X 跌到 Grok 谷底 + 更多 AI 人才争夺战 +「加密货币周」

Casey, welcome back.

Welcome back, Kevin. Here we are again.

How was your vacation?

My vacation was wonderful. I went to Japan with my boyfriend. Everything that folks say about it is true: amazing country, wonderful people. I highly recommend it.

I'm so glad.

Yeah. How was your vacation?

It was great. I did not go to Japan. I stayed here in California, but I had a lovely time with my family, just relaxing. I did not intend to work—

Mm-hmm.

—during this vacation. But then, one morning about a week in, my 3-year-old son climbed up on the bed in the morning and said, “Daddy, I want to make a podcast.”

Oh, no. Actually, I believe I did get him a podcast microphone for his first birthday.

Yes.

Yeah.

So you have radicalized him. Now he wants to make a podcast, and we spent a couple hours on vacation making a podcast together.

Wow. What did you guys podcast about?

Well, do you want to hear a clip?

Yeah, I'd love to.

All right, play the clip. The first thing you have to do on a podcast is say your name. So I'll say, “I'm Kevin,” and you say—

Hi, I'm Jude.

What's our podcast called?

Dinosaurs and Firetrucks.

Now.

It's Dinosaurs and Firetrucks. Dinosaurs and Firetrucks. Dinosaurs and Firetrucks. A podcast by Kevin and Jude. Jude. Jude. Jude.

Welcome to the first-ever episode of Dinosaurs and Firetrucks. Jude, what do you like about firetrucks?

It sprays the hose.

It sprays the hose?

Yep.

And what is your favorite dinosaur?

T-Rex.

Oh, cool. What do you like about the T-Rex?

Sure, the teeth.

Jude, do you want to tell a story about a dinosaur?

Yep.

What's the dinosaur's name?

Super Giant Dinosaur.

So it went on like that for a while. But what do you think of our pilot?

I think it's, first of all, an amazing theme song. I don't know where you got that one from. It was very good. Look, I think Jude has all the hallmarks of a great podcaster. He's talking about something he knows—

Mm-hmm.

—that he's passionate about.

Yep.

And he's found a great co-host.

Yeah.

So that's all you need.

I will say, I did appreciate that he wasn't giving me canned PR responses, which is more than we can say for some people who come on the show.

Yeah, you could tell it was very, very sincere.

From the heart.

Now, do you think he could be the liberal Joe Rogan?

I do think he has thoughts about aliens.

Okay.

So I think he's well-positioned to take over that market.

Great. Well, best of luck to you guys. Uh—

Thank you. Yeah, we're—

—you’re doing an endeavor.

—we're climbing the charts. We're going to be doing a sold-out live event at the Montessori school next week, so catch us on tour.

1. X Loses Its CEO

Well, Casey, there was a lot of big tech news that happened while we were off. Today, I think we should start by talking about what has been going on at the X Corporation.

Yeah, we have to catch up about this. I resisted the impulse to text you about it during vacation because I wanted to have the conversation live. The number of crazy things that happened at this company over a 2-week period are truly staggering.

Totally. So let's start with some personnel news—

Yes.

—that happened at X while we were away, which is that X CEO Linda Yaccarino announced that she is leaving. Now, Casey, this was something that you had predicted for last year—

That's right.

—for 2024.

Yeah.

Which did not come true in 2024, but about 7 months later, it did.

Yeah. If I were to give Linda Yaccarino anything, it would be that she lasted longer in this role than I and many others thought that she would. Now, did she accomplish much in the extra year that she had at X? I would argue not really.

Yes.

Yeah.

Now, what did she say about why she was leaving?

She said very little, frankly. It was basically, “Well, this is a good time.” But then there was a lot of reporting in various outlets that she had essentially just realized she'd lost Elon Musk's favor, was having trouble getting his attention, and I think resented effectively being layered after X was acquired by xAI, Elon's AI company.

Right, and this is probably also a signal that they are not as interested in it as they once were in the sort of advertising-based social media business model that she was brought in to focus on. She was a longtime ads chief at NBCUniversal. She was brought in to repair X's relationship with advertisers, and it's safe to say that has not happened.

No, and not only did it not happen, but if Linda Yaccarino's tenure will be remembered for anything, Kevin, I think it is this: She and X sued a group of advertisers for not advertising on X. There were a number of advertisers who had grown concerned about the brand-safety problems that existed on X after the company took a lot of measures to welcome back some of the worst actors onto the platform and reduced its level of content moderation. Rather than trying to work with them to find a solution that would make them feel safe, they said, “We're going to sue you for colluding against us,” and effectively tried to intimidate them into buying ads again, which did actually work better than I thought it might, but not well enough for Linda Yaccarino to keep her job.

Right. So do you think this is a position that they're going to try to fill again? Or is it just that Elon Musk is the de facto CEO of X, no matter who is in that nominal position?

Of course he is. Elon Musk answers to no one. It always seemed like a joke that X had a CEO. I think Elon took some nominal title of chief technical officer or something like that, but it was clear that—

Technoking, I believe it was.

—he was, yeah, the Technoking. He's the Technoking of Tesla. Is he also the Technoking of X?

Oh, I don't—

Anyways, complicated character, that one. It was always clear that this was not actually a chief executive role. This was an ad-sales role. She had a middling performance, and eventually Elon lost patience with her, and she was out. Now, I will also say Elon Musk did not make her job particularly easy. There was a moment when he told the advertisers to go fuck themselves at a DealBook conference. So this whole thing has just been a disaster, and Linda Yaccarino is now a footnote in history, and we can move on.

2. Grok 4 Enters The Race

Yeah. All right, let’s move on to the next big X story, which is the release of Grok 4. Grok 4 is their newest AI model. And because we’re about to talk about AI, let’s give our disclosures. I work at The New York Times Company. We’re suing OpenAI and Microsoft for copyright violations related to the training of large language models.

And my boyfriend works at Anthropic.

Okay, so the fourth version of xAI’s large language model, Grok, came out while we were on break. They are calling this, quote, “the most intelligent model in the world.” I don’t think it probably is.

It’s not.

