Latest / Chaos Agents / Neural Networks, Sparse Autoencoders, and a Fleet of Duck Robots

Transcript

Neural Networks, Sparse Autoencoders, and a Fleet of Duck Robots · Chaos Agents

• 0:00
  All right, welcome.
• 0:03
  What's up, Sarah?
• 0:04
  How's it going?
• 0:05
  How was your weekend, Becca?
• 0:07
  Weekend was great.
• 0:09
  Watched some football, hung out with my baby.
• 0:13
  What about you?
• 0:15
  Oh, I had.
• 0:17
  So I had surgery.
• 0:18
  So this is like the.
• 0:20
  This is hot on all the boy podcasts, but everyone gets a vasectomy.
• 0:26
  Can we make it hot on the girl podcast?
• 0:29
  For sure.
• 0:30
  That got a girl's actomy.
• 0:32
  Oh, my God.
• 0:33
  Yay.
• 0:33
  Girl's sectomy.
• 0:36
  I should do like a TikTok journey.
• 0:39
  Girl podcasts, listen up.
• 0:41
  We're starting this trend, everyone.
• 0:45
  Okay, so recovery is not that bad.
• 0:46
  It's not as bad as I thought.
• 0:48
  Okay, good.
• 0:48
  I'm glad you're feeling.
• 0:49
  Amending.
• 0:50
  Amending.
• 0:52
  Okay, so we.
• 0:54
  We have to do our tagline.
• 0:55
  Welcome back to Chaos Agents.
• 0:57
  Okay.
• 0:57
  Welcome to Chaos Agents, the podcast, the AI News podcast that is curated by an agent named Eddie and then questioned, debated, and occasionally yelled at.
• 1:09
  We love you, though, Eddie.
• 1:10
  By human hosts Sarah, Chips and Becca.
• 1:12
  Louie rarely yell.
• 1:13
  I wish he didn't pick that word.
• 1:15
  But he did it.
• 1:15
  So we're not gonna.
• 1:16
  Yeah, we're not gonna.
• 1:17
  We're not gonna yell at him that we're not yelling at him.
• 1:21
  It's helpful feedback for us.
• 1:23
  Yeah, it's whatever he thinks.
• 1:27
  Cool.
• 1:28
  Okay, so also we're streaming this live on X right now, so we'll see how that goes.
• 1:35
  Interesting.
• 1:37
  Interesting.
• 1:38
  We're doing it.
• 1:39
  Okay.
• 1:40
  Today was an interesting newsletter, Sarah.
• 1:43
  What did you think?
• 1:44
  It was really?
• 1:45
  I think we're both really excited about the last topic, but let's get through the first ones first.
• 1:51
  Yes.
• 1:51
  Or at least I am.
• 1:52
  I bet you are too.
• 1:54
  Oh, yeah, you know, I am.
• 1:56
  But the first one I thought was really interesting and I learned a bunch.
• 1:59
  So the headline is our biggest story today is a 10 year old CNN beat GPT6 Astra and clot fable 5.1 at diagnosing skin cancer.
• 2:11
  Yes.
• 2:12
  And this is a report by Dr.
• 2:13
  Magnus Lynch.
• 2:17
  Magnus lynch from Guy's Hospital.
• 2:20
  Why is it called Guy?
• 2:21
  I'm sure Guy's Hospital is like the.
• 2:25
  And for.
• 2:27
  For those who don't know, AKA me, before reading this, a.
• 2:33
  CNN is a convolutional neural network.
• 2:38
  Yeah.
• 2:38
  It's a specific type of neural network that's good at identifying images, essentially.
• 2:45
  Yeah.
• 2:46
  So.
• 2:46
  And they're about 10 years old.
• 2:48
  That's what I was reading.
• 2:49
  Or the one that they use is about 10 years old.
• 2:51
  And so 10 years ago, were we all doing the.
• 2:54
  Remember, it was like.
