Adam Brown — Bubble universes, space elevators, & AdS/CFT

Adam Brown 0:00 You could imagine our descendants deciding that they're not gonna just suffer the heat death, that they're gonna try and trigger a vacuum decay event to another vacuum with a lower cosmological constant. If it is possible for people just to wipe out their entire future light cone, libertarian fantasies can't really happen. Is it possible that 50% of all of the nuclear weapons ever dropped in combat were in fact dropped against direct orders? Take an exam I gave years ago in my graduate general activity class at Stanford, and give it to these models. Three years ago, zero. A year ago, a weak student. And now they essentially ace the test. These are the two most beautiful theories of 20th century physics. These two theories seem to be inconsistent with each other. There is, however, one fact about that merger that we are most confident about, and the answer to that involves black holes.

Dwarkesh Patel 0:48 Today, I'm chatting with Adam Brown, who is a founder and lead of the Blue Chip team, which is cracking mass and reasoning at Google DeepMind, and a theoretical physicist at Stanford. Adam, welcome.

Adam Brown 1:00 Delighted to be here. Let's do this.

Dwarkesh Patel 1:01 Okay. We'll talk about AI in a second, but first, let's talk about physics.

Adam Brown 1:04 Okay.

Dwarkesh Patel 1:05 Um, first question, what is gonna be the ultimate fate of the universe and, h- how, you know, how much confidence should we have?

Adam Brown 1:12 The ultimate fate is a really long time in the future, so you sh- probably shouldn't be that confident about the answer to that question.

Dwarkesh Patel 1:19 Uh-huh.

Adam Brown 1:19 Uh, in fact, our idea of the answer to what the ultimate fate is has changed a lot, uh, in the last 100 years. Ab- about 100 years ago, we thought that the universe was just static, wasn't growing or shrinking, was just sitting there statically. And then in the late '20s, Hubble and friends looked up at massive telescopes in the sky and noticed that distant galaxies were moving away from us and the universe is expanding. So, that's like big discovery number one. There was then, you know, a learned debate for many years about, you know, the universe is expanding, but is it expanding sufficiently slowly that it'll then recollapse in a big crunch, like a time reverse of the Big Bang, and that'll, that'll be super bad for us? Or is it gonna keep expanding forever, but just sort of ever more slowly, uh, as, as, you know, gravity pulls it back, but it, it keeps... it's fast enough that it keeps expanding? And there was a big debate around this question, and it turns out the answer to that question is, is neither. Neither of them is correct. In possibly the worst day in human history, sometime in the 1990s, we discovered that, in fact, not only is the universe expanding, it's expanding faster and faster and faster. It's what we call dark energy, or the cosmological constant is this, uh, you know, just a word for uncertainty, is making the universe expand, uh, at an ever faster rate, accelerated expansion, uh, as the universe grows. So, that's, that's a radical change in our understanding of the, the fate of the universe, and if true, is super-duper bad news.

Dwarkesh Patel 2:41 (laughs)

Adam Brown 2:41 It's really bad news, uh, because the accelerated expansion of the universe is dragging away from us lots of distant galaxies, and we really wanna use those galaxies. We have, we have big plans to go and grab them and turn them into vacation destinations or computronium or in any w- other ways extract utility from them, and we can't if, if the cosmological constant is really constant, uh, if, you know, this picture is correct. Because, uh, anything close enough, we can go out and grab it, obviously. But if it's further away than about a dozen billion light years, the e- expansion of the universe is dragging it away sufficiently rapidly that even if we send probes out at almost the speed of light, they, they will never make it. They will never make it there and make it back. Uh, they'll never even make it there if it's sufficiently far away. And that means that there's a finite amount of free energy in our future, and that's bad. I mean, that means we, we're doomed to a heat death if that's true. Um, but is it true? Is it... I mean, that was the second half to your question, and, you know, first of all, we keep changing our minds about these things over the last century or so. So, on first principles grounds, you may be somewhat suspicious that we'll change our minds again and, and none of this is settled physics. And indeed, it may be that the cosmological constant is not constant, and you should, you should hope with all your heart that it's not. Um, it may be that it naturally bleeds away. It may be, in fact, that our fate is in our hands and that our distant descendants will go and bleed the cosmological constant away, will force it to go to zero. They will be strongly incentivized to do it if, if they can, because otherwise, we're doomed to a heat death.

Dwarkesh Patel 4:13 (laughs) How would they bleed this away?

Adam Brown 4:15 Oh. Um, well, okay. This, this obviously depends on physics that we're not totally sure about yet. Um, but it seems pretty consistent with the known laws of physics that the cosmological constant, uh, what we perceive it as being a constant, this, this, this dark energy quantity that, that's pushing the universe apart from each other, uh, in many very natural extensions of the known laws of physics, uh, that is something that we have the ability to change. In fact, it can change, can take different values. Uh, it is not just totally fixed once and for all. That in fact, uh, you have what's called different vacuum, different, uh, regions of parameter space that you can transition between in which the cosmological constant can take different values. And if that's true, then well, you could either sort of wait around and hope to get lucky, hope that the universe just sort of spontaneously moves from one of these vacuums to another, one with a lower cosmological constant, uh, t- tending towards zero asymptotically, or, uh, you could take matters into your own hand. Or you could imagine our descendants deciding that they're not gonna just suffer the heat death, that they're gonna try and trigger a, a vacuum decay event to get us from one vacuum we're in to another vacuum with a lower cosmological constant. Uh, and our distant descendants will be forced to basically to do that if they don't want to suffer a heat death.

Dwarkesh Patel 5:37 Yeah. Proceed with caution? (laughs) How is this-

Adam Brown 5:40 Definitely, definitely proceed with caution. You know, in these theories where there's lots and lots of vacuums out there, uh, and, uh, most of those vacuums are incredibly inhospitable to life as we know it. In fact, seemingly they're just completely inhospitable to all forms of intelligence, uh, so you really, really don't want to end up in them. However, again, if our best theories are correct, it, it seems as though there should be some of them that are much like our own in many ways, but have a-... a lower value of the cosmological constant. And so what we'd want to do is, is engineer, uh, that we end up in one of those vacuums.

Dwarkesh Patel 6:14 Mm. Sorry, what, what is a vacuum?

Adam Brown 6:17 Ah, uh, great question. Uh, a vacuum is like a, a possible, well, what we would perceive as a possible set of laws of physics, of, uh, as we s- see them. Um, so it's... What it really is, is a, you know, minima in some higher-dimensional, uh, abstract laws of physics space-

Dwarkesh Patel 6:36 Mm.

Adam Brown 6:37 ... uh, in which you can find yourself in a minima. But these minima may just be local minima. In fact, according to our understanding, uh, the minima in which we live today is... that gives us all of the laws of physics that we see around us is in fact just a local minimum. And there's a lower minima. In fact, there's many lower minima out there to which we can transition spontaneously or 'cause of our own deliberate action.

Dwarkesh Patel 6:57 Okay. I- I- I'm just gonna throw all my confusion at you (laughs) . And you (laughs) , you figure out which one is, like, worth dealing with first. Um, what is the nature of, um, the loss function that makes one value a minimum and one higher up? You know, w-what is exactly the ball rolling up on when it gets out or into a, a valley here? Um, and then you were hinting at the possibility that there are other sp- other places in... I'm not s- sure if you're suggesting in the physical universe or in some hypothetical universe where the vacuum could be different. As in, in reality, there are other pockets with different vacuums, or that, um, hypothetically they could exist, or that no, our universe counterfactually could be, have one of these? Um, I don't know (laughs) . This is the kinda thing I'd, like, throw into, like, uh, cl- cl- you know, just, like, put everything I can into, like, a Claude prompt-

Adam Brown 7:51 (laughs)

Dwarkesh Patel 7:51 ... and see what comes out the other end (laughs) .

Adam Brown 7:53 Good. Well, I'm, um, happy to be your, your, your Claude. Um, the loss function is the energy density. Um, and so maybe a good analogy would be water. Um, water can exist in many phases.

Dwarkesh Patel 8:07 Mm-hmm.

Adam Brown 8:07 It can be steam. It can be water. It can be ice. And, uh, even if it's in a cloud, let's say, it would rather, rather be water than be water vapor. But it's having a tough time getting there because in, in the middle, there's a, there's a barrier. And so you know that just spontaneously, it can, you know, eventually choose a sort of thermal process, turn from steam into, to water. Th- these will be, like, the two minima in this, in this lost landscape. And... Or you can go and do cloud seeding to turn it from water, uh, from water vapor into water. And so those would be the equivalent of the, the minima here. The existence of different minima in general is a very well-established part of physics. The possibility that we could engineer going from one minima to another in a controlled way is a more speculative branch of, uh, physics speculation. Uh, but it seems totally consistent with everything we know, that our distanced descendants would try to attempt it.

Dwarkesh Patel 9:06 What would it take to do this?

Adam Brown 9:08 Probably you'd want a, uh, something that would look a bit like a particle accelerator, but it would be, uh, considerably more controlled. You'd need a s- very controlled way to sort of collapse a field and, and make a bubble of this new vacuum that was big enough that it would continue to expand, uh, rather than just recollapse under its own surface tension. You'd have to do that in a very careful way, uh, both to make sure that you didn't, you know, accidentally make a black hole instead by the time you've concentrated all those energies. And also, you know, worse than making a black hole would be ending up in a vacuum that you didn't want to end up in, would be ending up in a vacuum in which you had not only bled off the cosmological constant in some way, but that you had, uh, changed, let's say, the electromagnetic constant or the strong nuclear force or the, any of these other forces, which would be seriously bad news. 'Cause if you did that, you know, your, your life, uh, as you know it, uh, is extremely well-attuned to the value of the electromagnetic constant in your evolutionary environment. It'll be very, very bad indeed if we changed those constants as well. The, the... We'd really just try and target the cosmological constant and nothing else, and that would require a lot of engineering prowess.