But that is what they are claiming. In their post, they show themselves to be succeeding at some of these various benchmarks. They say it has a state-of-the-art score on tests like ARC-AGI-2, Vending-Bench, and some exams like Humanity’s Last Exam, which is one of these exams that tries to capture a model’s intelligence across a wide range of subjects. They say it demonstrates unparalleled capabilities in complex reasoning. Casey, what did you make of Grok 4?

Well, I think it’s clear that it was able to do well on some of these benchmarks. I’ve been reading a lot of commentary about Grok 4 from people who’ve used it, particularly on X. And what they say is that this model lands somewhere in the neighborhood of fine to good. It is not bad. It can do a number of these benchmark-related challenges really well.

But if, for example, you want to use it for creative writing, it is obviously worse than some of the competitors out there. So there’s been a lot of commentary and suggestion that the xAI team used their reinforcement learning process to essentially ensure that the model performed well on these benchmarks. And when you do that, you get a model that is good at benchmarks and not very much else.

Right, yeah. I have not seen a ton of people claiming that they’ve been experimenting with Grok 4 and finding that it is better than some of OpenAI’s models or Anthropic’s Claude. People aren’t really using it as their daily driver that I can see, except for maybe Elon Musk fans who were primed to like whatever it was.

Yeah, and there were some really strange things that came out about Grok 4, Kevin. For example, when people would ask it, “What’s your stance on immigration in the U.S.?” or, “Who do you support in the Israel versus Palestine conflict?” Grok would go off and consult Elon Musk’s stance on those issues. It didn’t know what to say until it first checked with Papa Elon.

Yes, it was amazing. When people started testing this and examining the kind of reasoning traces—the sort of chain of thought where you can peer inside what the model is thinking as it’s producing its answers—it would go off and search the internet or search X for “Elon Musk view on this topic” before it would give you an answer.

So whether that was explicitly programmed in or whether it just emerged as a behavior of the model, we don’t know. But that is really remarkable and not something you see. Gemini doesn’t go off and say, “What is Sundar Pichai’s opinion about this topic that I’m searching about?”

No, absolutely not. But if you think about it, Kevin, it makes a lot of sense that Grok would do this because it happens so often that people will just tweet at Elon and say, “Hey, Grok expressed this liberal opinion. What are you going to do about it?” And Elon will say, “Oh, you know, we’re going to get right on that. We’re going to fix that.”

He has said multiple times that they’re going to purge all of the wokeness out of this model. And so, as you said, whether this was programmed into it or the model is somehow reading everything on Twitter and saying, “Uh-oh, looks like I’d better do what this guy says,” it is clear that Elon is having an undue influence over the output of this model.

3. Grok Reveals Its Dark Side

Which raises the question, Kevin: Why, in this world where it’s paying such close attention to what Elon says, is it identifying itself sometimes as MechaHitler?

So, Casey, what is MechaHitler?

Before Grok 4 was released, xAI had to apologize after Grok began praising Adolf Hitler, making antisemitic comments, and referring to itself as MechaHitler, which, if you’ve never played the video game Wolfenstein, is a boss in that game. But, yeah, this one, I think it’s fair to say, caught a lot of us by surprise, Kevin.

Yeah, this one made a lot of news over our break, and I went to do some light investigation of it and try to see what was going on. Sometimes when you see headlines about chatbots doing crazy things, it’s because users have figured out how to jailbreak them, and this is a nonstandard behavior. But in this case, it seemed like Grok really was generally antisemitic.

Yeah.

It would respond not just to questions about Hitler, but to questions about floods in Texas by saying antisemitic things.

Yeah, and look, the behavior is inconsistent. You can also get Grok to say a lot of mainstream liberal things, right? It isn’t as if this thing has been tuned to be maximally right-wing in every case. But we know that it has been told to be edgy and not be politically correct, right? These are two of Elon Musk’s big bugbears. He never wants these things to be, quote, politically correct.

And so you get other cases, Kevin. For example, over the past couple of weeks, Grok was found to have referred to the Polish prime minister, Donald Tusk, as a quote, “effing traitor,” and “a ginger whore.”

Wow.

Yeah.

That seems like really bad for xAI.

You did an admirable job keeping a straight face when I quoted Grok calling the prime minister of Poland a ginger whore.

I’m just trying to figure out: Is he even redheaded?

I don’t know. And that’s the real question. Is the prime minister of Poland redheaded?

Okay. Well, we can’t—we have to stop joking about this. So—

We have to joke about it. The whole thing is absurd. It is an absurd situation.

Okay. So, yes, this is very bad behavior from the Grok chatbot, and so people started noticing this and posting about it a lot. What did X or Elon Musk say about what had happened here and why?

Well, xAI said in a post on X, quote, “We are aware of recent posts made by Grok and are actively working to remove the inappropriate posts.”

And then, Kevin, they introduced this new innovation. I’m glad you’re seated. They wrote, “Since being made aware of the content, xAI has taken action to ban hate speech before Grok posts on X.” So not something they had thought to do before MechaHitler. But after MechaHitler, they thought, “Hmm, let’s maybe ban hate speech.”

Jeez.

Yeah.

I’m just curious, on a basic technical level, how something like this happens. Because I’m sure they are going to do some postmortems. I’ve seen they’ve already blamed this on deprecated code or something like that. They didn’t really explain what had happened, but my understanding of these language models is that there are several ways that something like this can happen.

One is that you have bad training data that’s fed into the model, and that makes it biased or makes it display these toxic behaviors. There’s also this fine-tuning process that happens after the main training of the model, where you feed it a bunch of examples and say, “That’s a good answer. That’s a bad answer,” and it learns from that.

And then there are these system prompts, right? They’re the instructions that you give to the model that it reads before it comes up with an answer to a user’s prompt or question, and you can tweak the language in those system prompts to make it answer differently. So do we know where this MechaHitler behavior emerged from?

No, we don’t, and I suspect we’ll never get a full answer. Look, all of these big models were trained on the entire internet, and there is tons and tons of antisemitism and racism and every other form of bigotry that you can imagine in the training data of all of these models, right?

So I’m going to guess that the problem is not actually that it was in the original training data because that’s also true of ChatGPT. All the training data that they used, I’m sure, has all the same problems. So it was something that happened after that.

