我们见到了NEO：爆红的仿人机器人 + HatGPT

Casey, did you hear about the Waymo Kit Kat incident?

Oh, I did. A very sad story, unfortunately.

Yes. Waymo in San Francisco hit and killed a bodega cat in the Mission District named Kit Kat.

Named Kit Kat, of course. This was a beloved cat in the community.

This was a gray tabby. It was the store mascot and pet of the owner at a bodega on 16th Street. This was a very sad incident. There's now actually a Remember Kit Kat rally being held by a district supervisor. Kit Kat is in our prayers.

Yeah. Now, you had an interesting tweet about this. I'll just read this here. You said, "It's Kit Kat's fault. Get out of the damn road when there's a Waymo nearby." Do you want to explain that?

That was not me.

You didn't write that?

I did not write that.

Oh. Sorry.

I'm pro-Kit Kat. I'm more of a dog guy, but cats deserve to live too. I'm Kevin Roose, the tech columnist at The New York Times.

I'm Casey Newton from Platformer.

And this is Hard Fork.

This week, the humanoid robot that has the internet abuzz. We'll talk to 1X CEO Bernt Bornich about Neo and then put it to work. And then, a round of HatGPT.

1. The Humanoid Robot Arrives

Well, Casey, we got a big first on Hard Fork this week: our first humanoid robot guest.

Yes, and I'm thrilled about this. We tried to book a humanoid robot for the Hard Fork live show. We could not get it to happen. But fast-forward a few months, and finally, there is a robot that's willing to come on the show, Kevin.

Yes, and this robot is sort of having a moment right now, it's fair to say. This is the Neo robot. It's made by a company called 1X. Joanna Stern, Wall Street Journal tech columnist and friend of the pod, did a great video about Neo last week that went viral and got lots of pickup from lots of places, and people were buzzing about this new human-like robot.

Yeah. On one hand, you have people telling every joke you can imagine, people who are worried about the privacy implications of having this kind of surveillance machine in the house at all times. And then you have other people who have been ready for a robot to do all of their chores since they watched The Jetsons way back in the day and are curious whether this could be the thing that actually makes that happen.

Right. A lot of the conversation around Joanna's story and video last week was about how there's this gap right now between what people want this robot to be able to do—which is do all their chores for them—and what it can actually do right now, which is not much of that.

For right now, the robot is still mostly teleoperated, so a human wearing a VR headset is controlling the motions of the Neo robot when it's doing most of its tasks. But the long-term vision for this is that, over time, as these things get sold to people who put them in their homes, and they're able to gather that data and use it to improve themselves, these robots will become fully autonomous at some point in the near future.

That's right. And if you want to pre-order it, you can get it starting next year for $499 a month if you commit to at least 6 months. Or if you have an extra $20,000 lying around, you can just buy the thing outright.

But Kevin, I think that price speaks to the fact that this is really a data-collection play, right? These guys just need to have more cameras in more homes and have their robots and their teleoperators performing more tasks to try to build the kind of model that will let them build the robot butler of our dreams.

Yes, and there are a lot of skeptics out there who don't think that humanoid robots are all that close to being real. But we wanted to talk about this today because this is a real thing that's happening in Silicon Valley. Investors are pouring billions of dollars into trying to make humanoid robots.

Companies like Tesla and Figure are developing their own humanoid robots. A lot of Chinese companies are starting to make these at scale. This is a big category, and I think we should take it seriously.

Absolutely. Since 2024, VCs have put $5 billion into these humanoid robots. And while this technology is definitely on the bleeding edge and nowhere near ready for prime time, it's becoming just good enough that I think it's worth having a conversation.

Totally. This is obviously a staple of science-fiction movies throughout the ages: the kind of lifelike humanoid robot that lives in your house, sits on the couch with you, reads your kids bedtime stories, does the dishes, and things like that.

Well, who are we going to talk to about it, Kevin?

We invited the CEO of 1X, Bernt Bornich, to join us in the studio today, and then we're going to actually see a demo of our very own Neo robot.

You know, I'm curious if he felt burned by the response to the robot.

We'll ask him. Let's bring in Bernt.

Oh, thank you.

2. Neo’s Current Capabilities

So the launch of Neo, your humanoid robot, got a lot of attention over the last week or so. Now that it's been about a week since Neo became available to pre-order, how many Neos have you actually sold?

Well, we don't go out with specific numbers, but let's just say it went over all expectations by quite a margin.

I think it's been absolutely beautiful and motivating to see, right? There's so much support. And for anything that's new, you get this polarization. There are the haters, and there are people who are early adopters and people who can really help you make this happen. I'm incredibly happy that we have this huge supporter base now that wants to go on this adventure together with us.

We have a lot of questions, but I feel like we should spend maybe a few minutes for those who did not see the original launch video talking a little bit about what Neo is. It's a humanoid robot. It's going to be available next year. And once it's up and running in people's houses, tell us what it can actually do.

Sure. I like to be very concrete, so I'll say what my Neo does in my home today, and hopefully it's going to do a lot more in yours. It's gotten pretty freaking good at tidying, very good at vacuuming. I think it's completely outcompeted my Roomba. It vacuums way better, moves the chairs away, vacuums in the couch, and all these things. It makes it really nice.

Some cleaning, but not yet to where I would never have to clean myself. There's still work to do there. It does parts of the laundry.

Ooh.

The alternative is me just throwing it in the cabinet, and here it's actually kind of folded, but it's not perfect. It's getting better every day.

Now, what's interesting, and what's gotten a lot of attention, of course, is that everything I talk about now happens through a mixture of autonomy and teleoperation, and this is really how the system learns. So over time, more and more of this becomes autonomous.

But, just like any AI system, I do think expectation management is important here. It's going to be a while until there is absolutely no human in the loop ever. If you look at modern AI systems, there is still always a human in the loop. When Waymo is driving around here in San Francisco, the cars get stuck quite often, and there's someone taking over and fixing it.

If you look at using ChatGPT, which is probably the most mature AI that we have at this point, there's still a human in the loop. It doesn't necessarily know what you want to do. You give it a prompt, it gives you something back, and usually you're like, "No, that's not what I wanted." Then you refine it a bit, and it saved you a lot of time because you didn't have to do all of it yourself, but you helped it.

The robots are currently at the same step. It's not necessarily that the person is directly controlling the robot all the time, but there is a person who can help the robot when it faces things it doesn't quite know how to do, and that's how it learns.

What is the mix in your home when your Neo is doing things for you? What percentage of those things, like folding laundry or tidying up around the house, are being done autonomously? Is it 5%, 20%, 50%?