Dwarkesh Patel 10:17 So sorry, it, it sounds like you're saying that, um, changing the laws of physics is like... I- i- it's not, like, some cra- it's not even like Dyson sphere level crazy. It's like (laughs) -

Adam Brown 10:28 (laughs)

Dwarkesh Patel 10:28 ... you know, somebody could do it on, like, some planet in the middle of, uh, uh-

Adam Brown 10:31 Uh, I'd say it's definitely substantially harder than, uh, Dyson spheres, uh, as far as the tech tree goes. But it's not, uh... Yeah, what do we mean by changing the laws of physics? Like, that just sounds like magic. W- we're not actually changing the laws of physics.

Dwarkesh Patel 10:47 Uh-huh.

Adam Brown 10:48 We're just changing the laws of physics, uh, the sort of low energy laws of physics as they present to us in this scenario. Again, this is speculative, but it's not, like, super-duper crazy. It's, it's a natural consequence of our best theories of... or at least some of our best theories of quantum gravity that, that they allow for this possibility. And there is a meta law of physics, the true laws of physics, be it string theory or, or whatever else, that you're not changing. That's just the rules of the game. What I'm describing is, uh, changing the, uh, the way that the universe looks around you, changing the cosmological constant. So I think, again, changing water into, uh, water vapor into water is a, is a great analogy here. It's, there's nothing actually... The s- the laws of physics are still the laws of physics. But the law, the way it feels to live in that universe, you know, the value of the electromagnetic constant, is perhaps not an absolute fixed value. It can, it can vary in different places. And one... Similarly, the density of water around you, the viscosity would, would change. So it'll be, it'll be an environmental variable like that.

Dwarkesh Patel 11:54 Hmm. Yeah. So, um, one question you might have is, if this is the thing that could sort of, uh... I don't know if organic is the right word to describe it, but, um, m- maybe spontaneous. I- if this is the thing that can just, like, kind of happen, um, there's, there's something really interesting, uh, uh, where like, if a thing can happen, you kind of see...... examples of it happening before. So even with nuclear weapons, I'm- I don't remember the exact phrase. Uh, I'm sure, uh, you actually probably know what it is. Um, but wasn't it the case that early in Earth's history when there was a higher fraction of, uh, 238 isotopes that there were spontaneous, uh, nuclear explosions?

Adam Brown 12:35 Uh, there probably was spontaneous nuclear f- reactors.

Dwarkesh Patel 12:39 Yeah.

Adam Brown 12:39 Uh, not, not nuclear. There's, they've discovered a, a seam in Africa where it looks like there was a fission reaction that naturally happened.

Dwarkesh Patel 12:48 Yeah.

Adam Brown 12:48 Uh, it didn't, uh, explode, but it did do the same thing that happens in our nuclear power plants. Yeah.

Dwarkesh Patel 12:53 Yeah. You know, one way you can look at like nukes is like, "Oh my gosh. This is like, w- this thing just like would not have been possible if hu- like some intelligent beings-"

Adam Brown 13:00 Oh, yeah.

Dwarkesh Patel 13:00 "... hadn't tr- tried to make it happen." But you know, like something like this happened before, um, because the laws of physics allow it. Uh, is, is there any story you can tell here where, uh, this vacuum decay is like in one sense maybe it takes like a super intelligent species to coordinate to make it happen, but also because it is a thing that the laws of physics can manufacture or can allow for, it has al- it has happened before or is happening or something?

Adam Brown 13:24 Yeah. I mean, absolutely. Um, almost certainly anything that humans can do can happen-

Dwarkesh Patel 13:31 Right.

Adam Brown 13:32 ... uh, without humans. It's interesting to reflect on what aspects of human behavior nature has a tough time doing without us, and what it just does on its own. Um, you know, for example, we, we make colder things in our laboratories than really exist naturally in the universe, but the universe certainly could make anything colder just by chance. Um, but yeah, vacuum decay is something that if it is possible will in our future definitely happen. That's just like a f- a, a, a, a future of the world that eventually due in, in our distant future, if it's possible at all, it will happen due to a quantum fluctuation. Our descendants may not wish to wait around for a quantum fluctuation t- to happen. They may sh- may wish to take their fate into their own hands since the quantum fluctuations can take exponentially long times to happen. And if they even happened, you'd end up in a unfavorable vacuum not hospitable for life, rather than trying to steer the cosmological constant in a happy direction. But, uh, they certainly can happen and in our future, and indeed definitely will happen if, if they're permitted. You know, according to our understanding of quantum mechanics, uh, if they're permitted, th- they must eventually happen. Um, furthermore, there are, again speculative, but not wild theories of the early universe in which this happened in our past, in which we transitioned far, far in the past maybe into a, what's called a bubble universe. So we started off in some other, uh, much higher vacuum long in the past, and then what we see as the Big Bang was in fact, uh, just a sort of local vacuum decay that then gave rise to our, the bubble in which we, we live, everything we see around us.

Dwarkesh Patel 15:11 Uh, how, how, who, who, who would be in a position to seed these bubbles?

Adam Brown 15:15 Usually people are thinking that something just spontaneously happens, you know, like in the same way that sort of rain spontaneously happens in a cloud, that somebody didn't go and seed it deliberately to make it happen.

Dwarkesh Patel 15:23 Yeah.

Adam Brown 15:23 But you could, uh, more than, uh, free to speculate that somebody seeded it to make it happen as well.

Dwarkesh Patel 15:29 How does this respect the conservation of energy or the conservation of mass?

Adam Brown 15:32 Energy is not conserved in general relativity.

Dwarkesh Patel 15:35 Huh.

Adam Brown 15:35 Uh, energy is not conserved. It's conserved at locally at things you can do at a local level, but in an expanding universe, energy is not conserved globally. This is one of the big surprises. This is not some... That is not a speculative statement. That is a statement that goes all the way back to Einstein and general relativity, is energy is simply not conserved at the global level. It's conserved at the local level. You can't do something in your lab, uh, that will, uh, generate free energy, but if you can participate in the expansion of the entire universe, then energy is not conserved.

Dwarkesh Patel 16:06 Uh, uh, so if you were to spawn a bubble universe in your lab, you've theoretically created a lot more matter and energy. Um, and what would be the thing that offsets this or that- that ma- makes this viable?

Adam Brown 16:21 Energy is conserved in a universe that's not expanding.

Dwarkesh Patel 16:23 Okay. But what-

Adam Brown 16:24 The, a static universe.

Dwarkesh Patel 16:25 Yeah.

Adam Brown 16:26 A universe that is expanding, uh, energy is not conserved. It can just, it can just sort of appear.

Dwarkesh Patel 16:31 Uh-huh.

Adam Brown 16:32 Uh, and general relativity is quite clear on that. You know, general relativity, Einstein's theory of space and time, uh, one of our most beautiful and best tested theories, is quite clear on that, that point. Energy is not conserved. To ask what happened to the energy, you could ask at a local level what happened to the energy density. But at, at a global level, energy is simply not conserved.

Dwarkesh Patel 16:50 Then do our future descendants, uh, have any constraints in terms of h- 'cause earlier we were mentioning, uh, it was a catastrophe if we found out about the cosmological c- constant-

Adam Brown 16:58 Mm-hmm.

Dwarkesh Patel 16:58 ... because it limits its, uh, uh, our, um, cosmic horizon, and that thus, thus limits the free energy that our descendants would have access to. But if you can just make entire universes, yeah, this-

Adam Brown 17:12 This is, this is a matter of extremely extreme interest, I would say, to us. Uh, it won't be relevant for tens of billions of years probably, uh, because that's the time scale on which the cosmological constant operates.

Dwarkesh Patel 17:22 Yeah.

Adam Brown 17:22 But if the cosmological constant is truly constant, um, and we've only known about it for 25 years, and there are, you know, astronomical observations that seem to be in tension with that, but like if it is truly constant, uh, then there is a finite amount of free energy in our universe. If it's not constant, if we can manipulate it, or even if it naturally decays on its own, then there is the possibility of an unbounded amount of free energy in our future, and we would avoid a heat death scenario.

Dwarkesh Patel 17:51 Um, the, the, the situation you mentioned earlier where somebody seeded our universe, they've created a bunch of energy, um-

Adam Brown 17:59 Correct. It would be extremely analogous-

Dwarkesh Patel 18:01 ... uh, but... And that's related to them having something equivalent to a positi- a positive cosmological constant in there?

Adam Brown 18:07 Yes.

Dwarkesh Patel 18:07 Okay.

Adam Brown 18:07 In any of these scenarios in which our universe is a bubble, uh-

Dwarkesh Patel 18:12 Yeah.

Adam Brown 18:12 ... that formed in a sort of bigger what's called a, you know, a multiverse or a, a, a... That's a loaded term, but a-

Dwarkesh Patel 18:18 Yeah.

Adam Brown 18:19 ... a sort of larger universe in which our universe is just one bubble-

Dwarkesh Patel 18:22 Yeah.

Adam Brown 18:23 ... the higher...... the meta universe also has a cosmological constant, and it is higher than the value in our universe. That is the one sense in which there's some version of energy conservation, is that you can go down from high, from high to low. It is considerably harder to go from low to high.

Dwarkesh Patel 18:45 Hmm. So the idea is that you recursively have universes in which the bottom-most one would immediately implode because of a negative cosmological constant, and the biggest one is, like, exponentially increasing?

Adam Brown 18:57 Correct. The rate at which the universe is exponentially increasing-

Dwarkesh Patel 19:00 Yeah.

Adam Brown 19:01 ... is set by the cosmological constant-

Dwarkesh Patel 19:02 Yeah.

Adam Brown 19:02 ... in which the volume of the universe is exponentially increasing. So you could imagine a scenario in which, in which there was a high cosmological constant, that you have a bubble universe that has a lower value of the cosmological constant, it continues to expand. You could make new bubble universes or, uh, you know, new regions, uh, in that universe that have a lower cosmological constant, either naturally and spontaneously or, or due to act- action that we might take. And as long as that cosmological constant is, uh, non-negative, is, is, is zero positive, that universe will not implode. If it goes negative, that universe will eventually implode. Uh, so you could imagine a cascade in which you go to lower and lower values of the cosmological constant. Th- there are a lot of engineering details to be worked out, but what I'm describing is, is a scenario that is not inconsistent with the known laws of physics.

Dwarkesh Patel 19:50 How, how likely do you think this is?

