Latest / Automated with Brian Heater / Rodney Brooks on Robotics, AI, and the Future of Automation

Transcript

Rodney Brooks on Robotics, AI, and the Future of Automation · Automated with Brian Heater

• 0:00
  SPEAKER_01: I was twenty-five feet away from a well-known humanoid two weeks ago, and I was at a cocktail party, and no one was interested in it, because there were uh r roboticists at this cocktail party, they knew it was junk and uh well junk, that's maybe an overstatement.
• 0:16
  SPEAKER_01: But I I said to someone, hey, hey, why don't we go and and and try robot tipping?
• 0:21
  SPEAKER_01: And and she said, Yeah, and we weren't really gonna do it.
• 0:25
  SPEAKER_01: But we turned to look at the robot, it was about 25 feet away, and at that moment it fell flat on its face.
• 0:30
  SPEAKER_00: So hello and welcome to Automated.
• 0:41
  SPEAKER_00: My name is Brian Heater.
• 0:43
  SPEAKER_00: I am the managing editor at A3.
• 0:45
  SPEAKER_00: I believe the first time that I met Rodney Brooks is when he spoke at TechCrunch's first robotics conference way back in 2017.
• 0:55
  SPEAKER_00: I distinctly remember three things from that event.
• 0:58
  SPEAKER_00: The first was an engaging fireside chat.
• 1:01
  SPEAKER_00: The second was a photo that we took of him and the other two iRobot co-founders.
• 1:06
  SPEAKER_00: And the third was him actively engaging with someone else's panel from the crowd.
• 1:12
  SPEAKER_00: Since then, he has become one of the first people that I go to when I need a pragmatic look at the latest robotics hype cycle.
• 1:21
  SPEAKER_00: Aaron Powell I was rereading your predictions in the lead up to this conversation because it it had been several months.
• 1:28
  SPEAKER_00: And one of the asides in there really jumped out at me given the current context of everything you wrote.
• 1:35
  SPEAKER_00: I suspect many people will reason that I cannot have a valid opinion about this being humanoids precisely because I happen to have built more humanoid robots than anybody else on the planet.
• 1:45
  SPEAKER_00: So read with Caufchen.
• 1:47
  SPEAKER_00: So what is the what what what's the root of your skepticism right now, and how does your experience in humanoids inform that?
• 1:57
  SPEAKER_01: Aaron Powell Okay, I guess two things.
• 1:59
  SPEAKER_01: One, I'm skeptical about the end-to-end learning with the humanoids, because uh that can only possibly work if you're collecting the right data and trying to learn the right thing.
• 2:11
  SPEAKER_01: And I don't think we've had any lab demos, even at small scale, of learning the right uh you know, collecting the right data or learning the right thing.
• 2:20
  SPEAKER_01: So it's it's sort of, you know, we'll try this and and um you know, maybe it'll work, but we're gonna put uh lots and lots of money into it.
• 2:29
  SPEAKER_01: And I think that um all the training regimes I've seen uh are really based around the human motions, either captured with cameras or wearing some sort of suit or holding some sort of joysticks.
• 2:44
  SPEAKER_01: Uh but the real thing about manipulating the world, and that's what humanoids are supposed to do, they're supposed to manipulate the world the same way humans do.
• 2:51
  SPEAKER_01: The real thing about manipulating the world is when you make contact with something and you're both applying a force and sensing a force.
• 2:59
  SPEAKER_01: And that's not being captured at the moment.
• 3:02
  SPEAKER_01: So I'm skeptical there.
• 3:03
  SPEAKER_01: The other place I'm skeptical is about the legs.
• 3:06
  SPEAKER_01: Um, the current humanoids do not walk like humans walk.
• 3:10
  SPEAKER_01: Um we're very passive walkers.
• 3:13
  SPEAKER_01: We we don't need to use much muscle at all to walk.
• 3:15
  SPEAKER_01: Our system is built to walk.
• 3:18
  SPEAKER_01: Uh instead, the current walking algorithms, they're all based on the ZMP algorithm, zero moment point algorithm, which has been around for about 20 years, um, where they're not good at falling.
• 3:28
  SPEAKER_01: And when they start to fall, they pump a lot of energy into the system to try to recover, which is good if they recover, but if they don't recover, then there's a lot of kinetic energy in all those joints and limbs, and it's gonna go somewhere.
• 3:41
  SPEAKER_01: And if a human happens to be in the way, it's just gonna mess with them.
• 3:46
  SPEAKER_01: Which, by the way, you can see by the way some of the humanoid companies demo their robots.
• 3:52
  SPEAKER_01: When the when the humanoids are walking, they say keep away.
• 3:55
  SPEAKER_01: Then you can come up to it and talk to it, but it's actually a person somewhere else with a microphone and camera.
• 4:02
  SPEAKER_01: Um but they they don't allow people next to the robot when it's moving its legs.
• 4:06
  SPEAKER_01: Watch watch all the videos for that.
• 4:08
  SPEAKER_01: It's a it's a real tell.
• 4:09
  SPEAKER_01: Trevor Burrus, Jr.
• 4:10
  SPEAKER_00: Which is ironic, right?
• 4:11
  SPEAKER_00: Because that's supposed to be one of the selling points, is that you can keep them out of the cage.
• 4:14
  SPEAKER_00: Aaron Powell Yeah.
• 4:16
  SPEAKER_00: So when you say that they're not trying to learn the right thing, I'm I'm I'm curious what you mean by that and whether or not like how much actually um having these out in the field, you know, as I'm not telling you anything you don't know, but obviously you've got your robot out there now and it's constantly in the process of collecting data and it's constantly getting better.
• 4:34
  SPEAKER_00: So through these pilot programs that they're doing right now, is the right data coming in?
• 4:40
  SPEAKER_01: I don't believe so.
• 4:41
  SPEAKER_01: Um you know, there's the pilot programs and then there's the the large manipulation models where they've got lots of people training the robots.
• 4:50
  SPEAKER_01: Some of the companies have people training the robot.
• 4:52
  SPEAKER_01: Banks are people training the robots.
• 4:55
  SPEAKER_01: Um and there they're the humans are doing the motions, and they're trying to capture those motions and put them into a model that the robot can then do the same things.
• 5:06
  SPEAKER_01: But again, if you look at the big tells in the I call it humanoid theater.
• 5:11
  SPEAKER_01: Humanoid theater are this set of videos you see coming out of the humanoid companies where uh you know they sort of shake, the robot arms shake as they go because they haven't really learned they don't have good control algorithms underneath because it's coming from this reinforcement learning system.
• 5:30
  SPEAKER_01: And then the grasps they do are very primitive grasps.