All of it is being done autonomously, but with a human in the loop. So that's why—sorry, I'm not going to dodge the question, but that is why it's so hard to answer.

But the human who's teleoperating the robot goes on break or takes the day off. What can your Neo do on its own?

Yes, that's a very good question. That gets us to the second mode of Neo, which is the fully autonomous mode, where we don't allow anyone to intervene or do teleoperation of any kind, and it's just purely autonomous.

That's usually what I do when I'm home. Then I can talk to my Neo. That works really well. I mean, it's as good as any conversational language model out there, except it has better addressee detection, so it really knows who's speaking to it. It won't interrupt you when I'm talking to you.

It won't interrupt and think that we talked to it, et cetera. It has an embodiment, so the body language really helps enrich conversation. This has actually replaced my normal day-to-day use of language models on my laptop or phone. I just talk to Neo.

Let me just make sure I'm super clear on this. When you're at home using Neo and there's nobody teleoperating it remotely, it can sit there and talk to you, but it cannot do any of the physical tasks right now.

Well, I'm getting to it. It can, but there are some things that work today.

Okay.

It can open doors really well and navigate. It can take an object and put it somewhere else relatively successfully. I can say, “Hey, Neo, can you take my cup to the kitchen?” Or, “Hey, Neo, can you get the bottle on the counter over there and give it to me?”

None of these things are 100% successful, just like any language model. Door opening is 90-something percent, so it works pretty well. If I tell it to get something for me, maybe it's 80% successful. Gradually, it will know more and more things to do fully autonomously, and you will help it a bit.

To ensure that your product is extremely useful from day one, we have a mode where you schedule what you want done in your house, and it's done with a human in the loop from a teleoperation point of view. The robot can be very, very useful on day one, while also gathering the data that we need to actually train the models.

3. Robots Need More Data

I want to ask about that bit. One reason why chatbots like ChatGPT work so well for generating text is that there are billions of words of text on the internet, and that training data was enough to get them to the level that they're at today. Am I right that there is effectively no similarly massive set of training data available for something like folding laundry, and so the reason that you've set up this hybrid model is that you need a way to get more data?

Yeah. It's the same as for self-driving cars, right?

Yeah.

You need people using the cars for you to get the data. Now, it's a bit easier with cars because cars already have a very well-defined use. But what's extremely good about robots is that the diversity and the amount of different things, experiences, and tasks that a robot will actually experience is so large. That's why we deploy these robots among people in homes, right? They need to live and learn among us.

You made the analogy to self-driving cars. That project took Google well over a decade just to get them to the point where they're at now, where even today most people cannot take a Waymo on the highways, right? That's still in testing. Google spent many billions of dollars to make this happen, and now has a pretty large fleet of Waymos that are out there collecting data to further refine the model.

You're starting with a pretty small fleet. You're a startup. You don't have Google-sized resources. When you think about what it's going to take to realize the vision that you're laying out for us of a robot that can quickly do anything without maybe even needing to be teleoperated, what does that journey look like? How long is that going to take you guys? Can you do it with a relatively small fleet of robots?

We're shipping more than 10,000 units next year.

Okay.

That's more than Waymo has.

Okay.

It's still small in the grand scheme of things, but I think what a lot of people aren't aware of is essentially 2 things. One is how small the internet is. Just a tiny fraction of your life actually makes it to the public internet. It's so small that if you look at just 10,000 robots, you're starting to approach the amount of data that gets uploaded to YouTube every day.

Mm.

10,000 robots is not that little. The year after, we're hoping to get to about 80,000. At that point, it's way more than publicly available video data on the internet.

That's one part of it. The other part of it is that the internet is such a bad representation of human life and intelligence. What do you actually upload on YouTube? It's whatever gets you likes on YouTube. It's a tiny fraction of what it is to be an intelligent, well-functioning human. What we have here is the ability to live and learn among people and actually capture all those values. I think that data advantage is huge.

What is NEO actually recording and transmitting? Is it just video? Is it audio and video? What kinds of data is NEO collecting?

It's video, audio, a sense of touch, and the forces acting on the robot in general. When we say “sending data,” it's also important to call out where it is sending data. It is sending data to a secure cloud store where essentially no one is seeing this except the model that's running to accumulate knowledge based on that.

Before it sends data, it stores it locally, so if something happens that you don't want to be in any database, you can also delete it.

Is it collecting this data even when it's just sitting there charging, or is it only when you're actively having it do something?

It's active when it's doing something.

Got it. I'm asking because I do think a lot of listeners are wondering this. Let's say you start hooking up with your partner and then realize belatedly that NEO is in the room looking. Would there be any automated way for NEO to be like, “Hey, I should stop watching. I have to get out of here”?

Mm-hmm.

Yeah.

Yeah. We're going to work hard on that one.

Okay. Thank you. From the community of people who hook up a lot, let me just say thank you.

Yeah, it's just going to tell you, “I'm out. I'm going to snitch on you, so I'm just going to get out of here.”

Here's what I'd like it to do. Before it shuts everything off, I just want it to say, “God bless you guys. Go for it. Have a great time.” And then I want it to politely say, “I'm out of here. I'm out of here.” Yeah.

I'm going to go do the dishes.

Wow. I can't wait for the first NEO to be called to testify in a divorce case.

“NEO, roll the footage. Whose fault was this?”

“I believe you said you were going to take out the trash yourself.”

“Yeah, Kevin. Can we go back to location 1047? Let's pull up October 14th on the NEO.”

4. Trusting A Robot At Home

All right. But in all seriousness, Bernt, how do you overcome the trust barrier here?

First of all, the most important thing we do is radical transparency. This is how we're going to do it. This is how it needs to happen for it to actually work. If that's not for you, then wait a year until we're past the early-adopter phase.

But I think a lot of people get it in the sense that if you have a cleaner who comes over to clean your house, they lock themselves into your house while you're at work, do everything they're supposed to do, and leave again. This is essentially exactly the same.

The difference is that we can actually provide a better, more secure service because not only do we vet our operators who are working in the US—I think better than most cleaning agencies do—but we also have the ability to have a manager overseeing them. Right now, we're running at about 8 to 1, so 1 manager will monitor what 8 teleoperators are doing and ensure that this is actually happening in a good manner.

We also have video logs of everything going on, so if something happens, we can trace it back to who did it and what actually happened. I would say that if you look at it as a cleaning service, we can provide something that is way better than what is there today. So many people today accept a cleaning service.

