SpaceX’s Quiet CFO Ahead of Wild IPO, Anthropic Sounds Alarm on Recursive Self-Improvement

0:13Welcome, everyone, to The Information's TI TV. My name is Akash Pasfregia. It is Monday, June 8th, and the New York Knicks are in Game 3 tonight. We're going to be watching. Before we get to today's show, I want to highlight some reporting that the information published today. Google and NVIDIA are considering Intel as its backup chip manufacturer. This comes as TSMC is struggling to keep up with overwhelming demand. You can read more about that on our website. We're also tracking Apple's Worldwide Developers Conference today. We'll learn about how the company plans to implement more Apple intelligence features in its software ecosystem.

0:51We're also awaiting the Blockbuster public debut of SpaceX. The company plans to list on the Nasdaq this Friday. This Wednesday, June 10th, I'll sit down with the reporters who have been leading the information SpaceX coverage. Theo Waite, Belita Powell, and Corey Weinberg. The event is exclusive to subscribers. You can RSVP right now at theinformation.com slash events. Today on the show, we will dig deep into one of the most interesting angles of this IPO, the man leading SpaceX's finances, Brett Johnson. We published a profile about the CFO that we're going to talk about very shortly. We're then going to shift gears and talk about recursive self-improvement in AI, what it is, why it's important, why companies are pouring money into it.

1:37We're going to close out the show with a conversation with PwC about budget predictability in the era of AI. and token maxing. It's going to be a great show, so let's get right on into it. To get us started, I want to bring on our Elon Musk reporter, Theo Waite, to help set the scene for us. Theo, happy Monday to you. It is SpaceX week. How are you doing? I'm doing well. How are you? I'm doing well. Okay, so give us the rundown here. Now and Friday, there's a lot of time between then. give us the day by day what should we be watching for just go ahead

2:14Theo Wayt:so most of the week is kind of just sitting here nervously awaiting Thursday and Friday which is when most of the action will be the next few days you know everyone's getting their orders in and based on demand you know SpaceX will finalize the IPO pricing on Thursday and then and the actual stock will start trading under SPCX on Friday. And SpaceX has said the pricing will be$135 a share, but that could still change and fluctuate, and we'll know the actual number on Thursday. So, yeah, mostly we're just anxiously sitting around waiting. Okay. Well, while we wait, I want to unpack a profile that you published over the weekend.

3:01You did a deep dive into CFO Brett Johnson. He is the man leading the IPO from the financial perspective for SpaceX. Tell me a little bit about Brett Johnson. I mean, how does he differ from Elon Musk?

3:18Theo Wayt:So he's a very under-the-radar figure. I mean, there's a reason that we were the first people to really write a profile of him, which is that he's the ideal CFO in that he's very understated. He's quiet. people describe him as as steady and you know reliable um which it doesn't you know obviously is a gigantic contrast with elon who gets all the attention um but he's he's been you know a survivor he's he joined way back in 2011 um when most people didn't know what spacex was um and he's been the first and only cfo of of the company um so he's you know an incredibly important figure and and he's about to be a billionaire too.

4:02Okay. So he's an important figure, but I mean, you know, you, you spoke to people who were in Brett Johnson's orbit. I mean, tell us a little bit about him. Is he, is he as, I don't want to say chaotic, but is he as extra as Elon Musk on every level? Or I mean, it sounds like he's a little bit less, less. I would say no.

4:23Theo Wayt:I would, I mean, he's willing to, to, you know, roll with the punches. is like when when elon musk says you know we need spacex to acquire xai which is losing a ton of money just a few months before we're going public i mean you know in a normal world that would be a very extreme uh thing to do and in elon musk's world it's just you know pretty typical and and he's you know shown willingness to go go along with that kind of thing and that's you know paid off. In terms of the actual IPO process, you know, my colleagues who cover the IPO and I have both, you know, heard about him being pretty, you know, willing to throw SpaceX's weight around when working with bankers.