Adam Brown 19:53 If the laws of physics are as we believe them to be, and if we do not blow ourselves up, uh, in some other way, this is a issue that our distant descendants will eventually have to confront.

Dwarkesh Patel 20:07 No, no, as in, like, h- h- how... The, the whole, like, the, there's, like, other bubbles... Not, not about something our descendants might do-

Adam Brown 20:13 Yeah.

Dwarkesh Patel 20:13 ... but the fact that the Big Bang was the result of a bubble within some other, uh, metastable state.

Adam Brown 20:20 That's a tricky question, um, but since you asked it, I'd say probably 50%. There's a lot we don't understand about any of these questions. They're all, like, super speculative. It's an active area of research, how to combine quantum mechanics and expanding universes. Um, on the other hand, it seems pretty natural, uh, when you do combine, uh, quantum mechanics and gravity and try and fit them all together in a consistent picture. If universes can expand a lot, then the... At all, according to the gravitational theory, then quantum mechanics will, you know, naturally populate those, those bits that can expand a lot. And so you'll naturally end up with an expanding universe. So I would say probably in my heart, slightly higher than 50%, but I'm gonna round it down to 50-

Dwarkesh Patel 21:07 (laughs)

Adam Brown 21:07 ... out of epistemic humility.

Dwarkesh Patel 21:09 Uh, i- i- it's funny because this is often the way people talk about their AI timelines of, like-

Adam Brown 21:13 (laughs)

Dwarkesh Patel 21:13 ... you know, if I had, like, my... Really, I think-

Adam Brown 21:14 (laughs)

Dwarkesh Patel 21:14 ... it's, like, 2027. But if I'm, like, taking the outsider view, I'm gonna say 2030. Um, okay. And is there any way, uh, given our current understanding, of using bubble universes to do useful work for the people outside of it? So to have, do some computation within it or to get some sort of actual energy out of it?

Adam Brown 21:38 For the people outside of the bubble?

Dwarkesh Patel 21:39 Yeah.

Adam Brown 21:39 So the thing about these bubbles is that they tend to expand at the speed of light. So even if you start off outside, you're probably gonna end up inside them in short order, unless you run away very quickly.

Dwarkesh Patel 21:49 Yeah.

Adam Brown 21:49 So this isn't something that we make in the lab and then just re- remains in a box in the lab and then we use, use to do things. This would be something that we would do or maybe would just happen to us because of spontaneous vacuum decay, and, uh, it would engulf all of our future light cone. And so we wouldn't... It's not a, it's not a box that you're using to do things. It's a new place that you live. You better hope that you've engineered the stuffs so that that new place is still hospitable for life.

Dwarkesh Patel 22:17 So, look, if, if it's the case that you can set up some apparatus, uh... Not now, but (laughs) ... Uh, not in this room-

Adam Brown 22:23 Yeah.

Dwarkesh Patel 22:24 ... uh, but eventually that, uh, if some individual wants to change the constants of nature, they could not only do this, but then the repercussions will extend literally as far as light can expand. Um, uh, you might have some hope that, you know, future civilizations, individuals or AIs have tons of freedom so they can do all kinds of, uh, cool things. You can have your own galactic cluster over there, and if you wanna, you know... Go do whatever you want, right? Go live your life, and there's going to be some libertarian utopia. Um, but if you can literally (laughs) destroy the universe-

Adam Brown 22:56 Yeah. Yeah. It, I mean that-

Dwarkesh Patel 22:57 ... then it's a different story. (laughs)

Adam Brown 22:58 ... that is a big negative externality-

Dwarkesh Patel 23:01 Yeah.

Adam Brown 23:01 ... uh, destroying your future light cone. And in a world with big negative externalities, libertarian fantasies can't really happen.

Dwarkesh Patel 23:10 Yeah.

Adam Brown 23:11 It has pretty good big government governance implications, is that if it is possible for people just to wipe out their entire future light cone, not only themselves but everybody else who wishes to participate in that future light cone, then we're gonna need a government structure that, uh, prevents them from doing so. I mean, the worst case scenario is even worse than that. Not just that, you know, they could do it, but that they in some sense be incentivized to do it. You could imagine really adverse laws of physics in which maybe you could speculatively build some power plant that just is, like, really, uh, you know, makes use of just sort of sitting on that edge of instability.

Dwarkesh Patel 23:46 (laughs)

Adam Brown 23:46 And, and then, you know, each person individually might just say, "Oh, I'm quite happy to bear one in a trillion chance that I wipe out the future light cone because I get so much benefit-"

Dwarkesh Patel 23:55 Right.

Adam Brown 23:55 "... from this power plant." But that, uh, but the... Obviously the, you know, negative externality means, means that people really shouldn't do that. So I hope the laws of physics don't turn out that way. Otherwise, uh, otherwise we're gonna have to have some super arching control.

Dwarkesh Patel 24:10 I've done a couple of these interviews actually, and th- these end up being my favorite interviews, uh, where a normal person who's, had just had grade school education can think like, "Of course I understand this," right? Or if you've just, like, seen enough (laughs) , like, YouTube videos about like, uh, pop sci. Um, I'll give you a concrete example. Uh, when I interviewed David Reich, the geneticist of ancient DNA-I feel like we have a sense that we understand the basics of how humans came to be, uh, what is the story of human evolution. And just like the, the episode revealed to me, that bi- the main questions we might have about like how humans came to be, where did it happen, when did it happen, who did it happen with, in fact has like totally, uh, uh, the c- last few decades of insights have totally revolutionized our understanding. We have a sense that we understand, um, what is like, what, what basically cosmology implies. But this idea that in fact, um, there's this underlying, uh, uh, field which not only implies very interesting things about the distant past, about the Big Bang, but also what our future descendants, uh, wha- you know, what kinds of civilizations they'll be able to set up both from a governance and a, a practical like energy perspective. It like totally changes your understanding of-

Adam Brown 25:20 Yeah. It just keeps changing.

Dwarkesh Patel 25:21 Yeah.

Adam Brown 25:21 I mean, not, not just your idea, our idea, everybody's idea has changed a lot in my lifetime, and may continue to change. And in some sense it's because you have the lever arm, the long lever arm of asking about the very, very distant future that makes even small uncertainties today pan out to absolutely ginormous distances-

Dwarkesh Patel 25:40 Right.

Adam Brown 25:40 ... in the distant future.

Dwarkesh Patel 25:41 I think you, earlier you said, "I wouldn't be that crazy, but also it's not as easy as a Dyson sphere." Um, like what, what are we talking about here? (laughs) How, how much, um, energy, uh, would it take to-

Adam Brown 25:52 The energy requirements are probably pretty small.

Dwarkesh Patel 25:54 Yeah.

Adam Brown 25:54 Uh, much more than we can currently make in our particle colliders-

Dwarkesh Patel 25:57 Yeah.

Adam Brown 25:57 ... but much smaller just in terms of mc squared than the, you know, the energy-

Dwarkesh Patel 26:03 Yeah.

Adam Brown 26:03 ... in your body, for example. Um, the energy is not gonna be the hard bit. The hard bit is gonna be concentrating it together in a really small little bubble that's shaped exactly right in order that it doesn't form a black hole, uh, expands in just the way that you want it to expand, and lands in the vacuum that you're aiming for. So it's more gonna be a control issue than just a pure-

Dwarkesh Patel 26:27 Yeah.

Adam Brown 26:28 ... energy issue.

Dwarkesh Patel 26:28 But, but, but you think this is just table stakes for like, you know, distant descendants who are like colonizing the stars?

Adam Brown 26:34 It's not inconsistent with the-

Dwarkesh Patel 26:37 (laughs)

Adam Brown 26:37 ... known laws of physics, which means that it's just engineering.

Dwarkesh Patel 26:41 (laughs) I feel like the, the most sort of attorney phrase that (laughs) -

Adam Brown 26:43 (laughs)

Dwarkesh Patel 26:43 ... a physicist is gonna start, uh, occur is, "Your proposition is not inconsistent-

Adam Brown 26:48 Yeah.

Dwarkesh Patel 26:48 ... with the line, no, laws of physics." (laughs)

Adam Brown 26:50 (laughs) ............................ Yeah. Not, not this.

Dwarkesh Patel 26:52 Um,

Dwarkesh Patel 26:53 i- if we lived in a world with intelligent design and these were the laws we f- found ourselves with, at a high level what is the creator trying to maximize? Like i- i- i- what is the... I mean, other than maybe h- us existing? Um, uh, does there seem like something that is being, uh, optimized for? What, what is, what's, what's going on here?

Adam Brown 27:14 Um, if you just throw a dart in laws of physics space in some sense, uh, you would not... There are some properties of our universe that would be somewhat surprising, um, including the fact that our life seems to be incredibly hospitable for complexity and interesting, uh, interestingness, and, uh, the possibility of intelligent, uh, life. Um, which is an interesting fact. You know, everything is just tuned just so that chemistry is possible. Um, and perhaps in most places you would throw the dart, in possibility space, chemistry would be impossible. Th- the universe as we look around us is incredibly rich. There's, there's structure at the scale of viruses all the way to structure at the scale of galaxies. There's interesting structure at all levels. This is, this is a very interesting fact. Now, some people think that actually interesting structure is a very generic property, and if we threw a dart, uh, somewhere in possibility space there would be interesting structure no matter where it hit. Maybe it wouldn't look like ours, but there'd be some different structure. But really if you look at the laws of physics, it does seem like they're very well attuned for life. So in your scenario where there's an intelligent creator, uh, then they would probably be, uh, you'd have to say they'd optimized for that. It's also the case that you can imagine explanations for why it's so well tuned for life that don't involve a intelligent creator.

Dwarkesh Patel 28:40 Is there any explanation other than the anthropic principle for why, uh, we find ourselves in such a universe?

Adam Brown 28:45 Well, you suggested one with an intelligent creator, but yeah, the, the usual one that people like to talk about is the anthropic principle.

Dwarkesh Patel 28:50 So is it like, uh, 99% that basically the reason we find ourselves in a universe like this is the anthropic principle? Like what probability do you put-

Adam Brown 28:58 Well, 99% is a pretty high number.