• 2:55
  What was it like, dogs or cereal?
• 2:57
  It was dogs or Cereal.
• 2:58
  It was like cats or not cats.
• 3:00
  Yeah, yeah, yeah, yeah, yeah, that.
• 3:04
  So this doctor took a bunch of images of moles that are possibly melanoma and sent them to both this 10 year old CNN called Inception v3, which is an image model Google released in 2015 and then.
• 3:23
  Also gave the images to I think GPT, Astra and maybe Claude Fable and was like, what do you guys think?
• 3:31
  Which of these are concerning?
• 3:34
  Yeah.
• 3:35
  And so on 200 of them, it got 62%.
• 3:39
  Right.
• 3:39
  On the CNN and then Astra and Fable only got 42%.
• 3:45
  So this is interesting.
• 3:47
  What are your thoughts?
• 3:48
  Well, so my first thought was like, that's what I would expect because a CNN is a neural network that is supposed to be particularly good at identifying images.
• 4:00
  It's only trained, it's trained for that very specific task.
• 4:04
  I guess what's interesting, the interesting takeaway here is like, even though general models are getting so good, they're not necessarily replacement level good for like a specially trained model like this type of cnn, which I think is not
• 4:27
  necessarily a bad thing because a smaller, cheaper model that is trained to do a specific thing and still outperforms these big models is more economically feasible.
• 4:42
  Yeah, exactly.
• 4:43
  And I would be interested to know if like you did like a rag lay on top of Fable and then specifically for melanoma type images or specifically for classifying this type of images, how it would perform against the
• 5:01
  cnn.
• 5:03
  Yeah.
• 5:04
  And my other question then also is at a point in the future where a general model does outperform something like a cnn, what happens?
• 5:18
  Yeah, what, when do doctors know?
• 5:21
  Are there still.
• 5:22
  Yeah.
• 5:22
  Is there still an argument to use smaller, specifically trained models for things?
• 5:27
  I, I, I haven't thought through the second order effects of that, but that's where my brain goes.
• 5:32
  Yeah, makes a lot of sense.
• 5:35
  Okay.
• 5:36
  The next thing in the newsletter, I was pumped because it's got one of our old buddies, Sarah, the sparse autoencoder.
• 5:43
  I know, I was like, every time I think of a sparse autoencoder, I have to be, I have to remind myself what a sparse autoencoder.
• 5:53
  So I'm gonna, I'm gonna get a.
• 5:59
  But, okay, so this is before I get the official definition.
• 6:02
  So my understanding of.
• 6:03
  Cause we Talk about this PowerSoft encoder all the time is a tool to help us understand how AIs think.
• 6:11
  Correct.
• 6:11
  That's, that's basically it.
• 6:12
  It's a tool that we are, we use to try and understand what' inside a neural network.
• 6:18
  Great.
• 6:19
  And what's been really interesting about all the papers we've been reading lately, I think is.
• 6:22
  First of all, they keep coming up.
• 6:24
  So people are doing a lot of work with sparse autoencoders.
• 6:26
  And either that or Eddie is weirdly obsessed with sparse autoencoders.
• 6:32
  But there was a paper we read last week which was really interesting, that it's like, is this even the best way to understand how these models work?
• 6:41
  So I think we're learning in real time along with everyone else.
• 6:44
  Totally, which feels cool.
• 6:46
  That's one of the reasons I, like, actually really like this practice that we have of, like, going through the newsletter and talking about this stuff out loud.
• 6:52
  Guarantee you sparse autoencoder would not be in my vocabulary otherwise, I think.
• 6:59
  Okay, so I'm gonna try and pitch to you my understanding of this.
• 7:08
  This headline.
• 7:09
  Yeah.
• 7:09
  Okay, great.
• 7:10
  And then you tell me if it matches your understanding.
• 7:14
  Okay, so.
• 7:16
  So the headline is, yes, virus auto encoders need way less data than we thought until you ask if the features are real.