5:07Theo Wayt:Like they, you know, everyone wants to be a part of this deal and therefore he can make pretty aggressive demands from them around, you know, splitting up responsibilities and having lower fees and that kind of thing. So he's a good conduit for Elon. in that way how did he even get this job in the first place this seems like a really tough job to get yeah so he you know in the 90s he worked at an accounting firm he worked at chipmaker qualcomm he worked at this digital media company for a second then he worked at broadcom for a long time in 2008 he lands at this chipmaker called mine speed um which he started that job like right at the beginning of the Great Recession.

5:51Theo Wayt:The stock went down like 75%. He did a bunch of layoffs and cost cutting and sold off assets. It was a pretty tough situation, but he kind of helped this company dig out of a pretty deep hole, essentially. Then in 2011, he gets a call from a recruiter for this company called SpaceX, which he's never heard of. And he thinks at first that it's a storage company, you know, offering you extra space. And the recruiter says, no, it's a space company. You know, we make rockets. And he says, oh, I don't really know anything about that. But I've seen Star Wars is the story that he's told about that first conversation.

6:35Theo Wayt:But he takes the job anyway and sticks around. And pretty soon, you know, the Falcon 9 rocket is working better. Starlink launches and, you know, SpaceX becomes a pretty gigantic company, you know, in the subsequent 15 years. So he gets hired as CFO right away. I mean, that's the only job he's ever had at SpaceX. Yeah, he's the first one. It's the only job he's had. It's, you know, compared to all the other gigantic big tech stocks that SpaceX wants to be, you know, traded alongside, it's pretty unusual. And what is it about him that Elon Musk loves so much? I mean, earning Elon Musk's trust, let alone for the biggest IPO ever, is not a small task.

7:24You know, did you listen in any interviews with him? You talked to people and know him. Like, what is it about him that is so special to Elon?

7:34Theo Wayt:I mean, I think he's just been willing to roll with the punches, basically. basically. Okay. And keep the books tight, I guess. One of the punches you just talked about, the XAI merger, how is that going? What do we know about how he's managing the integration there and finding synergies, as they say? Yeah. I mean, XAI has definitely been cutting costs. They've done some layoffs. They've lost a lot of researchers. We've covered that pretty extensively. Like the actual research lab part is not going so great. My colleague, Grace Kay, had a good story about that last week as well. But, you know, from Johnson's perspective, like he's coming in and trying to generate a ton of revenue as quickly as possible, basically.

8:28Theo Wayt:And that has meant these gigantic compute rental deals that SpaceX announced with Anthropic a few weeks ago or several weeks ago at this point and then there was also this deal with google that spacex announced on friday um which the anthropic deals 1.25 billion a month the google deal is 920 million a month you know these are obviously gigantic numbers and it will mean that spacex gets to report huge growth in ai revenue this year um but it's a totally different kind of ai revenue than OpenAI or Anthropic will be reporting. I mean, this is not coming from people paying a lot of money for Grok subscriptions.

9:11Theo Wayt:This is coming from XAI having poured so much money into building data centers and now trying to figure out what to do with them. And that's just, you know, it's totally different. And you've reported a lot on the turnover in the Musk empire. Do you think Brett Johnson lasts in this role for as long as he has been in the role already? I mean, he, you know, he's, like I said, he's been in the role longer than any, you know, Mag 7 CFO out there. He's been at SpaceX during a period in which Tesla has had four different CFOs, including one that left and then came back. I mean, it's unusual that he's been in the job so long already.

10:00Theo Wayt:So if he lasts like two or three years, that'll still be, you know, after SpaceX is public, that'll still be pretty, pretty gigantic. And, you know, some of the people I talked to for the story just made this point that, you know, being a CFO of a private company is a very, very different job than being a CFO of a public company. And when you're public, you have to do all these, you know, tightly choreographed earnings calls where millions of people are going to be listening because it's SpaceX. You have to go to these conferences where your every word is watched in a different way. So it's a totally different job.

10:35Theo Wayt:And, you know, I think we'll see if he's able to adapt or not. Right. I want to ask you about another story that you published over the weekend. You wrote about the SpaceX mafia or the alumni of the company who are no longer there, but still very much hold shares in the space giant. Tell me a little bit about who you focused on here, notable alumni who have gone on to found their own companies. Yeah, I mean, the SpaceX mafia is quite big at this point. Like you said, there's a ton of billion-dollar-plus companies founded by former SpaceX people. One of the people I talked to who I thought was interesting was this guy, Scott Morton, who worked on basically software for building and testing rockets at SpaceX and he now says, you know, we're going to apply the lessons we learned there to building software for a whole bunch of other things like, you know, energy companies and other aerospace companies.