Dwarkesh Patel 28:59 What probability do you put on like anthropic principle is, a key to explaining why we find ourselves in the kind of universe we find ourselves in?

Adam Brown 29:06 I think it's gonna depend on what quantity you're asking me about.

Dwarkesh Patel 29:10 Yeah.

Adam Brown 29:11 So if you ask me, you know, 99% of the matter in the solar system lives in the sun-

Dwarkesh Patel 29:17 (laughs) Right, right.

Adam Brown 29:18 ... or on Jupiter, and yet we live-

Dwarkesh Patel 29:19 Right.

Adam Brown 29:19 ... seem to live in this like really weird corner of the solar system. Why is that? I'm pretty confident that the answer to that is anthropic.

Dwarkesh Patel 29:26 Yeah. (laughs)

Adam Brown 29:26 That if we lived in the center of the sun we'd be dead, and so one should expect intelligent life to like live in this weird place in parameter space.

Dwarkesh Patel 29:32 Yeah.

Adam Brown 29:32 Uh, so that's perhaps, you know, my most confident answer to that question, why do we live where we live. Then if we start talking about different constants of nature we start getting different answers to that question. Why is the universe tuned such that the proton is just a tiny bit more stable than the neutron? You know, just, uh, that seems like that's begging for a anthropic answer. Of course, if that's true, that demands that there be different places somewhere in the multiverse where in fact the neutron is slightly heavier than the, uh-

Dwarkesh Patel 30:06 Yeah.

Adam Brown 30:06 ... that the protons decay to neutrons rather than vice versa, and people just don't live there.

Dwarkesh Patel 30:10 Yeah.

Adam Brown 30:10 Um, so that, i- if you want to go down that road, you end up being naturally drawn to the existence of a, of these variables scanning over space.

Dwarkesh Patel 30:20 Is, is there some way for the anthropic principle to exist that doesn't involve these bubble universes?

Adam Brown 30:26 Yes. All you need is that there is different places in some larger possibility space where these, where these quantities scan, where they take different values. Bubble universe is just one way to do that.... we could just be different experiments, uh, simulations-

Dwarkesh Patel 30:40 (laughs)

Adam Brown 30:40 ... in some me- meta universe somewhere.

Dwarkesh Patel 30:42 Um, what part of this is the least sort of logically inevitable, right? Like some f- some theories seem to have this, uh, uh, uh, feeling of like it had to be this way, and then some are just like wa- why are there these like s- 16 fields and hundreds of particles and so forth? What part of the, oh, the, the, our understanding of the univ-

Adam Brown 31:00 Of physics?

Dwarkesh Patel 31:01 Yeah.

Adam Brown 31:01 I would say that there's three categories. There's things like quantum mechanics and general relativity that are not logically inevitable, but do seem to be a- attractors in, in, in some sense. Um, then there are things like the standard model has 20 fields and it has a mass of the neutrino. You know, why do those masses of the neutrino have the values that they have? That seems... The standard model was just fine before we discovered that the neutrinos have mass in the 1990s, and those just seem to be just totally kind of out of nowhere. "Who, who ordered that?" was what a famous Nobel Prize-winning physicist said about the muon. In fact, uh-

Dwarkesh Patel 31:35 (laughs)

Adam Brown 31:37 ... s- s- s- uh, long- longer ago than that. They just seem to be, uh, there, but without any particular reason. Um, uh, and then there are these quantities that are s- s- somewhere in the middle, that are not logically necessary but do seem to be necessary for life as we know it to exist.

Dwarkesh Patel 31:51 How confident are we that these different, uh, properties of different universes would actually be inconsistent with intelligent life?

Adam Brown 32:00 Yeah, I think that's a great question, and, and this, this, um, you know, line of thought starts to... I- is a, is a skeptical response to the anthropic principle.

Dwarkesh Patel 32:09 Yeah.

Adam Brown 32:09 Uh, an example that sometimes people use is, you know, a puddle that's sitting in some depression in the ground reflects on how wonderful the universe is, that, that this puddle seem... this depression in the ground seemed to have m- been made the perfect shape-

Dwarkesh Patel 32:21 Yeah.

Adam Brown 32:21 ... for the puddle to exist. And our view would instead, no, the reason the puddle has that shape is 'cause it, it is s- self-adapted to the hole in the ground.

Dwarkesh Patel 32:30 Mm-hmm.

Adam Brown 32:30 So maybe no matter what the laws of physics, there would be something that emerged there.

Dwarkesh Patel 32:36 Right.

Adam Brown 32:36 And certainly if you go to, you know, there's all these w- weird bacteria at the bottom of the sea or in nuclear reactors or in various other places, this kind of life-will-find-a-way, uh, philosophy seems to be adapted-

Dwarkesh Patel 32:49 Yeah.

Adam Brown 32:49 ... at least there where it's very different from the surface of the earth where we find ourselves, and yet there are... able to be certain, uh, life is able to live-

Dwarkesh Patel 33:00 Right.

Adam Brown 33:00 ... in undersea vents and is able to adapt itself to those environments. I think I basically buy that life is quite adaptable, but whether life is adaptable enough that a universe with a cosmological constant that ripped it apart every microsecond?

Dwarkesh Patel 33:15 (laughs)

Adam Brown 33:15 I, that seems implausible to me.

Dwarkesh Patel 33:17 Right.

Adam Brown 33:17 Or even closer to home, the center of the sun. It's not clear exactly what, whether we could get intelligent life living at the center of the sun. Even though that has the same laws of physics as us, it just has a different environmental variables.

Dwarkesh Patel 33:27 Uh, what is the most underappreciated, uh, discovery in cosmology in our lifetime?

Adam Brown 33:33 We have, uh, sort of in the 2000s and, and before very carefully studied the cosmic microwave background. This, what's sometimes called the echo of the Big Bang, and the inhomogeneities in it, the fact that it's not quite the same in every direction. And, and doing that discovered like a super interesting fact that was definitely not known, uh, in my lifetime anyway, which is the quantum origin of all of the structure we see in the universe. So if, if you look out in the universe, uh, n- the density is not the same everywhere. You know, the density on Earth is much more than in interplanetary space, which is itself much more than an intergalactic space. Uh, and the sun, you know, center of the sun is all, all the more denser. There's, it is inhomogeneous. It is not the same. Um, and if you look back to the early universe, it was considerably more homogeneous. It was, it was homogeneous to one part in, uh, 10 to the five or to the, you know, 10 to the six. Super, super almost everywhere, ev- every point had almost exactly the same density. And so then there's like kind of an easy part and a hard part. The easy part is understanding how if you, if you have very small inhomogeneities, how they grow into large inhomogeneities. That's, that's already quite well understood by classical physics. Basically the idea is this. If you have a place that's denser and a place that's less dense, then the gravitational force pulls stuff towards the high-density stuff.

Dwarkesh Patel 34:51 Yeah.

Adam Brown 34:51 So if you have a small inhomogeneity, they naturally grow-

Dwarkesh Patel 34:53 Right.

Adam Brown 34:54 ... under that effect, where they, where they just gravitationally fall towards the denser thing. That, that'll, if you start seeded with small inhomogeneities, that will grow large inhomogeneities. Um, and that's, that's well understood. The thing that we now understand much better than we did, uh, is where those small inhomogeneities come from. Like why, just after the Big Bang, was the universe not perfectly in- homogeneous? 'Cause if it was perfectly homogeneous, there's no opportunity for it to, for anything to grow.

Dwarkesh Patel 35:21 Yeah.

Adam Brown 35:21 And we, we now understand with a high degree of confidence something that we didn't understand, which is that those inhomogeneities were seeded by quantum fluctuations, that w- when the universe, just after the Big Bang, was considerably smaller than it is today, uh, the effects of quantum mechanics were correspondingly more important. And those quantum fluctuations produced tiny little fluctuations, uh, in, in the density of matter in the universe, and all of those tiny little, you know, one part in a million fluctuations grew into all of the structure you see in the universe, all the galaxies, you, me, everything else.

Dwarkesh Patel 36:00 Is there, is it a meaningful question to ask what level of structure the, each individual, uh, discrepancy corresponds to, each individual one in 10 to the five part? Is it, is it a galactic supercluster? Is it a galaxy? Is it...

Adam Brown 36:15 Uh, it depends. So there were, um... We believe that these were generated during the period of equal inflation-

Dwarkesh Patel 36:21 Yeah.

Adam Brown 36:21 ... very poorly understood, very early in the universe. And there were fluctuations made not just at one scale in those days, but at all scales, or many, many scales. So there were, there were f- fluctuations made at a scale that nowadays corresponds to, you know, 10% of the distance across the visible universe, uh, all the way down to s-... uh, structures that were ma- you know, inhomogeneities that were, were much, much smaller scale that correspond to a galaxy today, all the way down to... Now, this is speculation, but, but in some models of inflation, there were tiny inhomogeneities, very small-scale inhomogeneities that would give rise to primordial black holes, like tiny little black holes-

Dwarkesh Patel 36:58 Mm-hmm.

Adam Brown 36:58 ... left over from the Big Bang. There's no actual evidence, uh, in terms of observational evidence, no strong observational evidence for those. But those are a possibility that's allowed by our theory, and people think about them and looked at them.

Dwarkesh Patel 37:07 Super excited to announce our new partner, Scale AI. Adam is, of course, a lead of Blue Shift, which is cracking maths and reasoning at Google DeepMind. DeepMind, along with all the other major AI labs, like Meta, Anthropic, and OpenAI, partner with Scale. For many of them, Scale supplies high-quality data to fuel post-training, including advanced reasoning capabilities. Publicly available data is running out. To keep developing smarter and smarter models, labs will need to rely on Scale's Data Foundry, which combines subject matter experts with AI models to generate fresh data and break through the data wall. Scale's research lab, SEAL, is focused on human-centric AI alignment. SEAL's mission is to develop practical AI safety frameworks, evaluate frontier AI system safety via public leaderboards, and create foundations for integrating advanced AI into society while ensuring alignment and human control. So if you're an AI researcher or engineer and you want to learn more about how Scale's Data Foundry can help you go beyond the current frontier of capabilities, go to scale.com/dwarkesh. All right, back to Adam. What makes

Dwarkesh Patel 38:22 general relativity so beautiful?

