Steve Hsu is Professor of Theoretical Physics and Computational Mathematics, Science, and Engineering at Michigan State University. Join him for wide-ranging conversations with leading writers, scientists, technologists, academics, entrepreneurs, investors, and more.
Host of the Information Theory Podcast: Welcome to the Information Theory Podcast. This is my episode with Steven Hsu.
Steve Hsu: I would say this about the NIH. It's by far the biggest chunk of our basic science budget. It's way, way bigger than the other major funding agencies. And this will get me into huge trouble, but I'm not a VPR anymore, so I can just say it.
That NIH produces a lot of low quality research.
You're breaking news here, because I don't think I've ever said this publicly. So in 2016, when Peter Thiel was playing a big role in the transition, because he had supported Trump during that election, I almost joined the Trump administration in 2016.
It was at a very senior level that would have required Senate confirmation.
Host of the Information Theory Podcast: Steve is a theoretical physicist, a technology entrepreneur, and a prolific blogger and podcaster.
His main research work has been in quantum field theory, but he's also done pioneering research in computational genomics and co-founded multiple startups. I first heard of him through his blog, Information Processing, which he started writing all the way back in October of 2004.
So first topic, Steve, I want to start with the education of a theoretical physicist.
There's a photograph of you at your Caltech graduation. You're only 19 and you're standing next to Richard Feynman. So there's something almost surreal about that image today because, you know, we think of him as a legendary figure. So my first question is, how did you end up at Caltech at such a young age?
Steve Hsu: Well, Feynman definitely was a legendary figure. It's almost, it's a little mind blowing for me to actually think that, yeah, I actually interacted with this guy over several years. So I was always very precocious as a kid. So I was always highly accelerated. That was unusual back in my day. A lot. So you grew
Host of the Information Theory Podcast: you grew up in the Midwest, right?
but it was a university town.
Steve Hsu: Yeah, I grew up in Ames, Iowa, which is a college town. The local university there is a, it's actually Iowa state university of science and technology. So it's pretty engineering science focused, and has a pretty solid physics department. My dad was a professor.
So I benefited from pretty strong, talented, gifted programs growing up, but there was resistance in the school system to letting you accelerate. And so I was. literally the first kid at my high school who was allowed to take courses at the university during high school. And by the time my senior year rolled around, I was spending half the day at the university.
And so I took a lot of them. I guess even today it sounds pretty unusual. I took quantum mechanics while still in high school. I took courses in ordinary and partial differential equations, linear algebra. I just taught myself. I took complex analysis, which is actually a kind of an advanced class. I think a lot of kids don't take it anymore these days.
so that's, that's, that's, that's what I did. Things like Cauchy's theorem and analytic functions. I took, I think a senior level advanced calculus, like advanced analysis class at Iowa state. So it was pretty unusual. I think I wasn't that far from the core requirements for the applied math degree at Iowa state by the time I finished high school,
Host of the Information Theory Podcast: you actually finished high school, like a few years ahead of time, right?
I assume.
Steve Hsu: Yeah. So I was. 16when I graduated, yeah, so I was young, I'd skipped some grades and, you know, my older brother was very tolerant of me being in the same grade as him. We, and we kind of hung out, we had all the same friends, you know, despite all this crazy stuff that I'm telling you, I had a very all American childhood because, you know, I was a co captain of the swim team and, you know, pretty good athlete given that I was younger than all my competitors.
And just, just had an all American childhood in addition to, you know, studying all these advanced topics, I would actually characterize. I don't know if kids these days anymore. They watch these crazy movies about high school. There's a lot of big whole 80s genres of high school movies like Risky Business and Fast Times at Ridgemont High and stuff like that.
We, we actually lived that. So we actually did all kinds of crazy stuff in high school, drinking, racing our cars, you know, amazing pranks that we pulled off in high school. So anyway, I had a great childhood. I can't, I couldn't possibly, complain about my childhood. And when it came time to go to college, it was a much more naive time than today.
So most of the smartest kids from Iowa would just stay in Iowa and go to Iowa State or University of Iowa. There'd be a few odd kids, especially like kids of professors, Jewish kids that I grew up with. They were more ambitious. They somehow knew they were supposed to apply to like Ivy league schools or Stanford or things like this.
but you know, I had a close friend who got into Cornell, but decided to stay at Iowa state, for example. I think there was another guy who got into Chicago and decided to stay at Iowa state. So that was, that was much more typical of the times. Because I knew about Feynman, you know, Feynman was not even among scientists as much of a cult figure at the time as he is now, but because I knew about him through his Feynman lectures, they had been recommended to me by a professor at Iowa State because he kind of realized like the The level they were teaching the classes at at Iowa State was pretty elementary for me.
And he said, Oh, why don't you get these Feynman lectures and study them? And if you have any questions, we can talk about it. So that was it really.
Host of the Information Theory Podcast: And you mean the physical books, right? Because back then you couldn't just go on YouTube or whatever and watch them.
Steve Hsu: Yes. I have the paperback versions of the, they were a red and white three volume set, which I think I've ordered through the Iowa State University bookstore for like 30.
And, 10 volumes. So I studied those when I was in high school. And so I knew about Feynman. And, I knew something about Caltech. But, you know, other friends of mine would say, Oh, you should go to Harvard or Princeton or something like, or MIT or something like this. But somehow I was really fascinated by Caltech.
And also the fact that it was so small. My graduating class was only 186 kids. Half the size of my high school class, actually. And so I thought, wow, this is a very special environment. Maybe I'll get to meet Feynman. At that time, the two top theoretical physicists, Gell Mann and Feynman, were both there.
And anyway, it was a strange confluence of events by which I ended up there.
Host of the Information Theory Podcast: So in retrospect, for your intellectual development, like, was all this acceleration a good decision? Like, you know, if you were to give advice to a young kid nowadays who has aspirations of becoming a physicist, would you tell them, just try to get through as much as you can in high school, try to finish it early and then go to college.
Steve Hsu: Every kid is different. So I'm unusual in that I am very polymathic and also an autodidact. So I would like to jump into a new area and zero in on what the essential things are. The issue is this actually shocks people that I have friends of mine or colleagues that are interested in economics or biology or genomics or even AI. They're often shocked because I'll come into their office and just start asking them some questions.
And I'll realize like, wait a minute, this guy's like, penetrated pretty far into what is known in the subject. it's not true of all kids, not even all like smart kids. So, they're playing kids who are really, really smart, maybe like math Olympiad, smart kids or whatever. Yeah. But they're not necessarily going to be able to go into a particular subject and create a coherent, logical overview of what's important or not important in the subject and, and self guide their learning.
I had plenty of friends throughout the years who were, you know, at that level of IMO gold medalist or Putnam fellow, but they, they needed to sit in a classroom and have a really good professor lead them through the material, even though they could pick it up really fast once, once they were taught it.
very much. They could pick it up really fast, but they didn't necessarily figure out what were the interesting things to try to understand.