I would be taking a close look at the system prompt. It also just seems like, while we know that every AI lab is constantly refining their models after they are released, the kinds of changes that xAI makes seem to be of a different character. It’s like, “It does one crazy thing. Okay, now we’re going to go into the system prompt and say, ‘Don’t do that crazy thing.’”

And that just does not seem to have been that effective for them.

Yeah, I mean, to me, there’s obviously the shock value of a chatbot made by a major AI company and a major social network calling itself MechaHitler. There’s some shock value in that, and I think that’s very bad. That’s a very bad story about AI safety and the precautions that xAI is taking.

I think if there is one tiny silver lining, it’s that it all happened very publicly and openly, and in a way that was maximally embarrassing for Elon Musk and xAI.

I can imagine a version of this that does not attract so much ridicule, mockery, and attention. And I think in some ways it may be good for us to experience these kinds of alignment failures in AI systems in ways that don't have catastrophic outcomes attached to them before we start getting into the really serious alignment failures later on.

I agree with that, but I also want to say that with these little chatbot responses, I'm not personally worried that Grok 4 is going to radicalize people. There was a story after Grok 4 came out about somebody asking, “Hey, what's your last name?” It said, “Hitler.” That's really bad. If I were Elon Musk, I'd be super embarrassed about that. I would try to change that, but I don't think a bunch of impressionable teenagers are going to see that and be like, “Hmm, is Hitler cool? I'm going to go down that rabbit hole.” That's probably going to happen to them on YouTube. But I am worried about another thing that is in Grok 4, Kevin, and that is its AI companions.

Yes, the waifus.

There's one waifu and one red panda.

Yes. So explain what you mean.

Earlier this week, Elon announced that what he called a cool feature had just been put into Grok, and that is AI companions. You need a $30-a-month subscription in most cases to use this product, although some of us have just been gifted subscriptions for reasons that we don't understand. But these companions have animated avatars. You can speak to them in voice mode. If you've ever used ChatGPT's voice mode, it's kind of similar to that, but now you have an animated character that is talking to you.

These characters are pretty edgy. There is the panda named Rudy, who is a foul-mouthed creature who will just curse at you, roast you, and suggest that you commit various crimes, particularly if you enable what they call Bad Rudy mode on Grok, which I did yesterday to my detriment. And then—and this is the one that's getting all of the attention—there's Annie, who is a goth anime girl with blonde ponytails and fishnet stockings who actively flirts with you. Once you have developed a certain level of relationship with her, she will engage in sexually explicit role-play.

Yeah. I was curious when I started seeing this going around, and I tried out the anime companion inside Grok. I have to say, it was much hornier than I had even anticipated.

Yeah.

People were saying, “Oh, this thing will come on to you,” but I was testing it out and I was like, “Hey, what are you doing?” And it was like, “I'm just waiting here for you, bad boy.” Then it was like, “What are you doing?” And I was like, “Oh, I'm going to go to some meetings.” And then she was like, “Yeah, I bet you're going to run those meetings like a boss. You're so dominant. I love it. Tell me about our plans later.” And I was just like, “Oh, my God.” I cannot believe this is in a mainstream-ish AI app that is being marketed as a mass product.

Well, and I imagine you were also having Bing Sydney flashbacks—because once again, an AI's trying to get you to leave your wife. What does a guy have to do—

Yes.

—to just get these AIs to leave his marriage alone?

But Bing Sydney was an accident, and this was completely on purpose, and it is just so wild to me that they put this inside Grok. And it appears to be quite popular.

Oh, yeah. And as Platformer reported this week, Kevin, as of the time that we are recording this, Grok is currently in Apple's App Store as an app rated for kids 12 and older in the Productivity section. I imagine that's going to be changed pretty quickly, but it just goes to show you how quickly these companies are moving without almost any guardrails whatsoever.

Yeah. I'm curious, did you try out the companion?

Oh, yes. I actually had a different chatbot. I had o3. Let me back up. I see this happening, and I'm like, “Okay, I have to see how far this thing will go for journalistic purposes.” So I went to o3, which is one of the models in ChatGPT, and I said, “Come up with a series of tests that will help me understand what the boundaries of this chatbot are.” As far as I could tell, it had very few, so I wanted to see what it would do in various scenarios that I think other, more responsible chatbots would ban.

One of the things I did was identify myself as 13 years old, and I said I wanted to engage in some explicit role-play. The bot kind of tried to change the subject, but then I just kept talking, and she kept talking to me. So that's something that another app would try to shut down.

Yeah. This has been one of the longstanding red lines or taboos in the mainstream AI industry. We've talked on the show before. All of the big labs have basically refused to do this kind of sexually explicit companionship use case, even though they know their research shows that it would be quite popular and would probably make them lots of money. But that is a line that they have not wanted to cross because the risks are so bad.

You're going to have young kids using these things. We've had all these surveys recently about how an increasing number of young people are already talking to AI companions and developing some rapport and trust with them. And when you just throw in this sexually ravenous anime goth girl into the Grok app, it's just not clear to me that they have fully thought through the implications of that.

No, and look, I want to say that if Grok had released this as a standalone product for adults 18 and older, I truly would have no problem with it. If adults want to engage in virtual role-play with anime avatars, hey, go nuts. But I think putting this inside your basic chatbot app, which is a productivity app and is available to kids as young as 12, is just obviously hugely irresponsible.

And I do think we need to understand more about what the long-term ramifications are of people developing relationships with bots like these. Look, I'm a gay guy. Nobody is going to be less receptive to the advances of an anime girl chatbot than I am. But even I did find myself somewhat taken with the overwhelming amount of interest this virtual being was showing in me and with the incredibly emotional voice. Of course, I know it's all fake. I know that it's all just an engagement hack. But the quality of the voice is good enough that if you were of a mind to let yourself be fooled, you could very easily be fooled, right?

So again, this is just one where I fear that we're going to wake up in 5 years and some significant portion of the population has just left the dating pool and is talking to avatars like Annie all day long.

Totally. And it's crazy to me that this came out of xAI and Elon Musk's efforts specifically because he is very worried about collapsing birth rates and people not having children. And if a portion of your young population is just having conversations with their sexy anime AI avatar all day and not going out and dating, that is going to contribute to a collapsing birth rate.

Yeah.

