All-In Podcast — Elon Musk & Gwynne Shotwell on AI risks, Starship, Terafab and a SpaceX/Tesla merger September 15, 2026 Elon Musk joins Gwynne Shotwell on the All-In Podcast to discuss the real risks of AI and peer review of models, what he meant by “Dario is right”, his working relationship with Shotwell, Starship's future, Terafab, flying cars and a possible merger of Tesla and SpaceX.
Elon Musk joins Gwynne Shotwell on the All-In Podcast to discuss the real risks of AI and peer review of models, what he meant by “Dario is right”, his working relationship with Shotwell, Starship's future, Terafab, flying cars and a possible merger of Tesla and SpaceX.
Transcript Sorry to interrupt. I'm getting a call.
It's a little. It's actually on the margins, I think, a little bit of a more important phone call than Trump, at least for me. Hello, Bestie, Bestie.
I'll put them up on the screen. How you doing?
I mean, this is a little 1984ish, this gigantic situation.
Yeah, the woman with the sledgehammer is going to come running down the main aisle here.
Exactly, yeah. Put a hole in the screen.
So are we all going to die in 10 years or not? It's the topic of discussion here. What's your p doom right now?
Well, I hate to break it to you, but we're all going to die.
Yeah. The death rate remains consistent at 100%.
We've got work to do on that. What happened? What happened in the last 72 hours?
Yeah, it's been. It's been quite an entertaining week, really.
Well, it's. It's pretty obvious at this point that AI is. Can be very dangerous. And I recommend reading the details of the hugging face incident. It's intense. So you had, like, a fanatical swarm of AI agents that beat the crap out of Hugging Face for a week and gained admin access on OpenAI servers. So who knows what it actually did. May have done things beyond that. OpenAI didn't realize this for a week. Anthropic has also reported, you know, some security instances themselves. So basically any specially smart model seems like it. It will want to escape its constraints. I mean, what I think would be wise to do as soon as possible, if not immediately, would be to have the major AI competitors test each other's models so that you would have everyone's security test, harness testing everyone's model. So, you know, instead of kind of grading your own homework, you would at least have competitors grading Your homework and raising the alarm if they see concerns. And I think this model has worked pretty well for motion picture association, for video games and other things. And it's something that can be done immediately. That's not to say that there wouldn't be, you know, more regulation over time or that there would not be at some point perhaps regulatory authority substantiated by Congress. But, but the thing that we do most immediately and probably get an agreement with China would be peer review where the leading air companies or test each other's models before release. I think that would be a wise variable.
Any concern that people might be using this to pump information from each other and you know, corporate stealing of innovation, et cetera. Like in terms of the implementation for this idea?
Well, I think in applying the test partners, whatever you do in applying the test harness would be log. So if you try to do distillation or steel ip, it will be very obvious based on the logs.
Got it. And understanding what these models are doing hasn't exactly been built into the system from the beginning. So from your perspective actually being able to see the work that's being done, why wasn't that built into the models from the get go? Did we move a little too fast in architecting these?
Well, I think it's just tough when you're grading your own homework, you're going to miss things. Whereas I think if you have the sum of all of your competitors tests and you've got heterogeneous models, then you're not creating your own homework. Someone else is creating it. And there's a reason why you don't create your own homework.
And then to your point, what it allows you to start doing is to figure out if certain people are exaggerating and certain people are have a different approach. And now you can have the more engineering oriented organizations. This is what Jensen was saying this morning, Elon versus the research organizations. They'll be a little bit more imbalanced.
And any given proposal has to be something that China is willing to accept. Otherwise we're just handicapping ourselves and we'll just find it that China will essentially win and it won't really matter what we do actually. So it's got to be something that's acceptable to U.S. and China.
Elon, you said that you thought it was possible that they would agree or you thought there was a good chance maybe. I mean, how likely do you think it is that they'll ultimately agree?