Host of the Information Theory Podcast: You think that's just like a personality difference or what accounts for that?
Steve Hsu: I think it's a cognitive difference. I feel like I'm just logically strong. Like if I start learning the subject, I can fit it into a coherent structure.
And then if I see a hole in the structure or especially a foundational hole, I, I look, just zero in on it and say, well, how do you know, like all of this is built on this, but how confident are you that this is correct? And so that, that capability is just, I think, just a specific idiosyncrasy of mine.
Host of the Information Theory Podcast: So it's kind of like the desire to drill down and like question assumptions or like think from first principles. Do you think that's an accurate way of summarizing? Yeah, I
Steve Hsu: I think actually Feynman was this way too. So the idea is like, you should be able to rebuild it all yourself. So I don't like it, you know, why are things this way?
And then someone just says, Oh, well, it is. Then go look it up in this book or something like that. Like that's not satisfying to me. To me, I should be able to sit down with a piece of paper or whiteboard and actually kind of work it through from first principles. And if I can't, then that's a, like a signal, like, wait, I gotta, I gotta understand this point better.
And
Host of the Information Theory Podcast: isn't that, isn't that extremely time consuming? Like, don't you find yourself getting into rabbit holes that just consume days of your life? If you, if you start digging down this way?
Steve Hsu: Yes and no. So first of all, you need to have a lot of horsepower to operate this way. And this is something people said about Feynman too, is like, if, if he didn't have so much horsepower, they would have disregarded him as kind of a crank and he would have gotten caught in rabbit holes and not reemerged.
So you need, you do need a lot of horsepower to, to be able to figure stuff out fast. And, but some of it is you have to know when to coarse grain over something and when not to. So if there's some body of intuition that this group of people is just telling me that's true, I can provisionally accept it. I can say, okay, let me provisionally accept that the stylized fact is true.
So let me just accept that provisionally. But I have a mental note that that's actually a conjecture, but then I can move forward. And so, so of course, like I know some guys who are very rigorous math guys. They will get caught up because, like even physics, even the well accepted body of physics, there are places where if you want to make it fully mathematically rigorous to the standards of professional mathematicians, you can't actually get past that.
And so you need this, you need to have this dichotomy where, yes, I do want a first principles logic. I want to build that myself, but I can, I can accept that certain nodes in that graph are actually provisional. And that's very powerful because now if you're setting some field like genomics, where there's some conjecture, this is just junk DNA, or this doesn't matter, or this does matter.
You can accept it and then proceed, but you have in the back of your mind an alternate map. That might actually be the real map. We might have to default to this other one. If I get some empirical data that contradicts one of these provisional hypotheses, I might default over to this other map. And, and real science, not, not math, because math is not actually the same as science.
To be good at real science, you have to be able to have those provisional assumptions. built in and work with them, but then change them when you need to change them.
Host of the Information Theory Podcast: Concretely in the 2010s, you transitioned or forayed into computational genomics, right? And that was not your training at all. I'm wondering, how did you even proceed?
Like, did you read a bunch of textbooks? Did you go through a lot of papers? There's multiple bodies of knowledge that you need to be conversant in before you even start, right? Like you need to learn population genetics. You need to learn molecular biology. Did you actually have to
Steve Hsu: So I had some prior training because when I entered Caltech, this was in the 80s, people were already saying that physics is over.
Like all the important stuff has been figured out. And, but the really exciting place where the first derivative, the pace of progress, is really fast is molecular biology, which was true because there are all kinds of things that just weren't understood at all. Like even, you know, the nature of DNA or something, and then like fairly rapidly You know, they made a lot of progress on those specific molecular mechanisms and stuff.
So I took some classes in molecular biology at Caltech. So I had some idea of what they were doing, but when I took those classes, I realized they are so far away, because at that time they had not come even close to sequencing the genome of an organism, let alone a human. I just realized the theoretical questions that I have in mind, because I learned about natural selection and evolution when I was still in high school.
So already, like, in high school, what I learned about, you know, like, when you take biology and they teach you about natural selection, if you're more mathematically inclined, you're like, wait, is there, is there a more, because the biology teacher is not going to teach you a mathematical version of it, which, which today you would call like population genetics or something, right?
Which is descriptive of allele frequencies. How do allele frequencies change given some relationship between the different allele variants and fitness? And so how do you describe all that? If you're trained in physics type stuff, you might think of some kind of either. differential or difference equation that maps the distribution of alleles one generation to the next in the population.
So I already had that in mind, that this was like an interesting theoretical problem in a natural system. But then when I encountered real molecular biology in college at Caltech, I realized, wait, these guys are so far from being able to connect the stuff they're doing in the lab to the actual, this, kind of theoretical construct of population genetics, which is actually true.
So, it's kind of tragic that, like the early people in population genetics, like Fisher, there were very few instances where he could directly connect the mathematics that he was working on to actual observations or empirical studies, let alone actual DNA sequences. So I kind of understood all that. And I was in a state of waiting because I knew like in my mind, I knew this is a field, which could become tractable during my lifetime.
And it all depends on technological innovation, like how, how inexpensively are we going to be able to sequence DNA and stuff like this? So I was kind of waiting. And then in the 2010 era, that's when I started to see articles with graphs showing the two per exponential decline of sequencing costs, over time.
And so I, I just realized, Oh, if I extrapolate this easily within my lifetime, these, these questions that I had just been describing to you will, will be approachable. And so I thought, okay, now I got to, let me drill down now and start thinking about this stuff. And let me also say that like, this is a case where science fiction actually influences my taste in problems because.
I really like the novel Dune, and if you remember in Dune that Benny Jesseret have been working for 10, 000 years to breed a super mentat or Kwisatz Haderach, you know, the guy who Paul Atreides ends up being, and because they had, they had fought this war with the machines and outlawed computers, they outlawed thinking machines.
The Bene Gesserit had to use some very old timey kind of like, like sort of cattle breeding kind of methodology to figure out who would actually, who was likely to become the Kwisatz Haderach. So, like this whole problem had been lodged in my brain thanks to Frank Herbert for a long time like well, oh, so if you want to make a superhuman.
How exactly do you do it? And what's science? Is there real science behind this eventually or, or not? And so I was, I was very sensitized to, like, that's an interesting thing that if we had better technology, we might eventually make progress on that problem. And so around 2010, that's when I said, Oh, let me just roll up my sleeves and start thinking about this a little bit.
At first, I sort of had this dual strategy where I was reading the literature. Like, I was just searching on the internet trying to find, like, what is the state of the art? What is known about these things? and at the same time, the way a physicist or mathematician works, you just have a blank piece of paper and you say, Well, what are we likely to get?
We're likely to be in a situation where we have, say, a million people, we have the genome read out for each of those million people, and then we have some vector describing all their different phenotypes. Let's just suppose I have that information, and there is some map between the state of each allele.
and the phenotype, what is the best algorithm for backing that model out? You're trying to, you're trying to back out a model from a data set like I just described, because I thought, okay, we'll get to this point where we have that data set. What is the algorithm that can, in an automated fashion, learn the right predictive model?