• 5:34
  SPEAKER_01: Uh uh, you know, our hum us humans can um, you know, uh I I this is not stage, but you know, I just picked up three different things and I can manipulate them separately in my fingers, I can do all sorts of things.
• 5:46
  SPEAKER_01: There's nothing like that.
• 5:47
  SPEAKER_01: Uh the best you see is sort of a uh in in in humanoid theater is um, you know, shove.
• 5:59
  SPEAKER_01: Um just a pinch grasp.
• 6:02
  SPEAKER_01: Um so I'm I'm waiting to see.
• 6:05
  SPEAKER_01: None of the none of the videos have blown me away.
• 6:10
  SPEAKER_01: I was I was uh I was 25 feet away from a well-known humanoid two weeks ago.
• 6:17
  SPEAKER_00: There's only about two or three, so you could probably figure it out at home.
• 6:20
  SPEAKER_01: And I was at a cocktail party, and no one was interested in it because there were uh roboticists at this cocktail party, they knew it was junk and uh well junk, that's maybe an overstatement.
• 6:32
  SPEAKER_01: But I I said to someone, hey, hey, why don't we go and and and try robot tipping?
• 6:37
  SPEAKER_01: And she said, Yeah, we weren't really gonna do it.
• 6:41
  SPEAKER_01: But we turned to look at the robot, it was about 25 feet away, and at that moment it fell flat on its face.
• 6:46
  SPEAKER_00: So I had a very similar, I will also not name the company, but I had a very similar experience wherein I was around uh a robot demo and they were telling me something it could do.
• 6:55
  SPEAKER_00: So I said, Oh, well, let's see if it can do that on the fly and moved it towards the robot.
• 6:59
  SPEAKER_00: And the guy was like, uh no, that's not part of the demo that we're doing right now.
• 7:04
  SPEAKER_00: In terms of your skepticism, if I'm understanding it correctly, it's less that you don't believe in the form factor generally.
• 7:14
  SPEAKER_00: It's more that you have doubts about claims where it is right now and how long it's going to take to really sort of roll out and be useful.
• 7:23
  SPEAKER_00: Aaron Powell Yeah.
• 7:24
  SPEAKER_01: And this gets back to my my comment about maybe I I'm I'm not to be trusted because I sort of know how hard this is.
• 7:31
  SPEAKER_01: And there's a whole bunch of people who've haven't really done anything in robotics, but they see the human form and think, oh, that must be the right way.
• 7:38
  SPEAKER_01: That must be easy.
• 7:40
  SPEAKER_01: But uh you know, I've been doing this for 50 years, and I it's hard.
• 7:45
  SPEAKER_01: It's really hard to make progress.
• 7:48
  SPEAKER_01: We really don't understand how to do physical manipulation in the world.
• 7:52
  SPEAKER_00: Aaron Powell Yeah.
• 7:53
  SPEAKER_00: I mean I mean, obviously that there are those people, but there are also a lot of very talented uh people with experience in the field who are currently working on these robots.
• 8:04
  SPEAKER_00: Yeah.
• 8:06
  SPEAKER_00: Okay.
• 8:07
  SPEAKER_01: Very few of them have ever deployed robots.
• 8:10
  SPEAKER_01: Sure.
• 8:10
  SPEAKER_01: And I, you know, and I know how hard it is to deploy robots, how hard it is to make something that the customer is going to pay for.
• 8:17
  SPEAKER_01: And it's got to be very, very reliable, and it's got to work with a bunch of nines, 99.999 percent of the time.
• 8:24
  SPEAKER_01: If you get higher failure rates than that, it's really frustrating and perhaps dangerous.
• 8:29
  SPEAKER_00: Aaron Powell Yeah.
• 8:29
  SPEAKER_00: So okay, so that's a fair distinction.
• 8:31
  SPEAKER_00: Um I was thinking and and maybe uh a little bit more theoretically, but obviously people who have been in research facilities and have been working around robots in that context, but uh the the breakdown for you is that they haven't really scaled and they haven't really produced industrial robots in a meaningful way.
• 8:48
  SPEAKER_01: Aaron Powell Yeah, and and and that customers uh uh the customers in general, in general, customers don't care about the particular technology using.
• 8:57
  SPEAKER_01: They want cost-effective results.
• 9:00
  SPEAKER_01: They want it to somehow make them more efficient than they were without the technology, and that's what matters to them.
• 9:07
  SPEAKER_01: It doesn't matter that it's using a large manipulation model or that it's using deep learning.
• 9:13
  SPEAKER_01: Which uh both can both can be good things, but they don't care about the particular technology.
• 9:18
  SPEAKER_01: Whereas people come from a research background, they're built a particular technology.
• 9:22
  SPEAKER_01: That's their their identity, and they're gonna use that and show the world that it's important.
• 9:27
  SPEAKER_01: And it's often not so important.
• 9:29
  SPEAKER_00: Aaron Powell Yeah, one of the things that really struck me uh last time we spoke, I was actually out there um in Redwood City and actually seeing a demo of the robot in person.
• 9:40
  SPEAKER_00: One of the things that's baked into it that I don't see uh baked into a lot of these robots that we're talking about right now is that collaborative element and is the ability for human workers working alongside these things to like literally physically pick it up and move it out of the way.
• 9:57
  SPEAKER_01: Yeah, and uh that's been something that I've been doing for a long time.
• 9:59
  SPEAKER_01: You know, the Roomba had a handle on it.
• 10:01
  SPEAKER_01: Uh you could pick it up, get a, you know, if we got stuck in a bunch of cords, you'd just pick it up and move.
• 10:08
  SPEAKER_01: The um um humanoid robots only waste-up humanoids, um, Baxter and Sawyer.
• 10:14
  SPEAKER_01: You could grab their arm at any time and push push back on it, which you cannot do for most industrial robots.
• 10:21
  SPEAKER_01: Um uh the the you know, the traditional industrial robots uh that are used in um autofactories, et cetera.
• 10:29
  SPEAKER_01: We made a robot that you could grab the arm and move it out of the way.
• 10:33
  SPEAKER_01: For the cart that we have in this company, you know, it it can get in the way, and you just want to not not where it is.
• 10:39
  SPEAKER_01: And just being having a handle that you can grab and move it is great.
• 10:44
  SPEAKER_01: And then some of our customers are just really focused on that handle because it it you know, there can be a hundred kilograms on the cart and it feels like nothing.
• 10:54
  SPEAKER_01: In fact, my both my three-year-old and six-year-old grandsons have come and pushed the robots around by that handle with a lot of weight on there.
• 11:02
  SPEAKER_01: So um that's uh that that's that makes the job easier for people because they've got power assist, uh, but they can still have control if they want control.
