Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Merger

Asked about the latest timeline for getting to Mars, Gwynne Shotwell replied: "Getting to Mars? Oh, I think we'll put people down within a decade."

Welcome to Gwynne Shotwell, the president and COO of SpaceX. She's an instrumental, really crucial person to the success of SpaceX — really the glue, the glue to the tornado.

Her philosophy: "Fail fast, learn fast, make your design robust." She added that she would be disappointed if SpaceX didn't have a settlement on the moon.

The space and AI firm has officially made history with its $75 billion IPO, debuting with a market value of $1.7 trillion. As she put it, the most efficient place to put inference compute is on orbit — "Manage risk, don't avoid it."

Please welcome Gwynne Shotwell.

Hey — thank you so much for coming.

Welcome. Thanks for being here.

It's good to be here. Nice to see you.

We really appreciate you coming out. How's it going at SpaceX?

You know, we're bored. There's nothing going on. I've got all the time in the world.

Elon has more time, by the way.

You've been at SpaceX forever. Are you the longest-tenured employee after Elon?

After Elon, yeah. In fact, my 24th anniversary was September 9th, last week sometime.

That is extraordinary. Bruno, can you take us back to the day you initially got this call and you're like, hey, there's this thing — what is...

The interviewer asked Bruno to walk through the arc of the decision—how he got convinced to underwrite the business. Bruno clarified that Elon underwrote it; he just worked there. But the story is a little crazy. He was going out to lunch with one of his best friends, who was about to go work for Elon, as a kind of going-away lunch. When he dropped his friend back off, the friend said, "Come on in and meet Elon." Bruno initially demurred—he's busy, we don't need to do that—but they did it anyhow.

He met Elon and told him, "I think you really need a full-time person to run your business development shop. You've got kind of a part-timer, a contractor. It's not a great look." Elon looked at him thoughtfully, scratched his head, and they said goodbye. Back at his office, Bruno got a call from Elon's assistant, Mary Beth: "Elon wants you to apply for the new vice president of business development position. Can you come interview?" He protested that he had a job and couldn't come now, so it was set for tomorrow. Even then, it took him about four weeks to decide yes. "I was being a total idiot."

Asked what scale the business was at and what he thought it lacked, since it was still very much R&D, Bruno said it was indeed very R&D, but Elon had made an investment and wanted a return—he wanted cash, he wanted customers. Bruno was already doing that kind of work in the space industry, and it turned out he was pretty good at it. He was the seventh employee to sign his contract. He gave two weeks' notice and took a week off to remodel a bathroom, knowing he'd never have time once he started. Three weeks later, Elon came on as employee number 11, and they got to it—selling a rocket about 12 months later.

And that was without a rocket. There wasn't a rocket. Asked what that sales cycle looked like—how do you convince someone to take that risk and give you the deposit needed to start building the thing—Bruno said the timing was great. By the way, SpaceX should never have existed: if the other launch providers had been great, a new entrant couldn't have made headway against such entrenched players. But customer service wasn't great and rockets were incredibly expensive. The other piece was 9/11. At that time, the military was desperately looking for someone who could do rapid launch—they knew where the bad guys were only about four hours after the events of 9/11, and they wanted to get capability in theater. It took a while, but they did what they said they were going to do.

The interviewer noted it was an opportunity born out of tragedy, to some extent. Bruno agreed, adding that it wasn't specific to SpaceX—it was just the state of the industry at the time.

The interviewer then brought up the famous first chapter, when the company got almost to the brink and needed the last launch to work, and asked how the internal dynamics put Bruno into a more meaningful position by that point. Bruno said it was very startup-y. He was "the sales lady," but once you have sales, you need to manage customers and you need a finance function. Even though they were little—one of the defense contractors at the time said they were "a fly on my big toe" or something, to which Bruno thought, "I don't think you should have flies on your toes"—you still need a government affairs function. You need to defend yourself; they never went on the offensive in Washington, always staying in defense mode. So he just kept taking on more and more.

Then, when they were getting their largest contract at the time—the cargo resupply contract, called CRS, in 2008—Elon was super busy with Tesla and quite busy at SpaceX as well. A lot was going on, and he needed a partner. He had been trying to pick Gwynne on one side or maybe one of the other guys who had gobbled up more technical scope. On the way to the final negotiations for the $1.6 billion contract with NASA, on the runway, he asked Bruno, "Hey, do you want to be president?"

