
  SPEAKER_02: Do you get do you get a sense of what the source of the skepticism was at the time?
• 12:03
  SPEAKER_02: It wouldn't scale, right?
• 12:04
  SPEAKER_01: That it it was too data inefficient, sample inefficient.
• 12:08
  SPEAKER_01: Okay.
• 12:08
  SPEAKER_01: So like the same things we're kind of talking about right now, it sounds like uh what changed is the amount of compute that we could throw at the problem.
• 12:15
  SPEAKER_01: But the the data problem is still a conversation we're having.
• 12:18
  SPEAKER_02: Aaron Powell The data problem for manipulation is a slightly different one.
• 12:21
  SPEAKER_02: You know, I think what those um the capabilities in locomotion were really I mean, I a couple magical things happened.
• 12:29
  SPEAKER_02: Yeah.
• 12:30
  SPEAKER_02: First of all, what we learned is if you can do the right amount of domain randomization and in simulation, take your robot model and walk over stairs and walk over bumps and just a handful of different things, somehow that is good enough to make a robot walk over almost anything in the real world.
• 12:48
  SPEAKER_02: That was um not expected that it would be that easy.
• 12:52
  SPEAKER_02: It works incredibly well.
• 12:53
  SPEAKER_02: It's just a huge result.
• 12:56
  SPEAKER_02: Separately, the simulation and GPU infrastructure has made that just absolutely and and you know it's now open source available.
• 13:05
  SPEAKER_02: Um that the fact that anybody can do that and that that recipe is kind of out there is just amazing.
• 13:11
  SPEAKER_02: Like that is that is fantastic.
• 13:13
  SPEAKER_02: So um I think I also do think that simple policy gradient style reinforcement learning worked with neural networks as the as the parameter, you know, parameterized model worked better than most people thought.
• 13:28
  SPEAKER_01: I am always fascinated when when people, when really smart people who work on these things and know these fields really well use words like somehow.
• 13:37
  SPEAKER_01: It you know, it sounds like there was this a a surprise, and at least at the time, you know, you weren't quite able to explain why it worked as well as it did.
• 13:46
  SPEAKER_02: That's definitely true.
• 13:48
  SPEAKER_02: I have I have used that analogy um several times in the last few years where I feel like machine learning success empirically has gotten far ahead of our ability to understand it theoretically.
• 14:01
  SPEAKER_02: And it's I feel like we've had to change from being engineers where we design everything on first principles to we're becoming more scientists, almost like behavioral scientists, where we're building things we don't fully understand, and now we have to go probe them to figure out what the heck just happened.
• 14:16
  SPEAKER_02: It was a fascinating transition.
• 14:18
  SPEAKER_02: How does that change the nature of your role?
• 14:22
  SPEAKER_02: Totally changes that.
• 14:23
  SPEAKER_02: I mean, you have to you have to really uh think more like a scientist that's trying to well, there's there's one view of the world which is we don't have to understand.
• 14:32
  SPEAKER_02: Let's just pour in more data, pour in more compute, you know, uh follow the recipe, and we'll get better and better things out.
• 14:39
  SPEAKER_02: I suspect that's not quite your view.
• 14:41
  SPEAKER_02: I think you have to do that also.
• 14:42
  SPEAKER_02: Okay.
• 14:42
  SPEAKER_02: I do subscribe to that.
• 14:44
  SPEAKER_02: But I also want to go through and try to understand what the heck just happened.
• 14:48
  SPEAKER_02: Uh maybe just for my own psychological well-being.
• 14:52
  SPEAKER_02: Possibly because the maybe the field will hit a wall.
• 14:55
  SPEAKER_02: Who knows?
• 14:56
  SPEAKER_02: Um and then we'll, you know, our deeper understanding will help us push through when that come time comes.
• 15:01
  SPEAKER_02: Uh, but I don't think we've seen the limits of scaling yet.
• 15:04
  SPEAKER_02: I think we've got more headroom to go.
• 15:06
  SPEAKER_01: Well, yeah, I mean, uh obviously, I don't know, not obviously.
• 15:09
  SPEAKER_01: Um, again, this is come coming at it from somebody who like very much is is not a scientist who is a uh you know a liberal liberal arts major.
• 15:17
  SPEAKER_01: But to me, it seems obvious that you would want to understand the machinations there.
• 15:23
  SPEAKER_01: You want to understand how they're working so you can recreate them.
• 15:26
  SPEAKER_01: So, you know, you can continue that research going on.
• 15:29
  SPEAKER_02: That is the goal of science, I think, is really to understand.
• 15:32
  SPEAKER_02: I mean, we have the opportunity to understand intelligence, you know, physical intelligence in this case, right?
• 15:37
  SPEAKER_02: It used to be we'd have to, I don't know, probe uh Drosophila or or you know, uh, you know, there's all kinds of ways you could get around it.
• 15:46
  SPEAKER_02: But the fact that we're creating things that are incredibly capable, uh, and we have the code, and we can we can measure activations and we can start asking fundamental questions.
• 15:57
  SPEAKER_02: We can change the data recipe, we can change the learning rates and and start to really study these things in an unprecedented way.
• 16:03
  SPEAKER_02: So that's a huge opportunity.
• 16:05
  SPEAKER_01: What do you mean when you say ask fundamental questions?
