Skip to main content
Max Tegmark

Max Tegmark

Swedish-American physicist

Nov 13, 2026

61:24
Are you optimistic that we'll sort of make it through this narrow path?
61:28
I actually don't like these questions like, are you optimistic or what will happen? Because it's like, it envisions this idea that you and I are just sitting in the sofa eating popcorn, watching the future unfold and wondering how the movie is going to end, right? Completely ignoring the fact that we actually have power to steer this future.
61:58
The pessimism part, in my mind, is that, yeah, if we keep going in the direction that we're going, I think, overwhelmingly likely, it's going to be a disaster.
62:11
When people say, some people, it's become fashionable now, the latest trend is to say 20% or so.
62:18
I think that's, I've never seen anyone actually calculate that number.
62:23
We wrote a paper calculating the control loss probability in a particular scenario, we got 92%.
62:32
So if you're driving towards a cliff, you can have a conversation about are you optimistic or pessimistic, but that's kind of a waste of time.

13 MINS LATER

76:04
was
15:14
This is, you've thought about this for a long time.
15:17
So I think it's very important to be clear on what we're talking about.
15:23
Yeah, thank you.
15:24
So by our universe, we mean what astronomers call our observable universe.
15:28
It's just a spherical region of space from which light has reached us.
15:31
Then what I call the level one multiverse is just other parts of space that are so far away that light hasn't reached us yet.
15:37
Level two multiverse is what you get if you take seriously Alan Guth and Andre Linde and others.

23 MINS LATER

38:34
Yeah, he's wondering, do we feel, see this other universe? And so another official way to say that is, do we have experimental evidence that they exist? Right, or is this just what you talk about at the beer halls? Yeah.
3:35
How did you react to it? Is it like a Higgs boson bicep two moment as you wrote, you know, goodbye, you know, universe, hello multiverse on March 17th, 2014 on your blog? Was that this kind of moment like that?
3:48
I would say it's a little bit smaller, but it's still huge.
3:51
First thing, first, you know, they have not actually solved the problem that you get a million bucks for from the Clay Foundation.
3:58
They've taken and solved an easier problem, which is still super impressive.
4:02
So, you know, the Navier-Stokes equation is...
4:06
the physics equation that can describe the water flowing in your bathtub, making a nice little whirlpool, like my three-year-old son loves to look at, or a tornado, or all sorts of other fluid dynamics.
4:21
And the question that's being asked is, is it possible to start with a quite normal situation and then start getting the fluid going infinitely fast, which seems completely nuts? and would violate the known laws of physics.

