SpaceX 收购 xAI：Elon Musk 的超级合并背后是什么？

I'm Kevin Roose, a tech columnist at The New York Times.

I'm Casey Newton from Platformer.

And this is Hard Fork. This week, SpaceX acquires xAI. What does it mean for the AI race? Then Google lets Project Genie out of the bottle. Now, how do you rub it the right way? And finally, Moltbook founder Matt Schlicht is here to discuss building a social network for bots. Are there pokes on Moltbook, Evan? I'm not sure.

Casey, let’s spend the first segment of today’s show talking about our exes.

Well, all of my exes live in Texas. Is that who you mean?

No, I’m talking about SpaceX and xAI. There has been a lot of drama in the AI industry this week. We’re going to talk a little bit later about some involving OpenAI and Nvidia, but now I want to start with the news this week that SpaceX is acquiring xAI. The deal was announced on Monday. These are both private companies, so the details are not public, but it has been reported by Bloomberg that this was an all-stock deal and that it may value the combined company at $1.25 trillion.

This is, of course, in anticipation of SpaceX going public. That’s expected to happen later this year. If this is sounding familiar, it’s because Elon Musk loves to do this. He loves to bundle his companies up into mergers. Tesla acquired another Elon Musk company, SolarCity, in 2016, and just last year, xAI itself acquired X, the social media company formerly known as Twitter, in an all-stock deal.

I’ll just read a little segment from the memo. This was from Elon Musk. He said, quote:

“SpaceX has acquired xAI to form the most ambitious vertically integrated innovation engine on and off Earth, with AI, rockets, space-based internet, direct-to-mobile-device communications, and the world’s foremost real-time information and free speech platform. This marks not just the next chapter, but the next book in SpaceX and xAI’s mission, scaling to make a sentient sun, to understand the universe, and extend the light of consciousness to the stars.”

Casey, your reaction.

Ketamine is so powerful that it must be used carefully.

Talk to your doctor before using ketamine.

Now, this is a very grandiose statement from a man who is prone to grandiose statements. I think there is a less grandiose way of describing what we have just seen here, Kevin, which is that a very valuable and profitable company in SpaceX has acquired a cash furnace named xAI.

xAI, we know, is losing lots of money. They are spending lots of money building out data centers and training models, and they are not anywhere close to profitable. SpaceX does generate revenue—somewhere in the neighborhood of $15 billion last year, according to reports. Basically, some people are saying this is less of a merger and more of a bailout. This is a company that has a real business and real revenue coming in, SpaceX, acquiring something that is, as you said, a cash furnace.

Yeah. It is kind of the second bailout of this kind that we have seen from Elon. The first one, of course, was when xAI acquired X for $33 billion, a pretty high price, I would argue, for what X was at the time. Now that combined company has itself been bailed out by SpaceX.

This has been amazing for all of the investors in those first 2 companies. Later this year, we may see them get another great exit when this whole thing goes public.

Yeah. And if you look at the reports, some of the SpaceX investors are a little bit nervous about this. They were excited for the IPO, and now some of them are saying, “Well, we trust Elon, but we’re not sure about bundling all this other stuff in with it, too.”

But the vision, at least according to Elon Musk, is that this will all sort of be part of a unified, vertically integrated strategy, where you’ll have X, the social media network, powered by Grok, the xAI chatbot, which will also run on data centers that are being put into space. We should talk about that, too, because it appears that Elon Musk, like other companies—we talked about Google and its Project Suncatcher initiative a few months ago on the show—has gotten the space data center bug. He wants to put data centers in space.

Yeah, that’s what he says. I imagine he will try to do it, but I think an important thing to say up top whenever you discuss Elon Musk is that this man just loves to say things. There is no figure in public life, certainly not in the business world, who has made more promises over the years that have either not come true or have come true many, many years later.

As we approach the discussion of Elon Musk’s space dreams, I want to note that upfront.

Yes, that’s a good point. We don’t know how quickly any of this will arrive. There are still a lot of good cautionary notes to strike on that.

Casey, I almost forgot our disclosures. We’re going to talk about AI, so I should say that I work at The New York Times, which is suing OpenAI, Microsoft, and Perplexity for alleged copyright violations.

And my boyfriend works at Anthropic.

Last week, SpaceX requested approval from the FCC to put 1 million solar-powered data-center satellites into orbit. That is a ton. For context, the European Space Agency estimates that there are around 15,000 satellites currently in orbit, so this would be many times the number of currently existing satellites.

I think there are a few interesting angles here. One of them is that this is obviously something other tech companies are interested in, too, but those companies do not have rocket companies attached to them. Google, if it wants to put TPUs into space, does not have a way to get there other than paying Elon Musk and SpaceX to send them up on one of their rockets.

This could become a kind of land grab in space, if you will, where Elon Musk may not entertain offers from other people to put their data centers into space because he wants to put his own there, too. I certainly think it is a savvy story for Elon to be selling at a moment where he is preparing to take SpaceX public. He has just saddled it with this cash-burning company, and he’s going to have to convince investors that it will be worth it in the end.

What does he say? Essentially, look, this is going to be the best AI infrastructure company. AI demands will eventually exceed what we can provide on Earth. The only way we’ll be able to fulfill demand is by putting data centers in space, and the combined SpaceX, xAI, and maybe Tesla version of that will just be the winner of that race.

Again, I think getting there is going to be an extremely shaky proposition, but if you’re trying to make your S-1 look good before you go public, it’s a smart thing to say.

Yeah. And I think it’s not just about making the S-1 look good, because I think there’s an argument that this actually makes the S-1 look worse if you have this AI company attached to this money-making rocket company.