• 16:08
  SPEAKER_02: You mean well, what type of fundamental questions one might want to ask?
• 16:11
  SPEAKER_02: Right.
• 16:11
  SPEAKER_02: So I think um there is a just an understanding of like why does it work at all?
• 16:16
  SPEAKER_02: Like why does something as simple as gradient descent on these super over-parameterized model um manage to find the things it seems to find, right?
• 16:25
  SPEAKER_02: So there's a whole theory of deep learning which has uh got incredibly talented people making uh lots and lots of more insights.
• 16:33
  SPEAKER_02: Uh, I'm actually super excited about how that field is progressing and the talent that that is in that field.
• 16:39
  SPEAKER_02: There's also sort of things that are, I think, one step closer to practical for a robotics practitioner, right?
• 16:44
  SPEAKER_02: Where we want to understand how to efficiently use data.
• 16:47
  SPEAKER_02: Like if you want to um you want to construct a data curriculum for your robots to learn as much as possible with as little compute as possible, let's say.
• 16:55
  SPEAKER_02: Um, you know, I think theory can potentially help with that.
• 16:58
  SPEAKER_02: Certainly extremely well-organized empirical science can help a lot with that.
• 17:04
  SPEAKER_02: So um understanding robustness, understanding privacy, understanding all, you know, there's all these sort of fundamental questions that you want to understand.
• 17:12
  SPEAKER_01: Yeah, and it's interesting.
• 17:13
  SPEAKER_01: I I've sort of up till this point, I've been really thinking about this in the context of, you know, ostensibly we're kind of talking about obviously you at TRI, and and you know, you're in the process of launching this startup, but for a very long time you've also been a professor and you've been in academia, and um, you know, it strikes me that I don't want to say it's difficult, but it's it's like uncharted territory to be uh teaching something that you don't like fundamentally understand.
• 17:44
  SPEAKER_02: Oh, that's interesting.
• 17:45
  SPEAKER_02: So um, but you can still give people I mean, people teach science.
• 17:48
  SPEAKER_02: Uh, you know, we teach about biology, even though our understanding of it is is uh via probes and experiments and a you know amassed understanding.
• 17:55
  SPEAKER_02: And so I think I do think maybe we have to teach it a slightly differently.
• 18:00
  SPEAKER_02: We give um people experiments that they can probe and and and the like.
• 18:05
  SPEAKER_02: You know, here's a here's a set of code that can teach a biped to walk.
• 18:09
  SPEAKER_02: Go mess with it, figure out why it works, you know, um, or build it yourself.
• 18:13
  SPEAKER_02: That's what's so amazing right now.
• 18:15
  SPEAKER_01: Yeah.
• 18:15
  SPEAKER_01: So a lot of it is really um like teach me how like how this thing works.
• 18:20
  SPEAKER_01: Yeah, there's a lot of that.
• 18:22
  SPEAKER_02: I mean, it's so I think it's both uh amazing and probably challenging to be a student in 2026.
• 18:28
  SPEAKER_02: I think you can do almost anything.
• 18:30
  SPEAKER_02: Like the world is at your fingertips.
• 18:32
  SPEAKER_02: You can start kind of like vibe coding almost anything, but at the same time, it's also things are moving so fast.
• 18:40
  SPEAKER_02: And uh, you know, how do you pick the right problems and how do you chart your own territory and where is the field gonna be?
• 18:45
  SPEAKER_02: How do you sort of skate where the puck is going when things are moving this fast and changing so so so quickly?
• 18:51
  SPEAKER_02: So um no, I think it's it's uh it's an incredible time for the field.
• 18:57
  SPEAKER_01: That that's something that I wanted to also talk about um in relation to running a startup.
• 19:03
  SPEAKER_01: Um who was I talking to?
• 19:05
  SPEAKER_01: I was talking to the CEO of uh Rhoda AI.
• 19:07
  SPEAKER_01: Right.
• 19:07
  SPEAKER_01: Um, and you know, he he just offhandedly said, you know, obviously, like we're following, and obviously you're following all of the research that's coming out.
• 19:15
  SPEAKER_01: And every once in a while something must come out and it must be like, uh-oh, we've got to, you know, we have to, we have to rethink.
• 19:21
  SPEAKER_01: So, you know, you are sort of maybe in a similar position to the students in that uh things are happening so quickly and things are are changing in such fundamental ways that it also must kind of be both an exciting and very challenging time to be running a physical AI startup.
• 19:37
  SPEAKER_01: Totally.
• 19:38
  SPEAKER_02: But um, you know, I think I think I've been doing it for a while, and and uh I think one has to be focused enough to not chase after every shiny bauble.
• 19:50
  SPEAKER_02: And I think setting super clear directions and helping um students or or uh other folks sort of know what's important, what's worth paying attention to.
• 20:02
  SPEAKER_02: Uh but also being nimble enough then when something amazing and new happens, then you recognize it as such and are willing to change.
• 20:10
  SPEAKER_02: Uh, I think that is the that is the skill set now is to be able to sort of walk both of those lines.
• 20:15
  SPEAKER_02: You know, keep the steady hand when appropriate, but pivot quickly when appropriate.