11:39Theo Wayt:So they just raised that, I think,$1.25 billion recently and it's a very young company. I also talked to someone from this company, K2 Space, which has a ton of SpaceX alums and is building satellites. You know, there's just this, in general, there's this gigantic appetite among investors to invest in former SpaceX people because they have, you know, just shown that you can do things in space that can actually potentially make a return. So there's just a ton of money being thrown at this area. Other than the big moonshot ambitions that all these people seem to hold, did you get any sense for any other through lines that connect them?

12:22Are there any personality traits? Do they all hang out together? Is there a specific type of founder that they have taken? Maybe they're all trying to be the next Elon. Like, is that what connects them?

12:33Theo Wayt:I mean, there's definitely this perception that they have a work ethic from being at SpaceX that other people from other companies don't have. And, you know, I think that there, yeah, there's this willingness to, you know, pull all nighters, fly to Texas randomly, you know, go be super scrappy and figure out a particular part. Like, there's definitely this, like, intensity and kind of single-minded focus on engineering problems that, you know, a lot of SpaceX people say they have and investors, you know, respond to. um you know i one person i talked to uh you know scott scott morton from rebel the software firm you know he kind of described this this bond that a lot of spacex people have um based on you know the experience of working together on a rocket launch like you you watch something take off and everyone has spent months you know grinding away at a tiny component or a particular tiny process and if any tiny part you know that that one person has been focused on fails the whole thing falls apart and there's this like collective bond that all these people have based on those intense experiences and a lot of them are you know even when they leave they they keep in touch they go to each other's weddings they're all on group text they go to happy hours like it's a very tight-knit community i i think and and you know as it gets bigger um We'll see if they, you know, continue that bond or not.

14:03Theo Wayt:But, you know, another question that I had when I was talking to a lot of these alums is, you know, whether the IPO is just going to make people so rich that they'll stop working. I mean, a lot of these people. What do you think? Did you get an answer there? You know, I think there's it seems like there's a minority that's going to just like buy a cool car or a sailboat or whatever and like hang out. but a lot of people were very insistent that they you know they didn't join SpaceX for the money it was less clear um you know years ago that it was going to increase in valuation so gigantically um and a lot of them insist that you know maybe the money will enable them to like do a crazier startup or you know have to raise less money and get to be more true to their own vision for something but a lot of them say you know they're not gonna they're not gonna stop working they're not gonna stop you know trying to solve a problem well and i will say i mean you know one of the folks you had on that you had on your list justin lopez he's the coo of base power and he's someone who we've had on the show and i mean look these ambitions are not small that they have embarked on themselves and so they very much have set themselves up for long journeys ahead but it was a great story and I encourage everyone to check it out.

15:24I want to thank you for coming on, Theo. I imagine I will be talking to you, if not every day this week, nearly that much. That is Theo Waite, our Elon Musk reporter here at The Information. One of the most important terms in AI that you might start hearing more about is called recursive self-improvement. My colleague Rocket Drew wrote about that in today's AI Agenda newsletter. I want to bring him on to tell us what it is and why it's getting so much attention right now. Rocket, welcome back to the show. It's great to have you here. Hey, Kash. Thanks for having me. Okay. Recursive self-improvement.

16:00This was not on the glossary of AI terms that you published, I think, a little over a year ago now. I think that one probably needs to be updated. What does this term mean, Rocket? Yeah, that's right. I think if we did one for 2026, it would be at the top of the list. I mean, it's a mouthful, but we're hearing it all the time these days. It basically means AI models that are capable enough that they can create their own successors. So say Claude Opus 4.8 could create Claude Opus 4.9. And at that point, humans would be out of the loop entirely on the research and development process that goes into creating AI models.

16:36We're hearing it a lot. Anthropic just released a big blog post talking about some of the evidence they're seeing that Anthropic is approaching or nearing this point of recursive self-improvement or RSI. There are a lot of startups that are raising funding, saying they're working on AI models to train other AI models, or even AI models to design the chips that will then be used to create other AI models. Of course, AI is making a lot of progress at writing code, but so far there are parts of the research process that have eluded the capabilities of AI models, where for now researchers are still better.