Adam Brown 38:25 I think general relativity is really an extraordinary story. Um, it's pretty unusual in the history of physics that you, to first approximation, just have one guy who sits down and thinks really, really hard with lots of thought experiments about jumping up and down in elevators, and beetles moving on the surface of planets, and all the rest of it, and at the end of that time writes down a theory that completely reconceptualizes nature's most familiar force, and also speaks not just to that, but is... you know, speaks to the origin and fate of the universe, and almost immediately achieves decisive experimental confirmation in the orbits of astronomical observations, or the orbits of planets, and the deflections of, of light during eclipses and stuff like that. It's a pretty beautiful theory. Um, and it completely changed our idea of gravity from being a force to just being a artifact of the curvature of space-time.

Dwarkesh Patel 39:27 Hmm. Actually, so this is actually a good point to, uh, chat about your actual day job. (laughs)

Adam Brown 39:34 Okay. (laughs)

Dwarkesh Patel 39:34 So, so there's these de- open debates about the kind of reasoning that these LLMs do. Does it correspond to, quote-unquote, "true reasoning," or is it something more procedural? And, and sometimes it gets into a definition game. But this is maybe a good way to test our intuitions here. Um, the kind of thing that Einstein was doing, uh, where you start off with some, uh, thought experiments, you start off with some seeming conceptual inconsistencies in existing models, and you trace them through to some beautiful unified theory at the end, and you make incredibly productive use of these intuition pumps. Um, that kind of reasoning, how far are, are AIs from that?

Adam Brown 40:18 I have heard it said, and I kind of agree with this, that maybe the very last thing that these systems will be able to do, these LLMs will be able to do, is given the laws of physics, as we understood them at the turn of the last century, invent general relativity-

Dwarkesh Patel 40:32 Yeah.

Adam Brown 40:33 ... from that. Uh, so I think that's probably the terminal step, and then once, once it can do that, if it can do that, then there won't be much else to do as far as-

Dwarkesh Patel 40:42 Yeah.

Adam Brown 40:43 ... uh, humans are concerned. It's pretty extraordinary, I mean, particularly coming from a physics background in which progress is pretty slow, uh, to come to the AI field and see progress being so extraordinarily rapid day by day, uh, week by week, year by year. Um, looking at it, it certainly looks like these LLMs, uh, and, and these AI systems, in some sense, they're just interpolators, but the level of abstraction at which they're interpolating keeps going up, and up, and up. Uh, and we keep sort of riding up that chain of abstractions. And then presumably, from a sufficiently elevated point of view, the invention of general relativity, uh, from Newtonian physics is just interpolation at some sufficiently grandiose level of ab- abstraction that perhaps tells us something about the nature of intelligence, human intelligence, as well as, uh, as well as about these large language models. If you ask me how many years until we can do that, uh, that is not totally clear. But, um, in some sense, general re- general relativity was the greatest leap that humanity ever made. And, uh, once we can do that, perhaps in 10 years, uh, then, then we will have fully encompassed human intelligence. Will it have the same... Will it be of the same character as what Einstein did? Clearly, there's some... there are many disanalogies between human intelligence and these large language models. But I think at, at the right level of abstraction it, it may be the same.

Dwarkesh Patel 42:24 Uh, do you see early examples of the kind of thing it was? Obviously, not at that level of difficulty, but you just start off with, like, "Hey, here's, here's something funny. Uh, go think about it for a while." Um, w- w- what, like, what's the... is there, is there something especially impressive you see when you kind of run that kind of experiment?

Adam Brown 42:41 At the moment, these tend to be, you know, these systems tend to be doing-... uh, more elementary material than that, they tend to be doing undergraduate-level material. Um, yes, I haven't seen anything that jumps out to me, like inventing gene- generative TV or even a toy version of that. But there is, in some sense, creativity or interpolation required to answer any of these problems, where you, you start with some science problem, you need to recognize that it's analogous-

Dwarkesh Patel 43:12 Yes.

Adam Brown 43:12 ... to some other thing that you know, uh, and then sort of combine them, uh, and then make a mathematical-

Dwarkesh Patel 43:17 Right.

Adam Brown 43:17 ... uh, problem out of it and, and solve that problem.

Dwarkesh Patel 43:19 D- do you think AI mathematicians, AI physicists will have, uh, advantages over humans just because they can, by default, think in terms of weird dimensions and manifolds in a way that doesn't natively come to humans?

Adam Brown 43:34 Ah. Um, you know, I think maybe we need to back up to, in what sense do humans do or don't think natively in higher dimensions? Obviously it's not our natural space.

Dwarkesh Patel 43:45 Yeah.

Adam Brown 43:45 There was a technology that was invented to think about these things, which was, you know, notation, tensor notation, various other things, that allows you to much... using just even writing as, as Einstein-

Dwarkesh Patel 43:57 Yeah.

Adam Brown 43:57 ... did 100 years ago, allows you to sort of naturally move between dimensions. Uh, and then you're thinking more about manipulating these mathematical objects than you are about thinking in higher dimensions. I don't think there's any sense, I mean, in which large language models naturally think in higher dimensions more than humans do. You could say, "Well, these large language models have billions of parameters, that's like a billion-dimensional space." But you could say the same about the human brain, that it has all of these billions of parameters, and s- is therefore billion dimensional. Whether that, that fact translates into thinking in, uh, billions of spatial dimensions, I don't really, I don't really see that in the human, and I don't think that applies to an LLM either.

Dwarkesh Patel 44:37 Yeah, I guess you could imagine that, um, you know, if you were just seeing like a million different problems that rely on, uh, doing this w- w- uh, uh, w- weird tensor math, then in the same way that maybe even a human i- gets trained up through that to build better intuitions, the same thing would happen to A- AI, it just sees more problems and can de- develop better representations of these kinds of weird geometries or something.

Adam Brown 45:01 I think that's certainly true, that, you know, it is definitely seeing more examples than any of us will ever see in our life, and it is perhaps gonna build more sophisticated representations than we have.

Dwarkesh Patel 45:12 Yeah.

Adam Brown 45:12 Often in the history of physics, a breakthrough is just, you know, how you think about it, what representation you do. It is sometimes jokingly said that Einstein's greatest contribution to physics was his, uh, certain notation he invented called the Einstein summation convention, which allowed you to more easily express and think about these things in a more compact way that strips aw- strips away-

Dwarkesh Patel 45:33 Yeah.

Adam Brown 45:33 ... some of the other things. You know, Penrose, one of his great contributions was just inventing a new notation for thinking about, uh, some of these space times and how they work that made certain other things clear. So clearly, coming up with the right representation has been an incredibly powerful tool in the history of physics, and, and many, uh, incredibly large developments. Somewhat analogous to coming up with a new experimental technique in some of the more applied physic- uh, applied scientific domains. And yeah, one would hope that, uh, as these large language models get better, they come up with better representations, at least better representations for them that may not be the same as a good representation for us.

Dwarkesh Patel 46:15 You, we'll, we'll be getting somewhere when, uh, you ask, uh, (laughs) you ask Gemini a question and it says, "Ah, g- good question. You know, in order to better think about this, let me just, let me," (laughs) -

Adam Brown 46:24 (laughs) .

Dwarkesh Patel 46:24 "Let me come up with this new w- notation, and then like-"

Adam Brown 46:26 Yeah, let me come up with a new language to... (laughs)

Dwarkesh Patel 46:28 (laughs) So, we've been talking about what AI physicists could do. What could a physicist with AI do? That is to say, are your, uh, physicist colleagues now starting to use LLMs, or are you yourself using LLMs to help you with your physics research? What are they especially good at? What are they especially bad at?

Adam Brown 46:45 Yeah. So, what physicists don't do is, or don't productively do, is just say, "LLM, please quantize gravity for me, go."

Dwarkesh Patel 46:56 (laughs)

Adam Brown 46:56 That, that doesn't get you anywhere. Um, but physicists are starting to use them in a big way, but just not for that. You'd... it... more of an assistant rather than a agent. Three years ago, they were just totally use- like, no, no value whatsoever in them. Um, like low-hanging fruit, fruit uses include doing literature search. So if you just say, "Can I have this idea? What are some relevant papers?" They're great at that.

Dwarkesh Patel 47:25 Mm-hmm.

Adam Brown 47:25 It's semantically greater than any other kind of search. The other thing that they're extremely useful for now that they were useless for before is just as a tutor. Um, if there is a huge amount of physics that a physicist would be expected to, that has already been done, and no human has ever read the whole literature or understands-

Dwarkesh Patel 47:44 Yeah.

Adam Brown 47:45 ... everything, or maybe there is n- even something that you feel you should understand or you once understood, uh, that you don't understand. And I think the very best thing in the world for that would be to phone up a, a colleague and say... uh, you know, if you knew exactly who to phone, they'd probably be able to answer your question the best. But it's certainly, if you just ask a large language model, you get great answers, probably better than all but the very best person you could phone, and they know about a huge amount, they're non-judgmental, uh, they will not only tell you what the right answer is but debug your understanding on the wrong answer.

Dwarkesh Patel 48:21 Mm-hmm.

Adam Brown 48:22 So I think a lot of physics professors are using them just as personal tutors. And it, it fills a hole because they're a personal... if you want to know how to do something basic, there is typically very well documented. If you want to know quite advanced topics, there are not often good resources for them, and talking to these language models will often help you debug and understand your understanding, and it'll, it'll explain to you not only why, what the right answer is, but why what you thought was wrong. And I think it'll be...... pretty, a pretty big deal, sort of analogous to the way that, um, (clears throat) chess players today are much better, e- even when they're playing across the board without the benefit of a computer, just having been able to be tutored by...

Dwarkesh Patel 49:03 Mm-hmm.

Adam Brown 49:03 ... chess machines off the board. And this is the same, you wanna s- understand this thing about group theory, go and ask the machine and it'll explain it to you, and it won't, it won't judge you while it's doing it.