• 20:20
  SPEAKER_01: Yeah, I mean I ask for uh for myself, and I don't know, maybe this is too abstract or too large a question, but um how do you determine when something is a shiny object and when something is like actually a real game changer?
• 20:35
  SPEAKER_02: Yeah.
• 20:36
  SPEAKER_02: Well, I've always tried to be more hands-on than you know, as I write code at night and um, you know, try to try to have a uh I worry about getting two arm's length away from it.
• 20:46
  SPEAKER_02: So I I really try to dabble a lot myself.
• 20:49
  SPEAKER_02: That helps a lot.
• 20:50
  SPEAKER_02: I surround myself with people that that are incredibly, you know, well-versed in this and I and I trust their judgment.
• 20:57
  SPEAKER_02: I remember it used to be that students would ask me what papers they should read.
• 21:02
  SPEAKER_02: And now it's like, you know, students are saying, uh sending me all the papers.
• 21:05
  SPEAKER_02: I'm like, wait a second, is this the one I have to read this week?
• 21:07
  SPEAKER_02: You know, like you gotta filter a little bit for me.
• 21:10
  SPEAKER_02: Um uh the there's a deluge of papers that that come at you, and you have to be selective with that.
• 21:16
  SPEAKER_02: But I also think over the years you you you can pretty quickly uh recognize um you know what are the key results and when when they happen.
• 21:25
  SPEAKER_02: So yeah, it's uh it's an art for sure.
• 21:28
  SPEAKER_02: I don't know that there's a playbook for it.
• 21:30
  SPEAKER_01: I I think it's helpful too when you when you like you recognize half the people and maybe a quarter of them were your students at some point.
• 21:37
  SPEAKER_01: That's just saying I'm old.
• 21:39
  SPEAKER_01: But uh I mean, you know, we were discussing it before that like uh obviously and and and I think that this is you know really one of the kind of wonderful things that I've noticed uh in covering robotics is that like obviously it's it's big and it's getting larger, but there is and always has been a way in which it feels like a really kind of like small world and everybody knows each other and people largely and Certainly, this applies to you.
• 22:01
  SPEAKER_01: People seem to be very supportive of one another.
• 22:03
  SPEAKER_02: Yeah, yeah, yeah.
• 22:04
  SPEAKER_02: That's uh yeah, no, it's it's incredible to see to have had the opportunity to work with really a bunch of very special people.
• 22:12
  SPEAKER_02: Um that's something that I think MIT really gives me the chance to work with just some incredible people to see what they've gone on to do has just been it's one of the joys of my life is to see, you know, what they're what they're capable of, what they're doing, how they're changing the world.
• 22:25
  SPEAKER_02: And so yeah, it's there's nothing that makes me prouder.
• 22:29
  SPEAKER_01: So so there there is a way, you know, it's funny because again, I'm coming from like from from TechCrunch and I'm coming from the startup world where everything just feels like so competitive and and so cutthroat, but there is a way to uh to sort of like coexist in this world uh where where you can like root other people's breakthroughs on and still kind of in a way be a competitor with them.
• 22:51
  SPEAKER_02: Yeah.
• 22:51
  unknown: Yeah.
• 22:52
  SPEAKER_02: So let's talk about so so I think um, you know, people there's there's a the voice of skepticism in the air, there's a voice of enthusiasm in the air.
• 23:01
  SPEAKER_02: You know, I think in my mind, uh a number of incredible things have happened all at the same time, and it makes me extremely optimistic, right?
• 23:12
  SPEAKER_02: I think I think nobody doubts that robotics will almost nobody doubts that robotics will change our world dramatically.
• 23:19
  SPEAKER_02: I think the biggest question is when is is this the time, right?
• 23:23
  SPEAKER_02: I think we have a bunch of energy, a bunch of a bunch of momentum, certainly from GPT and and people getting excited about what AI is capable of.
• 23:32
  SPEAKER_01: And is this time somehow fundamentally different than the other times when it kind of felt like this?
• 23:36
  SPEAKER_01: Exactly.
• 23:37
  SPEAKER_02: Um the reason I I think it's different is I certainly I think some things incredible have happened in the technology.
• 23:44
  SPEAKER_02: But it's been it's more than that.
• 23:46
  SPEAKER_02: It's that um that has led to an influx of talent, like the number of incredible people that have come into the field from other disciplines and are working on robotics or physical AI now, the amount of investment that has come in because of this, uh, you know, the changes in the supply chain.
• 24:02
  SPEAKER_02: You know, there's I mean, it's incredible what's happened in China in terms of the ability to manufacture these very capable devices at at cost.
• 24:10
  SPEAKER_02: You know, the world needs this technology in certain ways.
• 24:13
  SPEAKER_02: I mean, the the changing expectations for work, the silvering society, the there's there's so many things that sort of have aligned to make me think uh that the field has an incredible opportunity.
• 24:25
  SPEAKER_02: You know, I feel like you know, there's no question we're we're shooting for the moon.
• 24:29
  SPEAKER_02: You know, do we have escape velocity?
• 24:31
  SPEAKER_02: I think so.
• 24:32
  SPEAKER_02: I think so.
• 24:32
  SPEAKER_02: I think this is the time that it's really different.
• 24:36
  SPEAKER_02: And for sure, I mean, if um you can you can decide whether you're you think we're gonna make it to the moon or maybe come back to Earth, but I I would rather be on the rocket ship, you know.