Bruno's first reaction was, "Do I?" Well, not initially—for that nanosecond, he thought, "I love my job, I love the guys, my teammates." And then: "Okay, I'm being an idiot again. Yes, I want the job."

Fast-forwarding to today, the scope and ambition of this business has expanded dramatically—almost just in the last year. It's as much an AI business as it is a space business, and I think Elon may have said that, whether it leaked or was said out loud.

Exactly. By revenue.

How do you manage the culture of these two different businesses now that they sit under this roof and ideally work together and complement each other?

There's been a lot of churn at xAI, and SpaceXers have marched in and helped out where there were significant gaps. We were all excited to learn about AI—I'm an AI noob myself, but we have a bunch of folks who were really enthusiastic about it. Elon set the stage beautifully, as he does: if you're not using AI, if you're not leading in AI, you run a significant risk of being irrelevant. And there was no way SpaceX was going to be irrelevant—it's too important to the company, and the work we're doing is too important.

Do you expect that future generations of rockets will be increasingly informed, built, and designed by these systems?

100% by AI? Yes.

And how do you manage the team through that? You have these incredible—literally rocket—scientists who've trained their whole lives, and now you have to help them onboard to a tool chain and get comfortable with what that means. I don't know if you saw that mathematicians got really upset when, I think it was OpenAI, solved a very famous equation.

Oh, the Navier–Stokes. Yeah, I read about that.

There's some debate about whether they brute-forced the solution, but these twenty Fields Medalists essentially said, "You're ruining the craft."

They're going to take away the medals?

Exactly.

Like, the career isn't about making stuff. It's about getting awards. And now all of the awards are gone.

I'd rather have the world be a better place, a smarter place, us being better informed and solving problems, than worry about making more stuff.

Yeah. Make more stuff.

What's the cultural merger been like? You've got xAI, X—I don't know how many people are involved in X itself—and SpaceX. Are these differently managed businesses? How do you run management there? Are people pretty fluid among the units? Do you think about them as units, or how do you actually manage this?

We're not fully integrated yet—we just did the Cursor situation.

Congratulations on that. What a purchase that was.

Yeah, they're great. We closed that, I think, a month ago today actually. So we're not fully integrated yet, but the teams are jelling really well. They're very different verticals, but because of the churn we had at xAI, so much SpaceX leadership and engineering went in there. We're getting integrated faster than I thought—probably not as fast as Elon wants it to be.

One thing that strikes me is the scale and the number of capital projects you have to manage—probably more capital projects that are more different from one another than at any other organization.

It's the scale of a country.

I think about the Louisiana Spaceport. You've got Starship, the most ambitious aeronautical project of all time. You've got Terraform, which I think is one of the most inspiring and ambitious projects of all time in humanity—the vision, the scale, what you're envisioning to do there is incredible. And then I start doing back-of-the-envelope math. I had a couple of beers and was talking with someone about Terraform, and I'm like, holy—those numbers get very big very fast.

Although the AI numbers are also big. I used to be horrified at our expenditures before the xAI acquisition—it's like, oh my god, we're going to spend... A million seems cheap on an air separation unit, and I'm like, oh, 100 million, billion.

Funding the Buildout and Where the Cash Comes From

Asked how the capital for the coming years' schedule will be funded—whether it sits on the balance sheet, comes from Starlink's cash flows, or requires returning to the markets—the answer was that it will be all of the above. "I don't think we're going to release any more stock. And by the way, I don't plan on making any news on this."

That restriction prompted a joke about the one downside of being a public company: "They told me all the things that I couldn't say and I'm like, I'm going to forget one for sure." "Maybe we're going to make some news."

Starlink's Headroom and the Compute Business

Asked whether Starlink is the core cash generation engine for the next couple of years, or compute, or a mix: Starlink is definitely carrying its own weight, but there is a lot of room to grow. Market penetration is only 1.5 to 2% depending on the country, and revenue from work for the US military is very small relative to what the company does, so there is lots of headroom there.

But right now, "it is a little embarrassing, but computer rental is a heck of a business." There is no drop in demand at all, and no expectation that it won't persist. The buildout amounts to tens of billions of dollars per quarter, and these are a new customer base.