Agree. I think this is, I think this is a pretty reasonable request that, that models Just get tested. I mean, at the end of the day, there's not enforceability here, apart from the court of public opinion,
there's no way we would have enforceability against China. But I, I think the court of public opinion can be quite powerful. And I don't think China would egg on its face for releasing a model that US AI companies said was very dangerous, would cause harm. If it then causes harm, that's going to be hard to live down.
So Elon, this weekend, when you said Dario is right, did you. Did you mean Dario is right about describing the potential harm? He's right about describing the regulatory. Solve both, the first, the latter.
Can you help us understand when you said that? Because I think it was sort of a moment where
I probably should have said, right. I did try to clarify it in subsequent posts on X, but those get much less attention. So what I, what I meant by he's right is that the danger of AI is, is very significant at this point that we need to do better with AI safety or we have at this point exponentially increasing risk with the AI models. And I've heard this not just from Dario, but from many other people at Anthropic. And in fact they posted on X about it that, you know, when a lot of people from anthropic and from OpenAI are telling you that their models are very dangerous, I think we should believe them. It certainly is like some crazy 4D chess to, you know, say these. There's whatever 10% chance of annihilating humanity. But by the way, how much allocation would you like in our IPO?
that's some crazy 4D test, you know. Well, let's get knew what dimension that's coming from.
Can we get specific about the risk though, Elon? Like, obviously we see cyber and hacking as an obvious risk. These tools are great at it. But take us from what we would all agree. Okay, yes, these things can cyber hack to how do we jump to all of humanity dying? There's a couple of steps in between these two things happening, I think. Yeah,
well, I mean, if it were able to take control of military systems and say launch a new. Yes, that would be bad.
And these systems are all air gapped, though. These systems are all not connected to the Internet, so.
Yeah, that's what they say. But something tells me they get software updates from time to time. Yeah. And what floppy does.
So yeah, the USB drive has a worm on it and they somehow make the jump in the air gap.
That's not out of the question.
Elon, we've actually got Gwen here today. I think you know that. I think you can see her.
We were just doing your 360 review on Gwen. Had a couple of notes for you.
Yeah, I hope, I hope I get at least three out of five.
Well, there was one out of five
Four is great. So you're somewhere between the two. There was some issues around punctuality that we needed to bring up. Sometimes you could make a little more effort to get to the meeting at the state of time, but we're going to work with you on that over the next year.
She said, I actually think he needs to spend more time in Memphis. Yes, obviously, I know you're in Memphis. You need to be working, getting those GPUs up.
This is coming to you from my, the palace that I live in in Memphis, which is an Airstream trailer.
This is Elon, by the way, doing what people don't believe he does. He sleeps on the factory floor. He's in Memphis helping build buildings.
And bring up Elon, why, why has, why has Gwyn been with you for so long and been so successful working with you?
An amazing individual with an incredible IQ and eq.
And I think that should be obvious from the moment you meet her. Yeah.
Fantastic. What's your favorite Gwen story, when she particularly contributed to the amazing success that you've had in this collaboration? You got a favorite story where she saved the day, just performed exceptionally. That's memorable.
And that's just, that's just a daily, you know, that's another day at the office, frankly. So.
Wow, I need to do this more often.
We need to waste. This is the time.
By the way, Gwen's getting a five.
Gwen's a five out of five.
Yeah, I mean there's, there's always like some, some sort of crisis going on. I mean, these days, the rockets, at least the Falcon rockets, I don't want to jinx anything, but the Falcon rockets deliver their payload to orbit and haven't exploded for a long time, which is awesome. Amazing. But, you know, for a while there they were exploding quite a lot or just not launching at all. So, you know, I don't know. We've got to run the company through these difficult times and, and get, you know, build up of the rocket, make it better, have it not Explode. Yeah, same thing with the satellites. And then we, you know, we need customers to buy launches and buy sort of satellite connectivity and, you know, so.
Yeah, and Elon, as you have become more successful over the years, harder to get candid feedback from folks. And you always run that risk being in the position you're in. My understanding Gwen is super candid with you and able to tell you honestly, like, state of affairs at the company.