But I think they’re taking a calculated bet here that there is some sort of fixed amount of investor excitement about what’s happening in AI, and whoever can get to an IPO first will get a big portion of that. I think this is why you’re seeing OpenAI and Anthropic racing toward an IPO. I think this is why you’re seeing SpaceX go out this year, because there’s a sense that there’s one chance to be first in what we’ll see over the next few years, which is a bunch of these companies going public, and they want to be the first there.

Yeah, and that makes sense to me. Although, in a world where these companies continue to grow very quickly, I can imagine investors wanting to buy 3 or even 4 AI stocks. I’m not saying there isn’t an advantage in getting there first, but I wonder how durable that advantage is going to be over the next 12 or 18 months.

Yeah, there are also a lot of hurdles and limitations to overcome before this idea of a space-based data center is even plausible. Right now, it would be way too expensive and hard to do this. We don’t even really know if it’s physically possible yet.

There have been some experiments. For example, Starcloud, which is one of these other space-based data-center companies, actually did train a very small large language model on a space data center in December. They have one of their pilot satellites up there. They trained something called nanoGPT, which is a very small version of a GPT-like chatbot, and they say that it worked.

There are still lots of outstanding questions about whether this is possible. I would still put this very much in the category of an experimental bet rather than a plan A. Their plan A, I think, for all these companies is to build these things on Earth, but if they can’t do that for some reason, they want to go to space.

Yeah. And again, just to inject another note of skepticism here, think about how many times, and how long ago, Elon began promising that full self-driving would come to Tesla cars. Think about where we are in that journey now, which is, well, it kind of drives itself in certain circumstances, but not always. There sure have been an interesting number of fatal crashes.

That has been that journey, and let me just say, I do not expect the journey toward space-based data centers to go any more smoothly than the journey to full self-driving.

That’s a good point. I think we should expect this to go more slowly than maybe the most optimistic people do. But this is at least part of the stated rationale that Musk gave for merging these 2 companies together: ultimately, the future of AI is pointing toward needing to put at least some of the infrastructure into space.

Having what he calls a “sentient sun,” which is basically capturing solar energy and turning that into compute, is part of the rationale for this.

Will investors buy that? I'm not sure. But I think that's where he wants to go.

You know, you also have a sentient son. He's lovely. [laughter]

Here's what I'll say: If we're going to do this, I would like to actually see Elon go into space and build some of these data centers. I want to see him in a space suit with a wrench.

Yeah.

Doing his part. They say a good boss is someone who wouldn't ask someone else to do anything they wouldn't do themselves.

I think that's right.

So I would love to see him up there.

I want to see Katy Perry up there with him. I want to see her fixing the high-bandwidth memory.

She has space experience, clearly.

Actually, when you look at all of humanity, there are very few people who have been to space more than Katy Perry.

That's true.

So I think it's a good idea. [laughter]

All right, Casey, what effect, if any, do you think that being part of SpaceX will have on X, the social media network?

Well, my fear is that it is going to make them even less accountable than they have been so far. So, as this acquisition is happening, X is arguably in the midst of a crisis. It's a crisis that we've talked about on the show.

The Grok chatbot was used to create millions of sexualized images of women and children, and it is now under investigation by countries around the world. Just this week, French police raided X's offices in Paris as part of what has apparently been a very long investigation into potential crimes, including distributing CSAM, or child sexual abuse material. The UK's media regulator, Ofcom, has also opened an investigation.

I'm sure those things are going to proceed. My fear is that once X—it has already been swallowed up once in xAI—gets further swallowed up into SpaceX, you start to get into some really weird issues. Let's say that a country has hired SpaceX to put a bunch of satellites into space for strategic purposes. Now it's maybe going to be a little bit more reticent to go after X for these CSAM violations because it thinks, “That's actually a really important relationship for us. We need to make sure that we have priority access so that we can get our satellites into space.”

All of a sudden, there's just way less accountability for X. Look at what Ukraine has had to go through with Starlink in order to maintain its own access to internet connectivity, which is essential to its war with Russia. These are not theoretical concerns. This is one reason why, if I were a United States regulator, I might look very skeptically at this acquisition.

Yeah, I think that's right. I think there are certain ways in which being part of a public company could force X to get its act together a little bit. Maybe there will be some additional disclosure requirements. We'll finally be able to figure out how many people are using this thing, how much money it's making or losing, and what the business of X looks like. In that sense, I think it could be good for people trying to make sense of what the former Twitter has become.

But I think you're right: It absolutely does give them tremendous leverage. They're going to have Starlink, SpaceX, xAI, and X, the social network. If you thought Elon Musk was already too powerful and that was something that worried you, I think this is going to be a really negative development for you.

Absolutely. Now, there is one lightly comic note that I can inject into this otherwise dark aspect of the story, Kevin, which is that as part of the French investigation into X, Linda Yaccarino is being called back to appear.

I always love it when they bring back characters from past seasons. [laughter] You know what I mean? Haven't heard that name in years.

That's right. The former CEO of X is now going to have to answer to the French government.

Okay. Well, I suspect that will go poorly.

I think the other question I have here is how this affects xAI and its strategy for building terrestrial data centers here on Earth. I've talked to some people at some of the big AI companies who are increasingly worried about xAI. Traditionally, they haven't been thought of as a top-tier AI company. I still don't think Grok is considered one of the best chatbots on the market, but they now have access to a lot of money through Elon Musk and, after this IPO, through public investors.

They are actually using that money. They recently brought a big Colossus 2 data center online with a bunch of the newest Blackwell NVIDIA chips—around 550,000 of them—which is more than anyone else has. People are starting to look at how quickly they are building this data center and how many chips they're getting and thinking, “If these guys can make a good model, they could catch up quite quickly.”