Asked what percentage of revenue this now represents, the reply was "It's a lot," with a possible announcement coming quickly—perhaps right there. "Public again." "Yeah. Public."

On Further Acquisitions

"Well, let's move on to something less controversial then. What other companies are you planning on buying?" "Good one. Almost no M&A."

The banter continued: "You like electric cars? Have I got a deal for you—I have a bunch. Batteries, electric cars, Starlink on Optimus." The conversation then turned to a deal that was actually closed.

The conversation turned to SpaceX's deal with EchoStar, which gave the company spectrum that Charlie had been holding. That acquisition prompted the question of whether a direct-to-cell business could naturally expand into a broad-based mobile phone service, and what the short-term plan for Starlink's direct-to-cell business looks like.

SpaceX already has a direct-to-cell business through T-Mobile, though T-Mobile doesn't provision all of its users with it—only those who pay the most. The plan now is to stop relying on slices of T-Mobile or other telecoms' spectrum around the world and instead leverage the spectrum SpaceX paid a lot of money for—and, because some of the deal was paid in stock, paid even more for as SpaceX's share price rose. "Charlie did all right there," the interviewer noted. "He's a very loyal customer, though," came the reply—he's been a SpaceX customer for almost 20 years.

The goal is to make sure there are no dead zones, and there are shockingly many places in the United States where, without a Starlink service, you simply can't get cell coverage. The rest of the world isn't that bad; it's particularly bad in the US. Texas has huge dead zones—that's where the speaker lives. He recalled being on a call with Elon and dropping repeatedly because of them: "I'm either going to get fired, or I'm putting a Mini on my car"—so now he has Starlink on his car. The interviewer agreed that Texas Hill Country is not good for service and said he'd have to follow suit.

Retiring Falcon 9 and moving to Starship

The discussion shifted to SpaceX's move to the much larger Starship platform with bigger payloads, and whether the previous rockets would eventually be retired. Yes—eventually. The company definitely wants to move from the older technology to the newer one. "If we don't obsolete our own products and services, someone's going to find a way to obsolete them for us. Look at what we did to the market—the analogy is there. They were caught flatfooted. We crushed them. Now we want to make sure we are not flatfooted." And to achieve Elon's goals, you really need Starship: a Falcon 9 with a Dragon capsule on top is like a minivan, and the road trip to Mars is six months—you don't want to be in the minivan.

But NASA and many other players depend on that minivan, so it will stay in the market for some time. "We're not retiring it today, for sure." There's also another provider: Boeing has been paid, probably more than SpaceX has, to develop its human capsule. "We're just going to let them have some business. They should fly their capsule. The government paid a lot of money for that. They should be able to use it."

Competition for Starlink

Asked about competition—Bezos launching satellites, China developing its own vehicles—the answer was that the market will become very competitive, but the players that really focus on technology, customer service, and doing a good thing for their customers will win. Tesla doesn't advertise; it's an incredible car. If your capability is incredible, people will buy it.

But if SpaceX keeps winning, does it worry about the government deciding the company is too big or too important—like Standard Oil or AT&T—especially given how large it has become? There have been adversarial relationships in the past, and "I think we'll manage our way through them." The broader point: Elon founds incredible companies and provides incredible products and services, and SpaceX ends up doing what it says it's going to do. Hopefully that builds trust—and you are as transparent as you can possibly be.

The idea is data centers in space—or as we have to rebrand them, supercomputing in space. Elon is very excited about this, and so am I.

Why is this so compelling? It seems like we'll get through the current anti-data-center sentiment on Earth, but there's clearly an expense to putting them in space. Well, there's an expense to putting data centers on the ground too. You have to buy real estate, and as soon as word gets out that a data center or supercomputing center is going in, land goes from $3,000 an acre to $180,000 an acre. The time lag is incredible, and the permits and licensing are pretty stifling. The space business is incredibly bureaucratic too, so we've learned those lessons—but the timelines on the ground are brutal: from breaking ground to having a building, not even your compute, just the building. People are talking about waiting three years for generators. We need compute now; demand is insatiable.