• 24:46
  SPEAKER_01: You describe a phenomenon too that's really interesting to me.
• 24:50
  SPEAKER_01: Um uh we we we're we were just uh speaking with Claire Vue uh right before this, and and she was saying that uh when they were doing VAO, um, you know, at the time a lot of the people they were getting were uh you know either coming over from like automotive from self-driving cars, obviously, or uh medical devices.
• 25:07
  SPEAKER_01: And something that I've noticed in robotics is that um that as a discipline and when it comes to bringing people in, is that it doesn't seem to be particularly rigid.
• 25:18
  SPEAKER_01: And that's kind of amazing, you know, that I don't see quite the same way in other fields that as long as they're sort of smart people, like you're you will accept them and you will find a way to kind of integrate them into the system.
• 25:31
  SPEAKER_02: I do think we draw from many professions and um and many people who did not do robotics before can are certainly welcome.
• 25:40
  SPEAKER_02: I also I think Boston is an incredible place to be building a robotics company.
• 25:45
  SPEAKER_02: I think the ecosystem here with mass robotics and and um and the just the number of incredible robotics companies here, I think it's an incredible place to be doing that.
• 25:55
  SPEAKER_02: Um yeah, I do think we draw in many different disciplines.
• 25:59
  SPEAKER_02: People who weren't previously excited about robotics see what might be coming and want to be involved.
• 26:04
  SPEAKER_02: Uh, that's awesome.
• 26:05
  SPEAKER_02: That's one of the reasons I'm excited.
• 26:06
  SPEAKER_02: I think that uh we do have incredible people coming into the space.
• 26:10
  SPEAKER_01: You talked a little bit before about um, you know, coding every night, you know, to kind of like keep your skin in the game a bit.
• 26:16
  SPEAKER_01: And I'm curious, you know, what uh you know, at at TRI and now in your new role, um, you know, what what what is it that you're real really are doing most of the time?
• 26:26
  SPEAKER_01: Aaron Powell It depends on the day.
• 26:28
  SPEAKER_02: Sure.
• 26:28
  SPEAKER_02: But um I mean I um maybe not this month.
• 26:32
  SPEAKER_02: This particular month is better.
• 26:34
  SPEAKER_01: Well, so so just to like back up, so you told me that you yesterday you just moved into a new space.
• 26:39
  SPEAKER_02: Yeah.
• 26:40
  SPEAKER_02: Well, let's just take teaching for example.
• 26:41
  SPEAKER_02: So when I teach, I try to provide a lot of examples, worked examples.
• 26:46
  SPEAKER_02: Uh I use, I make demonstrations for class, which involve building up in uh intense sort of big pipelines of software engineering, sort of to get students to the point where they can very quickly experiment with an idea, or it shows an idea like on a slider or anything like that.
• 27:01
  SPEAKER_02: That is a really um, you know, that brings all of the components together into uh open course notes.
• 27:07
  SPEAKER_02: You know, I think behind that, so I have been working hard on this software simulation and and model-based design software called Drake.
• 27:15
  SPEAKER_02: I kind of joke that it's my whore crux, you know.
• 27:18
  SPEAKER_02: My I've put a piece of my soul in that uh that software package.
• 27:22
  SPEAKER_02: I still contribute code, you know, into the into production for Drake, you know, and uh uh that's important to me.
• 27:28
  SPEAKER_02: It keeps my my skills sharp.
• 27:31
  SPEAKER_02: It keeps me thinking the way I want to be thinking.
• 27:34
  SPEAKER_01: So so TRI, so that was a big part of your role at TRI was Drake.
• 27:38
  SPEAKER_01: Um and then obviously, like you're the uh large behavior model guy.
• 27:42
  SPEAKER_01: Something like that.
• 27:44
  SPEAKER_01: Gil Pratt came up with the name.
• 27:46
  SPEAKER_01: Okay, but but you know, and you know, it obviously, like, at least for now, as it's in, you know, it's in the company name.
• 27:52
  SPEAKER_01: Yeah, yeah.
• 27:53
  SPEAKER_02: No, it's uh we're super proud of that work, and that um it changed my view of of what's possible in the field.
• 28:01
  SPEAKER_02: So and uh, you know, I think TRI had an incredible role to play there of trying to do uh the science of maybe of LBMs at a level that wouldn't immediately be motivated um to be done in uh a startup, but and and maybe wouldn't also be accessible given the resources in academia.
• 28:20
  SPEAKER_02: So I think there was a particular role that TRI could play sort of on the science of LBMs, understanding the initial scaling laws, but doing it with just a lot of experiments that were required to dig in and understand rigorously like what the what those scaling laws could look like.
• 28:36
  SPEAKER_02: Very proud of that work.
• 28:37
  SPEAKER_02: Would you have called it something different?
• 28:39
  SPEAKER_02: No, no, no.
• 28:39
  SPEAKER_02: I thought it was a great name.
• 28:40
  SPEAKER_02: Okay.
• 28:40
  SPEAKER_02: I thought it was a great name.
• 28:41
  SPEAKER_02: I think some people um I like to try to clarify for people the difference between a VLA and an LBM or something like this, right?
• 28:49
  SPEAKER_02: So large behavior models in my mind are any model that takes sequences of images in and output actions.