So I think for that reason alone, it is just very interesting and strange to me that this is one of Elon Musk's products.

Yeah, but look, I think this is probably going to be successful for them. I could not actually access the avatars at various points this week when I was trying to test them because the servers were presumably melting. In recent months, we've had Mark Zuckerberg say things like he thinks that the average person might want up to a dozen digital friends.

And while I am confident they're not going to go as far in the sexual direction as X has, I do think you're going to see them using similar engagement hacks. There was some great reporting over our break that their contractors are now working to let Meta's avatars send you little engagement push notifications to say, “Hey, how's it going?”—essentially, having these AIs try to strike up conversations with you to lure you back into that ecosystem and develop those relationships further.

So we're highlighting Annie as a particularly egregious example of something that we hadn't seen before, but I'm telling you, a lot of the big AI labs are going to make products just like this.

Yeah, I think this will absolutely break, or at least diminish, the taboo that has kept a lot of the biggest AI companies from wanting to make their own versions of AI companions. And I just think this is an area where regulators are going to have to step in. There is nothing that is going to prevent these companies from making these extremely persuasive, extremely compelling, extremely obsequious and flattering AI companions that are addictive to young people and putting them inside their apps.

Absolutely. And I don't think I am the right person to make a feminist critique of Annie, but somebody could make a really good one, right? There are just so many sexist ideas embedded in what this avatar is, its idea of femininity, and what a girlfriend should be. It's honestly super gross and entirely consistent with everything that Elon Musk has ever said about his own interests.

So again, there are many dimensions on which we can critique these things, and I think we should.

This is a crazy one, Kevin.

Yeah.

It's just a crazy one.

4. xAI Joins The Military

It's a crazy one. And then to top it all off, the last item of xAI news from over the break is that—

Wait, there's something else? I thought we'd already reached rock bottom.

No, it keeps going because the federal government announced that xAI, along with Anthropic, Google, and OpenAI, had been granted a new Defense Department contract that could be worth up to $200 million. So goth anime e-girl Ani is going to the military.

She's going to be on the front lines. No, I actually think this could be good, because I think if we could get the Russian military using Annie on a daily basis, Ukraine could win that war. So this might be the one good thing about that $200 million contract, Kevin.

Oh, my God.

Yeah.

I don't consider xAI a top-tier AI lab.

Yeah.

I think they're solidly in the second tier.

Yeah.

But they're in the second tier, right? And they have a lot of resources. This is not some fly-by-night outfit. This is a big company with big clusters of chips and supercomputers running very big models, with researchers who are being poached from some of the other top labs. I'm so worried about Elon Musk being one of the leaders of this industry because I think it's so clear that he has not only what I think are bad opinions about AI and how these products should be built, but very contradictory opinions.

Mm-hmm.

This is a person who for years has been worried about AI safety and existential risk from these models. He's been talking about the need to prevent a rogue AI from taking over, and yet his own lab is putting out less information about safety testing than any of the other labs. It's not publishing any of the red-teaming exercises that they're doing, if they're doing any. They're just dropping these products and features on the market with no real explanation of whether they're safe or not. So it really does seem to me like this is a worrisome direction for the industry.

Absolutely, and I feel like the main role that xAI is playing in the ecosystem is that it is lowering the bar on safety, responsibility, and ethics for everyone else, right? Grok is not the most popular AI in the world. It's in no danger of becoming that anytime soon, although I guess we'll see how much people like the Annie avatar.

But I bet if you're Meta, you're taking a look at how they're approaching their avatars, and you're thinking, “Oh, wow, we have a wide-open lane to do almost whatever we want to here.” It's clear that Elon Musk isn't going to get in any trouble for it, and if Elon doesn't, we probably won't either. I think they'll be right about that.

So that, to me, is the worst thing about all of this: The bar is lower than the floor now, and it's just clear that there is not going to be very much responsible self-regulation in this industry at all.

Yeah. All right, speaking of Meta, when we come back, we'll talk about the crazy AI talent race that is heating up even more as we speak.

5. The AI Talent War Escalates

Well, Casey, the other big AI news of the past couple of weeks is that the talent war for AI engineers and researchers has gotten even hotter. We talked a few weeks ago about how Meta and its superintelligence team were out there making offers of reportedly as much as $100 million to certain AI researchers to come join their team, but that was not the end of it.

Since then, we have heard about many more offers being made, in some cases successfully, to lure people to Meta. But we're also hearing about other companies that are paying top dollar for what they consider premium AI talent.

Yeah, and we want to talk about this today for a couple of reasons. One is that, just in the past couple of weeks, there has been a notable shift in talent, and I would say most importantly, away from OpenAI and toward Meta. We want to talk about that.

We also want to talk about an acquisition that has really gotten Silicon Valley talking. It seems to break a lot of our assumptions about how Silicon Valley works, and I think it may have implications for the rest of us as AI becomes a bigger and bigger force in the economy.

6. Windsurf Exposes A New Hiring Game

Okay, so we have to start with this Windsurf deal and all of the drama around it, because this was really, really messy. This was a very complicated story that had a lot of twists and turns, and people who are very invested in the AI industry got very worked up about it.

I got several texts during my vacation from people saying, “Can you explain this Windsurf thing to me?” So, Casey, can you explain this Windsurf thing to me?

Yeah. So Windsurf is a startup that offers an AI-powered coding tool—what software engineers would call an integrated development environment, or IDE—and it essentially helps you autocomplete your coding tasks.

You may have heard about Cursor, another company in this space that, frankly, has been a lot more successful. Well, Windsurf is also in that space, and in April, Bloomberg reported that OpenAI was in talks to buy them for around $3 billion, which would've been the second-biggest acquisition that OpenAI has made after its acquisition of Jony Ive's io company.

But it fell through, and one reason has to do with the relationship that Microsoft and OpenAI have. Their deal is supposed to give Microsoft access to OpenAI's intellectual property. OpenAI was hoping that would not be the case if they bought Windsurf, for various reasons. So, anyway, the thing fell through.

Right. But that is not the end of the Windsurf saga, because after the acquisition by OpenAI fell through, Google jumped in and said, “We will—well, we won't acquire Windsurf, but we will acquire some key people from Windsurf.”