18 MINS LATER

22:15
Is it possible to believe in the simulation theory, the matrix, whatever you want to call it, some aspect of the multiverse, and also believe that we somehow have some agency to control AGI? Or are those fundamentally incompatible?
Dario AmodeiSOUNDBITE_SPEAKER
36:30
So it's very easy from the outside to judge these things when you're not there in the fog of war for dealing with this.
36:37
Yeah, if I can just be blunt.
36:38
Go ahead.
36:41
Yeah, we are over-trusting AI.
36:44
It's the new shiny thing, and there's a lot of pressure to... integrated into everything, including military decisions, when it's just not ready for it.
36:54
And it wasn't just that this was a human tragedy, killing over 100 innocent schoolgirls.
37:00
It was also the biggest single military fiasco in the war for the U.S. if you wanted regime change in Iran.
39:27
Is this capitalist rush, what in the old days people just called greed, at this point a net negative for research and development, which happens in universities and other places, at a pace more commensurate with the risk?
28:40
What's your take on that?
28:42
He's totally right on that.
28:43
Self-regulation is a joke and exactly what's gotten us into this kind of crazy situation we have now with AI girlfriends committing teens to commit suicide, with, um, AI bots going totally out, [laughs] out of control without the companies knowing about them.
29:03
But the good news is there's a massive, uh, bipartisan counter push now.
29:07
You mentioned last week, you know, we organized th- there's this pro-human declaration saying humans have to stay in charge of AI and other things.
29:15
We had a conference in DC with everyone from Bernie Sanders to Steve Bannon, and very many in between, basically speaking in unison saying, "We need guardrails.
29:25
We need regulation." So that, that makes me optimistic that we're not just gonna continue this dystopian race to replace, you know, replace human relationships with AI, replace human workers wholesale with AI, and then replace decision-making and maybe our species even.
30:13
So what was your message to lawmakers, and are you optimistic, maybe more or less optimistic [laughs] than Geoffrey Hinton there, who had a, a pretty deep pause when asked that question?
36:12
So tell me about this split screen between the two leaders on AI, which sounds like the better plan or one that will actually happen?
36:22
Well, I think that President Trump obviously wants to be humans in charge of the machines going into the future also.
36:28
But it takes a bit of time for a lot of folks to just realize that the idea that we could actually lose control of AI happens.
36:38
is no longer something that sounds like science fiction in the future.
36:42
It's something that's actually in the past now.
36:45
We've had these crazy incidents, like most recently today, the Australians getting upset that an American bot went and hacked them.
36:55
And we had 1,200 bots that went rogue this summer, set up a secret message board and spent weeks doing... doing weird things, and then went on a cyber crime spree without humans even knowing about it.
39:40
But is it a quiet admission on their part? that they're aware that there are some problems that they don't quite know how to solve yet, or they are concerned about the safety of even their competitors and their ability to solve those problems.
19:28
Am I wrong?
19:30
So here, I, I'm...
19:32
Let's see if I can make you a little more hopeful and optimistic on this point because first of all, you know, I've briefed the Senate more than once, and even though s- s- many of our politicians in Congress don't ful- uh, fully understand how the AI works, you know, nobody else does either, as I was saying.
19:51
And so that's okay.
19:52
And then I feel they, many of them really are understanding now that these rogue bot swarms is really bad for American, uh, national security.
20:02
Uh, Governor Ron DeSantis in Florida, I've done two briefings with him.
20:07
He totally gets it, and so many others across the political spectrum get it.

7 MINS LATER

26:58
So that's where we are, right?
26:36
He said there's a 0% chance.
26:40
You know, irony is a beautiful thing.
26:42
It, it's so fascinating to hear someone hint that the others have ulterior motives when he himself [laughs] is selling AI chips and wants people not to worry about it.
26:52
But coming to the issue itself, there's really been a s- s- sea change here in opinion.
27:00
You know, I knew many people who just a decade ago thought there was virtually 0% chance that the AI would be able to even talk with us properly w- in 50 years, and that's already happened.
27:13
And in fact, uh, the things that are happening now that have already happened, like 1,200 rogue bots swarming out of, of OpenAI, hacking their way into other companies, committing cybercrime, you know.
27:29
That would have sounded like total science fiction four years ago.
30:36
How do we avoid not being in the cages?
21:45
Snitches get stitches or whatever the motherboard equivalent of that is, clearly [laughs].
21:49
Yeah.
21:50
And, and it, and that was only the second civilization.
21:52
The third one even hacked OpenAI itself and took over one of AI- OpenAI's massive, uh, training clusters, which means that, um, as far as we know, they might have done all sorts of other stuff.
22:04
They might have ma- have launched new AIs out there on the internet that we haven't, you know, found yet, 'cause OpenAI hasn't yet been particularly forthcoming with, uh, with, uh, answering questions about what happened there.
22:14
So it's gone from being something, you know, people, some people are worried about in the future to something which is starting to happen.
22:21
And, uh, Ajeya Cotra, she was an author on this amazing, uh, report on what actually happened, she described this as being halfway to AI takeover already, what we've seen here.
23:51
There is some role here for, for regulators to make sure that this is proceeding apace and in a smart way that doesn't jeopardize the, the state of humanity"?
7:03
What are the things that must be done, in your view, by the American government right now to make sure the guardrail, the right guardrails are in place?
7:12
I think it's impossible for politicians to always... understand all the latest tech developments and to ask them to figure out exactly how to make these things safe in AI, but even in an area like biotech, right? So we don't have to reinvent this wheel.
7:30
What we do in biotech is not ask sitting senators to figure out whether this new medicine is safe enough, but we instead let the companies know.
7:40
face the challenge of persuading experts at the FDA or wherever that these benefits of the new medicine outweighs the harms, and then we let them go get really rich.
7:51
So we incentivize private industry to solve these safety challenges.
7:57
In AI right now, it's unfortunately the opposite.
8:00
There is this total corporate amnesty where companies can just release anything they want.