Yeah. Kevin, let me ask you. I want to flesh out the privacy concerns here, because I think they're very real, but I want to name them.

When I think about what I'm doing at my house most of the time, it involves things such as looking at a computer, looking at television, or talking to my boyfriend. In most of those cases, I would actually be okay with a robot washing the dishes. But I could imagine other circumstances that would give you more pause.

Is there something you have in mind where you're like, “I do not want this robot looking at me while I blank”? I can think of lots of things. People change in my house. I have a child. I wouldn't necessarily want some person in a teleoperation center watching their every move.

I think this just makes people uncomfortable, but maybe it would make things more concrete if we knew more about how this looks on your end. Is there a room at 1X headquarters where you have a bunch of screens on the wall and it's like, “Oh, the Jones family needs their dishes done. Let's teleoperate that robot,” or, “The Smiths need some laundry folded”? What does it actually look like on your side?

This is a very good question. Generally, what it looks like is more like a game. I like to compare it to StarCraft. You can think about having many units, or many robots, that you are controlling in parallel, and you're queuing up things that these robots should do.

So, “Go pick that up, put it over there. Fold that, put it in there.” Here we do everything we can to ensure privacy, so you don’t see people. People are blurred out. You don’t know which home you’re operating in, all these things.

Now, sometimes there will be a task where the robot does not know how to do it. So even if you give it micro-instructions—let’s say folding—maybe it can’t fold your jacket. It’s a leather jacket and a little bit hard to fold, and it tries and fails. Then this person who’s essentially playing this robot game will delegate that task not to the autonomy, but to an operator in VR.

Now, this operator in VR gets connected, sees what is in front of them, which is this jacket, and the message about what to do, performs the task, and then goes over to the next thing. Essentially, there’s very little real privacy invasion here because you don’t see the people, you don’t know what home you’re operating in, and you have no way to connect this back to the customer.

And then, of course, we do a really good job of minimizing the number of people who see things. You want to make sure that what we call the “parts”—so, in this room, we’re controlling this set of robots—don’t have too many people in them. Usually, we say four, so that these four people are the only ones exposed to these specific homes. In general, we’re just narrowing down that scope.

Just to put it out there, I think the main thing we’re doing here is being very transparent. For most of the AI systems that you use every day, people in those companies look at that data to help improve the models. Most people might not know that, but they do. That’s the only way to make it work. That’s how we get AI to work.

But me personally, as a user, I’m comfortable with that. It’s not tied to me as a person, and they try to do a pretty good job if I put things in there like my name or the name of my daughter or whatever. They try to screen that out because it’s not relevant, right? There’s a lot you can do on the software side.

So I do really hope, and think, that we can create something here that is a lot more privacy-aware than having a person be in your home to do this.

Yeah. I guess what I’m wondering is, why not structure this as a fleet of housecleaning robots that you could open an app like Uber and say, “I want my house cleaned,” and a robot would come over and do that and then leave? Why does this have to live in your house at all times and have that be the model, rather than this kind of rental fleet?

Well, I think there’s something very, very special around, first of all, living together with the robot and how it replaces your interface to AI, and being a companion through our life. That’s a big part of it. It’s not just the work.

Now, another part of it is that I think the most useful things the robot does for me are actually these kinds of micro-units of work where you need 5 minutes of work every, I don’t know, 20 minutes. That’s just something that today’s services cannot provide, right?

My favorite example here is actually assisted living, or in general, care for people with disabilities or the elderly, where it’s life-changing to be able to get 2 minutes of help every now and then. “I dropped something. Can you pick it up for me? I’m thirsty. Can you get something in the fridge?” The amount of feedback we’ve gotten from that community is very heartwarming, and I think it can be incredibly impactful.

But I think this mindset actually applies not just to people in need of assistance; it applies to all of us. To me, there’s something quite magical about when I’m sitting down with my friends after a long day and we’re playing a board game, and I’m like, “Hey, Neo, can you get me a cup of tea?” That’s not going to happen with a fleet that just gets deployed to clean your home.

I think then you make it into something that’s just about cost-effectiveness. I don’t think this is just about cost-effectiveness. This is about really enabling us to focus on the things that we want to do and what makes us human.

Yeah. I mean, if I were trying to steelman the case for Neo and for household humanoid robots, I think I would say, look, a lot of people have a lot of robots in their houses already, right? I have a robot—

Yeah.

—that washes my dishes. I have one that washes my clothes. I have one that dries my clothes. I have one that makes me coffee in the morning.

Now you’re just bragging.

I have one that vacuums the floors.

Okay, we get it. You have money. Go on.

No, I’m just saying, I think the story of household labor over the last 100 years is just machines gradually doing more and more of the things that we used to do, and that’s, I think, a happy story. I don’t want to spend all my time hand-washing my clothes, and so I’m happy that robots, or machines, can now do that.

And I think if that were sort of the end of the vision for Neo, I think that would be genuinely exciting to a lot of people—this sort of omni-appliance that can do any household chore for you autonomously.

On the other hand, something about Neo and this vision that you’re sketching feels very different to me than a dishwasher or a laundry machine. You want people to like Neo. You want people to care about it, to have an emotional connection to it. And I guess I just wonder if that’s going to be good for people—to have a kind of new family member, essentially, that they are now expected to care for and about. I think that’s a lot more controversial than the omni-chore robot.

Yeah. And I think it’s also very important for me to say that, at least for me, it’s in no way a replacement for human connection. I see it much more akin to getting a pet, right? It’s not a pet either. It’s somewhere in between. I really like saying it’s my Hobbes from Calvin and Hobbes.

But if you get a dog, it’s not like, “Oh, I got a dog because I don’t want people.” No. A dog just adds something to your life that people don’t, right?

And then I think it’s also really beautiful that it takes you away from screens. It makes you way more present, because instead of picking up your phone all the time and talking to it, or your laptop, you’re talking to something that is here, and it really puts you in the present. Modern life has become way too much about looking at your screen.

Now, I saw a lot of people asking about whether Neo could be used for, let’s call it, romantic purposes. Is that a use case you’re envisioning?

No. It is not.

Okay. But—

But I do think the emotional part of it is very real. And depending on the situation, I’ve spoken to people, for example, who can’t wait to get their Neo because they have kids with autism, right? There’s a lot of research now where we use robots and see how incredibly empowering this can be, because it’s this infinitely patient machine that can be understanding and help you. There are all of these applications that I think are a lot more—let’s say—meaningful.