• 7:24
  Okay, tell me what that means.
• 7:26
  First, my understanding of a feature is like.
• 7:31
  Like, if a.
• 7:34
  If a sparse autoencoder is essentially, the analogy I like to think of is like, a brain, a human brain, and you're reading electrical signals on a brain scan, and if I showed you a picture of an alien and
• 7:52
  these 10,000 neurons in this particular area light up, you might be like, oh, that part of Becca's brain is about aliens.
• 8:02
  Or that part is related to aliens.
• 8:05
  And at a incredibly simple, overly simplified level, that's kind of what a sparse auto encoder is.
• 8:12
  And alien is like, the feature is a.
• 8:15
  It's called, like, a feature.
• 8:16
  And my understanding here is, like, what they found is that you can give a sparse autoencoder, like, activity inside of a neural net.
• 8:30
  And it might be like, oh, cool.
• 8:33
  Okay, so this feature maps to Golden Gate Bridge.
• 8:37
  This feature maps to women's basketball.
• 8:43
  And thus, here are all the things we understand about what it's thinking, and then, scrap that.
• 8:50
  Do the whole thing over again.
• 8:52
  And it might be like, okay, this feature maps to suspension bridges, not Golden Gate Bridge.
• 8:58
  And this feature maps to women's sports, not basketball.
• 9:02
  And here's what we think the model is thinking.
• 9:05
  And the output of, like, here's what we think the model is thinking might be the same in both of those cases, even though the features are different.
• 9:13
  And so the question is, one, is a sparse autoencoder, like, really understanding?
• 9:23
  Like, is it really discovering what these concepts are?
• 9:29
  And then my question after that was like, does it matter.
• 9:31
  Like, why does it matter if the.
• 9:33
  What it's essentially telling us is kind of the same?
• 9:39
  Yeah, I think that's.
• 9:41
  I think that's a really good question.
• 9:42
  And I think that's kind of my understanding of how they work as well is like, it's like.
• 9:49
  It's like mapping.
• 9:51
  And so it's like, to your point, mapping how these neurons are working.
• 9:59
  And so for this specific study, it seems like they figured out how to do it with less.
• 10:09
  Everyone's.
• 10:10
  All of these are like, less tokens.
• 10:12
  Less tokens.
• 10:13
  Do you feel like you walked away?
• 10:15
  Yeah, to your point, you said you feel like you walked away being like, okay, now what I did.
• 10:21
  But then I, like, asked.
• 10:23
  I, like, tried to answer that question for myself.
• 10:26
  Like, okay, why should we care about this?
• 10:28
  Yeah, why does this matter in the field?
• 10:30
  And my.
• 10:35
  Where I landed is like.
• 10:37
  Or what someone who works in machine interpretability would probably say is, like, it matters because the promise of mechanistic interpretability isn't like, give me, like, an approximate understanding what's happening.
• 10:54
  It's like, I want to know why the model is behaving the way it does.
• 10:59
  And so that's why the specifics of what we think we are understanding inside the model matters.
• 11:11
  There was a really good analogy that I.
• 11:16
  That as I was discussing this with ChatGPT that it gave me, that helped this click in my head that I'd love to share with you.
• 11:27
  Yeah, yeah, yeah, okay.
• 11:28
  That would be great.
• 11:28
  So one of.
• 11:30
  I guess one of the ways that, like, we use sparse auto encoders or, like, the data from it to try and, like, understand what's going on is, like, not just mapping things out, but then, like, based on our
• 11:41
  understanding of what we mapped.
• 11:42
  Trying to recreate the same thing to see, like, oh, is this.
• 11:47
  Were these the components that it actually was using?
• 11:49
  Can we create the same thing the machine did?
• 11:52
  So, like, it's kind of akin to, like, listening to a song and you might be like, okay, this song uses piano, drums, violin, and, like, a synth.
• 12:07
  And then you.