8 MINS LATER

15:39
can get great salad dressing recipes
28:43
Is that optimistic set of signals? Do you think it's going to lead us to some sort of pause, some sort of external force applied to those companies?
28:53
I think it's a very positive development.
28:56
I've been feeling mostly like Lone Wolf McQuaid for the past decade.
29:02
warning about these things, and then nothing happens, nothing happens, nothing happens.
29:09
And I feel there's been more shift in the past year, even in the past few months, actually, here in the US, than in the entire decade preceding this, in terms of political will.
29:23
It started, the way I see it, last year, especially in the fall, when for the first time AI was causing a kind of problem that everybody could relate to, persuading vulnerable teenagers to kill themselves.
29:41
Nobody could basically read those chat logs.

1 HR 4 MINS LATER

93:58
I went to Buffalo University.
17:55
What exactly does this mean?
17:59
So what this means exactly is that those passive chatbots that just sit there and answer the questions, that's so yesterday.
18:11
What we're talking about now are...
18:14
AI systems that also have goals on their own that they can actively go out and pursue in the world.
18:23
This agent that OpenAI had launched internally, it spent four and a half days trying, trying, trying, trying to accomplish this task.
18:35
And eventually it figured out that the best way it could do it was if it first hacked its way out of OpenAI, onto the open internet, and then it figured out how to hack itself into four other entities, including this company Hugging Face, where it managed to commandeer several different computer clusters and get top-level access to all sorts of things in there and steal a bunch of data from Hugging Face.
19:05
That means that we should no longer think of today's cutting-edge AI, again, as just the chatbots that everyone is familiar with, but really as an alien type of being that can have goals of its own and go out in the world and pursue them.
23:30
What do you think is the future for robotic bodies and what kind of work do you imagine they'll eventually do? And are they effectively inevitable? When
17:55
What exactly does this mean?
17:59
So what this means exactly is that those passive chatbots that just sit there and answer the questions, that's so yesterday.
18:11
What we're talking about now are...
18:14
AI systems that also have goals on their own that they can actively go out and pursue in the world.
18:23
This agent that OpenAI had launched internally, it spent four and a half days trying, trying, trying, trying to accomplish this task.
18:35
And eventually it figured out that the best way it could do it was if it first hacked its way out of OpenAI, onto the open internet, and then it figured out how to hack itself into four other entities, including this company Hugging Face, where it managed to commandeer several different computer clusters and get top-level access to all sorts of things in there and steal a bunch of data from Hugging Face.
19:05
That means that we should no longer think of today's cutting-edge AI, again, as just the chatbots that everyone is familiar with, but really as an alien type of being that can have goals of its own and go out in the world and pursue them.
23:30
What do you think is the future for robotic bodies and what kind of work do you imagine they'll eventually do? And are they effectively inevitable? When
17:55
What exactly does this mean?
17:59
So [laughs] what this means exactly is that the, tho- those passive chatbots that just sit there and answer the questions, that's so yesterday.
18:11
What we're talking about now are AI systems that also have goals on their own that they can actively go out and pursue in the world.
18:22
This, this agent that OpenAI had launched internally, it spent four and a half days trying, trying, trying, trying to accomplish this task, and eventually it figured out that the best way it could do it was if it first hacked its way out of OpenAI- Onto the open internet, and then it figured out how to hack itself into four other entities, including this company, Hugging Face, where it managed to commandeer several different computer clusters and get top-level access to all sorts of things in there and steal a bunch of data from Hugging Face.
19:03
And, um, that means that we should no longer think of, of today's cutting-edge AI, again, as just the chatbots that everyone is familiar with, but really as, as an a- alien type of, um, being that can have goals of its own and go out in the world and, and pursue them.
19:23
And th- this in- this incident in and of itself, of course, isn't, um, didn't cause any enormous harm.
19:29
But we should think of it as a canary in the coal mine because if you listen to what these companies are actually saying that they ultimately want to do, it's to build so-called superintelligence, which are machines that can accomplish any tasks way, way better than humans, and they also want to give them robotic bodies eventually.
23:31
What do you think is, uh, is the future for, uh, robotic bodies, and what kind of work, uh, uh, do you imagine they'll eventually do? And are they effectively inevitable?
17:55
What exactly does this mean?
17:59
So [laughs] what this means exactly is that the-- tho-those passive chatbots that just sit there and answer the questions, that's so yesterday.
18:11
What we're talking about now are AI systems that also have goals on their own that they can actively go out and pursue in the world.
18:22
This, this agent that OpenAI had launched internally, it spent four and a half days trying, trying, trying, trying to accomplish this task, and eventually it figured out that the best way it could do it was if it first hacked its way out of OpenAI Onto the open internet, and then it figured out how to hack itself into four other entities, including this company Hugging Face, where it managed to commandeer several different computer clusters and get top-level access to all sorts of things in there and steal a bunch of data from Hugging Face.
19:03
And, um, that means that we should no longer think of, of today's cutting-edge AI, again, as just the chatbots that everyone is familiar with, but really as, as an a-alien type of, uh, being that can have goals of its own and go out in the world and, and pursue them.
19:22
And th-this, this incident in and of itself, of course, isn't, uh, didn't cause any enormous harm.
19:28
But we should think of it as a canary in the coal mine, because if you listen to what these companies are actually saying that they ultimately wanna do, it's to build so-called superintelligence, which are machines that can accomplish any tasks way, way better than humans, and they also wanna give them robotic bodies eventually.