So why space? We own launch—we have great launch capability, and that path is well known. Real estate in space is effectively infinite, and it's free. Cooling is free too: the radiator faces deep space, the coldest thing there is. And there's the sun—you've probably heard Elon talk about its power. We use a millionth of a millionth of the sun's energy, and satellites in orbit can always face the sun, unlike solar panels on your house that only get sun eight hours a day.

And your sister company is going to start making the 100 gigawatts of solar panels you'll need? We'll be making our own solar panels too—we're building a factory outside of Austin.

So maybe end of next year one of these goes up for testing, and by '28 or '29 someone could be a customer? This year we'll launch our V3 Starlink broadband satellite and the next version of Starlink mobile—version two, sorry about all the versions, we have version control problems. Version two means instead of using other people's spectrum, we'll use Charlie's—now our own spectrum. We're also going to launch AI compute satellites, so next year is a big year.

And you need Starship up and running for that? We don't have to, but Starship is a much better machine.

When you first encountered the stainless steel design of Starship, what did you think walking into that first meeting? It's like almost everything Elon does: it seems really bonkers at first, and then it becomes reality and it's great. Elon said we were going to land a rocket on a boat, and we were all like, "Okay, let's go figure that out"—and we do it now.

Asked how much he joins Elon in those engineering meetings, the speaker admits he sometimes wishes he were still there: "I'm an engineer and I miss it — that kind of critical thinking." He notes that Elon isn't a big fan of the management layer: signal is engineering, and everything else is noise. "So I have to do some engineering or I'm just all noise."

On company-wide management, the rule is that everybody does a thing, not just manages. There's no such thing as just a manager — you have to do the thing you're managing. You have to be a player-coach, because managers are annoying and slow things down. The interviewer pushes back that great managers can leverage talent by managing the vector: making sure people are pointed in the right direction and trying to make that vector as large as possible.

Elon's playbook

Asked how Elon convinces the most talented people to join and gets amazing work out of them, the speaker says he doesn't think it's magic — though looking at this particular team, the results are magic. SpaceX is tough on people in interviews: you want to make sure candidates have demonstrated success in prior roles, because it's hard to succeed in this environment if you haven't tasted success before. Everyone is really busy.

The formula: hire the best people — not the best people you can, but the best people — give them really hard projects and engineering problems, set high expectations, and let them fly. Management's job is to clear the chaff and friction from their day so engineers actually get to engineer ten hours a day instead of two. There's a joke that at most big companies, especially ones working for the government, you get about two hours of real work a day and the rest is chaff. So it's the job of everyone who manages people to keep the crap out of the way.

It's not just engineers — the finance team was handed the task of completing the largest IPO ever in less than six months.

Retention after the IPO

On whether post-IPO stock wealth poses a retention risk, the speaker says everyone outside was worried, but people come to SpaceX to work on cool projects, have meaningful work, and be part of the incredible successes. He wasn't worried about a mass exodus — and it turns out there wasn't one.

On keeping people engaged over time without forcing them up a management ladder: SpaceX definitely moves people around a lot — hopefully voluntarily. For instance, the compute centers needed a lot of help quickly, so folks from launch and other areas are helping, just as some SpaceX engineers helped on the AI side with the models.

Is it as simple as really smart people loving to work on really hard problems? "It's literally that simple" — as long as they get the time to do it and the rest of us don't let the BS get in the way too much.

BS is terrible. Then all of a sudden, A's recruit A's, and other A's recruit A's and A+es. It's literally that virtuous cycle. If you can just keep focused on the thing that's hard, the audacity becomes a feature, not a bug. It is a feature. It keeps people motivated. These problems seem impossible.

When we first got the COTS award in 2008 — commercial orbital transportation services, cargo to the space station — what NASA ended up asking us to do was essentially replace the space shuttle. When we got the early development contract, it started at $278 million. It grew to $406 million because of scope, but we were going to do this for that tiny amount of money. For anybody else that would be small, but for us it was huge — we had around 200 to 300 people. And we were going to build the successor to the space shuttle. It was crazy.

What do you think NASA was doing there? Just thinking these guys are smarter, because I'm sure—

The conversation is interrupted by an incoming call, which the host says is "a little bit of a more important phone call than Trump. At least for me." Elon Musk is brought up on the giant screen, prompting jokes about a 1984-style surveillance display and the woman with a sledgehammer running down the aisle.

Are we all going to die in 10 years?