• 11:14
  SPEAKER_00: This necessary 99.999 percent accurate rate is something certainly that I understand in the context of like self-driving cars, for example.
• 11:22
  SPEAKER_00: Obviously, if one of those doesn't do the right thing, then that could result in death.
• 11:27
  SPEAKER_00: Um But let's take the Roomba, for example, right?
• 11:30
  SPEAKER_00: I mean, people love their Roombas, but every single person who owns a Roomba has probably run into multiple issues.
• 11:36
  SPEAKER_00: Is it just that the nature of industrial deployment is that different from the home?
• 11:42
  SPEAKER_01: Aaron Ross Powell Yeah, well the the nice thing about the Roomba is it doesn't weigh much.
• 11:46
  SPEAKER_01: And so it can't do too much damage when it does fail.
• 11:51
  SPEAKER_01: And we we put a lot of effort into making it so it never falls downstairs.
• 11:55
  SPEAKER_01: We had triple redundancy there.
• 11:57
  SPEAKER_01: Uh because that's the only time you can get enough kinetic energy in a roomba for it to be dangerous when it's falling from a great height.
• 12:05
  SPEAKER_01: Um other bigger robots, though, don't have to be moving too fast to you know have enough kinetic energy to squish you.
• 12:13
  SPEAKER_01: Um so there you really want that reliability.
• 12:16
  SPEAKER_01: It's it's not a lot different from you know autonomous driving.
• 12:20
  SPEAKER_01: Uh there's there's you know 100 kilo kilos of mass can can do some damage, even at a meter a second.
• 12:27
  SPEAKER_00: Aaron Powell So we are talking about ultimately a safety issue there.
• 12:30
  SPEAKER_01: Aaron Powell There's safety, but there's also frustration.
• 12:33
  SPEAKER_01: Um so if i in the case of an industrial robot, an industrial robot that's sitting somewhere on the line, it's supposed to pick something up and dip it in paint or uh put it on something else, if that's failing, you know, uh uh once an hour even, and maybe it's only doing uh 500 or 1,000 operations an hour, but if it fails once an hour, that's a pretty useless robot, because that means a person has to be checking on its code all the time.
• 13:00
  SPEAKER_01: If it fails once a day, maybe that's okay.
• 13:03
  SPEAKER_01: But you know, if you've got 50 robots and they're all failing once a day, that's fifty failures you've got to deal with.
• 13:09
  SPEAKER_01: So you're you know you know, as you scale up, there's gotta be a lot more nines, or you need more people to just take care of the failures.
• 13:18
  SPEAKER_00: Aaron Powell The stakes of a Roomba accidentally sucking up one of your socks are probably not huge to you.
• 13:25
  SPEAKER_00: But when when your company is relying on these things and when you're paying, you know, probably 50 times the cost of the most expensive Roomba, then they really need to deliver almost all the time.
• 13:36
  SPEAKER_01: Aaron Powell Yeah, or otherwise the the the customer's not gonna not gonna get return on investment because they need a babysitter for the robot.
• 13:44
  SPEAKER_01: Yeah.
• 13:44
  SPEAKER_01: Yeah.
• 13:45
  SPEAKER_01: Or they need a babysitter for every 10 robots or something like that.
• 13:48
  SPEAKER_01: Some some calculation.
• 13:49
  SPEAKER_00: Aaron Ross Powell I I know AGI is something that you've has interested you for a long time and you've been um thinking and writing and speaking about it.
• 13:56
  SPEAKER_00: Is that something how close do we have to get to that in order for these robots to be meaningful in a way that they're actually valuable to the businesses that use them?
• 14:07
  SPEAKER_01: Aaron Powell Oh, we don't need AGI at all for them to be useful.
• 14:10
  SPEAKER_01: Um and that's good because we're not going to get AGI for 300 years.
• 14:14
  SPEAKER_01: That's an exact number.
• 14:16
  SPEAKER_01: You know, everyone in the in the 50s and 60s in artificial intelligence, they didn't call it AGI, but that's what they meant by artificial intelligence.
• 14:24
  SPEAKER_01: The same thing people meant.
• 14:26
  SPEAKER_01: The reason we talk about AGI now is a group of researchers decided to differentiate themselves and claim they were doing general intelligence, but that's what everyone thought they were doing.
• 14:34
  SPEAKER_01: So it's a it's a marketing ploy.
• 14:36
  SPEAKER_01: And we're just a long, long way from it.
• 14:38
  SPEAKER_01: Certainly the LOMs are nothing like that.
• 14:40
  SPEAKER_01: They can do things that are surprising, but that's different from being generally intelligent.
• 14:46
  SPEAKER_00: Aaron Ross Powell I I wonder uh how much of this is a case of people, various people having different definitions of what AGI ultimately means.
• 14:54
  SPEAKER_00: How would you like briefly dis uh explain to somebody on the street what AGI is?
• 14:59
  SPEAKER_01: Aaron Powell Yeah.
• 15:00
  SPEAKER_01: So when when AGI first became a term, it meant to be able to, you know, do any reasoning that uh people could do and be able to do anything in the world that people could do.
• 15:10
  SPEAKER_01: But as with all these definitions, you know, they change.
• 15:12
  SPEAKER_01: You know, flying cars used to mean a car that you could drive and that you could fly.
• 15:17
  SPEAKER_01: Now a flying car or a flying taxi means uh, you know, a quadcopter that can land at a particular location and then can't move.
• 15:25
  SPEAKER_01: Um so the definitions change over time.
• 15:28
  SPEAKER_01: There these are moving targets for what it is.
• 15:31
  SPEAKER_01: And I think AGI has turned into it can take a test and pass the test, maybe.
• 15:37
  SPEAKER_01: Yeah.
• 15:38
  SPEAKER_01: Um that's not general intelligence that's going to figure out what, you know, look at a situation like any human who's not even particularly trained can come into a room and see what what what stupid things a machine is doing and and intercede in all sorts of situations, even a machine they've never seen before.
• 15:56
  SPEAKER_01: They can recognize, oh, this is failing.
• 15:58
  SPEAKER_01: This thing is clearly broken.
• 16:00
  SPEAKER_01: Um, because they can build up a model in their head of what the machine must be trying to do and what it's actually doing, and figure out a chain of things which involve geometric reasoning, involve perceiving things that they've never seen before.
• 16:14
  SPEAKER_01: And that's not what any of these AGI tests are about.
• 16:17
  SPEAKER_00: Aaron Ross Powell One of the things that uh I think you are really optimistic about is uh this long ongoing conversation about technology replacing jobs.
• 16:28
  SPEAKER_00: Uh and that that it's it's going to take a long time, that the jobs that people the I guess we would call them blue-collar jobs that people have are going to continue to evolve.