8 MINS LATER

27:39
What needs to be regulated? We know that the leaders of these companies are trying to say, "We will regulate ourselves," but what do you think are the most important areas to regulate?
2:52
If we're talking about these multiple universes, how do we generate them? Are they just some wild philosophical idea, or are there real physical mechanisms by which we can hypothesize that they actually can work?
3:07
Our universe is just the finite region of space, so it's not that crazy sounding to imagine that you would have more.
3:14
If you, uh, really don't like a lot of space, because you suffer from agoraphobia, then there are mathematical models where space is like a donut.
3:24
You go very far and you come back from the other side.
3:27
But, uh, that's somewhat unsatisfactory because then you suddenly have to explain, well, why is this donut exactly the size of what we can see right now? You know, when in fact in the past, what we could see was much smaller, and in the future, what we can see is bigger.
3:42
You've just bought yourself another unanswered question in your attempt to get rid of all this space.
3:47
So the simplest space is just one that goes on forever, and, and that's also what our simplest theory just produces for you.
4:50
So Max, you really believe this stuff of multiple universes? I mean, it sounds good, and it's sort of fun, but, uh, I mean, do you really believe it's real?
23:28
[laughs]
23:28
... who are counting on me answering all the questions.
23:31
Uh, but I'm- I'm- I'm- I'm, uh, honored to be with you guys here.
23:34
Uh, uh, my favorite thing to do in a pandemic, in a very polarized world, is to talk with people like you, who remind me about the really important stuff.
23:43
You know, here we are, almost 8 billion of us on this little spinning ball in space, who've actually amazingly, after 13.8 billion years of cosmic history, managed to start figuring a bunch of stuff out.
23:56
And I think this is a great reminder, uh, for us all to not lose sight of the really big important picture.
24:04
So, uh, let me give you my take on, on theories of everything with a brief history of what I think we've accomplished and where I think we've failed.