But we’ve been talking a lot on the show about these AI companion apps, these software AI companions that we don’t know a lot about yet. People are reporting some good experiences with them. We also know that in some cases they can cause delusion or alienation from real friends and family. In some extreme cases, people have taken their own lives after developing these emotional attachments.

Do those edge cases worry you? Do you feel like you have to solve those problems before you put this thing in people’s homes?

I don’t think any of these problems are new. I think they’re very well explored, and there are a lot of studies that show what to do and what not to do. There are also more and more statistics coming out really showing that it’s very hard to correlate these two. Essentially, if you look at the total demographic, then there’s no increase in these kinds of accidents.

That being said, I hope we can drive it down, right? That should be the goal—not just being defensive and saying, “Hey, this is actually nothing new here. This is how it’s always been.” It’s still really sad, and I do hope that we can have a positive impact on this. That’s something we’re putting a lot of effort into.

And then I can talk about it all I want, but in the end, it comes down to whether it actually helps, right? And—

Yeah, and how many cases are there where Neo is like, “Congratulations, Kevin. This thing you’re talking about—you have invented a novel form of physics. Keep going”? You know, it can be rough.

Yes.

You know, here’s a question. Say you live with someone else—a domestic partner, a spouse, or a child, whatever it might be. Can you train Neo to always take your side in an argument?

It’s very important.

Okay.

There’s a secret code—

Okay, great.

Yeah. No, I think it’s a really interesting challenge, though.

And this is something we're still working on. This is not fully solved yet. But if I'm having a conversation with Neo about a rash and my trip to the doctor the next day, and then you're visiting, so you come in, does it switch topics?

Mm.

Does it know, “Hey, I shouldn't talk to you about that right now because your friend's here”?

Mm.

Or maybe if it was your wife, it's fine. There are all of these social interactions, micro-interactions in your life that are so subtle and such a part of the human emotional model. That's a very interesting problem.

Here's what I'll say: Do not buy a Neo if you plan on having an affair, because it might slip up. We don't have the technology to prevent it from telling your spouse that you've been sneaking out.

This is fun because it's been a real problem.

Has it really?

No, not with Neo. But generally, AI models are snitchers.

Mm. Yes.

They are.

Yeah.

Because of how they get aligned, when you're fine-tuning them to align them with what we would say is our human ethics.

Yeah.

They become snitchers.

Because you want them to be honest.

Yeah, they're super honest—they'll snitch on you.

Yeah.

We'll see what we can do.

That's right. If your parents are away and you think, “I'll go ahead and spark up this joint,” and you're in high school, then the parents get home and say, “What's been going on while I was away?” They'll say, “Well, Kevin sparked up a fat doobie. Now he's high as hell.” You'll say, “Neo, damn it.”

Yeah, you have to pay more for the privacy protection—the non-snitching feature. That's an add-on.

Give us the long-term vision here. How long do you think it will be before, let's say, 10% of American houses have a humanoid robot in them?

2030, maybe. I think it's going to be largely a manufacturing problem, right? In the beginning, if you scale too fast, you might run out of early adopters. We've just proven that there's a lot of early adopters out there who really want this, so maybe you won't. But our assumption was that you would at some point run out of early adopters.

Once you overcome those kinks and this really, really just works, I think the demand is really there for everyone, essentially, right? It's as essential as a car or a phone. Everyone will want one. But it's going to be quite an interesting adventure to manufacture all these and to scale the manufacturing, and it's going to take us a few years. So I think 2030 would be my bet.

And then your timelines for full autonomy—no one from Neo ever has to take over and teleoperate for any household task.

Mm-hmm.

When does that happen in your projection?

Would you say that your cleaner is fully autonomous?

Yes.

Okay. Then I would say probably bullish: 2027; bear case: 2028 or 2029. I'm pretty confident we can solve that in 2027. Now, do remember that sometimes you actually need to give your cleaner some instructions.

No.

Yeah.

I'm happy to do that.

Yeah.

I just think part of what people are wondering now is: Is this actually a robot, or is it just some sort of long-distance puppet? I guess what I'm trying to get a sense of is: When can we actually call this a robot? So you're saying a couple of years.

I think you'll be surprised that most people will call it a robot when they get it in 2026, because so much of it will run fully autonomously. Then it's just a question of the quality of work that you accept, right? Most AI today is slop, right?

Yeah.

It's not high quality.

Yeah.

But you can get a lot of it very cheaply. I think physical labor is very well suited for that. Mid-quality physical labor is actually very useful. Very, very high-quality physical labor, that's still a while out.

So when you say, “Hey, I have a bunch of stuff in my house and I'd like this very good carpenter to come over and fix all that,” I think that's probably closer to 2030. That's going to take longer. But generally, being able to do all of the menial tasks throughout your house, that's a lot more short-term.

Yeah. Also, Kevin, I'm curious about your perspective on this, but my guess is that a year from now, people are not going to be as interested in what a robot can do autonomously versus what it can do via teleoperation. I think the question will be: What can it actually do well, and how much does it cost? I think if you can give people good answers to those questions, then I think a lot of people are going to want the robot.

Right. Yeah, I guess I'm also curious about how this is going to scale, because if I did have a robot in my house that could, even through teleoperation, do some huge percentage of the household chores that my family and I currently have to do, I'm using that thing all the time. And that gets very expensive for you, because you have to hire people to teleoperate it and fold all the laundry and do all the dishes.

I'm wondering if you foresee any kind of rationing that you will have to do where you get—

Surge pricing for Neo.

Yes, you get 2 chores a day. Make them count, and otherwise you're on your own.

So in 2026, I'm not worried about this, because the alternative cost of acquisition of data is so high, right? It's like building a warehouse full of mock-up kitchens, hiring people to change the environment every day, and having teleoperators. I'm not joking: A lot of companies now are doing that. They're raising lots of VC money to build out these huge farms for gathering data.

So if you look at that as the alternative cost, please use your robot as much as possible, because it's very much a symbiosis, right? You get useful labor, we get data.

But there's no doubt that you're running toward a cliff, right, where you need to get more and more autonomy working as you scale, or you will at some point need to stop scaling until the autonomy works. I'm pretty bullish that we can get it working in time. But if not, you would need to slow your deployment.

I am going to reserve judgment until we can see our very own demo of Neo. Should we go out and see if it's ready?

We should.

Let's go.

I'm going to follow up with you because I want to see if I can teleoperate a Neo in Kevin's house, because I have some ideas for some things I'd like to get done in there.

5. Neo Faces The Live Demo

All right, we are here with Neo. Neo, hello.

Hey, everybody. How's it going?