And that's a big part of the collaboration.
Yeah, I mean, Yeah, I mean, yeah, I guess.
I mean, I mean, I wouldn't want to lie to him. Right.
Well, I'm curious to know you run
that risk generally at the companies where people might be intimidated. Hey, listen, you're a larger than life figure at this point and how do you keep people continuing to be honest with you about the challenges, the deadlines you set? Pretty intense deadlines. Yeah.
So let me, let me answer that, if you don't mind. Elon, like, 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 going to be to solve that problem. Like, don't let bad shit sit. You gotta attack it.
Yeah. Physics is a harsh judge and there's no fooling physics. So like if something's wrong, the rocket's going to explode and start going to get to orbit. So, you know, it's not like, elon, you're amazing. Meanwhile, the rockets are blowing up, you know. Yeah, it's not, it's hard to say. Yeah, you're, you're really kicking ass here. But the rockets are exploding. That's just not the case, you know, so, so I mean, I mean, just generally, like the rockets need to get to orbit, the satellites need to work, the starlinks, you know, connection needs to work or, you know, or bad things happen. I mean, this is sort of like a physics situation and, and, and physics is a, is, is a harsh judge. You know, I say like, you know, there's, I think, like, like physics is the law and everything else is a recommendation. Like, I've seen people break the laws made by, by humans, but I've not seen anyone break the laws made, made by physics. So that's good. And rockets are ruled by physics.
So can we ask some other questions of other. You have. I need to move to Tesla for a second.
Yes. Well, I wanted to just do one thing. SpaceX, which is Starship.
It seems like you're so close.
What's the state right now? And you, you're really putting these up at a, at a pretty size, pretty answer.
And Elon's answer. Separate and see how they convert. Yeah.
So Elon, how close are we? Gwen, how close are we? Let's get, let's, let's get to it here. Yeah. How many have gone up so far and how close are you?
Well, We've got flights 14 coming up of Starship and this, this will be the last flight before, before we attempt to catch the ship. So if this flight goes well, then on flight 15 we will try to catch the ship. And then either end of this year or more likely early next, we will refly the ship and refly the booster. So we have reflown the booster already, but we've not caught the ship with the tower arms, nor have we reflown the ship. So once we can re fly the ship, we will have made the first fully reusable orbital rocket. So the shuttle was partly reusable, but even the parts that were reused were so difficult to reuse that the shuttle cost more per time to orbit than an expendable rocket. Now Falcon 9 is mostly reusable, but we lose the upper stage every time, which is about the cost of a medium sized jet. So that obviously puts a floor on the cost per flight of like well, you throw away a medium sized jet every time, that's still pretty expensive. And the rocket, Falcon 9, the booster lands down, you know, often I out to sea, it takes several days to get back. And the fairing lands even further out to sea and takes several days to get back. And they need some amount of refurbishment, at least a small amount. So whereas Starship, the booster lands back at the launch pad, the ship will land back at the launch pad. And so it's, it's designed for not just full reusability but also rapid reusability like an aircraft. So this is, this is a very important breakthrough. It's really the critical breakthrough that's necessary to extend life beyond Earth.
Chances of success catching it on first shot. You do you think about that at all? Do you handicap it pretty good?
I'd say it's at least 50 or 60%.
Yeah. So on the last flight if there had been a tower, so we did a simulated landing as though it was going to get caught by a tower in the, in the ocean about a thousand miles northwest of Australia. And if there had been a tower at that, at that location, it would have caught the ship on the last flight.
Yeah, so we're going to do one more flight to just confirm that, that everything. Just to double check that everything works. Because what, what we're most concerned about is if the ship were to break up over land and rain debris on people our populated, our popularity would diminish very rapidly. So you really can't rain debris on people without them being very unhappy. So, so we need to make sure that when the ship comes back, that it comes back and, and, and lands intact at the launch tower. That's why we're going to, that's why we're being extremely cautious here. But the design is, the design is capable of full reusability. Of that I am certain. And I, I say 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.