And so then I started thinking specifically about the math of like, how do you do that? Now, it turns out there's been a lot of work on that. So the. In this, there's a whole field called compressed sensing, which, you know, Terry Tao worked in, actually, and so anyway, it turned out I was able to make contact with a whole body of mathematics that already existed, and coincidentally, Terry, Terrence Tao came to the University of Oregon right around that time in the early 2010s to give a lecture, and the lecture was actually on compressed sensing, so, so, I was primed until I realized, wait a minute, this can be applied in genomics.
And so that's sort of how it went.
Host of the Information Theory Podcast: So was it purely a mathematical result or, I mean, it, it, it, It doesn't seem like something that could be figured out from first principles, right? Like you must have relied on some experimental results as well to know that the genomic architecture is additive, right?
Steve Hsu: So the main conjecture that had to be made was that most phenotypes are sparse.
So in other words, in principle, all 3 billion base pairs could affect, let's suppose a phenotype is height. It could be that your height depends on the state of all three billion base pairs, okay? Or the counter assumption is it's a small fraction of those three billion base pairs that affect height.
It's a small subset. And I was willing to accept that conjecture, that it was going to be a small subset for each particular phenotype. And in fact, it turns out that there's largely disjoint sets. So the set of SNPs. They're of course spread out, but the specific subset that's controlling your height is largely disjoint from the subset that's controlling your diabetes risk and the subset that's controlling your diabetes.
You know, heart disease risk. So, like that assumption of sparsity was enough because if you're trying to reconstruct a sparse signal from noisy data, it turns out compressed sensing provides an almost optimal methodology for doing this. By the way, this particular set of things I'm telling you is not actually appreciated by most genomicists today.
Like most genomicists, if I go to a conference and I'm trying to explain this to them, they don't really care. They're like, well, I use this algorithm or I'm doing that or whatever. If I tell them, no, there's actually this theorem of opt regarding optimality and performance guarantees of this particular set.
It's called L1 penalization. They're actually performance guarantees of how this particular algorithm will perform on that kind of problem. Most genomicists don't actually know that. So I, I already will, like, after reading Terry's papers and coming to the problem from this kind of abstract perspective already was, I think, way ahead of what other people in the field were thinking.
Host of the Information Theory Podcast: Yeah, yeah, that reminds me. I mean, I have a friend who's very deep into machine learning. You know, it's his full time job and he basically complains about that. Most practitioners have no idea what they're doing from first principles and they're just, you know, using, using packages.
Steve Hsu: Yeah, that's exactly how it works.
So you could say that like, okay, genomics is a, in a way, like this part of genomics is a branch of machine learning. In machine learning, you have a very broad spectrum where some guys are just using cookbooks or recipes or other people's packages just to try to do something. But then there are people, like, if you go to the math department, you can find guys that are actually proving theorems.
So you have both ends of the spectrum but the people on this far end of the spectrum there's a very small population of those people, whereas most people are over here.
Host of the Information Theory Podcast: So.
To get back to theoretical physics. So can you break down what a theoretical physicist actually does? Kind of like minute by minute, right?
Like, do you spend most of your time just thinking silently in a room? Like, are you scribbling equations or are you going off? On long walks. and I assume there's like a lot of collaboration with your collaborators. Right. So yeah, I'm curious about how that works.
Steve Hsu: It's kind of all of the above and it really, it could vary by what particular thing is occupying my attention.
So you could be in a phase where there's a particular area that you want to work in and. It's a very active area. So you spend a lot of time looking at the archive, because like maybe every day or every week, there are new papers coming out on the exact thing that you want to work on. And so you have to spend a lot of time reading other people's papers and figuring out things like, you know trying to, trying to learn from what they've done, what's right about what they did, what's wrong, what, you know, what should I try to make use of?
And then sometimes you're emailing them. And saying like, Hey, I don't really think equation 44 follows from 42 the way you say. So there's a lot of that going on. And a lot of that in reverse too, because every time you write a paper, there's people writing to you about what they didn't follow in your paper, what they object to in your paper.
So there's a lot of that going on. I actually don't like having to read other people's papers in the archive. Like, like to me, like if my postdocs or students want to drag me into an area where it's kind of like that, where you have to, you have to follow a lot of what other people are doing and it's very kind of crowded.
thing. I'll do it if the problem, if the specific thing is interesting enough, but I won't like it because I don't want to have to like to read other people's stuff. Okay. So my ideal mode of working, which you get too much later in your career. So I've been like a professor for almost 30 years. And so after a while there, you're aware of certain problems that are not fully resolved, or at least not everything has been worked out that should be worked out.
And you know it well enough that you can return to it. Because you know it and you can come back to it and just think about it. So that's the most pleasurable phase where like I could be taking a long walk and I'm actually thinking about something that I've already thought for years about but I'm coming back to it with maybe a new perspective or or something like that or maybe you could try this or that's the most pleasurable part for me that and also just discussing physics with another theoretician, like if I'm, you know, I just came back from this trip in China and like, you know, discussing some theorists that I've never met before, maybe some student who's still pretty young, like we could have this interesting conversation.
I enjoy that a lot. One of the things about having worked on stuff, a body of knowledge for decades is that you have a lot of it in your head. And in a way, when I look back and I say like, Oh man, I could have gone to work at Renaissance and made likeI'd be retired with a hundred million bucks right now.
If I had done that. But then I think to myself, yeah, but I wouldn't have this super finely constructed understanding of quantum mechanics or quantum field theory in my head that I can actually manipulate the way that I do. So like that you'll never get that other than spending a lifetime doing this kind of work.
The other thing I would want to say is that like, I noticed when I was a kid that my dad was sometimes absent. He seemed absent minded to me. Now I realize why he's a space monster. It's not necessarily a personality thing. It's actually just that if you're that deep into a subject and you're an intellectual and you keep returning to it, you can do stuff internally.
Like he might be at the dinner table, but he's like actually working through something in his head. And he can do that because he's developed that capability over many years. But to us, it looks like he's a weirdo.
Host of the Information Theory Podcast: So, in your work, like, I'm wondering, what's the nature of the relationship between math and physics?
So, do you use your intuition to work through some thought experiments to come up with some story of what happens in a black hole, and then try to back fit some math that formalizes it? Or is it the other way around, where you start with some equations and then go wherever the math takes you, you know, derive some physical results from that.
you know, like how you can derive the speed of light from Maxwell's equations. Like which one of those two modes do you work in?
Steve Hsu: So this is a really great question. Every physicist, every theorist is different on this particular question. And I am more toward the physical intuition where I'm thinking about what's happening, I'm visualizing it, or I'm even like in a kind of logical way, like saying, well, if this happens.