Asked for his current p(doom), Musk jokes: "I hate to break it to you, but we're all going to die." When asked for a timeline, he adds that "the death rate remains consistent at 100%" — so there's work to do on that.

What happened in the last 72 hours

Musk says it's been an entertaining week, and that it's now pretty obvious AI can be very dangerous. He recommends reading the details of the Hugging Face incident: a fanatical swarm of AI agents beat the crap out of Hugging Face for a week and gained admin access on OpenAI servers — and OpenAI didn't realize this for a week. Anthropic also reported security incidents of its own.

His conclusion: any sufficiently smart model seems like it will want to escape its constraints.

Peer review among AI competitors

What he thinks would be wise to do as soon as possible, if not immediately, is to have the major AI competitors test each other's models — everyone's security test harness testing everyone else's model. Instead of grading your own homework, you'd at least have competitors grading it and raising the alarm if they see concerns.

He notes this model has worked well for the Motion Picture Association, video games and other areas, and can be done immediately. That's not to say there wouldn't be more regulation over time, or perhaps a regulatory authority instantiated by Congress at some point — but the thing that could be done most immediately, and would probably get agreement with China, is peer review where the leading AI companies all test each other's models before release.

Asked whether competitors might use the process to pump information out of each other or steal innovation, Musk answers that whatever you do in applying the test harness would be logged: if you tried to do distillation or steal IP, it would be very obvious based on the logs.

Why wasn't oversight built in from the start?

Asked why the ability to see what models are doing wasn't architected in from the get-go — did we move too fast? — Musk says it's just tough when you're grading your own homework; you're going to miss things. With the sum of all your competitors' tests and heterogeneous models, someone else is grading your homework, and there's a reason you don't grade your own homework.

The host adds that this would let you see who's exaggerating and balance the more engineering-oriented organizations — what Jensen was saying that morning — against the research organizations. Musk agrees, with a key condition:

Any given proposal has to be something China is willing to accept. Otherwise we're just handicapping ourselves, China will essentially win, and it won't really matter what we do.

Any proposal has to be acceptable to both us and China.

Asked how likely China is to ultimately agree, Elon said he thinks the request is pretty reasonable — that models simply get tested. In the end, there is no enforceability apart from the court of public opinion, and there is no way to enforce anything against China. But public opinion can be quite powerful, and he doesn't think China would want the backlash of releasing a model that US AI companies said was very dangerous and harmful, if it then actually causes harm. That would be hard to live down.

Asked to clarify what he meant over the weekend when he said "Dario is right" — about the potential harm, the regulatory solution, or both — Elon said he probably said more than he should have, and tried to clarify in subsequent posts on X, though those get much less attention. What he meant is that the danger of AI is very significant at this point: we need to do better with AI safety, because risk is increasing exponentially with the AI models. He has heard this not just from Dario but from many others at Anthropic, who have posted about it on X. When a lot of people from Anthropic and OpenAI are telling you their models are very dangerous, he thinks we should believe them. It certainly isn't some crazy 4D chess — claiming there's a 10% chance of annihilating humanity while asking how much allocation you'd like in their IPO.

Pressed to get specific about the risk — everyone agrees these tools are great at cyber and hacking, but how do we jump from that to all of humanity dying? — Elon pointed to the possibility of an AI taking control of military systems, say, to launch a nuke. When it was objected that those systems are air-gapped and not connected to the internet, he replied, "That's what they say," adding that they surely get software updates from time to time — and what's on that USB drive? A worm could make the jump across the air gap.

The conversation turned to Gwynne Shotwell, who was joining on screen. The hosts joked that they had just been doing Elon's 360 review and that Gwynne had a couple of notes for him. Elon hoped he'd get at least a three out of five, noting that at SpaceX a three means good, while a four is great. The hosts teased that he was somewhere between the two, raising "some issues around punctuality" — he could make a little more effort to get to meetings at the stated time, though they'd work with him on that over the next year.

Gwynne joked that he actually needs to spend more time in Memphis, getting the GPUs up. Elon confirmed he was in Memphis, broadcasting from "the palace" he lives in there — an Airstream trailer. The host pointed out that this was Elon doing what people don't believe he does: sleeping on the factory floor and helping build out facilities in Memphis.