17:10Human researchers are better than the models. This is actually a fun connection with the SpaceX topic you were just on. So Theo's point about people working on these rockets, spending a lot of time obsessing over little components that could break. The reason the Challenger space shuttle famously exploded is because the O-rings failed. This very simple component on the rockets, they failed in the cold, the whole rocket blew up. And this is where we get the term O-ring automation, which is an econ term for a process, a production process, that's bottlenecked by the weakest link, effectively. So people talk about whether we'll see O-ring automation here in recursive self-improvement, where sure, maybe the AIs can write lots of great code, but maybe the research process is bottlenecked on this one step, which is humans coming up with the best research ideas or deciding which research areas to explore.

18:01But the idea behind the research self-improvement is something that's the problem to do in all themselves. Okay, so let me jump in here. So I want to recap a couple things. Number one, it just hit me. Your name is Rocket, and you're not a Rocket reporter full-time? I mean, did you ever think about that? Oh, what the heck? Okay, well, put that aside. I mean, I'm glad to hear that you seem to know just about everything you need to know about Rockets. So I think if you ever want to expand, there is a beat waiting for you. But let's go back to Recursive Self-Improvement. So the idea is that can we get the models to a point that they are good enough to then train the next generation of model or basically design the next generation of a new chip on their own?

18:43That's what this recursive self-improvement threshold is. Is that right? That's exactly right. And there's a sense that once we hit that threshold, the overall pace of AI progress could explode. I mean, I think everyone's familiar that AI has been going very fast already. But imagine you take all of these slow humans that make mistakes and you replace them with AI models that run very quickly, make fewer mistakes, and can accomplish a lot of tasks in a way that's coordinated. But in parallel, you could imagine that the pace of AI progress goes even faster than it's been going to say. But have we hit that threshold yet?

19:15No. The answer is definitely no. They're writing a lot of code, but they're not writing all the code. And the code is the place where they're making the most progress, right? So the overall research process involves coming up with hypotheses, designing the experiments to test those hypotheses, implementing the experiments. That's where the models have made the most progress so far. But then interpreting those results and using that to update your hypotheses and this whole loop. So this is the alarm. So Anthropic is sort of sounding the alarm in saying that, hey, we are getting to that threshold where the models could design the next generation of models.

19:50We're not there yet. but once we get there, there might be some implications here. That's right. They think it's in view. They think it's on the horizon. Now, I think some people would disagree with that, but depending on if it's a matter of one year away, two years away, or even 10 years away, I think people have a sense that there is something kind of alarming about this idea. There's something kind of terminatory about it. It's like, wait, so humans aren't going to be involved at all in the process of creating the next model. We're just going to trust our current generation of models to train their successors, and then those will train their successors?

20:23Like at what point are humans getting involved and steering this process? So that's the reason Anthropic is concerned about this possibility. They say you could end up with one model that has its own goals, its own desires, and then maybe the next generation of models inherits those same goals or desires or ones that are even weirder and the whole process could kind of run away from us. So as a result, one of the things that they're advocating for in this post that they put out is that they would like the ability, at least the ability for AI labs to do some kind of coordinated slowdown or pause and the progress of AI capabilities.

20:58They're not saying we have to pump the brakes. They're saying, you know, the brakes should be in there. So that was how likely do you think that, I mean, Anthropic is advocating for that, but are other labs on board with this so-called pause? Yeah, yeah. I want to be clear that like the idea of a pause is not something that Anthropic just came up with. People have been talking about it for many years, and people have been pushing Anthropic specifically to adopt this position, which they're now taking, which is that, well, if everyone else is willing to pause, maybe we would be interested in pausing.

21:27Actually, Demis Asabas, the CEO of Google DeepMind, said something similar earlier this year at Davos. He was asked in an interview, if everyone else paused, could we do some sort of coordinated detente in the AI race? And he said that he was open to it as well. OpenAI, for their part, is also working towards recursive self-improvement, having an automated AI researcher. But plenty of their researchers and even some of their leadership are at least interested in the idea of the pause. Though, like I said, there's disagreement about when such a thing would actually be needed. Well, and it seems like there's a bit of game theory here.