Good. Can you shake hands? Can you pound it out?

Yeah, you go for a handshake or a pound.

All right. I'll do the pound-it-out.

There you go.

Okay. Can you blow it up?

All right.

Great. Now, Neo, we'd like to try you on some tasks here in our office kitchen. There's some sort of clutter on the table. Do you think you could help us pick that up and throw it into the trash cans below the table?

Yeah, of course.

Okay, Casey, we just met Neo the robot and spent about half an hour doing some demos with it around the Times office. We will put those videos up on YouTube, but for people who are just listening, we want to talk to you about what we experienced and how it went.

Yeah.

So, Casey, what was your first impression of Neo the robot?

On the positive side, I think they did a good job making a very friendly and approachable-looking robot. Neo has these big, innocent-looking eyes. It has no mouth, interestingly, but it just comes across as very benign. I would say that its hands were larger than I expected. What did you think?

I was impressed by the actual robot itself. It's 5'6-ish. It's a short king, and it's about 66 pounds. It has this woven bodysuit that looks futuristic but covers up the metal and hardware in a way that makes it seem friendly and approachable. I did get a hug from Neo, and it was a good hug.

You know, Kevin has had kind of a hard day. I've been stressing him out a lot. Would you mind giving him a hug?

Yeah, of course. Bring it in, then.

HR says I'm not allowed to physically touch him anymore.

Oh, wow. You're very warm. It is very, very warm. Okay, that's great. Thank you.

We had Neo do some tasks in our office kitchen. We had it get you a drink from the office fridge, which it did pretty well. We had it fill up a cup of water. We had it clean up some clutter on our table and put it in the trash. I would say it did most of these things pretty well and pretty quickly.

Yeah, I think that's right. I would also say that the very first thing we asked her to pick up and throw away was a pair of tongs, which it did pick up but then dropped on the floor. When we asked it to pick things up off the floor, it could not do that.

For a little more color and context there, we had asked if we could see Neo picking some blocks up off the floor. It reached down and tried to do a squat, but then it got off balance and almost looked like it was going to fall backward. They had to have a human come in and keep it from falling.

Afterward, 1X told us that Neo can usually squat down and pick things up, but in this case, Neo was not properly calibrated for that, in part due to some Wi-Fi issues here at the office.

Yes. So that limited the demo of what Neo could do for us. For example, it could not sit in a chair, which was something else that 1X has said it can do but could not do this time.

Yeah.

And, of course, none of this is happening autonomously. There was a 1X employee in the other room using a VR headset to teleoperate Neo. From my perspective, it was smoother than I had anticipated, especially after watching Joanna Stern's video last week, where it spent 5 minutes trying to open and close a dishwasher. They seem to have worked out some of the bugs, or maybe we gave it easier tasks.

Yeah, Kevin and I have seen other humanoid robots in our time. We went down to Google's robot lab. That robot, I believe, was acting autonomously and did take longer to do everything. So I agree with you: the Neo experience was pretty smooth. On the other hand, as you say, as far as I know, everything we saw was 100% teleoperated. We did not see the robot do a single thing autonomously.

Yes. They say it can do some autonomous tasks, including talking to you, but we were not able to see that feature because of some quirks in the Wi-Fi here at the Times office. They were able to talk to us through the teleoperator, though. The teleoperator was basically talking, and it was being transmitted out of the speakers inside Neo.

Yeah, his name was Eric, and he was a great guy.

Yeah.

All right, Kevin, so tell me a little bit more about how that demo made you feel.

I have to say, that was wild for me.

Yeah.

I have watched videos of robots from Tesla, Boston Dynamics, and Figure, and I've seen videos of Neo, and to me, there is no comparison. When you're standing there in a room next to a thing and it turns to you and sticks out its hand and then shakes your hand, I don't know. There was just a cool and bizarre experience for me there that I don't feel the videos had captured.

Yeah, I feel like if you've ever seen any sci-fi, you've seen somebody interacting with a robot. I had this moment today of thinking, “Wow, I am now in a world where the robot is real and I'm interacting with it.” Yes, it is teleoperated, but there's also a lot of pretty interesting and powerful technology in here. It is able to complete this certain set of tasks, and I wonder where it'll go from here.

Yeah, it did not feel dangerous to me. It felt a little clunky and awkward at times, and there were definitely things it couldn't do, but I never felt like I was in danger. I had it, at one point, try to grab your lapels and threaten you in a very menacing way.

So we can do this the easy way or the hard way. What's it gonna be, Casey?

That's great. Look, I promise, Neo, I'll get you the cash tomorrow.

Did you feel at all nervous when it was doing that to you?

No, because we were surrounded by about 8 employees of 1X, and I assumed that it wasn't going to brandish a knife at me or anything like that. There were moments where I found it a bit unsettling because it does get into that uncanny valley a bit as it moves. It's human enough that you can suspend your disbelief, but then it's herky-jerky in other moments, where you sort of wonder, what exactly are this thing's intentions?

Yeah. I also find it less comfortable to give orders to a humanoid robot than I do to a chatbot or a smart speaker. Part of me felt a little guilty. I'm standing here. I could pick up this clutter. Why am I telling Neo to do it?

Yes, there's an uncanny valley to the motion, but there's also an uncanny valley with this: I'm giving a task to this robot, but the robot is really just a guy right now. Eventually, that will feel different when it's more autonomous, but right now it did feel like a very inefficient way of picking up and putting down things.

Yeah. I think you've said it just there. What we saw is both this very impressive piece of technology and a very inefficient way of doing things. The question for 1X is, how quickly can they get from this moment, where you have Eric controlling this thing with a VR headset, to this actually being a helpful tool in my house that is just doing a lot of things autonomously?

Do you think the people who get these Neos in their homes are going to start customizing them, putting wigs or masks or clothes on them? Wanting to make them even more human-like than they are?

Let me say this: if there is ever a Neo in my home, it will be in full drag. I'm talking wigs, heels, lipstick—the works.

Well, Neo, thanks so much for stopping by. It's a real pleasure to meet you and your teleoperator, and I'll be seeing you in my dreams. Appreciate it. Thank you for having me.

Thanks, Neo.

Take care.

6. Neo Needs More Time

After having seen the demo of Neo and interacted with the robot a little bit, do you want one of these for your home?

No, not anytime soon. You know, I think Marques Brownlee made a great video about Neo, where he basically just warned consumers that this company is making a lot of promises, and we don't really know if Neo is going to live up to them just yet. Barrett just told us, “Look, we're trying to be very transparent with you about what this thing can and cannot do.” And then I found, even in our demo, there were certain things I expected it to be able to do, like sit in a chair, that it couldn't do. So I do think we're in a real buyer-beware situation.