• 28:56
  SPEAKER_02: So we're up training uh large multimodal models.
• 29:00
  SPEAKER_02: A VLA, in my mind, is um a particular type of LBM.
• 29:05
  SPEAKER_02: Now, the people who talk about VLAs might, I hope they're not insulted by that categorization.
• 29:10
  SPEAKER_02: But one choice for building large behavior models is to uptrain a VLM, a video, a vision and language model into being a robot model.
• 29:18
  SPEAKER_02: And that's one architectural choice you can make.
• 29:21
  SPEAKER_02: If you make that choice, I would say you've built a VLA.
• 29:23
  SPEAKER_02: And that's one of the things we do a lot.
• 29:25
  SPEAKER_02: Uh but um that's there are other choices.
• 29:28
  SPEAKER_02: You could start with a video backbone or a world model backbone, which are uh similar, and you could uptrain that to be an LBM.
• 29:35
  SPEAKER_02: So I I just like LBM as kind of you know from LLMs to large behavior models as kind of a natural way to talk about the space.
• 29:43
  SPEAKER_01: Yeah, so so uh you so it sounds like um generally speaking, or you know, at at TRI and maybe the work that you're doing is that kind of large uh umbrella approach to training?
• 29:53
  SPEAKER_02: Yeah, I think the um the methods are you have to be willing to try lots of different architectures.
• 30:01
  SPEAKER_02: Although the architectures have have have been more converging.
• 30:04
  SPEAKER_02: There's still lots of room for improvement there.
• 30:06
  SPEAKER_02: But I think we kind of know now that that if you want longer context lengths, you should be uh starting with a video model as a backbone.
• 30:14
  SPEAKER_02: Multimodal models are getting better and better and better.
• 30:17
  SPEAKER_02: So you should absolutely set yourself up to be able to leverage the advances that the field is making there.
• 30:23
  SPEAKER_02: Um But I think you know, being having a code base, a maturity that you can sort of quickly take in the latest model is I think really important in this uh time.
• 30:32
  SPEAKER_01: Yeah, I I I I told you that I was I was reading uh uh up a bit on Generalists before this, and I'm sure that you've been following the work that they've been doing in.
• 30:40
  SPEAKER_01: Um you know, one of the things that they were saying in their work is that they, I guess, sort of scrapped the their own model and started from scratch.
• 30:48
  SPEAKER_01: Is that something that you feel is is necessary versus I guess kind of starting uh on top of an LLM?
• 30:54
  SPEAKER_02: So there's this whole narrative about data paucity in robotics.
• 30:58
  SPEAKER_02: That data, you know, we don't have enough data in robotics.
• 31:00
  SPEAKER_02: You've had a few guests recently that have sort of talked about that.
• 31:03
  SPEAKER_02: And um and it's interesting to see different teams sort of have different answers to that.
• 31:09
  SPEAKER_02: They say, okay, well, maybe I'm gonna learn everything from egocentric data, or I'm gonna learn everything from simulation, or I'm gonna learn everything from the universal manipulation sort of simple interfaces, or or from teleoperation.
• 31:22
  SPEAKER_02: I think that whole narrative misses there's something that I that I I feel like I think about it maybe just a little, or I would articulate it a little bit differently.
• 31:30
  SPEAKER_02: Okay.
• 31:31
  SPEAKER_02: And people also draw in robotics.
• 31:33
  SPEAKER_02: They say, here's the size of the data sets in robotics, and here's the size of the data sets used to train GPT or LAMO.
• 31:42
  SPEAKER_01: Yeah.
• 31:43
  SPEAKER_02: I think that story is misses a bigger picture, which is that we are in robotics, what we're doing is we're starting with a strong model and we're up training it to be a robot model.
• 31:54
  SPEAKER_02: So the data you need for robotics is not every piece of data ever to teach a robot common sense about the world.
• 32:03
  SPEAKER_02: I think we're building on models like video models that already have a lot of common sense about the world.
• 32:09
  SPEAKER_02: And the data you need is to build a bridge from that common sense to basically train the model to have one extra output, which is robot actions.
• 32:20
  SPEAKER_02: Which is just a little bit different than what people are saying.
• 32:23
  SPEAKER_02: And it it I think it positions you to go after uh uh architectures and data a little bit differently if you sort of embrace that.
• 32:32
  SPEAKER_02: So absolutely you should um should be using simple grippers, you should you be using egocentric data, you should be using simulation, but you should be thinking about that as building a bridge from the common sense that you know if I took a picture of my robot in a in a real application and gave it to the latest Gemini video model, VO3, let's say, um, and gave it a text prompt saying, make a video of my robot doing some dexterous task, it's gonna do a pretty good job, right?
• 33:02
  SPEAKER_02: So that suggests that there's already a lot of understanding in that model.
• 33:06
  SPEAKER_02: Well, you you don't have to give it that understanding.
• 33:08
  SPEAKER_02: What you have to do is teach it how to map that output instead of making pictures to map it into robot trajectories.
• 33:14
  SPEAKER_02: And that's just a very different uh posture.
• 33:17
  SPEAKER_02: So okay, back to general.
• 33:19
  SPEAKER_02: So so I would not throw that.
• 33:23
  SPEAKER_02: That that value of that base model is so great.