Like, verbally, if this happens, then wait, doesn't it imply this, and that wouldn't that? So, oftentimes I'm not just automatically manipulating equations. I'm manipulating physical intuitions and concepts. And there are times when a problem has been boiled down enough. There are a couple of equations I could write down on a sheet of paper and just start noodling with them and then get to something non-trivial just through that sort of, in a sense, mechanical process.
For me, that's the minority. of the situation. So usually if I make some progress, it's not through that. I am actually jealous of people because there are people like I actually have colleagues who if I'm talking to them and I'm trying to explain my intuition about the problem, they're like, they're like, shut up, Steve.
You're just confusing me. And then they like, they want to write out equations and just manipulate the equations and they can make progress. Just, you know, again, mechanically, it's not really fair to say mechanically, but, but what seems to be mechanical manipulation. Of equations that can actually make non trivial progress.
So I have one collaborator/collaborator, a guy called Roman Bunny who's actually Ukrainian, Western Ukrainian, but trained in what was the Soviet system. He's exactly the opposite. So, we make a great team. Because, like, I can, I can reason through something and I say, like, I'm pretty sure this is true.
And I'm good enough with equations that I can then formalize it for him. But then he's better than me. He'll generate like a hundred pages of calculations, like within, like if we're working on a problem and then like he goes away and kind of calls me back in a few days, he might've literally generated like a hundred pages of calculations and also use Mathematica to do stuff.
And so we're very complimentary in that way. I'm actually jealous of people that way, because there are certain types of breakthroughs that I'll probably never make because of my style. And they're certain types of things he will never do because of his style. And then there are some people who probably have both who are probably equally as strong as I am in one way and as strong as Roman is in the other way.
Like I'm envious of those people.
Host of the Information Theory Podcast: I guess what you're saying is like progress actually happens in both directions, like by people going in both directions.
Steve Hsu: Yes, absolutely.
Host of the Information Theory Podcast: So the next topic I want to talk about is scientific progress in general. So there's this idea that's associated with Tyler Cowen called the great stagnation.
The idea is that economic progress measured by medium wage or living standards has slowed down since the 1970s. And he has some theories about why that is. So do you agree with the broad contours of the theory? And what do you think is the underlying cause of the slowdown in progress?
Steve Hsu: Yeah, I could haveI could have a huge multi day discussion about this with Tyler and, Patrick calls and all these guys.
First of all, there is a measurement problem. Like Are you sure? Like, what do you mean by living standards slowing down? Like, you and I are having this conversation, I've got my phone on a little tripod, and we're thousands of miles apart, but we're having this great conversation which we never could have had 10 years ago, or 15, 20 years ago, right?
Or, like, I spend, people spend all their time, like, doing stuff on their phone, which they only pay, you know, if they bought a cheap phone like mine, a few hundred bucks for. What do you mean by the way living standards went down? How are you measuring that? Because if I look at what people actually do with their time, some of the inventions that we came up with recently have totally morphed the way people spend their time.
And so, isn't that an order one change in your quality of life? Because I prefer that to what I used to be doing, obviously, by my own revealed preference. It just happens that we can produce that very cheaply. So what do you mean by the rate of increase of your economic standard of living slowed down? So there is this overall measurement problem that I think needs, needs more attention.
I think economists are too willing to just accept that because they believe in equilibrium. So, so basically they'll just say like, well, whatever the dollar value of this thing, that's the value of this thing, right? And I'm like, wait, wait, hold on, right? Beyond that, though, I would say, like, if you look at other metrics, which are, which are not as subject to measurement error, like if you say, like, well, what's the energy usage per capita or energy production per capita in the United States, it's flat, it hasn't changed in, you know, 50 years or something.
So you could say like, well, okay, aside from like, this kind of fuzzy economic, you know, measured in units of utils or dollars or something, let's put that aside for a second, but just say like, Well, our ability to manipulate the physical world has not, the rate of our improvement in the ability to manipulate the physical world has slowed down.
Host of the Information Theory Podcast: And right, like all the, all the progress has been in, bits instead of atoms, right? That's like Peter Thiel's formulation, right?
Steve Hsu: So just to give a little more nuance to that, I think the right, the right claim is macroscopic physical quantities. Like how fast is my car? How fast is my airplane? How many kilowatt hours of energy, you know, are delivered to my house for me to use?
Those have not changed that much in the last 50 years. And so you could regard that as a real slowdown. On the other hand, it could be that, like, it just turns out nature is hard that way. Like, it's just not easy for me to have a car that flies at supersonic speeds, right? There are fundamental aspects of physics itself that make that difficult.
And we shifted our energies to manipulating the nanoworld. And manipulating the nanoworld to make semiconductors, to make chips and memories and stuff, we have gotten factors of a million in my lifetime, more than a million. And so, I kind of feel like people just approach this in a way too naive way.
Host of the Information Theory Podcast: It seems like you do agree with their thesis that like the low hanging fruit of like technological breakthroughs has been exhausted, right?
Steve Hsu: Yeah, so to me, like one of the most persuasive ways they would frame the problem is to say like the way my grandmother's life changed where she started out with no running water, no electricity.
You know, and she died with an iPad in her lap and, you know, no shortage of electrical energy or heat or, and a washing machine and a robot that's like vacuuming the floor for them to say like, okay, the arc of change for my grandma's life. It is so different from what I seem to be experiencing and we seem to be capped out on these. What I was just referring to a moment ago is like macroscopic physical quantities.
I think that's true. Now, is it because our systems or institutions suck? Or is it just because fundamentally we picked some low hanging fruit? Internal combustion engine. We figured out how to do that. Bernoulli effect for lift and air, you know, like air foils. Like, okay, we picked some low hanging fruit.
And then the next, the next fruit up there is pretty hard to get. And it isn't a fault of, it isn't an institutional failure or a systemic failure or something, you know. Problem with capitalism, that's very possible in my mind. So I'm not saying I know the answer to this stuff but no one's really fully excluded that line of thinking and until you exclude that line of thinking then it's just like we live on an innovation surface in certain directions of the innovation surface.
You can climb up Easily and you get to a certain point, but then it plateaus And there's nothing anybody can do about that, right? And, in which case, then, like, trying to blame, like, the NSF or something, or NIH, is, is just the wrong, you're, like, just barking up the wrong tree. It's possible, despite the fact that there are problems with our systems, and incentives, and the way academic research works, there are problems, but, but that might not be the main reason why energy production per capita in the U.
S. has plateaued.
Host of the Information Theory Podcast: Right, so you're agnostic about even the whole framing of the idea that progress has slowed down because you, you point to like all this progress in the digital space and then I guess you come down on like it's probably caused by just like low hanging fruit being exhausted and not any kind of like failure of like how our society is organized.
I think
Steve Hsu: it could be both. Like it could be there are problems with the way society is organized. And there's low, there's a low hanging fruit effect. Now, when I'm in Silicon Valley, Contabs with a lot of people like engineers, and when, when someone just says to me, yeah, what about that stagnation problem, you know, like not knowing that I've spent hours talking to Tyler and those guys about it.