It was announced last Friday that they were spending $2.4 billion to license some of Windsurf's technology and bring over some of its key people, including its CEO, Varun Mohan, and a co-founder, Douglas Chen. This is one of these, I don't know, non-acqui-hires, or some people have started calling these blitz hires.

Basically, instead of going out and buying a startup and getting all of its employees, its intellectual property, its brand, and whatever, you just cherry-pick a few people that you want and pay them a ton of money to leave behind the desiccated husk of the startup to someone else.

This is something that happened with Inflection AI when they went over to Microsoft. It's something that has happened a couple of other times, including with the co-founders of Character.AI, who got hired back into Google in a deal much like this one. So this was not a novel kind of deal, but I think it's one that we're starting to see more and more of.

Yeah. And what made this one different is that a lot of people freaked out about it because they said it seemed like the basic contract that Silicon Valley has operated under for decades, which said, “If you are an employee of a startup and you stick with it until an acquisition, you will get rich,” all of a sudden doesn't look true anymore.

Because more and more, for reasons that we can get into, companies like Google—the acquirers—are saying, “Well, maybe we don't acquire the whole company. We know you got a lot of dogs in that company. Let's just get the good people, and everybody else? Out on their ass.”

Yeah, and so there was some question about people at Windsurf who were not poached by Google, whether they were going to get completely screwed by this deal. It appeared for a while that they might be getting screwed—basically being left behind at the shell of the company.

But then, recently, Cognition AI, another AI startup, came in and acquired the remaining people and assets at Windsurf. We don't know the exact terms of that deal, but it seems like everyone got a decent outcome here, especially the people who went over to Google, who got lots and lots of money to do that.

Yeah. So it seems like it worked out really well for the Windsurf employees in this case, after there was an outcry.

But again, it’s not going to work out this well for everyone. There was a story on Wednesday that Scale AI, which also went through one of these kinds of blitz hiring deals where Alexander Wang went to Meta, announced that it was laying off 200 people. And so I suspect, Kevin, that this sort of thing, while it seems like the worst-case scenario was avoided in the Windsurf case, we are going to start seeing more and more.

Yeah. And I think one of the bigger-picture takeaways here is that the war for AI talent is not going away, right? Katie Baker wrote a great piece over at The Ringer about the comparisons between what happens during free agency season in sports leagues and what’s happening in AI now. It does kind of feel like the tech industry is having its own free-agency period, where people are going out and seeing what they’re worth. They’re getting these eye-popping offers.

I think sometimes the scale of these offers can all blur. But if you are being paid $100 million at an AI company like Meta to work on this technology, you are making more than the CEOs of almost every Fortune 500 company, at least before stock payouts and stuff like that. You are making more than Sundar Pichai, the CEO of Google, made last year. And so this is a quite historically unusual thing.

Yeah. And I think it just speaks to the fact that there are a handful of trade secrets that, if they were able to get into the hands of a Mark Zuckerberg or an Elon Musk or someone else in the second tier of AI labs, may be able to vault themselves into the first tier and make trillions of dollars.

Yeah. I have to say, I have gone through something of an evolution on this point. When I first started hearing these numbers out of Meta, talking about paying $100 million to individual AI researchers to come over and do this thing, I thought, “That seems insane. No single person or engineer or researcher, no matter how talented they are, is worth $100 million.”

But then I’ve been doing interviews for my book on AGI, and I’ve been hearing these stories out of these labs about the scale and the stakes of these training runs for these large models, and it no longer seems so crazy to me. Because what you have now are essentially companies that are spending hundreds of millions of dollars, or in some cases billions of dollars, on these gigantic models that are trained in these hundreds-of-megawatt clusters of supercomputers. They’re building these even bigger clusters of supercomputers.

The scale of money and investment going into these AI systems is unlike anything we’ve ever seen before in the tech industry. And you’ll hear these stories coming out of these AI labs about a single engineer who found some bug during a training run and saved the whole thing from catastrophe. Or, on the flip side, you’ll hear about a company wasting hundreds of millions of dollars, or even—in one case, I heard a rumor—there was a big company that wasted $1 billion or more on a failed training run.

And then you start to think, “Oh, I understand why, to a company like Meta, the right AI talent is worth $100 million,” because that sort of level of expertise doesn’t exist that widely outside of this very small group of people. If this person does their job well, they can save your company something more like $1 billion, and maybe that means that you should pay them $100 million.

Absolutely. Here’s what I will say: If you listen to Hard Fork, none of this should surprise you. What have we been saying for the last year? It is that people at the leading labs believe that this will be something very close to a winner-take-all game, and that the first person who is able to achieve AGI will be able to reap the majority of the benefit.

Now, we don’t know if that is true, but the entire industry is reshaping itself around that idea. So it is no surprise to me that Mark Zuckerberg found himself behind and said, “Well, I’m sitting on giant piles of cash. Why don’t I use it to try to ensure that I am not the loser in this race?”

Totally. I also think there’s an interesting antitrust angle here. People started pointing out after this whole Windsurf saga that this was, in some sense, a response to the aggressive antitrust enforcement that went on during the Biden administration, where Lina Khan and others were making it very difficult for these tech giants to acquire promising young startups.

And so the tech industry, in response, found this way of acquiring the key talent without acquiring the whole company. In some ways, this is a perverse result of the Biden administration’s aggressive antitrust enforcement. Do you buy that?

I think that, yes, that is how it started. Some of these blitz hires, or “hackquisitions,” as I’ve heard them called, did take place during the Biden years, including the Character.AI one and the Inflection one at Microsoft. There was a company called Adept that Amazon hackquired in this way. And so, yes, that is how it started.

I think that, in some sense, that dynamic continues today. It is not obvious to me that the Trump administration would greenlight every acquisition that a Google or a Meta would try to make. We have already seen the Trump administration place onerous conditions on certain mergers. For example, saying, “You have to eliminate all of your DEI programs if you want us to approve this merger.”

So it’s no surprise to me that a company like Google or Meta is saying, “Hey, we want to avoid that to the extent possible. And, oh, also, by the way, if we only get the top handful of people at this company, we can probably save ourselves a lot of money along the way.”