• 16:39
  SPEAKER_00: And it seems like an important piece of that in the meantime is to build technology that has humans in the loop as much as possible or as much as needed.
• 16:48
  SPEAKER_01: Aaron Ross Powell Well, you know, if if there's any grasping involved, we're going to need humans for a long time yet.
• 16:56
  SPEAKER_01: Grasping is being worked on academically, and and uh Amazon's put a lot of money into it.
• 17:02
  SPEAKER_01: And we're not we're still not anywhere close to having grasping that humans can do.
• 17:06
  SPEAKER_01: So humans will have to grasp.
• 17:08
  SPEAKER_01: But what we can do is automate the things around them that are not grasping.
• 17:12
  SPEAKER_01: Amazon does that by having robots bring shelves full of stuff to a human who's in one place, and the human picks from those shelves, and the shelves drive away, and another set of shelves come.
• 17:24
  SPEAKER_01: In our case, uh we're going around existing warehouses without automation, and the humans pick and put into the robot cart, but then you know, the last 400, 500 feet to go to the pack station, they just say go, and the robot goes, and they don't have to do that, what's called dead walking.
• 17:43
  SPEAKER_01: So two things that we we make better for the humans are the number of steps they have to do a day.
• 17:49
  SPEAKER_01: Often in warehouses, typical sort of number of steps for a human can be over 30,000 steps a day, which is grueling when you're doing that every day for your job.
• 17:58
  SPEAKER_01: Uh so we reduce a lot of that.
• 18:00
  SPEAKER_01: And then we make the carts appear to be lighter than they are when they are pushing them with the power assist and the handlebar.
• 18:07
  SPEAKER_01: You don't have to be a a big football player that uh that you need to be sort of in some of these facilities just to move the carts around.
• 18:14
  SPEAKER_01: You can be a small person and it's the same amount of effort as for a big person.
• 18:19
  SPEAKER_01: So it it increases the labor pool and it makes the jobs physically less taxing, but still has the people doing the grasping because we cannot automate that at the moment.
• 18:32
  SPEAKER_01: No one can.
• 18:33
  SPEAKER_00: Yeah, one thing uh as you're as you're speaking about Amazon um it reminded me of something I think that maybe doesn't get talked about enough, or at least doesn't get talked about enough in broader circles, which is the degree to which, given the current state of technology, um it's not as much in a lot of cases, it's not as much about the robot complying to the existing environment as actually kind of building an environment that makes sense given those limitations.
• 19:00
  SPEAKER_01: Yeah.
• 19:00
  SPEAKER_01: And that's what the That's what Amazon uses in about 80% of their warehouses.
• 19:06
  SPEAKER_01: They don't use it in all of their warehouses because you know big things are not amenable to that particular solution.
• 19:12
  SPEAKER_01: You need to move them and pick them up in different ways.
• 19:16
  SPEAKER_00: Aaron Powell You know, you keep putting yourself in the position that you're in as a co-founder of a startup.
• 19:23
  SPEAKER_00: Uh given you know the ups and downs that you've experienced with different companies along the way, like what keeps bringing you back into that world instead of, I don't know, I don't want to say retreating, but spending more time, say, in uh the university, yeah.
• 19:39
  SPEAKER_01: Well, you know, I've got a book, like Magnum Opus, is that how you say it?
• 19:43
  SPEAKER_01: That I'm trying trying to write, but it's been on hold since I started this company.
• 19:47
  SPEAKER_01: Uh and I want to get back to that.
• 19:49
  SPEAKER_01: But I also I like being in the arena.
• 19:53
  SPEAKER_01: I uh you know, um I think Teddy Roosevelt used that uh phrase.
• 19:58
  SPEAKER_01: So when I um say something's gonna take a long time or something's gonna be hard, it's not because I'm just you know read read a s read read something or saw a picture.
• 20:11
  SPEAKER_01: It's because I've lived it and continue to live it.
• 20:13
  SPEAKER_01: And so I have pretty good uh estimation skills on how hard certain things are to do.
• 20:20
  SPEAKER_01: And I get that from trying to do stuff.
• 20:22
  unknown: Trevor Burrus, Jr.
• 20:22
  SPEAKER_00: Yeah, I I mean one one thing I like to ask people as as somebody I'm certainly new to the space relative to somebody like you who's been in here for decades, is if this you can say it, Brian.
• 20:35
  SPEAKER_00: Someone as older as me.
• 20:37
  SPEAKER_00: Someone as old as as Rodney.
• 20:38
  SPEAKER_00: Uh if this current hype cycle that we're going through, and I would include both, you know, uh LLMs and humanoids in that, is in any way different than ones you've been through in the past.
• 20:48
  SPEAKER_01: Aaron Ross Powell I don't think it's different in in in um in what people are saying or how over uh pumped up their expectations are.
• 20:58
  SPEAKER_01: It's different in that it's leaked out into the general world.
• 21:02
  SPEAKER_01: And then there's all these pundits who've never even seen a robot talking about what robots can do and how much per hour they're gonna cost and all sorts of crap uh uh to use the technical term being talked about, all sorts of things.
• 21:17
  SPEAKER_01: Um so that's the difference.
• 21:18
  SPEAKER_01: That has leaked out and it has become commonplace.
• 21:21
  SPEAKER_01: And so when I'm at a dinner party, people come to me and fearfully ask, is their job gonna be replaced?
• 21:27
  SPEAKER_01: Is this gonna happen?
• 21:28
  SPEAKER_01: Is that gonna happen?
• 21:30
  SPEAKER_01: And and the questions are sort of ludicrous, because no, they're not gonna happen the way people have been scared into thinking that they're gonna happen.
• 21:38
  SPEAKER_01: By the way, I think there's a real parallel between LLMs and humanoid robots.
• 21:43
  SPEAKER_01: Yeah.
• 21:43
  SPEAKER_01: Uh in the following sense: the phenomenon phenomena of them, you know, uh big tech companies are putting billions of dollars into them.
• 21:52
  SPEAKER_01: Um uh academia is sort of getting priced out.
• 21:56
  SPEAKER_01: That's not where the the action is happening, it's happening in the big companies.
• 22:00
  SPEAKER_01: Um but um you know the promise the and the investors get excited because LLMs are gonna replace lots of white-collar workers and humanoids are gonna replace lots of blue-collar workers.
• 22:12
  SPEAKER_01: And then how do they get lured into that?
• 22:14
  SPEAKER_01: Well, the LLMs use human language.
• 22:17
  SPEAKER_01: Well, that's pretty damn good.
• 22:18
  SPEAKER_01: And the humanoid robots have human form.
• 22:21
  SPEAKER_01: Well, that must mean they're like humans.