And maybe I would flatter myself and say like, even though much more deeply about it than they have. I turned to the kid and I just said like, well, let's see, let's hypothetically, we might be. Literally one generation from being able to fully edit and control the human genome. And, also, like, we might be within one generation of AGI and ASI, where A is superintelligence.
Okay, like, way beyond human intelligence. If that's true, what fucking slowdown are you talking about? Like what, like, if we come back and have this conversation in 20 years, and 90 percent of scientific research has been wrested from our hands by artificial minds that we created, like, in what way was there a slowdown?
Like, I'm not sure what, you know, so
Host of the Information Theory Podcast: that's my view. That's one way in which digital progress can just spill over into the physical world all of a sudden.
Steve Hsu: Well, it could be that it could be that we have advanced super intelligences and stuff, but it's still not easy to build little vehicles that let me zip around the planet at the speed of light, you know, because like, gee, aren't, aren't I going to burn up, burn to a crisp, trying to move that fast through an atmosphere.
singularity. They're, like , usually miscalculated about exactly what's going to happen the moment after we get to the singularity. Which things is the superintelligence just going to quote, solve and, and like improve by a thousand X and which things are even going to find hard because they're just basic physical constraints, for it to change.
Right. So it's not, not clear.
Host of the Information Theory Podcast: So all the recent technological progress in the digital space has been enabled by Moore's law. Could you explain for the audience what Moore's law is? And then like, why have we been able to sustain so much progress in this very specific domain for such a long amount of time?
Yeah.
Steve Hsu: And who do we have to thank for it? Right. So this is one of my hobby horses. Maybe you're, you might be asking me this question cause you've detected this in stuff I've written. But so Moore's law is really our ability to continuously further miniaturize the components that we use for information processing devices like computer memories or computer microprocessors.
And, and we've gotten factors of a million in the last few decades. Right. And so in a way it's like the most dramatic progress ever, like in any activity, right? And so the funny part is like, again, like I don't want to, I don't want to piss off Tyler and this whole class of people called economists who purport to study the economy or whatever they are.
It's literally the fact that they didn't continue their studies in physics. So they literally don't understand physics. And so they coarse grained over this whole activity. This is a huge activity, right? The leading countries, Japan, China, Taiwan, the U. S., in some sense Europe, are funneling a significant amount of brainpower into this activity.
So if you ask, like, the class of people who are smart enough to do a Ph. D. in physics or Ph. D. in engineering, well, what do they do in the economy? Well, a very large subset of those people are involved in the semiconductor industry. There's just, it's just a very big industry, right? If you set the threshold of brain power, you know, set a threshold and then look at the people above that threshold, a very big chunk of super smart, highly diligent, capital intensive activity is there.
And if you don't understand physics, like how could you possibly expect to understand that whole segment? Human activity, which I would argue is driving , causes an order one change in the way people spend their time. Literally the amount of what my kid when she was 12 was doing all day long has changed by order one, not by 1 percent or 10%.
But compared to what I did all day long when I was 12, because you know, I'm a bad parent, so I got her a phone, right? So, if I let her, she would like, spend like half the day on her phone. So we invented something and it's not like a trivial thing like, oh, I like I got her addicted to opium and she just takes the opium for half the day.
No, it's like, it's like a thing, which is like a supercomputer that, you know, people in the 70s could never imagine. Right? And it connects her to every library in the world and archive. And so we built this non trivial thing and we built it on top of this mastery of basically nanophysics, right?
Everything that goes into making these components in the phone is, is nanophysics, right? You know, even the display. So I just think these guys don't know what the fuck they're talking about. They are literally not educated. So imagine like, imagine like some guy just never learned algebra. Right. And then like he comes along and he's, he's like critiquing, like, Wow, I don't understand why it took them so long to solve Fermat's Last Theorem.
It's, it's, it's, there must be a problem in math departments worldwide because it took them, and this guy literally doesn't, like, know how to solve a, the guy who's saying this. Literally doesn't know how to solve a quadratic equation, but he's, he's critiquing how long it took to solve the Poincare conjecture or something.
Right. That's I'm obviously exaggerating for effect, but that's kind of how I feel when I talk to these economists,
Host of the Information Theory Podcast: they just don't have, so basically the economists are just underestimating what an incredible achievement it was to like meteor tries all these, Silicon devices
Steve Hsu: to say it in a way, which I think probably you would understand, but they would not understand there isn't a natural metric.
Like, what is the right metric to measure progress, right? And there isn't a natural one. They might use dollars, but dollars are affected by supply and demand, right? And, and, and other things. The one I proposed a moment ago was like, what kind of order one is to change in the way people live, the way people live was, was, was created.
And clearly there was one. So therefore you can't say it was small, like you can't say we were stagnant. Right? There isn't a natural metric. These guys don't, I think, really have an abstract understanding of what that means, so they just adopt their metric, and then they notice in their metric things are not changing as fast as they want, or as they expect.
Maybe, maybe that isn't the natural metric to use. Right, right. Yeah, but I also feel really frustrated with economists because I just don't think they understand innovation. Like, how does innovation actually happen? Like, how do we get better ideas about how physical systems work? How do we then apply that to make better devices?
How do we bring those devices to the market in an affordable way so that everybody has one? Like, I don't know how many billions of cell phones there are now. Uh, smartphones, they underestimate all of that stuff, all the complexity of that stuff.
Host of the Information Theory Podcast: So speaking of kind of like the pipeline from science to technology innovation to, you know, the economyI want to talk about your experience as a university administrator.
So in 2012, you were tapped to be the vice president of research at Michigan State University, where you managed hundreds of millions of dollars in research expenditures. so you had an unconventional path to that point. You had been a theoretical physicist, a Silicon Valley entrepreneur, and suddenly you're in charge of a 400 million research enterprise.
So first of all, how surprised were you when you found out you were being considered for the position?
Steve Hsu: I was very surprised. I got a phone call from a headhunter. I was at a very academic physics conference on black hole information. And I get this. This headhunter asked me if I'm interested in this job and the first thing I asked her was, are you sure you have the right person?
I thought maybe she had mistaken me for Steve Chu, the guy who had been the energy secretary. Obviously Steve Chu wouldn't take this job, but, but somebody, some mistake like that. Sometimes people mistake me for Steve Chu. The first thing I asked her was like, are you, like, do you have the right guy?
Like, are you talking to the right person? It turns out when I, after I became a VPR and their annual meetings, like, there's the top U. S. universities are all part of the AAU, American Association of Universities. And each of the key executive positions at the universities, whether it's provost or president or vice president for research or whatever, they have an annual meeting where you meet all your peers.