• 12:08
  You put in front of you pianos, drums, violin, and a synth, and you are able to recreate the exact same song.
• 12:15
  And you might think, oh, sick.
• 12:17
  I, like, actually was able to reverse engineer exactly how they made this song.
• 12:23
  And then someone else listens to the same song and they're like, okay, I actually hear, like, drums, a cello, an electric guitar, and, like, a cymbal.
• 12:36
  And they sit with those things in front of them and.
• 12:38
  And they're able to recreate the exact same Song, then you might be like, wait, wait.
• 12:45
  Okay, wait.
• 12:45
  But so what is.
• 12:47
  What was in it?
• 12:48
  Yeah.
• 12:49
  And like that's kind of the problem they're seeing here is like you can create the same song.
• 12:55
  Doesn't necessarily mean that you understood exactly what the instruments are inside.
• 12:59
  Yeah, yeah.
• 13:01
  Interesting.
• 13:01
  It's kind of like it's neat to your point.
• 13:03
  Like it's neat that we're watching these papers come out in real time and hopefully in like three weeks.
• 13:09
  But I don't know, like three weeks might be like we solved it.
• 13:12
  I know exactly how it worked, but it seems like there's a lot of people working on the same thing, which is how are these models working?
• 13:18
  You know, you're right.
• 13:20
  It actually is cool.
• 13:22
  It's like following a news story in real time in like a long term thread, which like I don't do with a lot of things.
• 13:29
  So it's kind of cool.
• 13:30
  Yeah, yeah, yeah.
• 13:31
  I'm really interested to see where we are in a month of this.
• 13:33
  If it's still like sparse auto encoders.
• 13:36
  More tokens, less tokens.
• 13:38
  Is this just going to become like a sparse auto encoder podcast?
• 13:43
  So neat.
• 13:44
  We need to have a sparse auto podcast on the.
• 13:47
  We're number one in the charts.
• 13:50
  Great.
• 13:51
  Perfect.
• 13:51
  Okay.
• 13:52
  Awesome.
• 13:52
  Okay.
• 13:52
  Interesting.
• 13:54
  Yeah.
• 13:54
  So now we have them working with less tokens.
• 13:56
  Let's keep moving.
• 13:57
  Let's keep moving.
• 13:59
  Okay.
• 14:00
  Rainbow say, great fun application of the day.
• 14:04
  Is this application called Rainbow say?
• 14:07
  That's a Mac menu bar app.
• 14:10
  I'm going to share, I'm going to share my screen to.
• 14:12
  Yeah, I want to see it because I haven't seen it.
• 14:15
  So I think this is so fucking cool.
• 14:20
  Basically what this does.
• 14:22
  I think my understanding, okay.
• 14:24
  Is like it helps you learn a language through like a TV show you're watching.
• 14:30
  So like let's say you're watching.
• 14:32
  I don't know, like, I don't know what's a tv, what's a show you love that is maybe in a different language or like you would watch Dark.
• 14:41
  Dark is like the best series, the.
• 14:45
  I think one of the best sci fi series of the last 20 years.
• 14:49
  And it's in German.
• 14:50
  Okay.
• 14:51
  Perfect example.
• 14:53
  So let's say you want to watch it in the original German on your computer.
• 14:57
  Yeah.
• 14:58
  What this will do is like help you learn German by watching it.
• 15:03
  So it'll like pause or it'll.
• 15:05
  You can pause it after something and it'll be like, here's what it said, here's what this means.
• 15:11
  Like, it'll like quiz you on stuff.
• 15:16
  Like, look at this, like here.
• 15:17
  Yeah.
• 15:18
  I just feel like I would learn language much better this way.
• 15:22
  Yeah, I've been really, you know, I've been really like facing this because I'm trying to learn Spanish and.
• 15:27
  You are.
• 15:28
  Yeah.
• 15:29
  And I've been taking lessons and feeling like I haven't been getting better.