• 33:27
  SPEAKER_02: Like people are spending immense resources compiling massive data sets in order to build these base models that have so much world knowledge.
• 33:34
  SPEAKER_02: I would not throw that away.
• 33:36
  SPEAKER_02: Yeah.
• 33:36
  SPEAKER_02: I would start from there.
• 33:38
  SPEAKER_02: And then I think you have to think about your data curriculum as building the bridge to teach that model how to act on your robot.
• 33:46
  SPEAKER_02: Um, I think generalist is doing that at a high level.
• 33:48
  SPEAKER_02: I think we're doing that at a high level.
• 33:50
  SPEAKER_01: Yeah, I mean, one of the things that one of the things that I did appreciate that they said, and and I I suspect that maybe um, you know, maybe you're on a similar wavelength, especially as we're talking about um uh large large behavior models being uh this kind of I I don't want to say catch all, but again, like an umbrella term um is you know, they're talking about sort of being uh like method agnostic, right?
• 34:14
  SPEAKER_01: Um and you know, the the the people that I've spoken with that make the most sense are the ones who say, hey, like all of these make, you know, well, we'll they'll all have their purpose for pre-training, for post-training, um, you know, even like teleop when it comes to like post-training, is that this is effectively where you stand?
• 34:34
  SPEAKER_02: Yeah, I think, you know, I've been doing it um with a super strong team for enough years that we've tried all of them.
• 34:40
  SPEAKER_02: Yeah, they all have their place.
• 34:42
  SPEAKER_02: I think what um being very thoughtful about pulling them in, understanding what they're good for.
• 34:48
  SPEAKER_02: Again, I think I like the analogy of building bridges.
• 34:50
  SPEAKER_02: So if you have simulation data of a particular, like a CUKA robot doing a particular task, then maybe you also want video of a CUKA.
• 34:58
  SPEAKER_02: And that allows you to take your video data set and your simulation data set and and and connect them together.
• 35:02
  SPEAKER_02: And that allows these models to transfer capabilities um more effectively.
• 35:07
  SPEAKER_02: Yeah.
• 35:08
  SPEAKER_02: So kitchen sink approach to data, but then filtering very thoughtfully.
• 35:13
  SPEAKER_01: Yeah.
• 35:13
  SPEAKER_01: I mean, are are are are we anticipating, though, that there's going to be another one of those like surprise moments as far as like dealing with kind of the edge cases or the long tail when it comes to actually putting these robots out in the world?
• 35:25
  SPEAKER_02: Well, that's the that's the point.
• 35:27
  SPEAKER_02: I think we've we've seen the beginnings of that.
• 35:30
  SPEAKER_02: We saw that in in the LBM paper that we published.
• 35:33
  SPEAKER_02: You know, we saw the beginnings of that where you saw robustness in individual tasks change in a substantive way by having done pre-training on other tasks.
• 35:43
  SPEAKER_02: I think that is the the magic of multitask pre-training.
• 35:47
  SPEAKER_02: Okay.
• 35:47
  SPEAKER_01: And but that that is um, you know, if I if I'm remembering the the work correctly, is is that the diffusion policy work?
• 35:55
  SPEAKER_02: Diffusion policy was the original sort of our um idea.
• 35:59
  SPEAKER_02: Large behavior models was the multitask version of diffusion policy is my in my vernacular.
• 36:03
  SPEAKER_01: So so so these but these are still kind of those like uh tasks in controlled environments.
• 36:09
  SPEAKER_01: Yep.
• 36:10
  SPEAKER_01: Yep.
• 36:11
  SPEAKER_01: Um as far as actually kind of I I guess again, like we're we're talking about really like we're talking about robots in the world, right?
• 36:19
  SPEAKER_01: Yeah.
• 36:19
  SPEAKER_01: Um, you know, as as far as more generalization and as far as moving outside of, you know, the uh obviously the TRI offices, but factories and warehouses.
• 36:30
  SPEAKER_01: Where does where does that magic come in?
• 36:33
  SPEAKER_02: This is I think the next major milestone for the field is to get robots deployed in the world.
• 36:38
  SPEAKER_02: I think the narrative of many people has shifted towards deployment.
• 36:43
  SPEAKER_02: Uh we have to to earn that.
• 36:46
  SPEAKER_02: The field has to earn that.
• 36:47
  SPEAKER_02: And if we get that, then we do get this virtuous cycle where more capable robots makes more robots in the field, makes more data, which makes more uh capable robots.
• 36:56
  SPEAKER_02: And I think we get a real virtuous cycle going there.
• 36:59
  SPEAKER_02: But I I think for me, I'm very excited now about taking that next step of sort of really deploying these things in the real world and understanding the right way to do that.
• 37:08
  SPEAKER_01: So so there uh there is a data gap from the standpoint of like obviously that there are just are not enough systems out in the world collecting that real world data.
• 37:18
  SPEAKER_01: That is a real issue.
• 37:19
  SPEAKER_02: Yeah, data makes things better.
• 37:21
  SPEAKER_02: What I um I only want to rephrase some of the, you know, the discussion says we we don't have as much data as a language model.
• 37:30
  SPEAKER_02: Like that's not the right, that's not the way I think about it.
• 37:33
  SPEAKER_02: Because we're starting with the language model.