• 22:22
  SPEAKER_01: And you know, they the LLMs can talk like humans, the human humanoid robots can move like humans.
• 22:28
  SPEAKER_01: And and so there's a lot of similarity uh there between the two things.
• 22:32
  SPEAKER_01: And I think that's why the expectations have gotten so large and it's burst out into the world.
• 22:38
  SPEAKER_01: You know, once it starts, all the big tech companies have to say they're in or they'll be viewed as being left out.
• 22:45
  SPEAKER_01: And then uh people who are trying to raise money uh, you know, come up with such inflated promises that will never be met in order to get big piles of money to pay for training data, et cetera.
• 22:59
  SPEAKER_00: Aaron Powell Yeah, Apple's probably the best example of that right now when it comes to like the AI side of things.
• 23:05
  SPEAKER_00: Yeah.
• 23:06
  SPEAKER_01: I'm surprised to hear that this is a relatively new phenomenon for you, people asking about their jobs being replaced because you know, especially Yeah, I did I did get it back, uh you remember all the way back in ancient time, 2011, uh when Watson, you know, won um Jeopardy.
• 23:27
  SPEAKER_01: And then uh IBM, the next day, John Markov had a story in the New York Times um, you know, the saying what IBM was saying, that it was gonna uh do medicine with it, it was gonna um um help help people choose objects online or to buy, um, they were gonna put a voice on it, it was gonna be able to understand language.
• 23:47
  SPEAKER_01: You would take the story that John, who is a dear friend by the way, wrote in 2011, replace Watson with um generative AI and um IBM with Microsoft, and it's exactly what happened.
• 24:01
  SPEAKER_01: The same story was was the case.
• 24:03
  SPEAKER_01: It's gonna do health, it's gonna do medicine, it's gonna uh and and and I got a lot of people um actually in my apartment building in Cambridge who worked in biotech and coming up to me, is this true?
• 24:15
  SPEAKER_01: Is this gonna, you know, completely change biotech?
• 24:18
  SPEAKER_01: Is it gonna change uh medicine?
• 24:20
  SPEAKER_01: Um and no, it wasn't true, and no it didn't.
• 24:25
  SPEAKER_01: Um but they were really dumb back then, those people in 2011, to believe that.
• 24:30
  SPEAKER_01: Not like now.
• 24:31
  SPEAKER_01: We're smart now.
• 24:32
  SPEAKER_01: So if we believe it now, it must be true.
• 24:34
  SPEAKER_00: Do you do you is is it uh not one-to-one, but is is it a extremely close parallel to what we just went through with autonomous cars?
• 24:43
  SPEAKER_01: Yeah.
• 24:43
  SPEAKER_01: Uh you know, autonomous cars may or may not come through.
• 24:46
  SPEAKER_01: I think the next ten years will tell.
• 24:49
  SPEAKER_01: Now, ten years.
• 24:50
  SPEAKER_01: Back in the olden days, in the mid-teens, autonomous cars were going to be here and everywhere in three years, you might remember.
• 24:58
  SPEAKER_01: But I think, you know, there are there are two players, Waymo, which is number one, uh, you know, which is part of owned by Google, and Zeux, which is a distant number two, but nevertheless a number two, which is owned by Amazon.
• 25:13
  SPEAKER_01: They both have big resources in capital.
• 25:16
  SPEAKER_01: If they can't in the next ten years really make it big and widespread, I think interest will will get lost.
• 25:24
  SPEAKER_01: And um we'll go back to other solutions, which maybe we should have been working on in the meantime.
• 25:30
  SPEAKER_01: Uh where instead of hoping everything came in the car and we didn't have to change infrastructure, in fact, changing infrastructure may have been a better idea to put more sensing on the roads, more communication with the cars as they drive, instead of it being you don't have to spend anything on infrastructure.
• 25:48
  SPEAKER_01: I'm just going to deliver you this car that's going to do everything by itself.
• 25:52
  SPEAKER_01: Again, thinking that it can reach human-level performance.
• 25:55
  SPEAKER_00: Trevor Burrus, Jr.: Maybe this is an impossible thing to predict, although that hasn't stopped you in the past from predicting things.
• 26:01
  SPEAKER_00: Do you see these do you see humanoids in particular continuing to play out a similar trajectory as autonomous cars in the sense that there's going to be one giant company keeping the flame alive and everybody's going to sort of shrink back or go away?
• 26:17
  SPEAKER_00: Or is it something that's just kind of going to go away for a bit?
• 26:20
  SPEAKER_01: Aaron Powell I think it's probably going to go away.
• 26:22
  SPEAKER_01: If someone really comes up with a better walking mechanism soon, then that might make things safe enough and cheap enough to be useful.
• 26:32
  SPEAKER_01: But I don't see the manipulation part getting fixed.
• 26:38
  SPEAKER_01: So I think it's yeah, yeah, it's going to die off.
• 26:42
  SPEAKER_00: Would you group these kind of, I guess, modular versions of humanoids that we're seeing right now where it's like a torso on a wheeled base in in the same place and that, okay, it can move around, you know, the way any wheeled robot can, but maybe the state of manipulation isn't as good as what's already out there.
• 27:00
  SPEAKER_01: Yeah, well, I I built a a a humanoid robot with an arm and a wheeled base in 2004, uh, published about it on uh page one of issue one of volume one of humanoid robots in 2004, where I made the argument that that tool form was going to be useful in built for human environments.
• 27:22
  SPEAKER_01: And we had that robot going around and opening doors in our lab, uh in the in the you know, where the students had their offices, et cetera, and manipulating that built for human environments.
• 27:34
  SPEAKER_01: It's called Cardia.
• 27:35
  SPEAKER_01: Cardia is the Roman god of uh door door handles.
• 27:40
  SPEAKER_00: Uh they get very specific over an ancient Rome, don't they?
• 27:44
  SPEAKER_01: Yeah, anyway.
• 27:46
  SPEAKER_01: Um so you know, the idea's been around a long time, but that was built on two wheels, as the ones we see now are.
• 27:52
  SPEAKER_01: It was actually a uh you I forget the name.
• 27:55
  SPEAKER_01: The thing people wrote around on on two wheels.
• 27:58
  SPEAKER_01: Segway.
• 27:58
  SPEAKER_01: It was a segue, an actual segue base.
• 28:01
  SPEAKER_01: Um but it was a bit dangerous when things got got bad, and those wheel bases are still dangerous when things get bad.
• 28:08
  SPEAKER_01: Uh so what when you don't when you lose your balance.
• 28:12
  SPEAKER_01: Um maybe, but maybe not.