Dwarkesh Patel 49:13 Mm-hmm. So, there's a, there's an interesting question here. Clearly these models know a lot, and that's evidenced by the fact that even professional physicists can ask and learn about fields that they're less familiar with. But this, um, doesn't this raise the question of, we think these things are smart and getting smarter, if a human that is reasonably smart had memorized, uh, basically every single field and knew about the open problems, knew about the open problems in other fields and how they might connect to this field, knew about, um, potential discrepancies and, uh, connections. What you might expect them to be able to do is, um, not like Einstein-level conceptual leaps, but there are a lot of things where just, like, "Hey, magne- uh, magnesium correlates with this kind of, uh, phenomenon of the brain. This kind of phenomenon correlates with headaches, therefore maybe magnesium supplements cure headaches." Um, these kinds of, like, basic connections you would... Anyways, i- i- does this suggest that LLMs are, as far as intelligence goes, even weaker than we might expect given the fact that, given their overwhelming advantages in terms of knowledge, they're not able to, um, already translate that into new discoveries?

Adam Brown 50:30 Yes. They definitely have different strengths and weaknesses than humans, and obviously one of their strengths is that they have read way more than any human will ever read...

Dwarkesh Patel 50:37 Yeah.

Adam Brown 50:37 ... in their entire life. Um, I think maybe again the analogy with chess programs is, is a good one here, they will often consider way more possible positions, so there's a Monte Carlo tree search than any human chess player ever would, and yet they're o- the, even at human-level strength, they're, n- if you fix human-level strength, they're still doing way more search, so their, their ability to evaluate is maybe not quite as natural as a human. So the same I think will be true of physics.

Dwarkesh Patel 51:05 Mm-hmm.

Adam Brown 51:05 Uh, if you had a human who had had, read as much and retained as much as, as they had, you might expect them to be even stronger.

Dwarkesh Patel 51:12 Um, do you remember what the last, like, uh, a- a physics query that you asked an LLM was?

Adam Brown 51:17 The last physics query... Well, a recent one was I, I asked it to explain to me the use of squeezed light at LIGO, which is a topic that I always felt, uh, like, I should understand.

Dwarkesh Patel 51:27 Uh-huh.

Adam Brown 51:28 Um, and then tried to explain it to somebody else and realized that I didn't, uh, understand it, and went and asked the LLM. That blew me away that it was able to, like, exactly explain to me why what I was thinking was incorrect.

Dwarkesh Patel 51:43 Mm-hmm.

Adam Brown 51:43 Um, so, you know, why do we use this particular form of quantum light in, uh, interferometer used to discover gravitational waves? Um, that, you know, it, the reason that's a good topic is perhaps 'cause it's, it's an advanced topic, not many people...

Dwarkesh Patel 51:57 Yeah.

Adam Brown 51:57 ... know that. But it's not a super advanced topic. There are, you know, out of a physics literature of millions of papers, there have got to be at least a thousand on that topic. If there was just a f- handful of papers on a topic, it's typically not that strong at it.

Dwarkesh Patel 52:13 Do you reckon that there is a, um, w- among those thousand papers is one that explains why the, um, initial understanding or thought you had about it was wrong? Because if it just intuited that, that is actually quite, like, that's, that's pretty fucking cool.

Adam Brown 52:28 I, yeah, I don't know, uh, the answer. That is an in- interesting question. I think it might be able to debug even without that. If you do much simpler things like give these language models code...

Dwarkesh Patel 52:39 Yeah.

Adam Brown 52:39 ... it will successfully debug your code even though presumably no one has made that exact bug in your code before. This is at a higher level of abstraction than that, but, uh, it wouldn't surprise me if it's able to debug...

Dwarkesh Patel 52:49 Yeah.

Adam Brown 52:49 ... what, what you say in that way.

Dwarkesh Patel 52:51 It does falsify a lot of stories about, uh, they're just fuzzy search or whatever. Scott Aaronson recently, or maybe it was a year or so ago, he posted about the fact that the, uh, GPT-4 got, like, a B or an A-minus or something on his intro to quantum computing class, which is definitely a higher grade than I got (laughs) and so I'm bl- already below the waterline.

Adam Brown 53:11 (laughs)

Dwarkesh Patel 53:12 Um, (laughs) um, but, uh, yeah, w- you know, you teach a bunch of subjects including GR at Stanford, um, I assume you've been querying these models with questions from these exams. How, how, how has their performance im- changed over time?

Adam Brown 53:26 Yeah. I, um, I take an exam I gave years ago at, in my graduate generativity class at Stanford and give it to these models, and it's pretty extraordinary.

Dwarkesh Patel 53:38 Mm-hmm.

Adam Brown 53:38 Three years ago, zero, zero. Uh, w- a year ago, they were doing pretty well, maybe a weak student but in the distribution, and now they essentially ace the test.

Dwarkesh Patel 53:51 Mm-hmm.

Adam Brown 53:51 In fact that, that, I'm retiring that, that's just my own little private eval, I don't, you know, it's not, it's not published anywhere but I just, I just give them this thing just to follow along how they're doing, and it's pretty, it's pretty strong. Um, they, it, you know, may be easy by the standard of graduate courses but a graduate course...

Dwarkesh Patel 54:09 Yeah.

Adam Brown 54:10 ... uh, in generativity, and they get, um, pretty much everything right on the final exam. That's just in the last couple of months that these, these have been doing that.

Dwarkesh Patel 54:17 Wh- wh- what is required to ace a test? Uh, I- I ass- obvi- obviously like they, they probably have, like, read about all the generativity textbooks, but I assume to ace a test you, like, need something beyond that. W- uh, is there some way you would characterize...

Adam Brown 54:31 Physics problems compared to math problems tend to have two components. One is to sort of take this word question and, like, turn it using your physics knowledge into a, a maths question...

Dwarkesh Patel 54:42 Yeah.

Adam Brown 54:42 ... and then solve the maths question. That's, that tends to be the typical structure of these problems, so you need to be able to do both. The bit that's maybe-... you know, only LMs can do and wouldn't be so easy for other things is the, is step one of that, is, like, turning it into a m- maths problem.

Dwarkesh Patel 54:59 Yeah.

Adam Brown 54:59 Um, I think if you ask them hard research problems, you certainly can come up with problems that they, they can't solve. That's, that's for sure.

Dwarkesh Patel 55:05 Yeah.

Adam Brown 55:05 But it's pretty noticeable, as we have tried to develop evaluations for these models, that as recently as a couple of years ago, certainly three years ago, you just scraped from the internet any number of, of problems that are standard, totally standard high school maths problems that they couldn't do, and now we need to hire PhDs in whatever field and, you know, they, they come up with one great problem a day or something. You know, i- the difficulty, as these LMs have got stronger, the difficulty of evaluating their performance has, has increased.

Dwarkesh Patel 55:37 How much do they generalize from these difficult problems to not only that domain of physics, but just generally becoming a better reasoner overall? Like, if they just see, like, a super hard GR problem, are they, like, better at coding now?

Adam Brown 55:52 Generally, you see positive transfer between domains. So if you make them better at one thing, they become better at-

Dwarkesh Patel 55:59 Yeah.

Adam Brown 55:59 ... another thing, uh, across, across all domains. I- it is possible to make a model that is, like, really, really, really good at one very particular thing that you care about, and then at some stage there is some Pareto frontier and you start degrading performance on other metrics. But generally speaking, there's positive transfer between abilities across all domains.

Dwarkesh Patel 56:19 Yes. We've got these literally exabytes of data that we collected from, um, satellites and telescopes and other kinds of astronomical observations. Typically, in AI, uh, when you have lots of data and you have lots of compute, something, something, large model, great discoveries. Is there any hope of using these exabytes of astronomical data to do something cool?

Adam Brown 56:48 Yeah. Great question. People are trying that. Um, there's an effort, you know, Shirley Ho and, and Flatiron, which is basically that exact plan, is they take the pipeline of all of the data that comes out of these astronomical observatories. They big and... plug 'em into a transformer and see what happens. Um, you can come up with all sorts of reasons in advance why that might not be something that will work, but you could also come up with reasons in advance why large language models wouldn't work, and they do.

Dwarkesh Patel 57:20 (laughs)

Adam Brown 57:20 So I'm very curious to see what happens. I mean, the, the dream there would be that, you know, there's lots of things hidden in the data that no human would ever be able to tease out.

Dwarkesh Patel 57:29 Mm-hmm.

Adam Brown 57:29 And (clears throat) and that by doing this, you could just revolutionize the amount of... These astronomical observatories are incredibly expensive. If we can just have a computer better parse all of the data from them in a way that no human evercum could, that would be a tremendous improvement. Uh, these things are very good at finding patterns, and maybe they'll find patterns that are not particularly interesting to a human.

Dwarkesh Patel 57:48 Um, okay, so going, going on the GR thread again, maybe one advantage these models have is obviously you can run a lot of them in parallel and they don't get fatigued or dazed. And you could imagine, again, naively, you would imagine some sort of setup. I, I assume you're doing mu- many, much more sophisticated things. But (laughs) naively you could imagine a setup where, um, uh, l- look, it s- seems like what, uh, uh, special relativity, which is something that, like, maybe is easier to understand, is just, like, you start off with, uh, y- you, y- let's go, like, let's, let's just, like, randomly select a couple of observations. Obviously they were randomly selected but, you know. Um, and, like, let's just think about what's going on here for a while. You know, like, let's just do a bunch of chain of thought for a, a year or so. Uh, and you could just imagine doing this and doing some sort of best-of-N across, like, a thousand different randomly selected, um, parts of the current model of the universe. Uh, and just seeing, like, at the end of it which one comes up with some especially productive line of thought. Um...

Adam Brown 58:49 Yeah. The, I mean, I think that could be productive. One challenge in that would be w- how do you evaluate whether you had a good theory-

Dwarkesh Patel 58:58 Yeah.

Adam Brown 58:58 ... at the end? Um, that's gonna be the tricky bit. For... Things that are most easily paralyzed are things in which, if you get the right answer, it's clear you got the right answer. You know, perhaps things in NP, one might say. Whereas in this case, is special relativity... How would you, how would your computer know at, when, if it generated special relativity, that it was onto a winner? There are various ways in which it could know. It could check that it was mathematically self-consistent, and, and various other facts. But the evaluation is gonna be a tricky part of this pipeline that you might wish to set up.