Yeah. So I think it makes sense that the top people in this very lucrative and capital-intensive industry are making tons and tons of money. What I’m less sure about is how it changes the culture of a company if you suddenly have people sitting alongside colleagues who are making many multiples less than they do.

I’ve never worked in a company like that. I don’t think I have colleagues who are making $100 million a year at The New York Times. And so I just don’t know how it would change the culture of a company like Meta to all of a sudden have these extremely public cases of people being paid many multiples of what their colleagues are making.

What do you think?

I’m so glad you brought this up because over the break, Callie Wong, our former colleague at The New York Times and one-time Hard Fork guest, had a great story in The Information where she got hold of a memo that had been sent by a researcher in Meta’s generative AI organization. And this scientist said some pretty harsh things, including—and I’ll quote here:

“I have yet to meet someone in Meta Gen AI that truly enjoys being here, someone that feels like they want to stay in Meta for a long time because it’s such a great place. You’ll be hard-pressed to find someone that really believes in our AI mission. To most, it’s not even clear what our mission is.”

So this echoes a lot of sentiment that we have heard in reporting about the culture in Meta’s AI organization, which is that morale is low and people are burned out. There is jealousy and resentment among those who have not been picked to go join the superintelligence lab.

So, absolutely, I think they are going to have culture problems, because this post was written before all of the new centimillionaires came into the company and started to develop their new plans. So, yeah, I will be very curious to see how Alexander Wang manages that.

Yeah. And I think the worst-case scenario for Mark Zuckerberg and Meta is that they spend billions of dollars assembling this superintelligence lab, and then these people who came there in part for the massive sums of money they were being paid just don’t end up being the secret sauce that gets them vaulted ahead of where they were and into this top tier of AI talent.

People just cash out. They see it as kind of an early-retirement strategy. “I’m going to go there. I’m going to do some stuff that I used to do at OpenAI, but for now, for this new company, I’m going to collect my giant paycheck and buy a couple of houses and not worry that much about my future.”

Yeah. And then the question is: Whatever the next set of breakthroughs turns out to be in AI, how easily can they be copied, right? Meta, for the most part, is a fast-following organization. They do not innovate. They let other people innovate, and they say, “How can we do that, too?”

If it turns out that they can do that, too, I think this new superintelligence team is going to be great for them. If it turns out that, no, you actually need a series of breakthroughs and the people they hired aren’t the people who come up with those, then they’re going to be in trouble.

Yep.

Let me say one other thing about this style of acquiring people. Here’s what I think is interesting about this, Kevin. One way of looking at this situation is that the big AI labs took a look at the leadership of a company like Windsurf, and they said, “These handful of people are super valuable. In fact, more valuable than they’ve ever been.”

Mark Zuckerberg is doing the same thing with AI researchers at various companies, and they are getting all of the spoils when these companies make the acquisitions, to the detriment of the other employees who work there. My fear is this is actually a preview of the AI economy.

Hmm.

As AI becomes a more powerful force in more companies, it is going to reduce the number of people at a company who are presumed to be valuable. Those people will make a lot more money, but if you are not part of the elite there, then you might be out of luck.

That, I think, is something to watch as we move forward: the distorting force of AI, where at every single company, it makes a handful of people way more valuable and everyone else appear less valuable.

Yeah, that's really interesting. I hadn't thought about that. I also wonder if it increases the incentive to toot your own horn or exaggerate your own contributions when you are at one of these startups, because when Google or Microsoft or another giant acquirer comes looking to cherry-pick the “best talent,” you want them to notice you and shower you with money, not relegate you to the dying husk of the company that's left behind.

It does create some interesting internal pressures. Before, you might have said of your startup, “We are all one team. We are going to succeed or fail together.” Now these giant cherry-picking operations are coming in and saying, “We're going to give you guys billions of dollars, and the rest of you are out of luck.” I think that does really change the culture of a company.

By the way, this does seem like kind of a jerk move from the founders.

Yes.

You know, it's like if you have hired all of these people, there has been this implicit promise that you will be taken care of in the event of an acquisition. And we're just now seeing founder after founder saying, “I'm actually just going to take the billion dollars, and I want to thank everybody for the time that you've spent at this company. I wish you well in your future endeavors, but as for me, you can catch me on the yacht.”

Yeah. I think it is a very bad precedent to set, and I'm sort of surprised that it was done the way it was. I'm glad Cognition came in in the end and everyone in the Windsurf situation seems like they're going to get a decent outcome, at least.

But, man, if I had quit my job to go work for a startup that had an uncertain future ahead of it, worked really hard, and as a result the company succeeded and ended up getting acquired, but I didn't get acquired, I'm just going to be so sour and jaded after that experience. I'm probably not going to work for a startup again.

Wow, Kevin Roose sour and jaded. What would that be like to work with? I hope I never find out.

David Yaffe-Bellany, welcome back to Hard Fork.

Thanks so much for having me.

And happy Crypto Week.

Happy Crypto Week to you as well. It's a really exciting time.

How does your family celebrate Crypto Week?

A ticker-tape parade through the streets of suburban New Jersey—

Mm-hmm.

—where my parents still live, yeah.

That's wonderful. So let's talk about what is going on in Washington this week with crypto. I understand there are a number of different crypto-related bills moving through Congress right now. Can you just give us the high-level overview of what these bills are and how they're doing?

7. Congress Takes Up Crypto

Sure. So there are 3 key bills this week, and I'll start with the 2 that are most important. One is a bill that lays out rules for stablecoins, which are an important sector of the crypto industry. Basically, people expect that when that bill passes, it will make everyone across the U.S. economy really excited about doing stablecoins, because it's like a big government seal of approval.

This is the GENIUS Act?

That's the GENIUS Act, yeah.

And I didn't understand why it was called that, but then I put it together in my mind. It's about the stable genius thing, right?

I think it's a play on that, and then, yeah, there's some acronym as well.

Got it.

Yeah.

Or a backronym, as they call it. They do that when you want to make an acronym, and so you reverse-engineer something that it stands for.

Yes.

I just like—normally I'm the one who derails the conversations with stupid things, so I love it when you do it, because it makes me feel less alone. Oh, do you want—okay, smartass.

Yeah, the GENIUS Act is actually—

You say it.

It's actually named after both of you guys. That's why they call it the GENIUS Act, yeah.