I think if you are an extreme, bleeding-edge early adopter and you like having really weird and pretty expensive experiences with technology in your house, this may be something that you want to consider. But I would really manage my expectations about this. I think you might get Neo expecting to use it as a butler and find that what you actually have is an intern that needs a lot of your attention and needs to be trained up. You're going to find that you are working for this robot in a way that you did not expect.

Totally. That was what kept coming up for me: this thing is just going to need so much hand-holding and oversight, especially in the initial phase, where it is mostly being teleoperated. It just feels like giving yourself a new chore rather than solving some of your existing chores.

You know, here's where I will give it a bit of credit, and I wonder if you would agree with this. Recently on the show, we were making fun of Johnny I for not actually having a plan to make AI hardware. I should say, as far as I know, that's a joke. They probably do have some sort of mockups of some sort of device that will eventually ship. But had Johnny I come out at that developer conference that we went to recently and said, “I'm making a humanoid robot, and you're gonna stop looking at your screen,” I would've been like, “Well, at least that's a plan.”

Yeah.

That’s something that I can visualize in my mind. I do want to give these folks a bit of credit for taking a big swing, right? There’s a huge gap between where they are today and where they want to go. But at the end of the day, if you could make this thing work, a lot of people would really get a lot out of it.

Totally. As we were interacting with Neo, I was flashing back to my first trip in a Google self-driving car more than a decade ago. At the time, there were safety drivers in the car in case anything went wrong. They could only go along certain routes. They had not been cleared for full autonomous driving.

The first ride I took was in a mall parking lot.

Yes.

That was what they had to do in order to keep you safe.

Yes. It was very limited. There were all these caveats you had to make if you were going to praise this thing. But at the same time, I had the feeling then that I had testing Neo: This thing is going to get good.

Yeah.

I don’t know when. I don’t know if it’s going to take 3 years or 5 years or 10 years or 20 years.

Or if it will be this company that does it.

Exactly. But this technology is clearly something that people are devoted to, that they are working on, that they are spending real money on, and that they are pursuing. Maybe it won’t take off, but maybe it will.

The most famous AI paper, period, is probably “Attention Is All You Need,” which is the paper that gets us the transformer. It creates this insight, which is that if you want to make LLMs better, you just feed them absolutely massive amounts of data, and so far, that just continues to scale.

What is interesting to me about Neo is that they are trying to take that insight and apply it to robotics, and they’re saying, “We are just going to create a system to gather as much robotics-relevant data as we can to build the robot butler that you actually want.” It might be that we can’t actually get to that point unless we have these things in people’s homes, picking up things off their floor, falling over when they try to squat. So it’s just going to take a lot of trial and error and data collection to make something that is useful.

Yeah, it was also striking to me as I was looking at Neo and testing it out: There is no way that people are not going to fall for these things, right? You have this thing. It kind of looks like a human. It kind of acts like a human in certain circumstances.

It has an incredible ass. It’s caked up as hell.

You heard it here first.

I took a picture of it. I’m not even joking. I did.

I just think there is no universe in which people do not develop feelings for these things, and that worries me.

That’s true, but aren’t there already series on TLC about people falling in love with a roll of paper towels or whatever? Paraphilias are a real thing. They’re actually going to be on the show next week.

Yeah, and I think that will probably end up being some upgrade that you have to pay for. The same way that there’s regular Grok and spicy Grok—

Mm-hmm.

—there’ll be regular Humanoid and spicy Humanoid, and that’ll be just a premium add-on.

Yeah. Well, that’s a fun thing—

But—

—to think about.

Yeah, I identify as an early adopter. I have all manner of gadgets and gizmos in my house. I think I’m not going to reserve a Neo just yet because, among other things, I would need to clear that with my wife and child, who would also need to live alongside this thing, and it would be gathering their data as well as mine.

And so I just feel like you kind of all have to be in on this. But I’m intrigued, and I can totally envision a world in which, 2 or 3 or 4 versions from now, these things have gotten much more autonomous, and I would actually want something like this in my house.

Yeah. I think here’s the Holy Trinity: If this thing can quickly and effectively do the laundry, run the dishwasher, and take out the trash, I think people actually would pay $500 a month for it.

There’s some other cleaning stuff in there that’d be really nice to have. There’s maybe a couple other things that I’m not thinking of, but I think if you nailed those three things, and you do them, and I truly just never have to think about them ever again, you can take your $500, Bernt.

Totally.

You know what would be the worst part of this if this actually gets good, Kevin?

What?

I’d have to fire my actual butler, Alfred. He’s been with me for 30 years. I’d really miss him a lot, honestly. I’d miss the hell out of the guy, but a deal’s a deal.

Yeah, you’d have to put on your own pants.

Yeah, it’d be a shame.

7. HatGPT Brings The Headlines

Well, Kevin, one thing that robots still haven’t yet been able to automate is HatGPT.

This is our favorite game. Sometimes we like to take headlines from the week, throw them on slips of paper, and put them in a hat, and then draw them out at random and talk about them. And when we’ve had enough of the other person talking, we say, “Stop generating.”

Do you want to start?

I’d love to start.

Okay, let me shake up the hat here.

This first story comes to us from Alex Roizman at The Atlantic: “Common Crawl Is Doing the AI Industry’s Dirty Work.” Alex reported that the nonprofit Common Crawl Foundation has basically scraped the entire internet, often without publisher permission. When publishers have gone to them and said, “Hey, why don’t you take our copyrighted material out of your data set that you’re providing to all of the big AI labs?” they actually have not been complying.

I love this story. I thought this was very interesting. I hear about Common Crawl all the time. It’s one of these very widely used data sets that was used to train GPT-3 and has been used to train many other models. I thought it was some esteemed nonprofit—Common Crawl, it’s for the common good. It sounds very august.

Yeah, so this story got a lot of journalists talking because Common Crawl’s executive director, Rich Skrenta, said, “The robots are people too and should be able to read the books for free.”

Yeah, and I didn’t actually know the history of Common Crawl here. It turns out it was started by a former Google employee who was kind of a techno-libertarian. This was during an age when people were saying things like, “Information should be free,” and he just decided, “You know what? It’s not good if companies like Google have this sort of snapshot of the internet, but no one else has. So I’m going to scrape the internet and make it available to anyone who wants it.”