• 15:33
  So maybe this is one good way to try it of now I need to find a Spanish show that someone recommends so I can get into it.
• 15:43
  Okay.
• 15:44
  Yeah, the DMS are open.
• 15:46
  Let us know what we should watch.
• 15:48
  Yeah, yeah.
• 15:49
  Maybe I'll pick a language to learn too.
• 15:51
  What?
• 15:51
  Okay, quick aside.
• 15:53
  Yes.
• 15:54
  Like, what made you decide to learn Spanish?
• 15:58
  So I want to retire to Mexico City.
• 16:03
  You're like, I want to leave America.
• 16:07
  And tacos.
• 16:08
  Okay.
• 16:09
  Yeah, it's really beautiful.
• 16:11
  Yeah.
• 16:11
  I really like traveling to South America and both South America and Mexico and different Spanish speaking countries.
• 16:19
  It's just beautiful.
• 16:20
  I love the warm and like trying different cuisines and different places.
• 16:27
  And my daughter is also learning Spanish at the same time.
• 16:30
  She learned English because she's six months old, so she doesn't know anything.
• 16:35
  Yeah.
• 16:35
  She's learning both and she's going to be like, oh my God, thank you.
• 16:41
  Like, how nice would it be to just grow up just bilingual.
• 16:46
  Like you just know a couple languages you didn't even have to try.
• 16:50
  Yeah.
• 16:50
  I'm afraid she's going to eclipse me.
• 16:52
  So this is a good tool.
• 16:53
  I'll try it.
• 16:54
  So basically you're competing with your six month old daughter and you want to move to Mexico City.
• 17:00
  I love this.
• 17:01
  And I would like to come visit you.
• 17:03
  You're welcome.
• 17:05
  Okay.
• 17:05
  Incredible.
• 17:06
  Okay.
• 17:08
  Okay, so for the one cool thing, we have to show the videos.
• 17:12
  Okay.
• 17:12
  From this weekend.
• 17:14
  Okay.
• 17:16
  Do you want to share them?
• 17:19
  The Nvidia ones?
• 17:20
  The Nvidia CEO is playing with Microdecks.
• 17:24
  Oh, from this, from on stage.
• 17:25
  Yes, yes.
• 17:27
  One cool thing while Becca brings it up is the Microdeck.
• 17:30
  It's an 11K star, of course, open source biped duck robot that walks.
• 17:37
  It's on Pollen robotics, microdoc on GitHub and it is very cool.
• 17:47
  I love open source hardware.
• 17:48
  I think there's like a whole new world coming for open source hardware.
• 17:52
  Oh, here we go.
• 17:54
  Look how cute they are.
• 17:55
  They're so cute.
• 17:57
  They're so cute.
• 17:59
  Can you see this?
• 18:00
  Is it full screen now?
• 18:02
  Yeah, it is.
• 18:02
  Oh my God, look at that.
• 18:03
  They're literally adorable.
• 18:05
  Look how cute.
• 18:06
  There's yoga.
• 18:06
  Why they fall over?
• 18:07
  Don't fall over.
• 18:09
  Are you okay.
• 18:10
  Are you okay?
• 18:13
  Oh, wait, Revive.
• 18:16
  Oh my God.
• 18:16
  I want a whole fleet of them.
• 18:18
  I'm obsessed.
• 18:20
  Okay, so they're so great and it's interesting and I imagine you can.
• 18:31
  Given all we've been learning about like how you can fit tiny models on hardware.
• 18:35
  I imagine people are throwing models on different models on these little guys.
• 18:42
  What would you do with yours?
• 18:44
  That's.
• 18:45
  That's the question, right?
• 18:46
  Like I think that.
• 18:48
  Oh my God.
• 18:48
  Sorry.
• 18:49
  I'm.
• 18:49
  Now I'm just.
• 18:49
  I need to turn the video off.
• 18:50
  They're so cute.
• 18:51
  They're doing summer songs.