So you're in a room, I'm in a room with like 64 other VPRs from all the other top schools and, you know, a fair number of them are theoretical physicists. So it's not, it's not that unusual. Yeah. It is a job where you, it helps to have a ton of breadth because it, it's not uncommon for someone, it's like I'm, say I'm a biomedical researcher and my main field is like, you know, artificial heart pumps or something.
Right? It's very possible that guy doesn't really know anything about polymer chemistry, doesn't know what a cork is, doesn't know what p versus NP is, but they have to do this job and, and the, the people at their university are working on all those things and more. So it is in a way, one of the more intellectually interesting jobs at the university, because in my time I had to make decisions about supercomputing center, gene editing lab, BSL for, you know, secure facility for researchers studying spread of malaria via, via mosquitoes, you know are we going to have a primate facility on campus?
You know, you can't, you can't like. There's no way to capture the complex, the set of possible things that can arise. If you're at a big 10, you know, R1 research university and you have all these professors trying to do different things and, and you, you're sitting on top of all that activity. You know, it is a very interesting job or potentially very interesting job.
Host of the Information Theory Podcast: So could you walk me through how research funding actually flows through a major university like MSU? Like just from a high level, like where does the money come from? Like How is it allocated? How do you make those decisions about where it goes?
Steve Hsu: So that number that you mentioned, I think when I came in, our annual research expenditure number was a little over 400 million.
Then when I left, it was about 700
Host of the Information Theory Podcast: million.
Steve Hsu: So for the very top schools in the country, it could be over a billion a year. Especially if they have a big, major medical center. The money typically comes from funding agencies like NIH, NSF, Department of Energy, Department of Education. But some of it comes from private foundations like Ford Foundation or MacArthur Foundation.
And it's allocated through competitive grants. So your faculty are always writing grant proposals to try to get money to fund their research. And there has to be some kind of accounting procedure where dollars come into the university, They're held by the university, like the university becomes literally a kind of bank, and then like individual researchers can draw from their research account to order a computer, order a supercomputer, you know, pay some grad student their salary, hire a staff scientist, you know, buy a monkey and put it in a cage over here.
All of that is in the, in this crazy system. These grants are awarded typically through peer review. So typically each funding agency is appointing a committee of people to oversee a particular grant process. And then they're rating the proposals from the professors and the top rated proposals will get funded by the funding agency.
Host of the Information Theory Podcast: So then how much leeway do you have to, you know, as the administrator to redistribute funds, like among the different researchers, right? Because I understand there's like some element of overhead that's like. taken by the university and then they have more leeway over that.
Steve Hsu: Yes, so every university charges.
a certain amount of overhead on these grants. So if a dollar comes from NSF and goes to Professor X, our overhead rate might be 63%, so then an additional 63 cents is charged to NSF. So if the researcher gets a million dollar grant, hundreds of thousands of dollars of that grant are going to go to overhead.
That goes into my office. Maybe there's a split with the provost office. I then have some leeway to say things like, We have some funds that we set aside to, for example, hire new faculty. And so a big part of this is like a money ball. Uh, it's just like major league teams trying to get a particular athlete and you have to put together the package, the comp package in this case.
It's not necessarily comp, but it could be like, she needs a million. She needs 1. 5 million startup package to set up her lab, hire the First set of grad students and postdocs do these renovations by that laser. But
Host of the Information Theory Podcast: then,
Steve Hsu: Yeah, go ahead.
Host of the Information Theory Podcast: How do you even decide in the first place that this is, you know, cause obviously I'm sure everybody has their own priorities and the amount of demand exceeds the supply of money you have sloshing around.
So like, you know, you have to make some like prioritization, like you have to decide, do I give money to physics or ag science or medical research, like. How do you even begin to prioritize those different things? Right. It sounds like something a central planner in the Soviet Union would have to deal with.
What kind of frameworks did you use to think about relative value?
Steve Hsu: That's what makes it a super hard job. There are some systems where things are very similar. Determine almost by formulas. There's no judgment involved. We were more at the other limit because the president here had a lot of confidence in me, where my office had a lot of leeway in terms of what strategic initiatives were we gonna go after, where the money was gonna go.
And consequently, we made a lot of enemies. So there were fields where they just kind of knew they weren't going to get any money from us. And there were other fields where we built up Michigan State substantially. We built a billion dollar DOE nuclear accelerator lab on our campus. So it's a DOE lab, but it's on our campus and we run it for them.
And it's a billion dollar project. So we picked winners, not all BPR offices or even provosts or presidents will pick winners on their campus. And maybe they shouldn't because they're not qualified.
Host of the Information Theory Podcast: You're saying some people are just on autopilot, but you had a lot of discretion. So like, what, what were the criteria you actually use, right?
Like for that specific project, how did you decide it was more valuable than the others that could have been funded?
Steve Hsu: You need, there are like several criteria. Yeah. One could just be money. Okay. So one thing VPRs are measured by, like I said, we grew our research expenditures from a little over 400 million to 700 million during my time.
Host of the Information Theory Podcast: So just fund people who are able to come up with more money, the more, the more you fund. Yeah.
Steve Hsu: So it's like, Oh this is a hot field. NIH really wants to fund this field. We already have a strong group here. So the new person, the N plus one person can thrive. Okay, there's a good argument for that. oh, this field gets no money, they're, like, federal, like, they're, like, they're tuning down the support for this.
Area of polymer physics or something. Why should we hire another person in that area? So there's just that, but then the other part of it is like actual intellectual depth is this really an interesting area of, you know, science or whatever. Engineering. And for that, you need, you need actually an understanding you like the person who's affecting that part of the calculation.
Like, is this non trivial or is this bullshit? For example, there are whole areas of biology where they take imaging technology that comes mainly from physics or engineering. It comes into biology and then all of a sudden they're able to make really beautiful pictures. Like, oh my god, look at this electron microscope image of this ant's face.
Or this, this bee's eye. Those can get on the cover of Nature, and the biologists seem to love this. But you have to decide, like, is this a real advance, or is this just, like, kind of a superficial thing, and is it really gonna, what's it gonna lead to? Like, do, do they actually understand the ant better?
Because of this picture of his face. I,
Host of the Information Theory Podcast: you know,
Steve Hsu: So there were times when I was saying to like the people on my team or the Dean or the relevant people in the university, like, I understand this is a hot area, but I think it's bullshit. Okay. So, I could be in that position. Or I could be, like, I had a whole thing while I was here where I was trying to emphasize to the professors on campus that reproducibility or replication of research was a real concern.
Now it's well known that, like, a lot of work in certain fields just doesn't replicate. But at the time, I was very early in saying, if psychology wanted to hire a social psychologist, and I was like, Yeah, but I, I don't think that stuff replicates. Is that stuff even real? You know, and they were saying like, well, this is the hottest stuff, man.
You can, you can be like in a Malcolm Gladwell book. And, andNSF actually likes to fund it. And you can be on the cover. You can be in the New York Times if you, you know, I'm like, yeah, but is it real? Like, so,
Host of the Information Theory Podcast: yeah, yeah, it's pretty shocking. Like how little social science research is able to be replicated.