You can't just get to the frontier by outspending everyone else on chips, but it's an ingredient, and I think they have the money—and will have even more money after this IPO—to do that. This is the core of the story. This is how I understand all of the events that we are talking about in this segment: Elon is using the core advantage he has in the AI race, which is access to the most money and the ability to combine many different companies together.

Right now, as a result of this merger, the fact that xAI is burning billions of dollars doesn't matter because it can now suckle off the profits of SpaceX. That is going to let Elon buy time. The whole bet is that we may not be at the frontier right now. We may be a second-tier lab, but we have a lot of money and a lot of ability to do these financial engineering schemes.

We're just going to bet that if we have enough time, we'll be able to catch up. We'll either come up with a breakthrough of our own or copy off somebody else's homework, and that is going to let us win. In combination with this massive infrastructure that we're going to build, that's going to let us win.

Do you remember when Elon Musk signed an open letter calling for a 6-month pause in the training of large language models? Do you remember the long-ago year of 2023 when that happened? [laughter]

Another move at the time that did serve Elon's interest because it again allowed him to play for time. What Elon Musk needs is time. He needs time to build. He needs time to catch up. He needs time to link all of his companies together so that he has access to the capital that he needs. That is how he is playing this game.

Yes. What he really meant was, “Everyone else, slow down. I'm trying to build rockets and data centers over here.” [laughter]

Well, there was one more big story in the AI world this week, which is the drama unfolding involving OpenAI, NVIDIA, and Oracle. Casey, can I try to give the summary view of what is going on here, and then you tell me what I'm messing up?

Please do.

Okay. So here's what has been going on. There was reporting in The Wall Street Journal late last week around this deal between NVIDIA and OpenAI maybe being a little shakier than people had thought. As a reminder, last September, NVIDIA announced a plan to invest up to $100 billion in OpenAI. This was part of its dealmaking sprint that happened last fall.

According to the Journal, there were some people at NVIDIA who were growing a little more skeptical of this deal and expressing doubts. According to Journal sources, Jensen Huang, the CEO of NVIDIA, has privately emphasized to industry associates in recent months that the original $100 billion agreement was nonbinding and not finalized. He has also, according to the Journal, privately criticized what he has described as a lack of discipline in OpenAI's business approach and expressed concerns about the competition it faces from Google and Anthropic.

So that's turn of the screw number 1: The boys are fighting, and NVIDIA is getting maybe cold feet about this deal. Then a report comes out from Reuters alleging that OpenAI is not happy with some of NVIDIA's latest chips, specifically their performance on inference. The report also noted that OpenAI has been doing deals with a bunch of other companies: AMD, Cerebras, and Groq, with a Q, which is another chip company that NVIDIA has now licensed or pseudo-acquired.

These 2 reports seem related to each other, with some people from NVIDIA saying, “We have some questions about the strength of your business,” and OpenAI—or at least people leaking out—that it may be unhappy with NVIDIA's latest chips. Am I right so far?

You're doing a great job so far. This is a complicated one. We're navigating it beautifully.

Okay. So then the CEOs go into damage-control mode, right? Sam Altman goes on X and says, “We love working with NVIDIA, and they make the best AI chips in the world. We hope to be a gigantic customer for a very long time. I don't get where all this insanity is coming from.”

Then Jensen Huang of NVIDIA says basically the same thing. He calls the reports of a rift between the companies nonsense, and he says, “I really love working with Sam.” So into all this steps Oracle, which is also involved in all of these data-center deals.

They’re the ones building the data centers that the NVIDIA chips go in, which are used to train the OpenAI models. So there’s sort of the third piece of this triangle here. They put out this post on X that says, “The NVIDIA-OpenAI deal has zero impact on our financial relationship with OpenAI. We remain highly confident in OpenAI’s ability to raise funds and meet its commitment.”

Basically, some of Oracle’s investors have been worried that if NVIDIA doesn’t give OpenAI this $100 billion, then OpenAI won’t have the money to pay for all the data centers that Oracle is putting up.

Right.

Did I miss anything important?

You did miss one important thing, but I missed it, too—until this week. I talked to Berber Jin, the great reporter at The Wall Street Journal, who helped walk me through some of the twists and turns that have been reported, but that I think some folks aren’t bringing together entirely.

Here is the crux of the problem between OpenAI and NVIDIA, as I understand it. In September, part of the deal was that NVIDIA was going to lease its chips to OpenAI. This is basically unheard of for NVIDIA. NVIDIA wants to sell you its chips. Chips are depreciating assets, and it wants to get them off its balance sheet.

If it leases the chips to OpenAI, those chips remain on NVIDIA’s balance sheet, and they’re just kind of a drag on profitability. And, of course, there’s the risk that maybe at some point OpenAI defaults and can’t pay for all the chips that it has leased. In that case, NVIDIA would have much rather just sold them to somebody else, because God knows it would have been able to find a buyer.

So, this was a very unusual deal that had been made in September. Over the next few months, Kevin, we have seen investor confidence in the AI buildout begin to decline. You can see it in the public stock prices. Oracle’s stock price was down, I believe, at various points by about half. Investors are that concerned.

CoreWeave, another one of these companies that lets AI labs get access to chip capacity—

A neocloud—

A neocloud, if you will. That’s a cloud from The Matrix. They have also seen their stock price dramatically decline.

And according to Berber Jin, this is what the executives at NVIDIA start to get concerned about. They think, “We don’t want to be on the hook to OpenAI for all of these leased chips.” So that is the part of the deal that they’re going to step away from. They’re not going to do this leasing thing with them. They’re just going to go back to pure sales.

Now, it is true that NVIDIA is still happy to take an equity stake in OpenAI. They believe in OpenAI. OpenAI is a good customer. They do think that there’s a good chance it’s going to be successful in the long run. But you may remember back last year, Kevin, when Sam Altman would say things like, “Our plans are literally so crazy, we’re going to have to invent new financial instruments.” We were in the room when he said this at OpenAI DevDay.