22:00Like, you know, somebody has to actually pause, you know, for any of this to really work out. And I mean, it sounds like they just all need to get in a room together and like collectively decide we're actually going to do it. Right, right. But they all have to decide. And so there's work going into making sure that they could detect if any of the other ones defected and verify that the other ones are actually following along like they said they would, which is easier to do when it's just, you know, anthropic open AI and demine. It gets more complicated if you have government in the picture and they're also trying to coordinate a pause with China and with the labs in China, which any, you know, truly comprehensive pause would have to do, of course.

22:40So let me ask you this. The term AGI kind of drifted out of focus in the conversation, certainly as Microsoft and OpenAI have basically removed that from the definition around AGI has been removed from the negotiating table in terms of when the royalty payments stop or keep going, stuff like that. is this just the new AGI? Like that's what it sounds to me is it's just a term that we don't really know when we'll reach it. We're kind of working towards it. It's something people are kind of fearful of, but is it just open to interpretation or is there actually like a measurable thing that no, we have hit recursive self-improvement.

23:25This is dangerous now. Yeah, that's a great question. Okay. Let me say one thing to defend, I guess, the merits of this term recursive self-improvement. And that's that when people said AGI, for a long time, what they really meant was recursive self-improvement all along. So in some sense, this is a refinement of AGI. Let me explain why that is, right? The idea of AGI was a model that's capable of doing all of the cognitive tasks that a human can do. That necessarily includes all of the tasks that go into creating an AI model, right? That includes all of the steps of that AI research process that I outlined.

23:54So AGI implies recursive self-improvement, which is part of the reason why AGI had so much oomph or people were paying so much attention to it. People thought that if we reach that threshold, that might be what kicks off this recursive self-improvement loop. Now people realize that intelligence is jagged and there's all sorts of things that AI can't do very well now and maybe won't be able to do well anytime soon. But the recursive self - improvement milestone could be something we hit sooner. So I think that's why people in some sense, you can think of it as a refinement. Another way to defend this is that AGI kind of served its purpose.

24:28It was a useful term when we were like a few years out, like a few years ago, when it still seemed more distant. And we just sort of like roughly needed to gesture at like very capable AI systems. And now that, you know, AI systems are much more capable on some domains, you know, rival human performance or exceed human performance at some domains, maybe that term has served its purpose and is no longer useful. So we need to refine it to something more specific. Now, that's my defense of it. I think to your point, it is absolutely true that this term is really fuzzy. and we're going to realize again as we get closer to that threshold that we're going to need to refine it further.

25:03It's going to turn out that, okay, sure, Opus 4.8 created Opus 4.9, but there were a lot of humans involved in creating the chips that went into that model and humans are still involved in creating the chips that are going to run for the next generation model and maintaining the data centers where these things... When you take a more holistic view, there are still going to be humans involved in... Which is what you mentioned is one of the key human bottlenecks to all this is the chips. The chips is very much something that you need humans to develop. So the last question I have for you, Rocket, then, is how close are we then to recursive self-improvement?

25:40And Dropik has put out this blog post. Do we have any kind of measurable criteria saying we are a year away, three years away, two months away? It's a great question. I feel like estimates differ so wildly that I don't want to stake a particular position. My cop-out answer is that it's clearly going faster. Even if we don't hit that discrete point, the trend line is clear that models are capable of doing more and more at all stages of that research process. And the complexity and the time horizon of the task that they can accomplish is getting longer and longer. So I think we see that pretty clearly.

26:18and there's no sign yet that that's going to stop, though it could in the future. Great. Well, Rocket, I want to thank you for coming on. As always, that is Rocket True, our AI and robotics reporter here at The Information. Our next guest has previously sold a company to Amazon Web Services. He is now working on another startup, Eon, in the cloud backup space. The company was last valued at$4 billion. I want to bring on Ophir Ehrlich. Ophir, welcome to the show. It's great to have you here. Thank you very much. It's great to be here. So we were just talking about recursive self-improvement, which I'm sure is a topic that you also are geeking out about nowadays.