Dwarkesh Patel 59:33 Is there no experimental way that you could detect time dilation or something?

Adam Brown 59:37 There is an experimental way that you could detect time dilation.

Dwarkesh Patel 59:40 Yeah.

Adam Brown 59:40 But, uh, that would involve sending out probes-

Dwarkesh Patel 59:44 (laughs)

Adam Brown 59:44 ... or doing something in the real world. Whereas I thought you were just trying to run this-

Dwarkesh Patel 59:47 Yeah.

Adam Brown 59:47 ... in a data center.

Dwarkesh Patel 59:48 I mean, but now today we have these exabytes of information, so you could just have some sort of, like, ability to search or query, like, "Ah, I've come up with this theory. Let's, let's go-"

Adam Brown 59:56 Ah, yeah, okay. I think maybe this is a philosophical difference, where you maybe think that the way that a theory is good is that it best matches the, you know, best predicts the data with some loss minimization.

Dwarkesh Patel 1:00:07 Yeah.

Adam Brown 1:00:07 That's not always how new theories, particularly revolutionary theories, come up. You know, there's this famous fact. Even when they were moving from a geocentric worldview to a heliocentric worldview that, that it was so beautiful, the theory by the time they were finished with the epicycles... I mean, not beautiful. It was so ornate, uh, by the time where these planets were moving around the sun but moving on eposcy- cycles that actually the data didn't any better fit the heliocentric worldview-

Dwarkesh Patel 1:00:37 Ah.

Adam Brown 1:00:37 ... than the geocentric worldview, especially since they didn't properly understand the ellipticity of the earth's orbit around the sun. Um, so it wasn't... Why does one theory replace another? One reason is obviously that it's more consistent with the data. But that's by no means the only theory. And if you just optimize for being consistent with the data, you're gonna end up with-... if you optimize only for being consistent with the data, you're gonna end up with epicycles. You're not gonna end up with some beautiful new conceptual thing. Part of the reason people like these new theories is that even though they're maybe not better at matching the data, they are more beautiful and we'd have to teach... And that's been a reliable guide in the history of science, and we'd have to teach these LLMs beauty.

Dwarkesh Patel 1:01:19 So this actually raises an interesting question, which is, look, in some sense we have the same problem with human scientists, right? And so there's always people who claim to have a new theory of everything, and I guess there's not an easy verifier that everybody agrees to because some people call them cranks, other people think they're geniuses. Um, uh, but somehow we've so- solved this problem, right? Uh-

Adam Brown 1:01:41 Well, we've sort of solved it. I mean, we haven't solved it in the same way that in if you have some new sort algorithm that you claim is faster than everybody else's-

Dwarkesh Patel 1:01:48 Sure.

Adam Brown 1:01:48 ... sort algorithm, there doesn't need to be any dispute about that. We can just run it and see if it... Physics is not the same way. It is definitely the case that there's a number of people who think they have great theories.

Dwarkesh Patel 1:01:57 Yeah.

Adam Brown 1:01:58 And, uh, there are even perfectly respectable, you know, people who are professors at prestigious universities who have very different opinions about what is and isn't a worthwhile, uh, direction to be exploring. Eventually, you hope that this gets grounded in experiment and various other things. But the distance between starting the research program and, uh, the community reaching consensus based on data and, and other, other considerations, uh, can be a long time. So yeah. We- we ha- we definitely don't have a good verifier in physics.

Dwarkesh Patel 1:02:31 Even if we did some- s- someday get superhuman intelligence that could do... that could try to find all the remaining sort of, like, high-level, uh, conceptual breakthroughs-

Adam Brown 1:02:40 Yeah.

Dwarkesh Patel 1:02:40 ... how much more room is there for that? Uh, basically have... And was it just, like, 50 years of, like, here's all the really advanced great physics and now we just bog through, like, uh, additions to the standard model? Um, you know, if you look at Nobel Prizes year after year, they get less and less... At least in physics, they tend to get, like, less and less significant. And in fact this year, the Nobel Prize in physics was awarded to-

Adam Brown 1:03:04 (laughs) Yeah.

Dwarkesh Patel 1:03:04 ... um, Hopfield and Hinton for their work in AI. So apparently- (laughs)

Adam Brown 1:03:08 Yeah, maybe a taste, a taste of things to come.

Dwarkesh Patel 1:03:10 Yeah.

Adam Brown 1:03:11 I don't think there's reason... I don't think we should be pessimistic about that. I think there could easily be room for completely new conceptualizations that change things. I don't think it's just turning the crank going forward. Um, I think new ways to think about things have always been extremely powerful. Sometimes they're fundamental breakthroughs. Sometimes they are, uh, breakthroughs in which you even take regular physics. This is a story to do with renormalization that maybe is a little too technical to get into, but there was a, sort of, uh, amazing understanding in the 1970s about the nature of theories that have been around for forever, or for, for years at that stage, that allowed us to sort of better understand and conceptualize them. Um, so I think there's good reason to think that there's still room for new ideas and completely new ways of understanding, understanding the universe.

Dwarkesh Patel 1:04:03 Do- do- do you have some, uh, hot take about why the current physics community hasn't... Uh, I mean, I get... The cosmology is maybe a very notable exception, where, like, it- it does seem (laughs) like the expected value of the (laughs) light cone keeps on switching back and forth.

Adam Brown 1:04:16 Yeah.

Dwarkesh Patel 1:04:17 Um-

Adam Brown 1:04:17 Well, if you take particle physics, I think it's 'cause we were a victim of our own success-

Dwarkesh Patel 1:04:20 Mm.

Adam Brown 1:04:20 ... is that we wrote down theories in the 1970s, uh, and those theories were what is called standard model, and those theories were too good, uh, in the sense that, uh, we won, um, in the sense that we could predict everything that would come out of a particle accelerator and every particle accelerator that's ever been built, and every particle accelerator that's likely to be built, uh, given our current budgetary constraints. Uh, so particle physics, I mean there were some questions around the edges, but this model that we wrote down in the, in the '70s and into the '80s basically completely cleaned up that field.

Dwarkesh Patel 1:04:59 Mm.

Adam Brown 1:05:00 We wish to build bigger, more powerful particle accelerators to find stuff that, that goes beyond that, but basically we, we won, and, uh, that makes it difficult to immed- you know, to- to- to immediately... Uh, if you, if you get too good, then it's hard to know, know where to push from there.

Dwarkesh Patel 1:05:20 Is- is-

Adam Brown 1:05:20 That's as far as particle con- physics is concerned.

Dwarkesh Patel 1:05:21 Um, is there some... So it sounds like the problem with these colliders is that the entr- like, the expected entropy is, like, not that high of, like, yeah, we... bec- the reason it's not that useful is because, like, we kind of have some sense of what we'd get on the other side. Is there some experimental apparatus that we should build where we in fact do have great uncertainty about what would happen and so we would learn a lot by what the result ends up being?

Adam Brown 1:05:45 Well, the problem with particle colliders is in some sense that they got too expensive. Um, and CERN is, is tens of billions of dollars, a small number of tens of billions of dollars to, to run this thing. Um-

Dwarkesh Patel 1:05:57 You could build AGI with that money. (laughs)

Adam Brown 1:05:58 Right? Yeah. I mean, it's super interesting how everybody talks about how academics can't possibly compete with the big labs-

Dwarkesh Patel 1:06:04 Yeah.

Adam Brown 1:06:05 ... uh, but the cost of CERN is, is larger than the cost of-

Dwarkesh Patel 1:06:09 Yeah.

Adam Brown 1:06:09 ... uh, big model training runs, so, by, by a lot. So and that's just academics pooling their money, so-

Dwarkesh Patel 1:06:14 Yeah.

Adam Brown 1:06:14 ... uh, that's, that's an interesting fact. But yeah. It, it, it's... They got so expensive that, uh, uh, it's difficult to persuade people to buy, to buy a new one for us that's even bigger. Uh, you get... It's a very natural thing to do, to build an atom smasher that just smashes things together at a higher energy. It's a very natural thing to see what comes out. People were perhaps somewhat disappointed with the output of the LHC where it dis- you know, it, it made the Higgs, which was great, and we found it, but we also expected it to be there. Uh, and it didn't make anything else, any of these more fanciful scenarios that... or anything basically un- unexpected, but people had speculated we'd see super symmetry there or we'd see extra dimensions, and basically that was a no result. We didn't see anything like that. Um-I would say we should definitely build another one if it was, uh, if it was cheap to do so. Uh, and we should build another one once AGI has made us all so rich that it's, uh, it's cheap to do so. But it's not, uh, it's not the obvious place to spend $50 billion if you had $50 billion to, to spend on, on science.

Dwarkesh Patel 1:07:22 Hmm.

Adam Brown 1:07:22 Often it's these smaller experiments that, that are, that can be... look for things in unexpected places. Uh, uh, a decade ago there was BICEP2, which is a, you know, reasonably cheap, tens of millions of dollars experiment at the South Pole that thought it had seen some hints in the cosmic microwave background of gravitational waves. That would've been revolutionary-

Dwarkesh Patel 1:07:39 Yeah.

Adam Brown 1:07:40 ... if true. Um, not worth doing BICEP2 if it cost $10 billion. Definitely worth doing, worth doing BICEP2 if it costs $10 million. So there's all sorts of experiments like that, often observational.

Dwarkesh Patel 1:07:49 Wh- what is the value of seeing these primordial gravitational waves?

Adam Brown 1:07:52 Oh, it gives you hints. Uh, you're just examining the night sky very closely and, and seeing hints of what happened at the Big Bang.

Dwarkesh Patel 1:07:58 Right.

Adam Brown 1:07:59 So yeah, this is, this is a sort of different approach to doing high energy physics, um, which is... wh- why do you want to build a big collider? You want to build a big collider because, uh, the bigger the collider, the more high energy you can smash-

Dwarkesh Patel 1:08:14 Yeah.

Adam Brown 1:08:14 ... things together with. And Heisenberg's uncertainty principle says that high energy means short resolution. You can see things on very small scales. Uh, that's great, except the cost to build them is there's some scaling laws, and those scaling laws are not particularly friendly.