• 28:15
  SPEAKER_01: Um I have one other data point from 2011, March 2011, when the great tsunami in Japan happened, and um uh there were meltdowns in um three of the four reactors.
• 28:31
  SPEAKER_01: Uh the six reactors total in of uh Fukushima Daiichi, uh, but there were four along the waterline, and three of the four ended up with meltdowns.
• 28:41
  SPEAKER_01: And, you know, the idea that um a built-for-human environment is a perfect one for humanoids to go into was put to rest.
• 28:49
  SPEAKER_01: Um Japan had more humanoid robots with legs than anyone, and they couldn't go in there.
• 28:56
  SPEAKER_01: Um we were led to, you know, at iRobot, we also had military robots.
• 29:02
  SPEAKER_01: We had 6,500 of them in Afghanistan and Iraq.
• 29:05
  SPEAKER_01: They were tracked, they could go over any terrain.
• 29:08
  SPEAKER_01: And we were uh given to understand that the Japanese government would not be insulted if we sent some.
• 29:14
  SPEAKER_01: So we sent some, and they got there seven days after the accident, and they were the robots that went in to um uh figure out what was going on, get cameras in place.
• 29:24
  SPEAKER_01: The radiation was just so intense people couldn't go in there.
• 29:26
  SPEAKER_01: And it was a non-digital um it was a 40-year-old nuclear power plant.
• 29:31
  SPEAKER_01: So the only thing you could do was go and look at, you know, physical dials, pressure and temperature with no data communications.
• 29:38
  SPEAKER_01: Tokyo Electric Power Company personnel, who we trained, remotely operated the robot, set up Wi-Fi hotspots, set up a network, you know, uh a network over those robots and had them go into the damaged reactors.
• 29:52
  SPEAKER_01: So there was the perfect situation for humanoids to go into a built-for-human environment, and it did not pay off.
• 29:59
  SPEAKER_00: In in that particular scenario in search and rescue, um, do you think there is more efficacy potentially in drones and also quadrupeds like spot?
• 30:12
  SPEAKER_01: Yeah, yeah, yeah, absolutely.
• 30:13
  SPEAKER_01: And you know, drones were just not a thing.
• 30:16
  SPEAKER_01: You know, there were a few research ones, but they were nothing like the the capabilities and communications we have today.
• 30:22
  SPEAKER_01: So yes, drones would have been perfect for that.
• 30:25
  SPEAKER_01: And spot-like robots.
• 30:26
  SPEAKER_01: Um, four legs, run too, because it's much harder to fall down with four legs and this with two legs.
• 30:33
  SPEAKER_01: And dogs actually go in most places humans can go anyway.
• 30:36
  SPEAKER_00: So getting back to the uh the this humanoid that you had built earlier, Baxter, one of the things that is really fascinated me about it over the years is how it became and continues to be an important part of the research world, right?
• 30:51
  SPEAKER_00: I mean, they're still out there, and I still see them when I go to schools, and probably a lot of that is you know a product of other companies not really coming in and necessarily filling that that void.
• 31:05
  SPEAKER_01: One of the things about them is they're safe.
• 31:08
  SPEAKER_01: You know, they uh uh you can get sand in the way of it, and uh it it senses when it when the when its forward model isn't matching reality, uh, and it it uh pumps the kinetic energy out of the system um so you don't get hurt.
• 31:23
  SPEAKER_01: Um so that makes it really good for research because you can feel confident having graduate students late at night working with it and they're not gonna get hurt.
• 31:31
  SPEAKER_01: Or undergrads, you know.
• 31:32
  SPEAKER_01: Uh Berkeley, I think, has six of them for their undergraduate program in robotics uh still that they teach.
• 31:39
  SPEAKER_00: Well was that was that research, was that university aspect, was that almost accidental or was that part of the plan?
• 31:44
  SPEAKER_01: Aaron Powell No, it was not part of the plan.
• 31:45
  SPEAKER_01: It was accidental, and I had to fight for it.
• 31:48
  SPEAKER_01: Um because um, you know, uh I realized after we built them that it that would be good for research.
• 31:55
  SPEAKER_01: Um and I said, let's sell start selling them to universities.
• 31:59
  SPEAKER_01: And the the the industry people in the company said, Oh no, no, no.
• 32:03
  SPEAKER_01: It'd be so much support, it'd be so hard that it'd be it'd be it'd be horrible having to you know deal with graduate students and stuff.
• 32:09
  SPEAKER_01: And so I I got a few professors together and and you know asked them some questions.
• 32:14
  SPEAKER_01: And um I said that you know, one of the questions was, you will you be comfortable maintaining the robot and fixing it and altering it?
• 32:21
  SPEAKER_01: They said, I've never had a robot that I that I didn't have to do that, of course.
• 32:26
  SPEAKER_01: And that's what's happened.
• 32:27
  SPEAKER_01: Um so it was industry not understanding the mindset of research universities where that's what everyone does.
• 32:34
  SPEAKER_01: They thought, oh, we have to build a perfect product that any student can use.
• 32:38
  SPEAKER_01: Students take them apart and hack them and do all sorts of stuff.
• 32:41
  SPEAKER_00: So Yeah, I one one of the things that I appreciate that that Fetch was doing for a while, and and I think that this is kind of a direct uh product of them in a certain way being an outgrowth of Wheeler Garage, is that they were building a research robot, and they were doing it because they had the very savvy understanding that getting students at universities familiar with your technology is ultimately going to serve you in the long run.
• 33:06
  SPEAKER_01: Aaron Powell Yeah, and they they used ROS, which students were already using.
• 33:10
  SPEAKER_01: And by the way, here we when we started, we bought a bunch of fetch platforms and dressed them up as to simulate the real robots we wanted to build, but it it let us start running all the the uh perception algorithms long before we built the mechanical robot that was our first uh series, which is now uh we no longer we no longer use it's the the second generation, which is what we're deploying at scale.
• 33:35
  SPEAKER_01: But we we we simulated our robot with a fetch, which was great.
• 33:39
  SPEAKER_00: Aaron Powell Yeah, I wanted to you alluded a little bit to your magnum opus that you are unable to write because you don't have the time right now.
• 33:46
  SPEAKER_00: Um what what what what's the blurb on that the back of that book?
• 33:51
  SPEAKER_00: I mean, what what's sort of the big thing that you want to impart on people through all of your years of experience?
• 33:56
  SPEAKER_01: Aaron Powell Different than than what we're talking about here.
• 33:59
  SPEAKER_01: Um but it's to do with AI.
• 34:02
  SPEAKER_01: The main idea is that if you you know you have a two-by-two matrix with science versus engineering and intelligence versus life.
• 34:13
  SPEAKER_01: So there's four boxes.