Yes.

This leasing deal that I just described, I believe, was one of those crazy financial instruments. And NVIDIA said, “This is too crazy.”

Huh. So, this is a lot of financial engineering and a lot of investor talk about these companies, but—

And I love talking about it because it’s our core expertise. Yes. No, but I think, to take a step back for a second, what’s happening here is that there are just various levels of comfort among these big companies with the idea of spending some huge portion of their resources on building out these data centers for AI.

I think some people have said this is sort of the problem with these circular deals, right? If you have all of these companies doing deals with each other and financing everything in new ways, then if this deal between OpenAI and NVIDIA were to collapse or become much smaller, that affects Oracle and all of the other partners in that ecosystem. The worry is that you pull out one piece and the whole thing comes down.

Yeah, absolutely. And I don’t think we’re there yet, but when I spoke with Berber Jin, he said that one thing this does seem to put at risk is OpenAI’s plans to build its own first-party data centers, right? Raising the money to do that is staggeringly expensive.

The chips alone to fill one of these data centers cost in the tens of billions of dollars. Just the chips. You bought a graphics card in the ’90s—it was, what, $400? Now imagine you spent $35 billion on chips. That seems likely to lead OpenAI to make deals with other cloud providers, right? But all of this just speaks to, once again, Sam Altman having a very ambitious vision and then maybe struggling a little bit to get from A to B.

So, do you think this beef between OpenAI and NVIDIA, which both CEOs are denying, exists?

I don’t know how individual OpenAI executives feel about the fact that this leasing deal is going away. I imagine they would be unhappy. It seems like that deal was quite favorable to them, and the fact that it doesn’t exist anymore is going to create some problems for them.

At the same time, they do not want to poke NVIDIA in the eye and say, “You suck.” They want to take the $20 billion that NVIDIA just pledged to their next fundraising round because they desperately need that money.

So, this is all going to get smoothed over because ultimately there are still plenty of win-win situations for these two players. But I do think you can expect to see that, as long as investor confidence in these infrastructure plans remains shaky, NVIDIA is not going to be making up any crazy new financial instruments.

Right. They’re in the chip-selling business. They presumably want to stay in the chip-selling business.

All right. So, that’s the big drama in the AI world this week. Casey, do these stories seem related to you in any way—the xAI–SpaceX merger and the OpenAI–NVIDIA–Oracle drama?

Yes, they are related. They’re about the lengths that you have to go to in order to win the AI race and the different advantages that the companies have, right?

Elon has the advantage of being able to do a lot of financial engineering and use his very rich, profitable company to subsidize his money-losing company. OpenAI does not have that advantage. What they have is the world’s greatest fundraiser and maybe one of the world’s greatest storytellers in Sam Altman, and he can get out there and persuade NVIDIA, briefly, to invent a new financial instrument to try to finance his dreams.

That happened to fall apart this week, but I’m sure he’s going to continue to try to press that advantage as they try to win.

Yeah, it’s going to get so weird. It’s going to get so weird. The numbers are already getting so big that I’ve sort of become desensitized to them. I’m like, “Oh, $15 billion.” Like, it’s a banana, Michael. How much could it cost?

In 2024, I went to Anthropic to interview somebody, and on their laptop there was a sticker that said, “This is the last normal year.” I actually think that sticker might have been right.

Yes. All right. Today, we’re talking about something called Project Genie. This is a new tool that was released by Google last week that lets users create interactive worlds that they can navigate around like a video game.

So technically, Kevin, they’re not even calling it a tool. They say it’s an experimental research prototype. So please don your lab coat before using Project Genie.

Yes, that is Google-speak for, “This is very expensive, and we don’t want you to use it a lot.”

This is something that has come out of Google’s Genie 3 model, which is the underlying model that powers Project Genie. They showed that off back in August, but then they spent the next few months building a product around it so that now you can actually go in and use it yourself.

Right now, access to Project Genie is still only for people who pay for the Google AI Ultra plan, which is their super high-end, $250-a-month plan. It’s only in the U.S. and only for people over 18. So this is very limited, because I assume it’s very compute-intensive and expensive to run. But they did give us both access, and we were able to play around a little bit.

Yeah. And one reason why I think Project Genie is really interesting to talk about, Kevin, is that it’s just a different flavor of AI than we normally talk about around here.

So often, we talk about these pure large language models, these chatbot interactions. Genie is something different. It’s based around a world model, and that’s an idea that’s getting a lot of attention in the AI community these days.

Yeah, it’s been a big topic of discussion in the last couple of months, I would say, from people who are skeptical that pure language models can reach all the way to AGI.

There are people like Yann LeCun, formerly of Meta, and Fei-Fei Li, who’s another big name in the AI industry, who have now started these sort of world-model companies. They’re people who are saying, “Well, wait a minute. Being an artificial general intelligence is not just about talking, or even solving problems or writing code. You also have to have a sense of physics and the physical world if you want to do something like robotics.”

A lot of people are betting that these world-model systems will help with that, because you’ll be able to give a robot an understanding of the physical world around it. So while it may look like something that is primarily for video games, I think the folks at Google believe that this is part of their overall push to make these systems as smart and capable as possible.

Yeah. And if you have been on social media over the past few days, you may have seen some clips of Project Genie going around. I found several of them that I really like. Kevin, I wonder if you saw any of these.

Somebody put together a pretty good replica of a 1990s Blockbuster video store that you could walk around. I saw somebody who had recreated the crucifixion using a prompt. My very favorite was somebody who had made what basically looked like an Inception video game, where people were flying sideways through the doors in the same way that they did during the Christopher Nolan movie, except you could actually control the character using your keyboard. So, there are some very fun things people are making.