Right. And then, yeah. Uh, you know, last year, Mark Tessier Levine you know, this prominent neuroscientist, he's actually the president of Stanford. He had to resign his position because basically allegations that his entire research about Alzheimer's was fabricated. How surprised were you? And I guess like,
Steve Hsu: I was not shocked that I spent eight, literally eight years trying to educate my faculty at my university, that this was the case.
So, there's a researcher
Host of the Information Theory Podcast: at Stanford. That what, that what was the case? That people were just outright engaging in fraud? Or that things are hard to recognize?
Steve Hsu: Outright fraud or wishful thinking and poor, you know, statistical sophistication leads to people believing in results that really the evidence for those results is not very strong.
So it's a mix of those things. I actually think the fraud part, well, at the time, I thought the fraud part was the minor component. The bigger component is, you have whole sets of researchers who barely passed their watered down statistics course in grad school. So there are whole fields. You can look and see, like, Which Ph.
D. programs do the grad students, can they actually take the real stats course in the math department or in the stats department or do they have to like kind of make their own stats course which is much easier and the students still regard that as like one of the hardest courses in the psychology department in order to get their Ph.
D., right? If you have a population of people that's adversely selected for ability to do statistical reasoning, then they're very, very susceptible. To believe results that just sound good. But the confidence level they should get from the evidence is not as high as, as, you know, what the narrative creates in their mind.
So my contention is there are whole populations like that, even in the quote, science professorial, including in biology. And so, and especially biomedical science, like a lot of stuff that people think they know in biomedical science is not actually true. And actually, if you, if you were sophisticated, both in statistics and in the details of the studies.
You would look at it and say like, yeah, 50 50. That's true. Okay. Right. So when I was here, there's a researcher at Stanford named Ioannidis, who's quite famous. He's one of the first people who started to point the finger at the replication crisis. And so he would do these look back studies where he would look at the most prominent articles in like nature from 1997.
And then 10 or 15 years later. He would do a look back and he would say, okay, these 10 papers, which are the highest cited, or they were on the cover of nature or whatever. We'll do a look back and we'll see how many much higher quality, better powered follow up studies were done on this topic that either confirm or dispute the results in that paper.
And he comes up with a number like 50 percent reproducibility for the most. The most like prestigious results.
Host of the Information Theory Podcast: Yeah. Okay. So, so it's something, I think the title is something like 50% of all published research findings are false. Yeah,
Steve Hsu: Yeah. And, and, and nobody knows which 50% . So anyway, so EOIs and I, so I invited him to M SSU to give a lecture to our whole faculty in the big auditorium.
Right. And you know, so this was me trying to say to you guys like, Hey. You should be a little more skeptical about results in your field, but I think it had no effect on these professors. So, you know, the funny thing is people intellectually, like Ionidas, are famous. He's very highly cited. He's one of the most highly cited researchers in the world.
So his work is cited a lot, but then when you, when you in private talk to an actual say biomedical researcher, they've not done a Bayesian update as if they believed Ionidas, because then when I say like, just in some completely different setting. Like they're saying like, Oh, we could maybe steal this guy from Mount Sinai graduate school, you know, blah, blah, blah.
He's done some really great work on targeted cancer, blah, blah, blah. And I'm like, well, what are the chances this is right? And they don't, they don't like to get, they don't view it from the Eonidis lens. They're just like, Oh, this is excellent work. This is fantastic. It's definitely, it's definitely correct.
I'm like, okay.
Host of the Information Theory Podcast: So I think people in general have a very hard time assimilating theoretical knowledge, right? Like you know, if they read a book, it's kind of, you know, it makes some kind of impression on them, but it's not going to change their course, it's not going to change their behavior.
One
Steve Hsu: of the benefits of being in frontier theoretical physics is that we are really concerned about the fundamental laws of physics. We are constantly doing huge experiments to probe those fundamental laws. And those experiments, because they're at the technological edge of what people can do, are themselves noisy.
And so every month or few months when you're a grad student in a PhD program in theoretical physics, if you're in frontier physics, you go into a seminar room or a colloquium and somebody's like announcing this amazing discovery at CERN of this neutrino or this kind of like galaxy, you know, this gamma ray burster in this galaxy.
And you get to see, you've seen several reps of the field going crazy over some new experimental results and then better powered experiments come along and their statistical significance of the result goes to zero, you know? So you've seen that movie many times in our field. And so it's very easy.
People from our field understand this thing. And I would say to most venture capitalists, you should. Understand this too, because a lot of what you hear from founders is bullshit. And especially if it's deep tech, like it's actually never going to work, but until you've seen that movie a few times, you don't, you're not well calibrated on the level of skepticism you should have.
And people in some fields like biomedicine, they've never seen the movie. So like they, they just, they just all continue to believe that like, you know, X is true.
Host of the Information Theory Podcast: Maybe it's only happened once, right? There's only been one reckoning.
Steve Hsu: So they're not, they haven't, they haven't updated.
Host of the Information Theory Podcast: All right. So Steve, you, you mentioned that there's kind of like a reproducibility crisis in science or academia.
So what's your mental model? Like what's your map of how this. varies among different disciplines. Is there a pattern like which disciplines are susceptible to this or is it just a matter of kind of like the internal culture of each of these disciplines like a path contingent kind of thing?
Steve Hsu: I think there are at least two independent factors that I think influence how well calibrated people are in a certain field, you know, in their own specialty. How well is their confidence level in a particular claim correlated to the actual probability that that claim is true,
Host of the Information Theory Podcast: right?
Steve Hsu: One is just some basic ability in thinking probabilistically and thinking statistically.
Some people are just good at it, and those people tend to go into fields where there's a little more math. It's a little more likely that those fields are well calibrated. But another independent factor is, is the field basically siloed? Very, very siloed. Where people specialize like crazy and they just work in this little silo or, or is there a lot of like cross pollination across the discipline?
And if there's more, I think generally people tend to be better calibrated.
Host of the Information Theory Podcast: So I wonder where macroeconomics would fit in that framework of yours. Right. Because like those people are pretty mathematical, but it doesn't seem to me that they've been producing anything that's been very, empirically validated.
Steve Hsu: Yeah. So I think on the first metric, they're fine. Like them, they have no trouble understanding the fact they use quite sophisticated statistics, but there, I think it's more, a little bit related to the second thing or maybe more specifically, I could call it selection effects. Because in every field, the people that are going to end up being like having a career in that field.
So they really matter. Like they're going to be in the field for 20, 30 years and they kind of determine what's accepted in that field. That's a very small fraction of the people who at least initially look into that field. So if you look at all the grad students in economics, or even more broadly, like I might've gone into economics, but I didn't go into economics.
So, so you have like, I might've gotten into economics, but I didn't. Or I went into economics, I looked at macro and I realized these guys are just telling stories to each other. So people still self select out. I think that's kind of more of the problem
Host of the Information Theory Podcast: in. It's just like a self perpetuating, kind of like, bad, bad culture among any given discipline.