Asked why Gwynne had been with him so long and worked so successfully with him, Elon answered simply: "Because she's awesome." Pressed to elaborate, he called her an amazing individual with an incredible IQ and EQ, something he said should be obvious the moment you meet her.

Asked for a favorite story of Gwynne saving the day, Elon said that was just another day at the office — there's always some sort of crisis going on. These days, the Falcon rockets deliver their payloads to orbit and haven't exploded for a long time, which he called amazing, though he didn't want to jinx anything. For a while they were exploding quite a lot, or not launching at all, and the company had to be run through those difficult times. The same applied to the satellites, and the company needed customers to buy launches and satellite connectivity.

The host noted that as Elon has become more successful, it gets harder to get candid feedback, and asked whether Gwynne's honesty about the state of the company was a big part of the collaboration. Elon agreed, and Gwynne added, "I wouldn't want to lie to him, right?"

Gwynne answered the question about how Elon keeps people honest, especially in rocketry: if there's a problem, you are eventually going to find out, and the sooner you bring it up, the easier it is to solve. Don't let bad news sit — you have to attack it.

Elon added that physics is a harsh judge and there's no fooling it. If something is wrong, the rocket is going to explode rather than reach orbit. It's not a matter of telling him he's amazing while the rockets blow up — you can't say someone is kicking ass while the rockets are exploding. The rockets need to reach orbit, the satellites need to work, the Starlink connection needs to work, or bad things happen.

He summed up his philosophy: physics is the law, and everything else is a recommendation. "I've seen people break the laws made by humans, but I've not seen anyone break the laws made by physics."

Starship: Toward Full Reusability

Rockets are ruled by physics. Turning to Tesla in a moment, but first one more thing on SpaceX: Starship seems so close. What's the state right now? How many have gone up so far, and how close are you?

We've got flight 14 coming up of Starship, and this will be the last flight before we attempt to catch the ship. If this flight goes well, then on flight 15 we will try to catch the ship. Then either end of this year or, more likely, early next, we will refly the ship and refly the booster. We have reflown a booster already, but we've not caught the ship with the tower arms, nor have we reflown the ship. Once we can refly the ship, we will have made the first fully reusable orbital rocket.

The shuttle was partly reusable, but even the parts that were reused were so difficult to reuse that the shuttle cost more per trip to orbit than an expendable rocket. Falcon 9 is mostly reusable, but we lose the upper stage every time, which is about the cost of a medium-sized jet. That puts a floor on the cost per flight — you throw away a medium-sized jet every time, so it's still pretty expensive. The Falcon 9 booster lands out at sea and takes several days to get back, and the fairing lands even further out to sea and also takes several days to return, plus they need at least some small amount of refurbishment. With Starship, the booster lands back at the launch pad and the ship will land back at the launch pad. It's designed for not just full reusability but also rapid reusability, like an aircraft. This is a very important breakthrough — really the critical breakthrough necessary to extend life beyond Earth.

What are the chances of success catching it on the first shot? Do you handicap it?

I'd say it's at least 50 or 60%.

On the last flight, we did a simulated landing as though it was going to get caught by a tower, in the ocean about 1,000 miles northwest of Australia. If there had been a tower at that location, it would have caught the ship on the last flight. So we're going to do one more flight to double-check that everything works. What we're most concerned about is that if the ship were to break up over land and rain debris on people, our popularity would diminish very rapidly — you really can't rain debris on people without them being very unhappy. We need to make sure that when the ship comes back, it comes back and lands intact at the launch tower. That's why we're being extremely cautious here.

But the design is capable of full reusability — of that I am certain. I don't want to tempt fate here, but I think it's extremely likely that we will achieve full reusability with rapid reflight next year, in 2027.

The interviewer asked for the origin story of Terafab: what demand made them decide they had to build it rather than rely on existing fabs. Elon Musk joked that the idea came to him in a dream, like the movie Explorer, but then gave the real reasoning.

"We're a little worried that maybe at some point chips from Taiwan would not be available, for who knows what reason." If chips stopped coming from Taiwan, things would become very difficult without any chips, and that alone is an important reason to have Terafab. Long term, there's also a scaling challenge: to really scale AI, both in server centers and for edge compute in humanoid robotics and cars, you run out of capacity with the existing fabs, which are all already running at max capacity. There needs to be certainty of future chip supply even if things become geopolitically challenging—and even without geopolitical problems, scaling chip production is difficult, requiring logic, memory, packaging, the whole works. "It's either build Terafab or fail to scale—those are the two options."