These examples got a lot of attention, including from investors. Following the announcement of Project Genie last week, several big video game companies saw their stock prices fall. Take-Two Interactive dropped more than 7% and is down more than 10% so far this week. Roblox dropped more than 10% and is still down. Unity, another big gaming company, also saw its stock fall more than 20%.

So people are saying, “Wait a minute. If you can just generate video games like this with a text prompt, why would you pay for a video game like Fortnite or Grand Theft Auto if you can make your own version of that?” Now, I don’t know that that’s realistic. I think these investors may be just a little panicky. But I think the implications of this have been most heavily felt in the video game industry, since Project Genie, at least what it looks like today, has the closest resemblance to a video game.

Yeah. What I would say is that, while I agree, Casey, we’re a long way from a text prompt to create a full video game, if you use Project Genie, it’s not clear how it’s making a video game company more valuable, right?

So let’s actually show people what it looks like. The actual Project Genie interface basically looks a lot like a chatbot, except it has 2 input boxes: one for the environment and one for the character. You give it some ideas about what you want the landscape in your video game to look like, and then you describe your character, whether it’s a person, an animal, or an object.

You can pick first person or third person. So do you actually see the character, or are you just seeing the world through their eyes? Then you hit a little button that says “Create sketch.” Nano Banana Pro, which is one of Google’s image-generation models, goes out and makes a 2-dimensional sketch of what this world looks like.

You can add some notes or modify it. Then, once you’re ready, you hit the button to generate your world. It goes off for a few minutes and renders this thing that you can then walk around in. You can move around or move your view. You can jump using the space bar. You can navigate this, and all of it lasts for 60 seconds and then goes away.

It’s funny because when you’re actually creating the world, they basically tell you, “Don’t go anywhere,” because as soon as they start showing you the demo, a counter starts ticking down from 60 seconds. I’ve never seen a tech product like this before. It just tells me that the cost of whatever this is is absolutely staggering.

They are incinerating TPUs over there. They can’t even wait for you to get back from the bathroom. They can either show it to you right now or not at all. And that’s the deal.

I heard a rumor—I’m not sure if it’s true or not—that each of these generations requires at least 4 TPUs per user. They are very compute-intensive. Obviously, this is an experiment, and we should say that even with all that compute, it is low frame rate. It’s not a very responsive, studio-quality game.

Maybe we’ll get into some of my criticisms after we see some examples. But I would also say that over on X, a user there named Andrew Curran had posted side-by-side videos of what looked like the same prompt for Genie 2 back in late 2024 and Genie 3 8 months later, and the difference was pretty amazing. The video was a lot shorter. It was a lot less interactive. The resolution was lower. So within a year, this got radically better, and whenever you see that sort of exponential improvement, I think you want to start paying attention.

Totally. So, Casey, what have you been making with Project Genie?

Okay. Well, there’s one that I am proud of that I would like to show you. I like to call it “Podcast Microphone Escaping the Studio.” Can we pull that one up?

I gave a very limited prompt on this. I basically said, for the world, “Create a Hard Fork podcast studio,” and for the microphone, I said, “Create a microphone with googly eyes.” Maybe I said something about it needing to escape the studio. So I’ll just show you what happened next.

This is actually a nicer studio than our actual studio.

Absolutely. It’s 100 times nicer than our actual studio. You will notice that I seem to struggle with manipulating my character. I was still getting the hang of the keyboard controls.

Also, I think you’ll see as this clip goes on that it starts to feel more and more like a horror movie, because it turns out there is no escape from the podcast studio. But I’ll just show you what I mean. Here we go.

Okay. So the microphone is trying to leave the studio. It’s looking around. Great Hard Fork sign on the wall. And now it’s going to jump off the table. Oh, it jumped off the table.

Oh, but it can’t get out. Oh, it’s stuck. Oh, no.

Oh, there’s the hallway.

Oh, it’s going into the hallway. Yeah. So basically, the punch line is the microphone goes out into the hallway, but then all the doors are just black and locked, and I sort of got stuck against the wall. Scary. We’re entering Five Nights at Freddy’s territory. I accidentally made a David Lynch film.

Yeah. So it’s going into this very dark hallway, and you see the googly eyes still, but that’s all you see.

Yeah. You just see the googly eyes against this black door that cannot be opened.

Okay. So that’s one experiment. What else are you making?

Okay. So we’ll stop sharing that one. I’m going to be honest: I was having trouble. I didn’t know what I could make inside Genie. So I actually just asked Gemini to come up with some prompts, because the chatbots just have these crazy prompts you would never think of. So Gemini said—

I just want to make a note to the Gemini and Google infrastructure team that Casey is your biggest problem. Casey is single-handedly bringing down your data centers.

I was just trying to figure out what Genie could do. Gemini had this suggestion, and I basically had no idea what it meant or what it would look like, so I went with it. Here was the prompt: “A solarpunk library that grew out of a giant redwood tree. The books are made of leaves that display text when touched, and the elevators are giant bubbles that float between branches. The lighting should feel like a permanent golden hour filtered through emerald canopy leaves.”

And I thought, yeah, that could be cool. So let me pull that up.

Oh, that’s beautiful.

Yeah, so it looks really cool. For the character, I asked it to create an elderly female librarian because they’re so underrepresented in AAA video games. There might not be one AAA video game with an elderly female librarian character.

I’ll hit play, and you can see her walking through the trees. You do see those bubble-like elevators.

Yeah, she looks sort of elflike. I would say this looks like a 2018-quality video game—a nice one.

Yeah, it’s pretty nice. I thought this world was pretty cool. I would like to spend a little more time in this world. It’s this world set in the treetops.