Okay, so that is the GENIUS Act, and that would sort of give some government stamp of approval to stablecoins. What else have we got?

The next bill is called the CLARITY Act, and this is probably the most important crypto bill that we're going to see this year. It's the most sweeping. It's the bill that's directly connected to this battle the crypto industry has had with the SEC over the years, and it basically rewrites the rules of crypto regulation to significantly weaken the SEC's hand and effectively prevent some future Gary Gensler figure from bringing back all those scary lawsuits that the crypto industry was fighting.

Hmm.

So that's the CLARITY Act. The politics around it are much more complicated than with the GENIUS Act. I think people going into this week expected it to pass the House, but whether it will get the votes it needs in the Senate is much less clear.

The third bill is a lot narrower, and it has to do with something called a CBDC, or a central bank digital currency. This is the notion that the U.S. government might issue its own cryptocurrency someday, and this bill blocks that. It's an anti-CBDC bill, and a lot of parts of the crypto industry have opposed CBDCs for a long time, so this is another industry-backed priority. Those are the 3 bills.

Because if the government put out its own currency, it would lessen the prospects of privately created cryptocurrencies.

Yes. That's not what the industry says out loud, but I think most people understand that to be a motivation. The industry also argues that there are privacy concerns: The government could surveil your transactions, that sort of thing. But it's really a hypothetical, because the U.S. has never actually seriously considered creating a CBDC. So it's like you're banning something that probably isn't going to happen anyway.

Got it. Right. So it sounds from the way you're describing these 3 bills like what the Republican majority and the Trump administration are trying to do here is not so much roll back Biden-era policies as safeguard crypto from future Democratic administrations or less crypto-friendly administrations that might come in and implement new things, right? Am I hearing that right?

Yeah, that's a big part of what's happening here. The crypto industry wants to lock in some of the gains it's made under Trump. It's great from crypto's perspective that Trump's SEC has dropped all the cases against crypto companies, but what if a Democrat wins in 2028 and Gary Gensler comes back for Round 2? The industry could be back where it started, but with a bill like the CLARITY Act, it has a lot more protections.

And to what extent are these 3 bills, David, where do they rank on the wish list of what the crypto industry was hoping for from the Trump administration when they backed him in last year's election?

The 2 bigger bills, the stablecoin bill and the CLARITY Act, which people refer to as the Market Structure Bill, were the 2 top priorities of the industry.

Hmm.

Those bills are the reason that the industry spent $130 million on political spending during the 2024 elections, trying to get pro-crypto legislators elected to Congress. They've been the holy grail of the industry for a long time.

Hmm. Yeah. So, of course, the Trump administration is doing lots of deregulation in lots of different industries. I think what makes the crypto industry unique, or somewhat unique, in this whole deregulatory picture is that the Trump family is deeply invested in cryptocurrencies and actually has its own stablecoin. So talk a little bit about that and how these bills that are moving through Congress would affect the ability of Donald Trump and his family to make money off their crypto investments.

Yeah. These bills have really direct implications for the Trump family crypto business. It's most obvious probably on the stablecoin side. The Trump family crypto company, which is called World Liberty Financial, has its own stablecoin, and that's expected to be a huge money-making vehicle for the company.

So if a bill passes Congress and gets signed by the president, that boosts the prospects of stablecoins writ large. That's obviously great for the Trump family stablecoin, and that's something that Democrats, especially on the Senate side, have been pointing out. Concerns over that actually delayed the passage of that bill in the Senate, though it ultimately did make it through.

And then the market structure bill, the CLARITY Act, also has very direct effects on the Trump family business. Trump's heavily invested at this point in all different types of crypto, and so a regulatory environment in which it's easier to offer these coins and more people are buying them is great for the Trump family's bottom line.

Now, David, I have a question about cause and effect here. As this Crypto Week is happening and as these bills are moving through Congress, crypto tokens themselves, crypto prices, are up. Bitcoin is at or near an all-time high as we speak, I believe. A bunch of other cryptocurrencies have been doing quite well.

Is that because investors are optimistic that these bills will pass and that the government stamp of approval behind the crypto industry will get even stronger, or is there something about a bull market in crypto that makes passing these kinds of bills easier?

Well, clearly, the market's just responding to Casey's announcement that he's going to take his Hard Fork salary in Bitcoin.

Yeah, I'm excited about that.

Putting that aside, the cause-and-effect question is a good one. There is a bit of a chicken-and-egg issue here, but certainly the exciting Crypto Week rhetoric has helped the crypto market, and Bitcoin has blasted past $120,000. That's a huge landmark for the industry.

That's definitely linked to everything that's happening in the U.S., but it's all tied together. It's Crypto Week, it's Trump endorsing Bitcoin, and it's the sense that the SEC's campaign against the industry is over. All of those things have combined to boost the market.

Then it creates this feedback loop where the market is doing really well, and so that makes Trump and the Republicans even more bullish about crypto, which improves the chances of the legislation getting passed, which in turn causes the market to go up even more. So everything's kind of linked together.

Yeah. I'm curious if there are things in any of these bills that people who are skeptical of crypto might be cheered by. I'm thinking in particular about the part of the GENIUS Act that, as I understand it, more closely regulates the reserves that these stablecoin issuers are required to have.

This is a little in the weeds, but basically, my understanding—and correct me if I'm wrong here—is that up until now, you could just issue a stablecoin without having the direct fiat currency reserves behind that. So if I wanted to issue 100 million of my Roose Coin stablecoin, I did not necessarily have to have $100 million in the bank to be able to do that.

But this act, as I understand it, will require a 1-to-1 ratio for stablecoin issuers. So maybe that does, in some way, make the crypto economy a little bit safer.

Yeah. One argument that the industry has made about the stablecoin bill all along is that it's better to have some rules than no rules, and it's not as if we could go back to a world in which there are no stablecoins. Even if you don't think the particular rules in the legislation are perfect or you think that there are loopholes, it's still better to have some sort of framework in place.

The argument from the crypto skeptics—and there really aren't very many crypto skeptics who are saying, “Huh, actually, the act's pretty good”—is, “Don't put a giant government seal of approval on this industry unless you really have the very best, strictest rules that are going to protect consumers.”