And look, there are good reasons to want to preserve the internet. Large swaths of the internet have already disappeared, and I do think it’s an important part of our cultural heritage, et cetera. At the same time, when people come to you and say, “Hey, take our material out of your data set,” you should just do that.

Yes.

Yeah.

You should do that. All right.

All right, stop generating.

All right, this one comes to us from The Wall Street Journal. It is part of a story about Elon Musk and his AI companion at xAI. And it says that Elon Musk personally oversaw the design of a racy chatbot called Annie, an animated character with blonde pigtails and revealing outfits. Employees were compelled to turn over their biometric data to train Annie, and this is just wild.

Now, what do you mean by biometric data, and how is it being used to train a chatbot?

Well, according to this story, a company lawyer for xAI told a group of employees at a staff meeting in April that it would need to collect biometric data from them to train the chatbots on how to act and appear like human beings. They were apparently required to sign a form granting xAI a perpetual, worldwide, non-exclusive, sublicensable, royalty-free license to use, reproduce, and distribute their faces and voices as part of a confidential project code-named Project Skippy.

Just imagine that you are a woman who has the misfortune of working for xAI, and HR brings you into the office, and they're like, “Well, the good news is we think your face would be perfect for our new project.” The bad news is that it's a sex bot, and this is not optional.

God.

Yeah.

I cannot wait for the people involved at this company to start writing their gossipy tell-alls, you know? After every presidential administration, it's like, “Here's what I saw in the White House.” We need that but for AI labs, and we need it now.

Yeah, I think we might need something stronger than a tell-all to deal with this xAI situation. But, yeah, obviously, our condolences to anyone who's ever worked there for any reason.

Stop generating.

All right. This next story comes to us from The New York Times. President Trump says he doesn't know the crypto billionaire he just pardoned. Kevin, I imagine you saw this. Trump last month granted a pardon to Changpeng Zhao, the billionaire founder of the cryptocurrency exchange Binance, the bisexual finance company, who had pleaded guilty to money-laundering violations in 2023 and whose company struck a business deal in May involving the Trump family's crypto venture. But now the president claims he does not know who CZ is. Do you buy it, Kevin?

I do, actually. I think probably someone hands him a list of people and says, “You should pardon these guys. They're good guys.” And he says, “Okay.”

Imagine just waking up and someone hands you a stack of paper and says, “Here are your pardons for today.” And you think, “I ain't gonna read all that.”

Yes. I just find it fascinating. Obviously, this seems all kind of corrupt, and I'm sure we will learn more about how this pardon came to be. But I'm fascinated by pardons. I feel like presidential pardons are kind of the closest thing to magic spells we have in this world—

Mm-hmm.

—in which you can just say, “You go free. You go free,” and a person is just free.

Yeah, it's very Oprah Winfrey-coded. You know? It's like, “Look under your seat. You may have a pardon.”

Yes.

Interesting way to run a country.

Yeah.

Okay, this is a good one. “Coca-Cola injects ‘Holidays Are Coming’ ads with an upgraded dose of AI.” This comes to us from Katie Dayton at The Wall Street Journal. Coca-Cola, which made an AI ad last year for its holiday season, has taken another stab at AI-generated advertising. This new ad is called “Holidays Are Coming,” and Coca-Cola's chief marketing officer told The Wall Street Journal that the standard ad-making process used to take a year, and now with AI, they can get it done in around a month. Should we watch it?

Let's watch it.

Some holiday cheer, some lights, some critters.

Okay, I think we get the point.

Now, Casey, what do you think of Coca-Cola's AI-generated holiday ad?

Well, here's what I've learned: If you took a full year, you can make a good Coke ad, but if you only spend a month on it using AI, then it's bad.

Yes. Not the finest ad the Coca-Cola Corporation has produced, but I bet it was cheap.

All of the animals in the ad just wobble a little bit. They have no integrity in their animation, and so everything is just floaty and weird.

Yeah.

I will say I'm very worried about this because a lot of polar bears used to have jobs making ads for Coca-Cola, and what are they going to do now if AI takes those jobs?

These are some of the first animals losing their jobs to automation.

Are they supposed to starve?

Things weren't going great for the polar bears before this happened.

Yeah. They gotta unionize.

Yeah. You heard what's going on with those glaciers.

I love that all the comments on the video are like, “Remember when Coca-Cola used to make ads with soul?”

Yeah.

It's like, no. They're selling Coca-Cola.

Yeah. I expect more out of this sugar syrup.

Yes.

How dare they?

Okay. Stop generating.

All right. Oh, I'm actually gonna grab two. All right, Kev, this next one comes from Alexandra Marquez at NBC News. The White House has launched a spoof MySpace page mocking Democratic leaders over the shutdown. Did you visit this?

No.

Well, you can find it somehow, unbelievably to me, at whitehouse.gov/mysafespace. You haven't been to this?

No.

You have to pull this up.

Okay.

It's literally insane. So you pull it up, and it is a parody of MySpace, a website that has arguably never been more relevant than it is today. It opens to a satirical profile page for Democratic leader Hakeem Jeffries.

Okay, and in the “About Me” section, it says, “Hey, we're Democrats in the House and Senate. We love DEI, transgender for everyone, and handing out taxpayer benefits to illegal immigrants.” I think I get the point here.

If you asked me to imitate a 13-year-old boy writing mean things about Democrats, this is basically what I would come up with. Under Hakeem's friends in his top 8—because remember, on MySpace it would show your top 8 friends—this website literally, like, 17 years ago—

Yeah.

—was when this was relevant at all. Anyway, Hakeem's top friends include George Soros and Antifa. Also, for some reason, the Chucky doll from that horror franchise. Anyway, we wouldn't be talking about any of this except it's literally on whitehouse.gov.

It's so strange.

Yeah.

Yes.

The trolling is coming from inside the White House.

Yeah. This is gonna hit very hard with 41-year-old political staffers in Washington. I'm not sure this actually moves the needle for a lot of voters out there.

Yes, completely embarrassing and bad. Another one.

All right. Stop generating. Okay, this one is about Ilya Sutskever's surprising deposition. This comes to us from a little-known rag called Platformer and a writer named Casey Newton.

Ooh, sounds hot.

This is about a new 10-hour deposition in the legal proceedings of Elon Musk's lawsuit against OpenAI. The transcript of this deposition by Ilya Sutskever, the former chief scientist of OpenAI, became public recently, and it details his role in Sam Altman's 2023 firing from that company and some of the behind-the-scenes machinations that were going on at the time. Casey, this was a hot topic of gossip in the AI community this week. What did we learn from this deposition?