• 37:34
  SPEAKER_02: We're starting with the base model that has already been trained on the internet and we're adding to that.
• 37:38
  SPEAKER_02: That just I think reframes what you think you need from the data is what you really need is to sort of map that into the tasks you care most about or the customers care most about.
• 37:48
  SPEAKER_01: So here's where we sort of get into sort of the the sticky territory for you.
• 37:52
  SPEAKER_01: But you know, what insofar as you can kind of talk about this, what role do you see yourself and your you know, your company playing in that world?
• 38:01
  SPEAKER_02: Yeah, no, I think this is um so I do think we're on escape velocity.
• 38:08
  SPEAKER_02: I want to be on that rocket ship.
• 38:10
  SPEAKER_02: I don't think it's inevitable.
• 38:13
  SPEAKER_02: I think um it takes incredibly good execution with an incredibly strong team and a lot of focus when you know, not chasing every shiny, shiny bauble.
• 38:22
  SPEAKER_02: So, and it takes you have to be uh you have to have a good business plan.
• 38:26
  SPEAKER_02: You have to, you know, do a lot of things right to make this work.
• 38:30
  SPEAKER_02: But I do think the component technologies, uh the investment in the field, the talent coming into the space, everything has aligned to sort of give us a real chance.
• 38:41
  SPEAKER_02: So, you know, the company that I'm excited to announce uh, you know, soon is built to basically maximize the chances of us trying to get to the next level.
• 38:53
  SPEAKER_01: I'm I'm curious, and again, this is a bit of an abstract question, and this uh you know relates to my techrunch days, but as you're kind of as the team is coming together and as you're sort of pitching yourself as a team versus other uh I guess physical AI companies out there, obviously a big selling point is uh is a talent that you brought on, but beyond that, um, you know, is it is it is it IP or what what are sort of the special sauces that you're bringing?
• 39:23
  SPEAKER_02: No, I think there's um there's a lot of things that go into that to make that sort of um make us unique.
• 39:31
  SPEAKER_02: I mean, from from the way we think about data, the way we think about deployments, the way we think about operations, the way we think about our business.
• 39:39
  SPEAKER_02: You know, so I think there's a lot of things that you have to get right.
• 39:43
  SPEAKER_02: Uh I am pretty proud of the team, though.
• 39:45
  SPEAKER_02: It's an awesome team.
• 39:46
  SPEAKER_01: Yeah.
• 39:47
  SPEAKER_01: You you feel though that there are a lot of angles that you're looking at that are that that do differ a bit quite a bit from the way that Yeah, for sure.
• 39:56
  SPEAKER_02: Yeah.
• 39:56
  SPEAKER_01: I'm sure that there have been plenty of points in your work, in your career, where you've been approached to do something like this, where you've been approached to launch something, and I'm sure that you've thought about it in the past.
• 40:07
  SPEAKER_02: That's a good question.
• 40:08
  SPEAKER_02: So um I have been asked many times why would we want to do something new?
• 40:12
  SPEAKER_02: Why not join an existing effort or something like that?
• 40:17
  SPEAKER_02: Um I will tell you there's one reason that's maybe not even technical that I really want to do this.
• 40:23
  SPEAKER_02: First of all, I think there's a particular uh, you know, I think we're building a team and and a business and to to do it right to be to have a chance of success.
• 40:33
  SPEAKER_02: Um I do think if we're successful, the field is successful, then it's gonna change the very nature of work for people.
• 40:42
  SPEAKER_02: And it's gonna have a huge, hopefully hugely positive impact on the world.
• 40:47
  SPEAKER_02: Uh but uh it is gonna be a big change for people.
• 40:51
  SPEAKER_02: And I um actually working with Toyota over the years, uh really this notion of amplifying, not replacing people, about thinking about what are the what are the implications for the future of factory work, you know, for um how do we help people age in place more gracefully.
• 41:07
  SPEAKER_02: It has really shaped the way I think about this.
• 41:10
  SPEAKER_02: And I'm spending a lot of time these days talking to labor economists and spending time with people that I think could be potentially impacted and um just trying to build my empathy muscles and and build my understanding and build my um sort of both the macroeconomics sort of understanding of it and the the sort of in the small understanding of how it could impact people.
• 41:34
  SPEAKER_02: And I think uh the companies that are most thoughtful about this have an opportunity to do it, uh, you know, make even small changes as you go forward and and really help people uh get the best version of this.
• 41:49
  SPEAKER_01: That's a that's a very good answer.
• 41:51
  SPEAKER_01: Uh that's uh you know, something that I uh have not heard discussed uh a lot.
• 41:57
  SPEAKER_01: Um, you know, and and and to break it down a little bit, so that the Toyota piece is there, obviously um aging in place is is a big part of Toyota because it's been such a fundamental piece of just Japanese culture in general, right?
• 42:10
  SPEAKER_01: I mean it's it's an aging population.
• 42:12
  SPEAKER_01: That's part of the reason why robotics have been uh big over there.
• 42:16
  SPEAKER_01: Um and then as far as uh amplifying uh people, who's I oh, I think I was talking to Max actually at TRI when we were over there, and you know, I think he was I think what he uh he was saying is he was just talking about how uh deliberate the factories are and set up that like everything is there for a reason.