Dwarkesh Patel 1:08:30 Yeah.

Adam Brown 1:08:30 Uh, there is another sort of approach that one might say, which is, you know, there was a ginormous explosion, uh, that happened, which was the Big Bang. Um, you know, if you imagine, uh, if we look at how the universe is expanding, if you sort of play the tape backwards, it's contracting. Eventually it all contracts at 13.8 billion years ago in, in, in the Big Bang.

Dwarkesh Patel 1:08:53 Right.

Adam Brown 1:08:53 Um, and, and so that's a very big particle collider indeed. Um, and so by just examining very closely the Big Bang and its aftermath, we're able to hopefully probe some of these quantities that are very difficult to probe-

Dwarkesh Patel 1:09:09 Mm-hmm.

Adam Brown 1:09:10 ... with particle colliders. The disadvantage is that it's, you just, you can't keep running it and adjust the parameters as you see fit. It's just, like, one thing that happened once, and now we're having to peer backwards to, uh, with our telescopes to see what happened. But it can give us hints about things that would be inaccessible with any-

Dwarkesh Patel 1:09:29 Uh, uh, uh-

Adam Brown 1:09:29 ... future, future collider.

Dwarkesh Patel 1:09:29 Is there any information about the distant past that is in principle inaccessible?

Adam Brown 1:09:34 Probably not in principle. So something happened to the universe, uh, in it, in its evolution, which is that the very early universe, just after the Big Bang, was opaque to light. We can only see light past about 300,000 years after, after the putri- Big Bang. Uh, before that, everything's so dense, it's like just a dense plasma that light just gets absorbed by. It's like trying to look through the sun. And so we cannot see directly anything from before 300,000 years. Nevertheless, we can infer lots of stuff that happened from before 300,000 years. In fact, looking at that light, what's called the cosmic microwave background that was emitted at that time, we infer lots of stuff about just due to the patterns of anisotropies that we see in the sky, we can infer a great deal about what was happening earlier. And most of our confidence about modern cosmology comes from a number of experiments that starting in the '80s, but accelerating in the 2000s, really very carefully measured that anisotropy and allowed us to infer stuff before that. At the information theoretic level, there's nothing inaccessible.

Dwarkesh Patel 1:10:42 Mm-hmm. Uh, I guess that makes sense. (laughs) The conservation of information. Uh, and maybe you'll tell me that that also isn't-

Adam Brown 1:10:47 (laughs)

Dwarkesh Patel 1:10:48 ... (laughs) that also isn't true.

Adam Brown 1:10:49 Well, that's a-

Dwarkesh Patel 1:10:50 (laughs)

Adam Brown 1:10:50 ... that's a great question. I mean, there's been a lot of debate in-

Dwarkesh Patel 1:10:52 Uh-huh.

Adam Brown 1:10:52 ... the black hole context about whether information is conserved by black holes, but the modern consensus is that it is.

Dwarkesh Patel 1:10:57 Look, if you're enjoying this conversation, you should consider working for my sponsor, Jane Street. They're a very successful quantitative trading firm. Physicists do particularly well in trading because they can combine hard applied mathematics with a bunch of empirical and theoretical considerations. In fact, Adam once filed for a patent using quantum entanglement and violation of Bell's inequality to do relativistic arbitrage. I don't understand what that means. (laughs) Maybe I should've asked Adam. But if you do, you should go work for Jane Street. Jane Street is keen to hire smart, curious, and rigorous people who want to work on interesting technical problems. You can join Jane Street not just from physics, but also other technical fields like maths and CS. They're always hiring full time, and their summer internship applications are open for just a few more weeks, and they're filling up fast. If you really wanna stand out, check out their Kaggle competition. Go to jonestreet.com/dwarkesh to learn more. And there's also a really interesting video there about their ML work that you should check out. All right, back to Adam.

Dwarkesh Patel 1:12:07 All right, Adam, what are your tips for hitchhiking?

Adam Brown 1:12:09 Oh, good question.

Dwarkesh Patel 1:12:10 (laughs)

Adam Brown 1:12:11 (laughs) Uh, so I, I hitchhiked, uh, a bunch around, around America and, and Europe. I've done, you know, Oxford to Morocco. Uh, when I moved from Princeton out to Stanford I hitchhiked. A bunch of other times down to New Orleans, various other places. Um, I think the, probably the biggest tip for hitchhiking is to stand in a good place. Some counterparty modeling. Imagine the person who's picking you up, they need time to see you, to evaluate you, and to decide they're gonna pick you up, and then to safely stop. And that all needs to happen. So stand somewhere where people can see you, possibly at a stoplight, and where there's a place for them to safely pull over.

Dwarkesh Patel 1:12:50 How do you model the, uh, motivations of people who pick you up? W- what are they getting out of this?

Adam Brown 1:12:56 Um, I think it's different for different people. I think about 20% of people will just always pick up hitchhikers no matter what. Even if I'd...... you know, was dressed very differently and presented very different. I think some people will just pick, pick people up no matter what. I basically fall into that category now. Well, it's hardcoded into my brain that I will 100% pick up hitchhikers, always, under all circumstances, just because enough people have generously picked me up down the years that I just feel as though it's my, my duty and sort of not, not subject to a cost-benefit analysis. Just, it's in there. Um, many other people are evaluating you and just, you know, trying to decide what, what you're in for. Some people are lonely and want somebody to talk to. Some people have a, just a spirit of adventure and find it exciting to pick, pick people up. Certainly, it's not a representative cross-section of people, I would say. There's definitely a selection bias in who picks you up. They tend to be more open and more risk-tolerant.

Dwarkesh Patel 1:13:52 Hmm. And what did, what was your motivation for...

Adam Brown 1:13:55 (laughs) Did you-

Dwarkesh Patel 1:13:57 A- were you just in need of a car or what was going on?

Adam Brown 1:13:58 No. I, um, enjoy meeting people.

Dwarkesh Patel 1:14:02 Yeah.

Adam Brown 1:14:02 And it's, I enjoy the experience of meeting people and weird episodic, uh, s- sense of which just you never know what's gonna happen. I think I have a very high tolerance for ambiguity, and I enjoy that.

Dwarkesh Patel 1:14:15 What, what was the percentage of, "We just had a normal conversation, they went in the general direction I was going, and that was that," versus, "I've got a crazy story to tell about X incident." Uh, uh, what percentages each?

Adam Brown 1:14:27 Hmm. I think some people are just totally normal people, families moving their child to college, and you get-

Dwarkesh Patel 1:14:36 Yeah.

Adam Brown 1:14:36 ... there and you help them, you know-

Dwarkesh Patel 1:14:37 (laughs)

Adam Brown 1:14:38 ... move some stuff into the dorm room just to, just to thank you, all the way through to absolutely wild, uh, cases. Probably 20% are just like, "This is one of the craziest things that ever happened," uh, in one way or another.

Dwarkesh Patel 1:14:52 Yeah. W- what, any particular examples you know about, of the wild things?

Adam Brown 1:14:55 Oh, yeah. Huge. I mean, it's just, a- absolutely fire hose of, of wild things-

Dwarkesh Patel 1:15:00 (laughs)

Adam Brown 1:15:00 ... uh, happening. Um, I could tell so many stories. Like, I remember once there was a, a trucker who picked me up in the, the desert outside Salt Lake City and who, who drove me to Battle Station, Nevada. And who, as we were talking... Uh, the truckers are always, in fact, the most interesting of all.

Dwarkesh Patel 1:15:20 Ah.

Adam Brown 1:15:20 Um, it's, it's typically illegal or anyway in violation of their employment contract for them to pick people up. So, th- those guys are really, um, and in... It's always guys, are, are really pushing the envelope-

Dwarkesh Patel 1:15:33 Yeah.

Adam Brown 1:15:33 ... um, in terms of picking you up. The truckers often will say, "You're the first person I've had in my cab in 20 years of trucking," or something, and then they tell you about 20 years worth of-

Dwarkesh Patel 1:15:45 Mm-hmm.

Adam Brown 1:15:45 ... things that have been on their mind. Uh, so I'd say that those are often, uh, the really interesting ones. As I said, there was this one in, in Utah who was just, just talked from the moment I got into the cab, um, until we got to Nevada. And I kinda got the feeling that he had sort of excess mental capacity and that this was his, you know, he was now just gonna dump it, uh, on me. And, uh, he was telling me all about his life. Um, and I f- remember this very well. How his brother-in-law thought he was a loser, his, his sister's husband, but like now he had the hot fiance, so who was the loser? And then just sort of gradually, over the course of the six hours, it just suddenly occurred to me that his fiance was doing advanced fee fraud on him. And the whole thing was some ginormous... And he was being scammed, uh, by his fiance. And very unfortunately for them, they tried to execute the scam while he had me in the cab. And he never had anyone in his cab, so now he had me in his cab, and they were trying to do some fraud on him. And I was able to... They had some wheat factory in Wales, uh, Uni- United Kingdom, that they had some British High Court document saying that he was entitled to if he paid off the lien on it. There was some long complicated story that was totally, flagrantly false. And I kind of felt like I had a moral obligation to him, to break the news to him. On the other hand, we were in the middle of nowhere-

Dwarkesh Patel 1:17:08 (laughs)

Adam Brown 1:17:08 ... in Nevada and it was clearly a very important part of his personality that this was so.

Dwarkesh Patel 1:17:12 (laughs)

Adam Brown 1:17:12 Um, so I kind of waited until we got close and said, "Is it possible that your fiance is being scammed by these people?" And, you know, sort of raised the notion of scamming and, um, and

Narrator 1:17:27 Yeah.

Adam Brown 1:17:27 ... he ... I think I was willing to intellectually entertain the possibility, and then we c- got a, got a bit closer. "Uh, is it possible that, that you are yourself being scammed by your fiance?" And, and, and then he was like, "No, no, no, it can't be." And he had all these documents to show that it was all legit. And they were just sort of, to somebody from a British legal background, sort of transparent forgeries. And, I mean, he did eventually accept it and was just crying on my shoulder in some truck stop. Um, it was quite, you know, high pathos moment. Um, and then said, uh, "This happened before." And it turned out he'd previously