• 18:54
  So this is the thing that I think is interesting about robotics right now especially so you have two different.
• 19:00
  I'll call them more accessible, like robots from under the Hugging Face umbrella.
• 19:06
  Now you have Microduck, which is made by a company called Pollen that was acquired by hugging face in 2025.
• 19:12
  And then you have this robot arm called Le Robot, a name I know you would like.
• 19:17
  It's like Lechonk.
• 19:19
  Lechon.
• 19:20
  Le Robot.
• 19:21
  And oh yeah, you.
• 19:22
  You built it?
• 19:23
  Yeah, I have.
• 19:23
  Yeah, yeah, yeah, yeah, I have the robot arm.
• 19:25
  Yeah, yeah, Those are pretty neat.
• 19:27
  So you basically can make a robotic arm and a cute little microduck robot that's just like two little legs walking around, little cute little face for like, you know, under 400 bucks.
• 19:37
  And which is so cool.
• 19:44
  And I still think it's.
• 19:46
  Both are in the stage of really cool fun for like nerdy hobbyist people like us to like play around with.
• 19:54
  Not super clear exactly what the real world usage is or like, or the technology hasn't caught up to it to make that obvious.
• 20:05
  For example, like even the ly robot arm, it's not like you build it and like, oh, it's just like picking shit up.
• 20:11
  Like it's still hard to like even train it and get it to like pick up a block and like put it in something.
• 20:19
  My real answer is like, honestly, I'd probably like.
• 20:25
  As for listeners of the pod, if you don't know, Sarah has a six month old daughter.
• 20:29
  I have a six month old daughter and I probably would like like she's getting, you know, she might be crawling soon.
• 20:35
  Hopefully soonish.
• 20:37
  I probably just like have it walk around.
• 20:38
  Have her like it be like a thing that she like crawls after and like engages with her and it's like a fun thing.
• 20:45
  I don't know.
• 20:45
  That's a really fun thing.
• 20:47
  Yeah, yeah, yeah, I think, I think you make a really good point.
• 20:49
  Like where we are now versus like iRobot.
• 20:52
  You know like I think of like the iRobot world where we have like these androids that show up at our house and creep us out and probably don't kill us or we hope that they don't.
• 21:01
  Is, is good.
• 21:03
  But looking for my copy of I of the iRobot that you sent me.
• 21:08
  Oh yeah, that's somewhere over there.
• 21:10
  Yeah.
• 21:10
  Great.
• 21:11
  Yeah, like we're, we're not there and I don't know.
• 21:15
  I mean I think the thing that I think about the most when it comes to this stuff, which is very different purpose than the microdoc is like people that are differently abled like having the help that they need
• 21:28
  just because like I, it's so like the, in the U.S.
• 21:37
  i think the price of care is so high.
• 21:40
  Yeah.
• 21:40
  That like only people that have the means to have full time help can get that.
• 21:46
  And we don't have a culture of like we don't have a culture of family help.
• 21:50
  So like something like this is not going to help because it's tiny.
• 21:54
  It's not going to do much.
• 21:55
  But I feel like that could be like a first application of where we go with the robot armor and the microduct.
• 22:01
  But besides that, I really don't.
• 22:04
  I think so.
• 22:05
  A friend of mine had this robot come to their house this weekend to try to do a bunch of tasks.
• 22:12
  Cool.
• 22:13
  What kind of robot?
• 22:15
  No idea.
• 22:15
  I just saw this robot that was falling all over their house.
• 22:23
  And so I like we're not there and I, I think like it's one of those things like for me personally I don't like, I think about the like, you know, the robot massage things where you're like, I don't
• 22:33
  want a robot massage.
• 22:34
  Like there's something about it.
• 22:36
  Nice.
• 22:37
  It's.
• 22:37
  It's so funny you're, you're talking about that because this morning I was doing, I was, I've been trying to write more as like my, as like a thinking tool and I was going through the thought experiment of like
• 22:52
  what are the second order effects of like intelligence becoming more accessible and abundant?