• 42:35
  SPEAKER_01: So if you're gonna bring in a piece of automation, like it better be there for a reason.
• 42:41
  SPEAKER_02: Yeah.
• 42:42
  SPEAKER_02: Uh no, I think there's wonderful things about the Japanese culture.
• 42:44
  SPEAKER_02: I think uh lifetime employment is also a strong theme in in traditionally in Japanese culture.
• 42:51
  SPEAKER_02: And so the narrative from TRI is is real.
• 42:54
  SPEAKER_02: It is uh there's a real need to do this correctly.
• 42:57
  SPEAKER_01: Aaron Powell Again, big broad question, but what does it mean to approach something like robotics or physical AI with empathy?
• 43:08
  SPEAKER_02: I think you really okay, let's let's see how we've already seen it play out from from maybe many people in your audience have have experienced um the software engineering version of this, or maybe the graphic design version of this, right?
• 43:20
  SPEAKER_02: So uh we've seen the field of software engineering change, right?
• 43:25
  SPEAKER_02: I think people who are already expert software engineers are now superpowered by quad code or or whatever your favorite tool is.
• 43:34
  SPEAKER_02: And um, people who are already incredible artists can use Photoshop or Nano Banana or whatever in you know to be able to do their exceptional art very quickly, you know.
• 43:46
  SPEAKER_02: Novices can vibe code almost anything now, and uh and I can make logos, I can make art in a way that I certainly couldn't a couple years ago.
• 43:57
  SPEAKER_02: So, what is that gonna look like in the physical world?
• 44:00
  SPEAKER_02: I think we're gonna have um a lot of tasks that would have been very hard to automate before that are suddenly vibe codable almost, right?
• 44:10
  SPEAKER_02: Uh we'll have top craftspeople that could do magnificent things more capable, you know, more more quickly with with super tools.
• 44:21
  SPEAKER_02: Uh but how does that change uh a culture who distributes wealth through labor, right?
• 44:31
  SPEAKER_02: That um that people's sense of worth, sense of you know uh purpose often as comes through the job that they've taken.
• 44:39
  SPEAKER_02: Well, you have to, I don't know.
• 44:40
  SPEAKER_02: I think we have to think very carefully about that.
• 44:43
  SPEAKER_02: And you just have to spend time uh uh helping people understand what the tools are capable of, helping them uh adjust, using the best parts of what they can do and and amplifying that with proper tools.
• 44:58
  SPEAKER_02: Uh I think we have to solve this for artists, we have to solve this for software engineers, and we're gonna have to solve it for physical labor too.
• 45:05
  SPEAKER_01: And this and this is a role that um a physical AI company could ultimately play.
• 45:09
  SPEAKER_02: I think a a thoughtful physical AI company will could change the way that plays out.
• 45:14
  SPEAKER_02: Yeah.
• 45:14
  SPEAKER_01: And and I'm I'm also, you know, you you biked over here.
• 45:17
  SPEAKER_01: Um your your the shot of you, I think, on your uh prof uh uh academic page is sort of you standing, it looks like on top of a mountain.
• 45:28
  SPEAKER_01: Um does that this this the time that you spend out in the natural world, does that sort of impact to the way that you interact with technology?
• 45:36
  SPEAKER_02: That is interesting.
• 45:37
  SPEAKER_02: I do super value um my commute.
• 45:42
  SPEAKER_02: That sounds like a weird thing to say, right?
• 45:44
  SPEAKER_01: But but um Again, you don't have to drive in Boston traffic, which is a thing to value.
• 45:49
  SPEAKER_02: You know, but I I run or bike to work and um it takes time, but that is my time to um to decompress, to compile everything that happened through the day to to think more holistic thoughts.
• 46:05
  SPEAKER_02: You know, there's uh famous thinkers through the through the ages that have a lot of people take long walks.
• 46:12
  SPEAKER_02: Yeah, that that's been defining.
• 46:13
  SPEAKER_02: Yeah, uh yeah, Thoreau was a big thing.
• 46:15
  SPEAKER_02: Thoreau went for like four-hour walks every day and then wrote for four hours a day, you know.
• 46:20
  SPEAKER_02: And um it's a nice way to um, I think, think more deeply.
• 46:27
  SPEAKER_02: It's so easy to just answer our email all day long.
• 46:30
  SPEAKER_02: Um so uh yeah, I love my commute.
• 46:32
  SPEAKER_01: So you can you can vibe code, but sometimes you gotta shut it off and walk step away.
• 46:37
  SPEAKER_01: I think so.
• 46:38
  SPEAKER_02: I think so.
• 46:40
  SPEAKER_01: Well, Russ, thank you so much for the sure.
• 46:42
  SPEAKER_01: Yeah, that's great.
• 46:42
  SPEAKER_01: Thank you for your questions.
• 46:43
  SPEAKER_00: Yeah, thanks to Russ and the whole team at Mass Robotics for hosting us in Boston all week.
• 46:49
  SPEAKER_00: Thanks as always for tuning in.
• 46:50
  SPEAKER_00: Please like and subscribe and rate the show.
• 46:53
  SPEAKER_00: Check out the automated lose letter that hits inboxes on Thursdays and LinkedIn on Friday.
• 46:59
  SPEAKER_00: And with all of that, we will see you next week for another episode of Automated.