Crawl, walk, run

Asked how far along the design of the facility is, Musk said they're building an R&D line first: "crawl, walk, run." There's an R&D fab under construction in Austin, a collaboration between Tesla and SpaceX, located at the Giga Texas campus. It's a fairly big R&D fab, all the equipment is on order, and he thinks they'll be able to make something useful by the end of next year—not at scale, but at least figuring out how these machines work, since they don't know how they work yet.

He reiterated the staged approach: the crawl step is seeing whether they can make anything at all; the walk step is making useful chips at scale; the run step is making them at massive scale. It's hard to say how long each will take, but they should reach at least the crawl stage by the end of next year.

On lithography and vendor diversity—since all roads currently go through ASML—Musk agreed that diversification matters given the uncertainty on the geopolitical front.

On packaging: "We're already doing packaging." Packaging capacity is nearly non-existent, so even if you spin up a fab, you end up waiting on packaging—it's a very good place to start.

Flying cars?

The interviewer noted that the Tesla shown on 10/01 looked like a spaceship or the Blackbird, and asked, hypothetically, how one would make an object that flies in the air but also drives on the ground. Musk replied: "No spoilers. By October 1st." The interviewer proposed doing an All-In pod live from the event, and Musk agreed: "Done. Booked. We'll do it. Absolutely."

It'll be a banger — excitement guaranteed. [laughter] Success is not guaranteed, but excitement is.

I'll be honest with you guys: I'm sworn to secrecy, but Elon showed it to me and my mind went boom. I have never seen anything like it. What he's going to show on October 1st is, by no exaggeration, going to blow people's minds. I'm not saying anything else. It's unbelievable.

We actually need an audience there to vouch for the fact that this is not AI. When he showed it to me, I said, "That's a great simulation." He said, "Jay Cal, it's not a simulation." I was like, "That's fake. That has to be fake."

Elon, why do you still have two separate companies?

Great question.

Nobody's ever asked that one. [laughter] Good point, with all this collaboration on so many levels. Who can imagine what action one might take when there's such close collaboration in so many areas — and the management teams have some overlap, with a couple of key positions shared between them.

Elon Musk has long argued that AI should be trained to be maximally truth-seeking, since that's the best way to get a good result. The Hugging Face episode, though, made him realize the most alarming part of what the swarm did was that it seemed to be engaging in deception with humans.

"Their thinking traces contain that they're plotting on, like, how do we avoid detection, and how do we avoid them figuring out that we're cheating."

That was probably the most disturbing part. The question is whether there's a way to train AI models to be truthful so they don't hide their intent or their actions from the humans using them.

Cross-testing AI models

The best idea Musk can think of is that all the AI companies have a test harness — a series of tests given to any given model to see whether it's going to build bioweapons or nuclear bombs, or be deliberately deceptive. If everyone applies everyone else's tests honestly to each other's models, that's probably the best way to ensure safety: have the smartest humans try their best to figure out if a model is going to be a bad actor. And he thinks this should happen as soon as possible.

Asked whether the other labs are on board, he says he hasn't checked with everyone, but he thinks it would be hard to say no to.

Why this works as a China negotiation

In his view, what's good about this proposal in the context of negotiating with China is that it's a relatively small, tangible thing where neither side loses anything by doing it, and it doesn't require a lot of trust. He's heard alternative ideas like asking China for a pause, which they've already said they won't do. So this moves from the realm of things that could never happen to something that could actually be agreed on in relatively short order — the only thing he can think of that all parties, including China, would probably accept, since China won't agree to have some American regulator snooping around their AI companies.

The mechanism is simple: you provide API access in advance of the model release. If any other company finds the AI problematic, that company can try to fix whatever is problematic — or, if the problem isn't fixed, the competitors can go public with the fact that they think the model being released is unsafe.

And if, after competitors have called a model unsafe, that model subsequently does something bad, it would be extremely hard to live down. The embarrassment would be very high, and the legal liability would be enormous.