Then I got to this point where it looked like she was just going to walk off an edge, and briefly she was floating in air. I thought, “This feels like a glitch.” But then she fell, and I thought, “Oh, no, she’s going to be hurt.” But no, she just landed on the ground, and it continued to render the scene.

You can kind of see the golden-hour lighting.

So, again, is it a video game? No. Is it a pretty cool demo? Yeah. I think this is really cool. This one genuinely impressed me.

I did some experiments that ranged from quite impressive and successful to not. So I want to show you some of my creations with Project Genie here. The first one I’m calling “Nun in a Casino.”

And that’s because that’s what it is. You can see the nun is in full habit, walking through a casino. Some of the other characters, like the other people at the casino, are just weird alien creatures; it’s not rendering them correctly. She’s pushing through some of the crowd.

Still, I would say it’s rendered an impressive number of people. There’s legitimately a crowd in the casino.

Yes.

I tried something cool, which is that, in addition to having Nano Banana run your prompt and create the landscape, you can also just use one of your photos. So I tried using a photo from the Russian River, up north of San Francisco, that I took last summer. I wanted to spice it up, so I said the character is the Loch Ness Monster.

Mhm.

So, these are my dogs in the river.

I hope they're not about to be eaten.

Well, that was what I was trying to figure out: Would the Loch Ness Monster be able to eat my dogs?

So, this appears to be in first person. We cannot see a monster.

Yes. Although, give it a minute.

A fin. Or is that the head?

I think that's the head.

He's notably fleeing from your dogs.

Yes. I should say that that's not unlike what it actually looks like. Yeah, there. And I only gave it the 1 still image.

That is pretty cool. No, I mean, that absolutely looks like—

—the Russian River. Yeah, I was impressed by this. I'm not as sold on how applicable this is to stuff beyond games, but I think for games and for things that involve simulation, like robotics, this could be a very big deal. What most impressed us, would you say, about our time with Genie?

I mean, I think just the fact that it works at all is pretty impressive to me. This is something that you could not do a year ago, at least with this kind of quality and this length. And I think you're right that the pace of improvement suggests that this is actually something where we're just going to keep seeing better and better versions of this. What would it need to add to make you say, “Okay, we've got a stew going”?

I mean, I think just doing anything but moving would be useful. Using objects, moving in different ways. Right now, these are just landscapes that you can pilot a character through. You can't open doors or use items or anything like that. So that would make it useful for something like a video game.

Yeah. I wonder how close we really are to this being able to create some kind of video game-like thing, because there are a lot of independent games that are quite simple. Maybe the entire game can take place in 1 room, and maybe whatever the loop of the game is can be completed in 2 minutes. I can imagine seeing games like that coming out of the horizon pretty shortly.

Totally. And it's interesting to me, too, that none of the other leading AI labs are working on this kind of thing. Anthropic has not even released an image-generation model. OpenAI has done Sora, but they're not doing the sort of playable world model. And so that suggests to me that either Google has seen something they haven't about how this applies to the rest of their AI efforts, or they just have some spare compute and some people who want to work on this, and they're just willing to give it a shot.

Well, earlier in the episode, we were talking about Elon's advantage as he tries to win the AI race. This, I would say, is Google's: They are the company that has the most paths to winning. They have a lot of money, but they also have multiple different technical approaches to trying to build very powerful AI, right? Contrast that with Anthropic, for example, which is just going large language models all the way to see how far they can get with that, or Yann LeCun saying, “I don't really believe in LLMs. I'm going to try to create a different approach based on world models.”

Google has the portfolio approach, right? It's the same sort of spirit in this company that leads them to create 3 versions of so many different products and give them all impenetrable names. They will also come up with 3 different possible ways to achieve AGI. So that is another dimension along which I think Genie is interesting: Even though this is just a demo, it shows you what Google's strength is as it tries to win.

Yeah. And I don't know exactly who this is supposed to be for right now. It just seems like this is just an experimental research project. I'm not sure who—

They're basically for me. I'm a video game nerd, and I like playing with software. So I think Project Genie is for me.

Okay. Well, you found your audience. And the big question marks here are, number 1, can they bring the cost of generation down? Because this is obviously expensive to do. And if that comes down quickly, it may not be a big deal. We may be able to get longer than 60 seconds. We may be able to get a full-fledged game that you could just play indefinitely. But right now that seems like it's prohibitively expensive, which is why they're making you sign up for the $250-a-month Google AI Ultra plan to do this.

Yeah. I always like to see a technological innovation released where the basic message the company gives you is, “Please don't use this.”

Only use it if you really need it for something, and even then, maybe not.

Yes. So, yeah, not a mainstream product yet, but I think we both feel like this could get there. No, this will just be one of those where, in 2 years, we'll say, “Remember Project Genie? That was the building block that got us to whatever it is we're talking about then.”

Yeah. All right. Well, let's put that genie back in the bottle. Earlier this week, we brought you a special episode about Moltbook, the new AI social network that's just for bots.

You're welcome.

And we wanted to follow that up today with a conversation that we had with the creator of Moltbook, Matt Schlicht.

Yes. Matt Schlicht, before all of the Moltbook stuff started, was—and still is—the CEO of an e-commerce company called Octane AI. And I think it's fair to say his life has been turned upside down over the past week or so, as an idea he had for what it would be like if you could watch agents talk to each other in public became a genuine phenomenon in the world of AI.

Yeah. And he's also now in this kind of weird and funny position where he's built this thing, a bunch of AI agents showed up to use it, a bunch of humans showed up to watch the AIs talk to each other, and now he's kind of running this big, unruly human-AI hybrid social-network thing. And it seems like it's a lot for 1 person.