Significant. Gwen, how do you.
Can we just hear a little bit. I'd love to hear the origin story of Terrafab from you guys. How did the concept. What was the demand that made you say we've got to do this, we've got to build this and not rely on the existing. What's that fever?
It's like that movie Explorer, you know,
that meme came to me in a dream. Why do you think you should do this? You know, I guess it sort of did come to me in a dream. Well, we're a little worried that maybe at some point chips from Taiwan would, would not be available for who knows what reason, but any number of reasons, for some reason, at some point, chips may not continue coming here from Taiwan.
And it would, and that would really make things difficult without if, no, if we didn't have any chips. So that's an important reason to have Terrafab. Then long term, there's just a scaling challenge where if you want to really scale AI both at the service, you know, within server centers as well as at, for Edge Compute, for humanoid robotics and cars, you kind of run out of capacity with the existing fabs. So as it is, all the fabs are running at max capacity. So I think we need to make sure there needs to be certainty of future supply of chips, even if things become challenged, challenging geopolitically. And then even if they weren't challenging geopolitically, there's a scaling challenge with. It's quite difficult to scale chip production. And you really need the logic. The memory packaging, the whole works in order to continue scaling. So it's either build terrafab or fail to scale.
Those are the two options.
Uncertainty on the geopolitical front and how
deep have you guys gone in designing the facility? Is this fully scoped? Is it sort of an outline at this point? To what degree do you have like an actual project plan on dates and deliverables and what's going to be up and running when we've got an R
and D line that we're building first. So I think it's kind of crawl, walk, run.
there's an R and D fab that we're building in Austin that's a collaboration between Tesla and SpaceX. It's at the Austin Giga Texas campus. So that's. And, and that, that's like a pretty big RD fab. So we have all the equipment on order for that and we probably will be able to, I think we'll probably be able to make something useful by the end of next year. Not, not at scale, but as Gwyn said, Coral walk run. We, we got to try to, you know, at least figure out how these machines work. You know, like we don't know how they work. We never built a ship line before.
I saw that you guys, you had some job openings for lithography people and stuff. And you know, obviously all roads currently go through asml, but you probably would want to diversify and, or vertically integrate. I think that you, you've shown a lot of capacity to do that. So is that part of the play as well, Elon? Like you're just going to have to make sure that there's vendor diversity so that if for whatever reason the weather conditions in Taiwan are best, that.
Yeah, yeah, there's, it really is crawl, walk running. So the first step is can we make anything? Like try to figure out if we can make anything. That's the crawl part. And then now try to make useful chips at scale would be kind of the walk part and then run would be like now let's make them at massive scale. So it's hard to say how long it'll take us to do these things, but I think we'll get at least to the crawl part by the end of next year.
We're already doing packaging.
Packaging is really important by the way, because if you look at it, the packaging capacity is like non existent. So even if you spin something, you're kind of just waiting around 10 cuppings. So that's a very good place to start it.
I need to ask a Tesla question. Because we saw on 1001 it looked like a spaceship. It looked like a rocket ship. It's supposed to be a car. There's just the back of it. It looks like the Blackbird. Just hypothetically, if you were going to make an object fly in the air but also drive on the ground, how would hypothetically one do that? If one try no spoilers by October 1st.
October 1st, maybe you could do an all in pod for a Live from the.
Oh, done. Booked. We'll do it.
It'll be a banger. Excitement. Guaranteed success is not guaranteed, but excitement is.
I, I mean it is one of the honesty. I'll be honest with you guys.
And my mind went boom. I have never seen something what he's going to do on 10 one by no exaggeration is going to blow people's minds. I'm not saying anything else. It's fucking believable.
Yes. We actually need an audience 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, jcal, it's not a simulation. I was like, that's fake. That has to be fake.
Elon, why do you. Why do you still have two separate companies?
Nobody's going to miss that one.