• 23:00
  Like what are the, from a, from a labor market jobs perspective?
• 23:06
  Like what are the things that eventually are going to be likely machine.
• 23:14
  Machine dominated and what are the things that are probably going to remain human?
• 23:18
  And a lot of the stuff that I was putting in this category of either remaining human or like becoming.
• 23:32
  Like maybe becoming like a luxury good or like a premium because of the human component is some of the stuff you're talking about.
• 23:39
  Like I was thinking about like spas and like Massage stuff.
• 23:42
  Like, I don't.
• 23:43
  I was like, do I, do I care that it's a.
• 23:45
  Like, do I want a person to be massaging me?
• 23:47
  Do I, do I care if it's a robot?
• 23:49
  I was thinking about childcare and elder care.
• 23:51
  Yeah.
• 23:51
  I was like, that feels like something that I want to be human.
• 23:56
  Like.
• 23:56
  Yeah.
• 23:58
  But at the same time, like, accessibility of those things is really important.
• 24:02
  Like, imagine we, we don't live in a country where we have free child care, which sucks.
• 24:11
  Childcare is expensive.
• 24:15
  Like, imagine the economic class of people that it could unlock if we were able to somehow.
• 24:23
  I'm not saying like we have robots watching kids, but like, if the, if there was some impact of using technology to help expand our ability to provide things like that.
• 24:34
  Yeah.
• 24:35
  And I think also too, like, when we think about that stuff, I think like, there's parts, like there are parts of it that I feel like you could.
• 24:41
  Like when I think about like what it's done to programming so far, it's not that, like, I see, like there's no more programming work.
• 24:49
  It's now that we have instead of ten PRs to review, we now have a thousand PRs to review.
• 24:54
  Right.
• 24:54
  Right.
• 24:55
  So like, what is that?
• 24:56
  In like another world.
• 24:58
  And I can see like, for example, with child care, like, yes.
• 25:01
  I don't want, like, we don't.
• 25:04
  I don't envision this world where it's just like robots with kids.
• 25:06
  But like, do we maybe like teachers or caregivers don't have to focus on making lunch because like lunch is made.
• 25:14
  You know what I mean?
• 25:14
  Like, can actually focus on those interactions that are more important.
• 25:18
  Like the learning, the one on one.
• 25:19
  It doesn't have to sit there like cutting up carrots.
• 25:22
  That's exactly.
• 25:23
  It's like it's intelligence becoming abundant makes human, like, frees up human time in certain areas.
• 25:34
  And like, what can we then capture?
• 25:37
  Like, what's the value capture of that is, I think, exactly what you're talking about.
• 25:41
  I like that.
• 25:42
  Yeah.
• 25:44
  Or maybe we like have the coolest robot nanny in the future.
• 25:48
  Who knows?
• 25:49
  Do you remember that?
• 25:50
  Do you remember that, that Robin Williams movie called like Bionic man or something like that?
• 25:57
  Nope.
• 26:00
  No, rewatch that.
• 26:02
  Anyway, that's an aside.
• 26:04
  Okay, next time on the pod we review Bionic Man.
• 26:11
  Okay.
• 26:11
  Any closing thoughts?
• 26:14
  No.
• 26:15
  Besides, I want a microduct.
• 26:17
  Same in conclusion, we want in the.
• 26:20
  Living room just doing somersault all day, just somersaulting.
• 26:22
  And then we can get our microducts together.
• 26:24
  And then I want to walk around and they all follow me.
• 26:29
  That's what I want to do with it.
• 26:31
  I want a fleet of them that follow me around the house.
• 26:35
  Just summer partners will love.
• 26:36
  They'll love it.
• 26:39
  And when they see my wife, they just quack at her and follow me around.
• 26:48
  Amazing.
• 26:49
  All right, bye, everyone.