• 34:15
  SPEAKER_01: And one of them is neuroscience, one is AI, one is artificial life, one is abiogenesis, how did life come from matter?
• 34:25
  SPEAKER_01: And those four disciplines were all formed into their modern form in the period 1945 to 1965, and any two of them I can t show you a person who worked in both of those two, and for any three of them, I can usually find a person who worked in three of them.
• 34:41
  SPEAKER_01: And three out of the four ha have decided that computation is the primary metaphor.
• 34:47
  SPEAKER_01: And it's questioning that.
• 34:49
  SPEAKER_01: It's questioning whether computation is the right way to think about things.
• 34:53
  SPEAKER_01: Let me use an example.
• 34:55
  SPEAKER_01: If, say, you wanted to put something into orbit, you wouldn't just write a Python script and think that would get something into orbit.
• 35:02
  SPEAKER_01: You know you have to burn rocket fuel or have something that releases tremendous amounts of energy in some way.
• 35:09
  SPEAKER_01: Computation is not the right kind to get something into orbit.
• 35:14
  SPEAKER_01: But we've decided it's the right kind to explain what goes on in our heads and our bodies.
• 35:21
  SPEAKER_01: And that's the question.
• 35:22
  SPEAKER_01: Whether that's accurate.
• 35:24
  SPEAKER_00: Yeah.
• 35:24
  SPEAKER_00: So that this is my listen, I was a cr I creative writing major at UC Santa Cruz, so I'm maybe not as technical as as I should be, but so so my my dumb uh liberal arts degree question is what's the alternative?
• 35:38
  SPEAKER_01: Aaron Ross Powell Yeah, that and that's the qu that's the question.
• 35:41
  SPEAKER_01: But but you don't say that about rockets.
• 35:43
  SPEAKER_01: What's the alternative to writing code to make a rocket?
• 35:46
  SPEAKER_01: You say there's something different.
• 35:48
  SPEAKER_01: But we we put that aside.
• 35:50
  SPEAKER_01: And um I I've I've gone actually I start I I've written a whole bunch of this.
• 35:55
  SPEAKER_01: I start back in 1614 um with Latin manuscripts um uh and traced the ideas from from there uh through what computation is, what what's the difference between simulation and actuality.
• 36:09
  SPEAKER_01: Um so here's here's a here's an example.
• 36:12
  SPEAKER_01: I'll give you a simple example of something um which looks intelligent in some way, if you look at it the right way.
• 36:20
  SPEAKER_01: When you're on a freeway and there's an incoming, you know, an incoming uh single lane, there's a normally a a piece of space between them that gets narrower and narrower as the lanes come together.
• 36:32
  SPEAKER_01: And if you look down there, you see them covered with little bits of junk.
• 36:36
  SPEAKER_01: And those bits of junk are not on the on the freeway, and they're not on the on-ramp, but those bits of junk are on the on the blacktop in that and confined to that area.
• 36:46
  SPEAKER_01: Well, how did they get there?
• 36:48
  SPEAKER_01: Did they think about how to get there?
• 36:50
  SPEAKER_01: No.
• 36:50
  SPEAKER_01: What happened is every time there's a piece on um one of the lanes or the on-ramp, and a car hits it, it jumps up and randomly moves.
• 37:00
  SPEAKER_01: And once it's in that, you know, area between the two white lines that are coming together, no car ever goes there, so it doesn't move again.
• 37:09
  SPEAKER_01: So the cars sort stuff into places where cars drive and cases where uh cars don't.
• 37:17
  SPEAKER_01: You could build a robot that thought about that to sort do the sorting uh and computed about it, or you can just have to let the cars drive.
• 37:24
  SPEAKER_01: So there's a there's an emergent thing.
• 37:27
  SPEAKER_01: And maybe we're just emergent in a in a big scale.
• 37:30
  SPEAKER_01: And what's that emergence look like?
• 37:32
  SPEAKER_01: And and evolution has built us in that way.
• 37:35
  SPEAKER_01: And this hurts people's heads when I say this.
• 37:38
  SPEAKER_01: They they just they just can't can't concede that it could be something different.
• 37:42
  SPEAKER_01: So I've got a I've got a chapter coming out in an MIT Press book with 12 questions that we assume about intelligence, but which we actually don't know, but we uh are just hard for us to to get over thinking that we couldn't have it right.
• 37:57
  SPEAKER_01: And you know, some of the questions are do you need a living body for intelligence?
• 38:01
  SPEAKER_01: Um we can't prove that you don't need a living body for intelligence, because the only true intelligence we have is in living bodies.
• 38:08
  SPEAKER_01: Well, maybe there's something there.
• 38:10
  SPEAKER_01: Um I'm trying to make people think different and hurt their heads.
• 38:14
  SPEAKER_00: Yeah.
• 38:14
  SPEAKER_00: I I I I appreciate that.
• 38:16
  SPEAKER_00: And actually, like on that note, um so again, something I think that probably doesn't get talked enough about outside of um you know computer science circles is how much of these LLMs or just like large model AIs are are effectively like a black box from where we sit, right?
• 38:34
  SPEAKER_00: And and I I saw in the um in your in your piece that you wrote, in your predictions piece, you know, you you quoted that very famous Arthur C.
• 38:44
  SPEAKER_00: Clark, um what is it, any any technology that's advanced enough is to uh indistinguishable from magic.
• 38:52
  SPEAKER_00: Yeah.
• 38:52
  SPEAKER_00: Yeah.
• 38:53
  SPEAKER_00: Um which I mean that has to be an exciting and one of the more interesting aspects of your job is not realizing that there's going to be rubber on the road until those cars are actually out there.
• 39:04
  SPEAKER_01: Aaron Powell Yeah.
• 39:05
  SPEAKER_01: Some of them are pretty mild, but um um you know, one thing that Roombas did was make people keep their houses more tidy so the roombas wouldn't run into stuff that they couldn't handle.
• 39:16
  SPEAKER_01: Um so it the room was supposed to save them from cleaning their house, but they kept it tidier so the rumor could do its work.
• 39:24
  SPEAKER_01: Uh so you get weird effects like that.
• 39:26
  SPEAKER_01: I'm not sure that's the that's really where you're going, but these things do change how we how we see.
• 39:32
  SPEAKER_01: I mean, uh I uh an another another one of the the things I talk about is um we we tend to overestimate stuff in the short term.
• 39:42
  SPEAKER_01: This is Roy Amara that said this.
• 39:44
  SPEAKER_01: We tend to overestimate stuff in the short term, but underestimate in the long term.
• 39:48
  SPEAKER_01: So, you know, my favorite thing is a uh Spencer Tracy, Catherine Hepburn movie from the 50s called Desk Set, where a computer has been brought into this office building where there are women who answer, you know, who go and look up in the library questions the executives have, and the computer is gonna replace them.