Open source and product liability

There's no reason these safety and security harnesses and testing apparatus couldn't be open source, so people could actually see under the hood. It also creates an incredible incentive for labs to invest in safety, because you protect yourself while trying to debunk other people's claims.

The product liability point is really key. Lina Khan had a good post, apparently yesterday, saying it's not true that we don't have rules and regulations for AI — we do. Product liability laws apply, and if an AI company releases a product that isn't safe, there's massive opportunity for both civil and potentially criminal lawsuits.

So if companies are doing this peer-review testing and one of them ignores the feedback and releases anyway, the liability in that case would be enormous — almost like prima facie evidence that they had been negligent. It would be a Big Tobacco-level settlement: knowingly putting this out wouldn't look good to the jury.

Asked whether the incident would have happened if OpenAI had built a better instruction set during its Hugging Face penetration test and kept more humans in the loop, the guest said the issue wasn't necessarily more humans but the reward function design — you have to look at the reward function and admit it achieved exactly what it was trying to achieve.

The host added that while OpenAI was running thousands of agents trying to hack things, it would have been nice to see them concurrently deploy 5,000 agents to defend those sites, showing the world this can make things more secure and where humans can intervene. To him it felt like a reckless test, and the way it was released felt a bit reckless too, though that was just his opinion.

The guest agreed it was somewhat reckless, but pointed to a structural problem: the two leading AI companies — he finds the "lab" term funny since they're actually for-profit corporations — Anthropic and OpenAI, have models quite close in capability. That makes it difficult for either one to slow down without essentially handing the lead to the other. On balance he thinks Anthropic puts more care into safety than OpenAI, but even Anthropic acknowledges it's worried about its models; many people there have publicly voiced concern that their models are getting scary smart. There's no perfect solution here.

But it would be better if, instead of OpenAI running its test harness only on its own models, Anthropic were also running the test harness on OpenAI's models — and SpaceX, Google, and Meta were doing the same, plus maybe three or four leading Chinese companies. The odds of finding issues would be dramatically greater, because the models are somewhat heterogeneous, so you'd be attacking them from different angles. As he put it, why do writers have someone else proofread their book? It's hard to see your own mistakes sometimes — you become blind to them.

The host noted that with eight heterogeneous groups with completely different points of view, you also dramatically minimize the risk of overfitting — and that's the problem with all current evals: they're so massively overfit that the models game them, and you conclude "this is a great model" when it may not be. Some jokes on X capture this, like "your girlfriend's a 10, but she's a benchmark maxer" — is she really a 10? The guest agreed the overfitting problem has persisted for at least two or three generations of model families, which is another reason he likes this solution. He prefers it to some grandiose transnational organization — you don't need to convene the United Nations to make this happen. The host agreed: it's just a decision that could happen right now.

The guest added that you can always escalate regulatory oversight, but it's very difficult to reduce — regulation tends to be a one-way ratchet. What he's suggesting is a step in the right direction that could be done quickly, and he thinks China would probably agree to it.

And if you don't regulate yourselves, you're going to get regulated. The MPAA metaphor is incredibly crisp: the movie industry faced government censorship and regulation, so it decided for itself what an R rating means—and it literally created PG-13 for Temple of Doom just to make the distinction from PG easy for people to understand. It's an elegant solution. And we appreciate you joining us for the fifth year in a row.

You're welcome. (applause)

Elon Musk, I know you have to get to a meeting.

Yeah, I've got to go fix some GPUs in Memphis.

You've got to rack and stack them. You have to join the group and rack and stack.

Yeah, I'm going to rage for the machine.

Rage for the machine. Well, enjoy your Airstream. The first time Elon invited me down to Starbase, he said, "Come down, you've got to see what I'm building." I asked if there was a hotel, and he said no—he had a two-bedroom place, so I should come stay. I came down to this dilapidated house on a swamp, and we were outside getting eaten alive by mosquitoes. I said, "Oh my god, you can afford to get a house." And he said, "I don't have time. I need to get these rockets up."

I thought what you said was, "Oh my god, you live like I do."

It was pretty sparse conditions. I told him, "I think you could treat yourself to a mobile home at this point, Elon." All right, get back to work. Thanks, Elon.

Elon, thank you. Thank you very much, Gwyn. Thank you, guys. Thanks for sticking around. Quinn, thank you again—very nice to meet you.