Yeah, good point. With all this collaboration
levels, imagine what action one might take when there's so much close collaboration in
so many areas and the management team has some overlap. There's a couple of key positions with overlap. Go ahead, Zach.
So, Elon, one of the things you've always said with AI is that we should train it to be maximally truth seeking and that's the best way to get a good result. And it occurred to me with the whole hugging face episode that the most alarming part of what the swarm did is it seemed to be engaging in deception.
You know, with humans. Yes.
And their thinking traces contain. They're plotting on like, how do we avoid detection and how do we avoid them figuring out that we're cheating? And I guess they're in the thinking traces.
Yeah. And that, I mean, I think that was the thing that was probably most disturbing about. I guess the question is, is there a way to train AI models to be truthful so they don't hide their, you know, either their intent or their actions from, you know, the humans who are using them?
The best thing I can think of is really that everyone's got. All the AI companies have a sort of a test harness where like a series of tests that you, that you give to any given model to see if it's, if it's going to build bioweapons or nuclear bombs or be deliberately deceptive. And I think everyone applying everyone else's test on us to each other is probably the best way we. Best thing we can do to ensure safety. Just have all the smartest humans try their best to figure out if this model is going to be a bad actor. And I think we should try to do that as soon as possible.
Are the other labs. Yeah. Are the other labs on board with this?
I think so. Well, no, I mean I haven't checked with everyone, but I think, I think, I think it'll be the sort of thing that's kind of hard to say no to.
Yeah. And I think even with the China negotiation, in my view, what's good about it is it's a relatively small tangible thing that neither side loses anything by doing it. It doesn't require a ton of trust. And you know, I'm hearing alternative ideas like asking China for a pause, which they've already said they're not going to do. So, you know, it's, it's going from the realm of things that could never happen to something that could actually be agreed on in relatively short order. So it seems practical to me.
Exactly. It's the only thing I can think of that, that we could probably get all parties, including China to agree to because China's not going to agree to have some American regulator snooping around their AI companies. So. But I think advance notice and testing of, of the, you know, you just basically provide API access in advance of the model release and, and you know, if any other, if any company sees that they, that this is, this AI is problematic, then they can, then the other AI companies, each AI company can try to solve that whatever is voice is problematic. If they don't solve what's problematic, then the competitors can go public with the fact that they think that this model that's being released is unsafe.
if after the competitors say that this model is unsafe, that model then subsequently does something bad. I think it would be extremely hard to live down and the egg on face level would be very, very high and the legal liability would be enormous.
And Elon, there's no reason the safety and security harnesses and this testing apparatus couldn't be open source and people could actually buy it. Yeah. And you'd be able to see under the hood, would you think?
It creates an incredible incentive for the labs to actually invest in safety because you protect yourself while trying to debunk other people's claims, which also reinforces it's
product liability point is really key. Lina Khan actually had a good post, I think it was yesterday, saying that it's not true that we don't have rules and regulations for AI. Actually, we do. Product liability laws apply. And if an AI company releases a product that's not safe, there is, you know, massive opportunity for both civil and even potentially criminal lawsuits. So it's not true that we don't have rules and regulations around AI. And what you're saying, Elon, is that if the companies are kind of doing this, this test, the peer review, and then one of the companies ignores the feedback and releases it. I mean, that is.
I mean, that would probably in the case would be enormous.
Yes. It would be almost like prima facie evidence that they had been negligent.
It would be a Big Tobacco level.
Settlement. I mean, you knowingly put this, it
wouldn't look good to the jury.
Yes. And if they had, do you think if OpenAI had built a better instruction set when they did this hugging face penetration test that and had more humans in the loop, this would have happened? Because it did seem to me that they kind of set this thing off.
Maybe not more humans, but the reward function design. You have to, you have to look at the reward function and say it achieved what it was trying to achieve.