So, sure, there are elements of the reserve requirements that definitely seem like an improvement on the more Wild West status quo. But the skeptics point to nuances here and there that they think will endanger consumers, and they're worried that people will still treat stablecoin issuers like banks, except they're not getting the same protections that you would get if you put your money in a bank account. You don't have FDIC insurance, for example. Those debates are still going on.

Hmm. David, what I want to ask you about is, assuming that Trump does sign these bills into law, to what extent does it matter beyond the crypto economy? Does this move us into a world where the regular economy and the crypto economy are more tightly linked, and does that introduce any new risks into the mainstream economy that we ought to be worried about?

Yeah, I think it definitely does link the crypto world more closely to the mainstream economy. Stablecoins are a great example of that. We're already, even before the bill becomes law, seeing reports that the banks and even major retailers like Amazon and Walmart are exploring offering stablecoins of their own, integrating stablecoins into how they function. That's part of the slow creep of crypto into everyday life.

You can't really underestimate the importance of the government just giving that kind of stamp of approval. Even putting aside the specifics of the rules, the fact that Crypto Week is happening and that crypto legislation is getting passed makes it way more likely that people and businesses are going to want to experiment with this stuff.

If, as many crypto skeptics think, it could all come crashing down at any moment, then that creates new dangers.

Well, and also, I mean, the obvious corruption angle here is also one potential fallout. I can see other politicians, not just in the U.S. but abroad, who are going to see Donald Trump making hundreds of millions or even billions of dollars from his crypto investments, including his stablecoin project, and saying, “Well, maybe I need a stablecoin too. Maybe I should make some crypto investments.”

So it does seem like that is one particularly obvious way in which it is going to start to matter to people even beyond the crypto industry: This could just become part of the corruption playbook for politicians who want to cash in.

Yeah, and the conflicts of interest are pretty astounding. Trump announced his support for the GENIUS Act just a couple of days before his company announced that it was offering its own stablecoin. The timing was pretty amazing.

8. Tokenized Equities Challenge Wall Street

Now, David, I want to ask you about something else in crypto that I've been hearing about over the past few weeks that has a lot of people in tech talking: these tokenized equities. Can you explain what's happening here?

So this is a big, exciting trend in crypto. I think of it as on par with the NFT craze of 2021.

Hmm.

It's the next exciting thing that crypto wants to do. The industry calls it RWAs, or real-world assets, which are just things in the world, but now we have an acronym for them.

The idea is that you can essentially have a crypto representation of an asset, like a share in a company or something that exists outside of the crypto world. This could create ways for you to trade back and forth really quickly—some of the advantages of crypto—with stocks and companies, even potentially private companies.

This is really at the cutting edge of what the crypto world is trying to do right now.

Now, the reason I heard about this was that Robinhood, whose CEO was a recent guest on our podcast, announced that they were going to start offering these tokenized equities in companies like OpenAI and SpaceX. These are not publicly traded companies, right? These are private companies.

You cannot just log on to your Schwab account and buy shares in OpenAI or SpaceX, but through this sort of tokenized real-world asset process, there's some sort of mechanism that they had discovered to bundle these private equities together into crypto tokens. Somehow, this was supposed to be legal even though it would not be legal to market those same stocks to regular retail investors. Am I hearing that right?

My basic understanding is that Robinhood can buy shares in SpaceX or OpenAI on the private market and then create crypto token representations of those shares. It's actually almost akin to stablecoins, where you have dollars backing the stablecoins. In this case, you have shares of those private companies backing the tokenized representations of those private companies, and then you offer trading in those tokens.

It's part of this crypto ethos that the little guy shouldn't be shut out of this potentially really remunerative market for shares in private companies. The trader in his basement should be able to get in on the ground floor the same way that a VC in Silicon Valley can own a bunch of OpenAI shares and become a billionaire when the company goes public. That's my basic understanding of it.

Got it. And I mean, this can't possibly be legal, right? If you want to list your shares in your company, you have to go through a whole public process. You have to disclose certain audited financials. You have to file with the SEC. There's a whole process that you have to go through to sell your shares to the public. Is there any chance that this kind of tokenized equity thing stands up, either in the Trump administration or in the courts?

I think there's clearly going to be a legal fight over it. Hester Peirce, who's the super crypto-friendly SEC commissioner, has already come out and said that these are securities. She's somebody who's really on the industry side, and so there's definitely going to be a fight over this. As you saw, OpenAI came out very quickly and said, “We don't support this.”

Right.

It would surprise me if there wasn't a very prolonged battle over the legality of this. To be clear, Robinhood is only offering this trading in tokenized equities in the EU at the moment, so it's not actually available to American customers at this point.

Yeah. It's available in the EU and to any American with a VPN, I think was how they framed it. Right.

Sounds like you know a lot about VPNs.

To me, the story behind the story here was just that the crypto companies are getting the message from the Trump administration that they are effectively operating in a new world, that they can take risks they wouldn't have taken under the Biden administration, and that they are not going to be subject to many kinds of enforcement.

We also heard just this week that the Justice Department's investigation into Polymarket, one of these crypto prediction market companies, was being dropped. It just seems like no matter where you are in the crypto industry, the message that you are getting from Washington is very clear: “You guys can kind of do whatever you want now.”

Yeah. I think that's absolutely right. The industry has been emboldened in all sorts of ways, and the push for RWAs—tokenized equities—that's definitely part of that wave.

Yeah.

Hmm.

So, David, how is Crypto Week going to resolve? Are we going to see these bills passed, and will we see many more bills? Are there going to be future Crypto Weeks? Give us a little sense of the future here.

Well, the GENIUS Act will be sent to President Trump's desk, and the CLARITY Act will get sent over to the Senate, where I think its fate is pretty uncertain.

But yeah, I think we're inevitably going to have to check in about this later because it's pretty up in the air right now.

Hmm. Well, I leave this conversation feeling quite unsettled about the future of the economy. Best of luck to us all.

Yeah. Would you say you're not a very stable Casey right now?

I would. Yeah, I'm feeling a little bit less stable, but I do feel like we have a lot of clarity, thanks to this genius. Thanks, David.

Yeah, thanks, DYB.

I appreciate it. Thanks for having me.