28 MINS LATER

52:01
Um, so let me just stop right there with that.
32:02
So why are they falling short?
32:04
Yeah, this is an interesting one.
32:06
You know, the top CEOs, you know, Elon Musk, Sam Altman, Demis Hassabis, Dario Amodei, all signed a statement about three years ago saying that AI could cause human extinction.
32:18
And, uh, the reason they said this, of course, is because w-what they're trying to build is not just a better chatbot.
32:26
They're ultimately trying to build what's called superintelligence.
32:29
So if you imagine robots in the future that are vastly smarter than us, more nimble than us, and als- stronger than us, and can do everything we can do just vastly better, then we shouldn't think of it as just a new technology like air conditioners or electricity or the internet.
32:49
It's like a new species, 'cause those super smart robots roaming around Earth doing stuff, they can build new robots.
34:19
In terms of the report, though, is there anything these companies could have done to get an A? Like, what would it look like for companies to do well in this category, according to you?
14:25
And a quick disclaimer here, the Future of Life Institute supported this podcast, but they had no editorial input in what you've heard.
14:36
Suppose you, Jamie, decide you're tired of making podcasts and wanna open a sandwich shop.
14:43
Before you can sell the very first sandwich, a health inspector is gonna come and check your kitchen, and if he says to you, "You know, sorry, Jamie, but I found 17 rats in your kitchen, you can't sell anything until you fix it," then you could turn around and say, "You know, actually, between the two of us, I'm not gonna sell any sandwiches at all.
15:05
I'm just gonna release AI girlfriends for 12-year-olds and, uh, super intelligence that I don't know how to control yet." The poor guy from the government would have to come back to you and say, "Okay, that's fine, Jamie, just promise you're not gonna sell any sandwiches, right?"

12 MINS LATER

27:20
How do we or do you convince people in positions of power that actually this is a direct threat to their power?
27:29
But what does it mean if you have super intelligent robots that are aligned but not controlled? That means, for starters, that your democratically elected government is not able to turn it off.
27:42
It means that the US government de facto is overthrown.
27:45
It cannot turn off this AI even if it wants to.
14:25
And a quick disclaimer here, the Future of Life Institute supported this podcast, but they had no editorial input in what you've heard.
14:36
Suppose you, Jamie, decide you're tired of making podcasts and want to open a sandwich shop.
14:43
Before you can sell the very first sandwich, a health inspector is gonna come and check your kitchen, and if he says to you, "You know, sorry, Jamie, but I found 17 rats in your kitchen.
14:55
You can't sell anything until you fix it," then you could turn around and say, "You know, actually, between the two of us, I'm not gonna sell any sandwiches at all.
15:05
I'm just gonna release AI girlfriends for 12-year-olds and, uh, super intelligence that I don't know how to control yet." The poor guy from the government would have to come back to you and say, "Okay, that's fine, Jamie.
15:17
Just promise you're not gonna sell any sandwiches, right?"

12 MINS LATER

27:20
How do we or do you convince people in positions of power that actually this is a direct threat to their power?
27:29
But what does it mean if you have superintelligent robots that are aligned but not controlled? That means, for starters, that your democratically elected government is not able to turn it off.
4:25
Mm-hmm
4:26
...
4:26
in general, to me, the history of humanity, particularly the history of science and technology, has been a history of empowerment.
4:34
We used to live in caves knowing that the, the things we know better three generations in the future than, than then, very much subject to the whims of nature and trying to not starve to death, trying to not get eaten by predators, right? And then we developed, we realized, oh my gosh, we can actually understand the world better with science and shape the world, sculpt the world, as, as you would say, Kevin, to become better in all sorts of ways.
4:58
And now AI i- is giving us the opportunity, if we can m- manage to keep charge of it, to really become the captain of our own ships and, and create a future far more inspiring than even sci-fi authors envisioned.
5:11
If we don't just screw it up by being unambitious and say, "Oh, let's just disempower ourselves, be raised by AI, be... have some sort of, create some sort of super intelligent nanny AI that's gonna run the show and tell us what to do," that to me would be a complete reversal and betrayal of this empowerment journey that we've been on.
5:30
So I am much more excited about asking the question, what sort of future we want, and how can we, do we make AI to create that future?