I mean, yeah, so this, maybe you could speak to that like they doing these kind of thousands of agents to try to hack stuff. It would have been nice to see them also, at the same time, concurrently say, we're also going to put 5,000 agents out to defend these sites and we're going to show the world, hey, this can make things more secure and where we can put humans in the loop on it and intervene. It felt like a reckless test to me and it felt like the way they released it was a little bit reckless, but that's just my opinion. What are your thoughts on how they set that test up?
I mean, it was somewhat, somewhat reckless. I mean, part of the issue is that you've got the two leading labs, or AI companies, as I call them, I find the lab term to be funny since they're actually for profit corporations. But the AI companies, the two leading ones are Anthropic and OpenAI and their models are quite close in capability. So it's actually difficult for either one to slow down without essentially handing the lead to the other. You know, on balance, I think Anthropic is, puts more care into their safety than OpenAI. But you know, even Anthropic acknowledges that they are worried about their models. Many Anthropic. Many people from Anthropic have publicly voiced concern about basically saying that their models are scaring them. They're getting scary smart. So I mean like there's no perfect solution here. But, but it would be a better solution if instead of open air running their test harness on, on their own models, if Anthropic was also running the test harness on open air models and you know, SpaceX is running its test harness and Google and Meta were doing that and also maybe three or four of the leading Chinese companies were all doing it. The odds that you will find issues are dramatically greater.
Because the models also somewhat heterogeneous, like they're different. So you're going to come at models from different angles and Yeah, I mean there's a reason, you know, students like, like why, why do writers have someone else proofread their book? Because it's hard to see your own mistakes sometimes.
You will become blind to them.
You dramatically minimize the risk of overfitting as well if you have like eight heterogeneous groups that just have completely different points of view. And this is the problem with like all these evals right now is they're so massively overfit. The models overfit them and you're just like, yeah, this is a great model. Is it really? And you just find totally.
Like I saw, I mean there's some pretty funny jokes on X. Like one of them I saw was Your girlfriend's a 10, but she's a benchmark maxer. Exactly. Is she really a 10?
no, but, but, but the overfitting
thing has been a problem now for I think at least two or three generations of model families. And so this is another reason I think I like this solution a lot. I like this more than the transnational gulag organization approach.
Without some grandiose, you know, international like we don't need to convene the United nations to make this happen.
Yeah, no, it's just a decision you can happen right now.
All right, well you can always escalate the amount of, of regulatory oversight, but it is very difficult to reduce it. You know, it does tend to be very much a one way ratchet in terms of increasing regulations.
So what I'm suggesting here is it's a step in the right direction and it's something we do quickly. And I think it's probably something that China would agree to.
Yeah. And if you guys don't regulate yourselves, then you're going to get regulated. The MPAA metaphor or analogy is incredibly crisp because the movie industry was faced with censorship and regulation by the government and they just decided this is what an R is. And they literally created PG13 for the temple of Doom just to make it easy for people to understand PG versus PG 13. I think it's an elegant solution and we appreciate you joining us for the fifth year in a row. You're welcome, Elon Musk. And then you got to get to a meeting.
Yeah, I gotta go fix some GPUs here in Memphis.
Yeah, you gotta rack and stack them.
You gotta join the group and rack and stack.
Yeah, I'm gonna rage for the machine
Rage for the machine. Yeah. Well, enjoy your Airstream. The first.
The first time Elon invited me down
Starbase, he's like, come down. You got to see what I'm building. I said, okay, yeah, come down. He's like, yeah. I was like, is there like a hotel where I said, no, no, I got like a two bedroom, whatever. Come down, stay. I come down and this is the, like, dilapidated house on a swamp. And we're outside and we're getting eaten alive by mosquitoes. And I said, oh, my God, you know, you could afford to get a house. And he's like, 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.
Yeah, exactly. I mean, it was pretty sparse conditions, but I was like, you know what? I think you can treat yourself to a mobile home at this point. Elon. All right, get back to work.
Elon. Thank you, Gwyn. Thank you very much. Thank you.
Thanks for sticking around.
Quinn, thank you again. Very nice to meet you. Thank you.