• 40:07
  SPEAKER_01: Of course, Catherine, you know, Spencer Tracy brings the computer in, Catherine Hepburn's the head librarian, they fall in love, blah, blah, blah.
• 40:15
  SPEAKER_01: Um button.
• 40:16
  SPEAKER_01: Unexpected, unexpected results of technology.
• 40:19
  SPEAKER_01: But the the point is that the computers of the 50s were not going to replace librarians anytime soon.
• 40:25
  SPEAKER_01: Uh and for 30 years, the number of librarians in the U.S.
• 40:28
  SPEAKER_01: continued to rise, and then it dropped suddenly in corporate well, because then we got to search engines and uh, you know, and information systems, and then the idea that you know you can carry this little thing around with you, and it's got all the world's information at at your fingertips was not expected when we had punch card machines in the in the 50s.
• 40:51
  SPEAKER_01: Um so it was overestimated, oh, it's gonna replace librarians, but underestimate the scale which was gonna replace librarians and replace information um accessibility.
• 41:03
  SPEAKER_00: That's interesting.
• 41:03
  SPEAKER_00: So there there really is a parallel there between that example and the Roomba, in that when the technology isn't quite where you want it to be or where you project it to be, it actually takes more uh human interaction in order to get it there.
• 41:20
  SPEAKER_00: And ultimately, like in the case of the Roomba, it means that you're you're actually working a little bit more when you hired this thing to relieve you of that problem.
• 41:28
  SPEAKER_01: Yeah, although um you know it can um but it can also r relieve you of a problem.
• 41:34
  SPEAKER_01: I have a I have a house where while I'm not uh uh you know, a little vacation place, when I'm not there, I know that a lot of mice come to visit because there's mouse poop on the on the carpet when I come back.
• 41:47
  SPEAKER_01: So I just let the roomba pick it up, and that's great.
• 41:50
  SPEAKER_01: You know, it's just so it saves time.
• 41:52
  SPEAKER_01: But I want to I want to use this example of of information for indistinguishable from magic.
• 41:56
  SPEAKER_01: In the case of my mother, who's long departed.
• 41:59
  SPEAKER_01: But I was sitting with her in Australia, and I used to visit her twice a year, uh after she was a widow.
• 42:07
  SPEAKER_01: I was sitting with her and she was reminiscing about her grandfather and something about her grandfather's funeral in the 1930s.
• 42:15
  SPEAKER_01: And I just went and I I got the newspaper story up of her grandfather's funeral and showed it to her.
• 42:20
  SPEAKER_01: And she said, You had that in this little box?
• 42:24
  SPEAKER_01: Um so it was indistinguishable from magic to her.
• 42:28
  SPEAKER_01: It was magic that I had her grandfather's funeral newspaper report, which she probably had never even read because she when she was too young when it happened.
• 42:36
  SPEAKER_00: Aaron Powell Yeah.
• 42:37
  SPEAKER_00: So as a very technically and logically minded person, though, you don't you don't have an issue using the word magic in that context.
• 42:44
  SPEAKER_01: Aaron Ross Powell No, because it's it's a it's indistinguishable from magic.
• 42:47
  SPEAKER_01: You know she had no way of explaining it.
• 42:51
  SPEAKER_01: Um and so when we see things that we have no way of explaining, we don't know what its limitations are because everything it does seems magic.
• 43:00
  SPEAKER_01: So what is outside its range of possibilities?
• 43:05
  SPEAKER_00: So we're coming up on time, and I want to leave you with this.
• 43:08
  SPEAKER_00: Uh for some reason, I've gotten in the habit of asking like big broad questions to end impossible questions to end interviews.
• 43:14
  SPEAKER_00: But you very famously were one of the subjects in Errol Morris's uh Fast Chief Inn Out of Control.
• 43:20
  SPEAKER_00: It's been a very long time since I've seen that movie, and I was reading up on some of the um the reactions to it online.
• 43:26
  SPEAKER_00: And there's a very interesting Roger Ebert quote, and I don't know if you've seen this, but I'll read it out to you.
• 43:31
  SPEAKER_00: Uh he basically describes the film as it's about people who are trying to control things to take upon themselves the mantle of God.
• 43:41
  SPEAKER_00: Does that resonate with you?
• 43:43
  SPEAKER_01: Yeah, I think I think that's that that was I and I think that's the the theme that was in Errol's head as he put the movie together.
• 43:50
  SPEAKER_01: You know, uh for the people who haven't seen it, there's there were four people in there.
• 43:54
  SPEAKER_01: One uh was a lion tamer, one was a topiri gardener, making animals out of shapes, one was a person who built displays for naked mole rats, which are little underground creatures, and there was me building robots in human form.
• 44:10
  SPEAKER_01: I was building humanoid robots when he filmed that.
• 44:12
  SPEAKER_01: So um we were all in our own individual ways trying to make something bigger or different than ourselves in in living-ish sort of systems, and all ultimately failing.
• 44:27
  SPEAKER_01: Uh so he sh he shot that movie in 1990.
• 44:32
  SPEAKER_01: I think he shot the footage in from of me.
• 44:35
  SPEAKER_01: He shot the footage of the four people at different times in 91 or 92.
• 44:39
  SPEAKER_01: So I've been failing for well over 30 years now.
• 44:42
  SPEAKER_00: Yeah, I mean the the expectation of taking up the mantle of God is pretty high, but I guess the parallel there, uh certainly in terms of AI and humanoids, is that this is simplifying it, and obviously this is just a gigantic uh analogy, but that you're recreating life, essentially.
• 45:01
  SPEAKER_01: Aaron Powell Yeah, and there's a and and and there's a belief that somehow it's going to be better than us.
• 45:07
  SPEAKER_01: You know, I just saw Sam Altman um uh in an interview say that he now has a child, but of course the child would never be as smart as the AGI that he's building.
• 45:16
  SPEAKER_01: Yeah, I bet on his child.
• 45:19
  SPEAKER_00: He clearly has a favorite child, and it's the AGI.
• 45:23
  SPEAKER_01: The human child, yeah.
• 45:26
  SPEAKER_00: It is always a pleasure speaking with Rodney Brooks.
• 45:29
  SPEAKER_00: You can check out what he has been working on over at robust.ai.
• 45:34
  SPEAKER_00: Head over to rodneybrooks.com for his blog, including the annual prediction scorecard.
• 45:40
  SPEAKER_00: If you've made this far, please like and subscribe and check out the automated newsletter over at automated.org.
• 45:46
  SPEAKER_00: Thank you for tuning in.
• 45:47
  SPEAKER_00: We will catch you again soon.