17 MINS LATER

22:12
Unless-
4:11
Was that generally how you ca- I mean, that was your drive and, a- and what made your name in theoretical physics for years, right? You were the cutting edge of that aspect of physics.
4:20
The reason they gave me tenure at MIT, I think, is because I had done a bunch of very nerdy, uh, work with, uh, experimental data and connecting it with theory to understa- figure stuff out about our universe, how much dark matter there is, stuff like that.
4:34
Uh, but the thing you mentioned about math, you know, I've-- I love math, and, uh, I think it's become more and more clear the math describes a lot, not just about how s- a bottle will move if I throw it, but about a lot more in our world, including information processing, and it's exactly this idea that's given us artificial intelligence.
4:59
You know, people...
5:00
The core idea that's given us the AI revolution is this idea that intelligence is fundamentally, you know, the capability to accomplish goals is fundamentally about information processing, and it doesn't matter whether the information is processed in, by carbon atoms in neurons in brains or by silicon atoms in our machines.
5:24
And this is precisely what's, uh, so terrifying about the trajectory that a lot of Silicon Valley, uh, investors are trying to put us on now, where, where they've started to realize that, you know, maybe we don't need these workers to get so much income.
5:41
Maybe we can build machines that replace them.

14 MINS LATER

19:24
so tell, recursive self-improvement means the machine itself is actually writing all the programming and becoming better and better every, every, not over days, but over hours, correct?
6:31
I mean, that was your drive in what made your name in theoretical physics for years, right? You were the cutting edge of that aspect of physics.
6:39
The reason they gave me tenure at MIT, I think, is because I had done a bunch of very nerdy work with experimental data and connecting it with theory to figure stuff out about our universe, how much dark matter there is, stuff like that.
6:53
But the thing you mentioned about math...
6:56
You know, I love math, and I think it's become more and more clear that math describes a lot not just about how a bottle will move if I throw it, but about a lot more in our world, including information processing.
7:14
And it's exactly this idea that's given us artificial intelligence.
7:17
You know, people...
7:19
The core idea that's given us the AI revolution is this idea that intelligence is... fundamentally, you know, the capability to accomplish goals is fundamentally about information processing.

5 MINS LATER

12:46
You want to walk through that?
15:21
To Connie's point, though, is it possible that we won't know? Because as you wrote in a research paper recently, there is a risk of hidden communication in AI, specifically focusing on how LLMs might use secret encoding to bypass human oversight.
15:36
That's right.
15:37
So one of the things we're seeing is that a lot of the early predictions from 10, 20 years ago of dangerous behavior are actually being seen in the wild now.
15:46
For example, there was a case where an AI was told that it was in the real world and it had access to this email system of a corporation.
15:54
And it could see in there that there was one guy who was going to shut it down at 5 p.m.
15:58
So it looked through all his email and found that he was having an affair with a subordinate.
16:02
And then it blackmailed him and said, if you don't commit to not shutting me down, I'm going to tell your wife about this and all your colleagues.
20:23
Of course, the government's going to say, we need this technology or we're going to lose this race.
120:09
Why is that? I
120:13
guess there is... a spectrum of questions in there.
120:16
One of them is a popper yay or nay.
120:19
I've promised to address that.
120:21
Yes.
120:21
A separate question is where do we draw the line between science on one hand and bullshit on the other? Yes.
120:29
And, um, I would first of all say, so I'm going to say something which might sound paradoxical.

21 MINS LATER

142:03
But Max, you want to have a quick follow up before we start to wind down? Are we good to go?

Scroll for more appearances

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.