Well, look, in some sense, a lot of this was already known. We knew that Ilya had come to the board with concerns about Sam in 2023, and that had played a role in their decision to fire him. But in this deposition, we did get a lot more detail, and the one that just stuck with me was that Ilya brought to the board a 52-page document containing various office misdeeds that Sam was alleged to have done.

And it's important for this reason. Back in 2023, when we were doing emergency podcasts from the airport, the big narrative coming out of Sam's firing was, “These crazy effective altruist doomers are out of control. And look, they blew up OpenAI for no reason. They're so scared of their own shadow,” blah, blah, blah, blah, blah. And this narrative persists today.

You read this deposition, and what you learn is that the real story was that Sam Altman's own C-suite was rebelling against him in a very similar way that previous C-suites had rebelled against him in his previous jobs. So I think that, obviously, this ship has sailed. Most people are no longer paying attention to this story. But for us real AI heads, we now actually do have a more or less complete understanding of why Sam was fired, and it had a lot to do with Ilya Sutskever, Mira Murati, and a 52-page memo saying, “Here's what Sam did wrong.”

Yeah, it was just—my AI group chats were melting down about this. It was so messy. This whole period just feels like a fever dream now that we're a couple of years removed from it. But man, what a wild time that was. What a bizarre turn of events.

To now start to understand some of the plotting, scheming, and plan-making behind all of this, I think despite how silly I think parts of this lawsuit are, I do like this quality of lawsuits: They kind of surface interesting and valuable public information that would not have otherwise come to light. And so, for that reason, and maybe that reason only, I am glad that Elon Musk brought this suit.

And speaking of that, Kevin, my understanding is that Mira Murati was recently deposed in the same lawsuit, which means that we should have another deposition to look forward to reading sometime soon.

I enjoy reading transcripts of depositions because at least 30% of them are just lawyers arguing with each other about what their client does and doesn't have to answer.

This one has a lot of that.

A lot of that.

Yeah, the lawyers—

Very juicy.

…were not happy.

Screenwriters: a gift to screenwriters everywhere.

Yeah. Unfortunately, I think they've finished filming that movie, but—

Uh-oh. Okay. All right. Stop generating.

Mm-hmm. All right. This next story comes to us from Ashley Bellinger at Ars Technica: “Meta denies torrenting porn to train AI, saying the downloads were for personal use.”

I think I've heard that one before. Meta asked a U.S. district court to toss a lawsuit alleging that the tech giant illegally torrented pornography to train AI, Kevin. The move comes after Strike 3 Holdings discovered illegal downloads of its adult films on Meta corporate IP addresses, as well as other downloads that Meta allegedly concealed using a stealth network of hidden IP addresses.

Strike 3 brought a suit against Meta in July seeking damages that have been estimated at more than $350 million. Now, we should say that Strike 3 has been accused of being something of a copyright troll. The Guardian actually just had a big story about that this week as well. At the same time, Kevin, one does have to wonder: Why was all that porn being downloaded at Meta's headquarters?

Right. So Meta's defense here is that these were just employees, office guests, or contractors who were downloading these pornographic movies on the Meta network somehow. And Meta says that they are not training models on porn, and I believe them. But on the other hand, how do you think Nasty Nancy got so nasty?

I'm glad you brought it up, because I think Nasty Nancy, which is, of course, one of the very most popular chatbots on all of Meta's properties, has some pretty steamy things to say. And she had to learn it somewhere.

But look, obviously people are criticizing Meta over this, but I just want to say: You think creating the metaverse is easy? The people there need to blow off some steam. And sometimes that just means downloading porn on their work computers. So, if you want to have a metaverse, you're just going to have to allow for some of this.

I love the—

And I know you want a metaverse.

I—oh, I—oh, I do.

Yeah.

I love that part of Meta's defense in this lawsuit, trying to get it thrown out, is that this was so little porn. Basically, they're saying, “Look, if we were going to steal your copyrighted data to train our AI systems, we would steal a lot of it.”

Mm-hmm.

When you look at our previous copyright thefts, they've been hugely, hugely bigger by orders of magnitude.

Okay, stop generating.

All right.

Casey, this last one is near and dear to our hearts. This is a story that comes to us from my colleague Reggie Ugwu at The New York Times, and it is about podcasting and AI.

Ooh.

It's called “For Podcasters, a Voice Clone Is a Double-Edged Sword.” Reggie explores how AI voice clones from places like ElevenLabs are starting to be used by some podcasters and asks the question: Are AI replicas a boon for productivity or a betrayal of the bond with listeners? Casey, what do you make of this?

Well, I think it kind of depends. There are just certain cases where I think it might make sense to use a voice clone. I don't think we're going to be doing it on this show anytime soon, but look, right now I send out an email newsletter. It's available only in text. Readers have been asking me for years to send out an audio version. Would they object to me doing it via a voice clone? Some people have actually emailed me asking me to do just this, so I was glad to read Reggie's story, because this is something that's been on my mind lately.

Where do you think the line is for that? Would you be comfortable letting an AI voice clone read an ad in your voice?

I think that doesn't feel good to me. If you're reading an ad, part of what the advertiser is buying is your personal contribution to it. I could see using it on a podcast when we occasionally make a mistake recording the show. Usually, it's Kevin, but then we'll have to go back and fix it afterward.

Maybe I literally just had one or two of the wrong words. If you could use a voice clone to quickly replace those words, I think that's probably okay. You probably want to disclose that to people somewhere along the process, but in general, I think, as always, just disclose what you're doing and it'll probably be fine.

Yeah. I think this has potential, but less for replacing the voices of podcasters than for allowing users to swap in their own preferred voices. Maybe you like some of what Casey says on this podcast, but you'd rather it be delivered in the silky baritone of an Ira Glass-type figure. You can just swap in Ira's voice.

It's true.

Problem solved.

Actually, you want to hear a funny story about this?

What?

So I actually have been training a clone of my voice via ElevenLabs, just to see what it would be like to have it read my column, and I fed it a bunch of Hard Fork audio that our producer Whitney gave me for this purpose.

The problem is that my newsletter, I write in this typically somber tone throughout, and then in Hard Fork, I amp up my personality and sound like a crazy person. It turns out that when you have crazy-person voice reading a somber column, I just sound insane. So I'm not actually able to use that voice clone.

Yes, and if you want sad Casey in your AI voice pack, that's going to cost you an extra $7 a month.

Yeah, that's a premium feature.

All right.

All right. Is there any more? And we're out of stories.

And we're out of stories. That is HatGPT. Thanks for playing, everyone.