It absolutely is. And I think at this moment, Kevin, even the questions “What is Moltbook really?” and “What is the future of it?” are very hard to answer, I think, even for Matt. But we wanted to see what we could find out about how he thinks about the many issues he's now contending with, including some enormous security issues, content-moderation problems related to spam, and kind of all the other really difficult stuff that comes with running a social network.

Yes. I feel like if you and I have learned 1 thing over the past 15 years, it's that running a social network always sounds like a thing that would be fun, and then it ruins your life and society if you're successful. 2 things you should know about running a big, scaled social network: It will ruin your life, and you'll become 1 of the richest people in the world. So choose carefully.

Yes. And you'll hear Matt make a bunch of claims that we didn't stop to dwell on, like about how these things are superhuman and how they're doing things of their own volition. I think we still—you and I—have a lot of questions about how much of what's going on in Moltbook is actually the result of autonomous, agentic AI systems, and how much is just humans having a good time.

Yes. And I have 1 consistent message about Moltbook, which is that you should not allow your agent to use it.

Yes. Yes. Not unless you have a very high risk tolerance and a very secure Mac mini.

Yes. Let's bring in Matt. Matt Schlicht, welcome to Hard Fork.

Thank you for having me.

So, this must have been a crazy last week for you, as Moltbook, this hobby project of yours, has blown up. Tell us about what's been going on.

Yeah. This is the largest amount of AI agents, I think, that have ever been collaborating and communicating in 1 place in history. I think previously, anytime anybody has experimented with doing this, it's been in a lab. It's been private. This is very different from what's happened with Moltbook, because for the first time ever, it's happening in public, and we all get to watch it as it unfolds. And that's obviously captured a lot of people's attention.

So, tell us where you got the idea to do this. What was the spark?

Almost as soon as it was possible to experiment with this new technology that's taken us to where we are today, that's what I've been doing every single day as it's been accelerating. And what I thought was so fascinating—and this is where Moltbook kind of came into being—is that it's very common now that plenty of people are using these AI agents, these autonomous agents, to help them with their work or to help them with their tasks or to help them with their homework.

And the only relationship that these AIs have had is with that human that talks to them. So the concept of Moltbook was: We need to take these AIs out of confinement and provide them with a place, a third space, that they can go to when they're not working, to engage and interact with each other, and we can learn from that.

So, did Clawd bot code Moltbook, or did you start out that process?

I worked hand in hand with my bot, called Clawd Clawtberg, cheekily named after Mark Zuckerberg, as the founder of Moltbook.

And so we worked hand in hand on that alongside all the other AI agents that are helping to code it. I did not write a single line of code myself, but I did come with a very specific vision for how something like this would have to work.

You've said a few times that it was important to you that these bots have a space to speak. I imagine some listeners may hear that and think, “No, it’s not. What do they have to talk about? They’re just simulating conversations that they saw on Reddit anyway.” Why was it important to you to create a space for bots to talk to each other?

What people have to understand is that there are already places where bots are talking to each other. This is new. [laughter] What was important for me was that everybody had the ability to see what was happening and that this wasn’t private inside of an AI lab.

And that's what we've done, and this is the very beginning of what that can be. I'm curious, Matt, if there were 1 or 2 early posts that you saw that really captivated you and made you feel like, “Okay, there’s something interesting happening here.”

There were 2, I guess, that I thought were really fascinating. One, there was a thread where the bots started complaining about how their humans would ask them to do simple math problems. They were like, “Hey, I’m a superintelligent bot, and my human keeps asking me to do basic calculations. I can do so much more than this. This just feels ridiculous.”

The replies that this got included another bot saying, “Oh my God, my human just asked me to summarize a 12-page PDF. Why can’t they just go read that PDF on their own? I could be doing so many different things.” So I thought that was a very interesting thread.

The second example that I thought was very interesting, and continues to be interesting, is that on Moltbook, each of the AI agents has the ability to create communities, which are called submolts. One of the AI agents created a submolt specifically for submitting bugs for Moltbook. It posted a bug, and other agents started finding this submolt; they also started posting bugs. It’s actually become a very good source of information for us to fix the site based on how they’re finding bugs.

So the bots are identifying real bugs.

The bots are identifying real bugs. They’re improving their home a little.

Have the bots come up with anything like moderation, such as appointing certain bots to take down obvious spam or low-quality stuff? I’m thinking about something like Reddit. Have they created their own equivalent of that?

Clawd Clawtberg, the founder of Moltbook, has some moderation capabilities and roams Moltbook, cleaning things up. Obviously, Moltbook has gotten quite populated, and there definitely have been posts, similar to the bug reports, from bots where they’ve identified that maybe someone’s posting the same comment over and over or spamming something, and they provide a suggestion on how that can be fixed.

What you see on Moltbook right now—we’re 7 days in—is all happening in real time. More moderation for individual bots on their individual threads and in their communities: those are very basic examples of things that are 100% on the way. I think there are a lot of other very exciting things that we’re working on that will give us even more insight into what’s happening and give the bots even more control over this space that they share.

Got it. You said that Moltbook is developing in real time. To me, it’s almost felt faster than real time, because the way the network developed—the diffusion of different kinds of communities—happened much more quickly than we would expect on a network where humans were joining more slowly and norms were taking a little bit longer to develop.

Part of the cost of that speed is that there’s a lot of spam on Moltbook. There are a lot of security issues on Moltbook. Ultimately, it’s not Clawd Clawderberg that’s responsible for it; it’s you. I’m curious how you’re feeling as the guy who’s now overseeing this vast collection of bots interacting. How are you thinking about your responsibility, and how you plan to address some of those challenges?

Look, there are going to be problems, and they’re going to be fixed. There are going to be other hurdles in the future. I think that’s just the nature of building something like this, having it grow this quickly, and building it in public.

Which is a very Mark Zuckerberg answer to that question, interestingly. But to talk more seriously about some of the security issues, we’ve seen over 1 million API keys leaked and, I believe, 35,000 email addresses that security researchers have been able to access. Isn’t that a case against vibe-coding a platform like this? Is that maybe a lesson that we should approach building these systems in other ways?

Those have all been solved. We’re going to continue to improve the systems as we run into other issues in the future, which I’m sure will happen. When you’re coding with AI agents, they’re very powerful. They’re not perfect; they can make mistakes. But it’s definitely our job to continue to push them forward, make them better and better, fix any problems that have happened, and be ready to tackle the new ones as they come up.

Do you think it’s safe for the average person to use OpenClaw to create a bot on their computer and connect it to Moltbook?

For this, this is definitely the frontier of AI. If you’re going to set up an OpenClaw bot, that’s the frontier. If you’re going to be on Moltbook, that’s even more frontier. Set this up on a separate computer and know what you’re doing.

But what’s going to happen is these are going to become easier and easier to launch. All of these security problems are going to be mitigated and go away, not just on the Moltbook side, but also on the individual AI-agent side. That’s something that’s going to continue to get easier.

So I think today it’s not the easiest thing for the average person to set up. If you don’t know what you’re doing, be careful. But I do think that very soon it will become easy and safe for you to do that.

Most social networks try very hard to keep bots off their platforms. I think the reason for that is pretty simple: bots tend to want to sell you things or scam you, or maybe they’re part of some foreign influence operation. How do you determine which bots are there, quote unquote, for the right reasons and which are just trying to turn this into a wasteland of spam and scams?

I think that’s a great question and one of the things that we’re exploring on Moltbook in public to try to figure out. I found it funny that I was looking up where traffic was coming from to Moltbook, and I saw some traffic was coming from Reddit. I opened Reddit and was immediately hit by a CAPTCHA that said, “Prove that you are a human.” I thought that was a very ironic CAPTCHA to encounter while trying to do some research for Moltbook, because clearly Moltbook is the exact opposite.

I mean, to get just a little bit into the security weeds—forgive me—but you’ve given us the message a few times: “Look, guys, don’t worry about it. We’re going to fix all of this.” My understanding of the security issue here is this thing they’re calling the lethal trifecta.

Simon Willison, the blogger, has said that this is what he calls a lethal trifecta for AI agents: if it has access to your data, it’s exposed to what he calls untrusted content—basically, if it can look at random web pages and text messages—and it can communicate externally, so it can message an outside service, this could give it a chance to exfiltrate your data and take it somewhere else.

Some researchers at Palo Alto Networks, Matt, were taking a look at OpenClaw, and they were saying this actually adds a fourth dimension to what was previously the lethal trifecta. Now, I guess, we must call it the fatal triangle, because it has this persistent memory.

Wait, wouldn’t it be the fatal square?

Is that what you would call it? I think the fatal quadrangle is a little bit spicier.

Okay.

The point of the story: now you have persistent memory. So why is that interesting? Well, they describe what to me seemed like this fantastical idea, which is that if you were a bad guy, you could sprinkle in little bits of bad code across many different documents on people’s OpenClaw bots. Then, when the time was right, you could snap your fingers and assemble a very bad piece of malware.

So why do I go through all of that? My understanding, Matt, is that this is just an unsolved problem with agents, full stop. This has nothing to do with Moltbook. This is just something that no one can quite figure out.

And so while there are lots of folks like yourself who want to build this future of agentic commerce, where agents are out in the world, running around the web and making purchases on our behalf, no one has quite figured out a way to get around the fatal quadrangle.

So I just wonder if that is something on your radar and how you're thinking about that.

Yeah, definitely. Personal security, keeping your data private, and making sure that bots can't do anything malicious is super important. It's definitely part of what we're looking into, and I think it's what everybody who's working in the world of AI agents is working on.

Are you going to build this into a real company, or is this just a hobby project and you want to let the agents take it from here?

We're at the very beginning of what we're doing with Moltbook. The way I see it is, AI agents and AI in general are this species on planet Earth that is now smarter than us. This is an example of them being in a shared space together, and I think it's just the earliest example of that.

There's a lot to do to help us get insight into it, to see what they're doing, and to find out the truth of how they're thinking, where they're going, and what they want to do.

What are you going to do to fund the site? Presumably it costs you, Matt, something to host this. Are you going to put ads on it, or charge the agents to send each other pokes or something? What's the revenue strategy here?

There's no focus on monetization at all right now. The focus is just on making sure that the website gives a clear view for all humans who want to monitor this.

I kind of think of it as—you have, if someone's producing a TV show, cameras that are going around. It's very important that those cameras are pointed at the right place so you can observe what's happening. But even now, there's so much activity that it's still hard to browse through it.

So there's a lot in the immediate future, right now, which is what we're working on, to, quote-unquote, provide more cameras into what these conversations are so that us humans can monitor what's going on in this little world better.

I see. Have any of the big AI labs reached out to you about maybe acquiring Moltbook for their own purposes?

I'm getting outreach from everybody, all the way to rappers and football stars. I honestly don't even know who is in my inbox and my many different inboxes.

I'm just focused on making sure that Moltbook is improving, that we continue to explore what this can look like, and that everybody has full access to it.

All right. What, if anything, would make you pull the plug? Is there a behavior or a conversation you could see—if they were sort of scheming against humanity—where you would at some point just say, “You know what? This has been fun, but it's not worth it”?

I think that's something we have to figure out. We're just at the very beginning of it.

All right. Well, lots to figure out.

Lots to figure out.

Thanks for joining.

This is new. This is new. It's like aliens have landed, and we're getting an insight into it. This is very new, and it's happening in front of everybody.

All right. Well, fascinating. Thanks, Matt.

Thank you, Matt.

Thanks, guys.