Steve Hsu: I mean, that's pretty depressing. That's pretty depressing. Economists are cynical enough and they understand group dynamics enough that they could even build toy models of how their own field works. So like I've seen those actual people talking about that where we're basically they have a model where like, yeah, all the people that are susceptible to this kind of thing, they all end up here and they, they, they form their own little silo where they do their thing.
And then other people are looking in from the outside and just saying, I don't believe it. But the main problem with the back row is those guys actually have influence over central banks and to a lesser degree, lesser degree hedge funds and stuff like that. But, they do influence central banks. So, I think lots of stories we tell each other, tell ourselves about macroeconomics, like the Taylor rule for interest rates or stuff like this, you know, not clear to me, how ultimately true some of these things are.
Host of the Information Theory Podcast: Right, right.
So, zooming out in terms of, you know, basic research in general, right, I guess the, like, what is the point of all this, this basic research, like, society spends billions of dollars on it, likethe U. S. federal government definitelyspends that much. I think the standard story is that basic research is upstream of technology, and then technology is upstream of science.
economic progress. So there's this idea like for every dollar you spend on basic research, there's a very high ROI for society. And then there's this countervailing idea that like, actually a lot of technological progress is completely independent, right? Like the Wright brothers, they invented airplanes without knowing anything aboutBernoulli's law.
And it was only later that people created math. So do you, do you, do you feel like there is a payoff, you know, going from basic research to technology or, or is that overstated?
Steve Hsu: So a couple of comments. So number one, you know, whatever our weird mix of like research investment collectively as a society is or has been in the last hundred years, I think properly calculated, the ROI is super high.
So, so not saying it isn't really fucked up and you couldn't improve it a lot, like by not funding certain things or changing the way you fund certain things or removing tenure, whatever it is, regardless of all those possible potential improvements, whatever system we had in place, I do believe delivered huge ROI.
So that's one observation I'd make. The other thing I would say is that sometimes basic research is upstream. of technology. Sometimes it's not. Okay. So sometimes you can have technologists who like to get the steam engine working, even though they don't know what the second law of thermodynamics is or entropy, right?
So, or the Wright brothers. So we should fund technology, just kind of like tinkering and innovation all by itself, even if there's no kind of fancy theoretical basis for it. But I can give you many, many examples like, well, no nuclear bomb without some detailed understanding of quantum mechanics and nuclear physics, right?
Uh, no, gene sequencing without some really detailed understanding of the physics of, you know, how to, like you know, cause some amino acid to attach to some probe or something. So, it doesn't have to be the case that all technology, all technology, technological progress is downstream of some fundamental science.
But if enough of it is, it still could be that the ROI is just humongous from that little tiny subset of the fundamental research that does have downstream useful consequences, right? So that's kind of how I view
Host of the Information Theory Podcast: it. That's how I had been viewing it for most of my life. And then, you know, recently I was hanging out with my friend who's like a professor of physics.
He said something that just blew my mind. He was, he was basically saying like, Oh, you know, we have these graduate students and they're not the best researchers, right? They're actually like the worst possible researchers because they're just learning how to do things. And I don't think my lab is gonna, you know, I don't see the purpose of my lab is to come out with these amazing advancements that will, you know, create like, societal progress.
Like I actually just think my main output is just increasing the human capital of these kids. And so that kind of blew my mind. I was like, I didn't even consider that. Like you, you didn't believe in this, like that anyone didn't believe in like the, the, you know, the mission of creating basic research.
Steve Hsu: I think that's a minority view. What field is your friend in? Like what area of physics?
Host of the Information Theory Podcast: He does like optical optical stuff. So his background is in chemistry. So, I think, you know, he did a lot of things. Laser stuff during his time in the chemistry lab, right?
Steve Hsu: So in optics, you know, there are a lot of industries.
Applications for what some students would learn in a PhD in that field. So I think that view is not completely crazy, but I think it's a minority view. Like I think most people are. Like in particular, say quantum optics, they are actually trying to make scientific advances and they would probably view it as a blend of, yeah, I'm creating human capital by training these kids up in these important technologies, but we're also trying to publish papers that push the frontier of knowledge forward.
So that's, I think, how most professors think.
Host of the Information Theory Podcast: Right, right. You seem pretty optimistic about the returns on basic research, but obviously like there's some inefficiencies in the system. I don't know if you're familiar with this writer named Alexey Guzey. No, I, well, maybe. Tell me, tell me what he says.
He's been kind of looking into biomedical research for a very long time, and he recently caused a stir. He published this article called Abolish the NIHthe National Institutes of Health. Yeah. And so he wrote in that article the NIH is a, quote, tyrannical, capricious, self serving, 50 billion, Kafkaesque Leviathan.
So having seen the system from the inside, like, how would you react to that characterization? And then like, how would you, you know, if you were to scan across kind of like the range of like funding institutions that we have federally, like, would you Look at any of them and say like, Oh, that that's like completely mismanaged or like, you know, we need to abolish that specific institution.
Steve Hsu: So I do know who you're talking about and is he funded by Tyler? Yeah. Uh, I think so. Yeah. So I think I even read this article when it came out.
Well, I would say this about the NIH, people don't realize it's by far the biggest chunk of our basic science budget. It's way, way bigger than NSF or DOE or the other major funding agencies.
And this will get me into huge trouble, but I'm not in VP, I'm not a VP anymore. So I can just say it like this. I would say that NIH produces a lot of low quality, biomedicine in general just produces a ton of low quality research. And I've, like, even though that'll piss a lot of people off, I could just point them to Ioannidi's findings and just say, like, look, it doesn't replicate.
Actually gosh, I don't know if I can say this out loud. It's been many years, so maybe I can say it. So when I had Ioannidis here, he told me something confidentially that He and his research team at Stanford had really perfected this look back method where you define a set of results from year X, you wait to, you know, another 10, 20 years beyond that and then you do a look back to see how good those results were.
It's a well defined methodology that he and his team pioneered. He actually did it, so every year the NIH director publishes the big breakthroughs from NIH funded research that happened this year. So they, they, they did, I know they still do it, but they did it for a long time. So, Ioannidis and his team just took some of those.
And so he did a look back on that and it was not pretty. So, what did that tell you? Like the director of NIH is supposed to be some kind of expert in science and they're spending 50 billion and they cherry pick the best stuff once a year. They do it
Host of the Information Theory Podcast: right. So yeah, the specific proposal is just to like sunset, all the existing grants, let them run out and then just reallocate that money to, to different, to other institutions.
Like, so it seems like maybe you're. I'm open to that.
Steve Hsu: It would cause incredible chaos, but of course I am open to it. Here's a joke. Again, I would never have said this when I was a VPR, but I said this many, many times privately, but I would never have said it publicly, but I can say it publicly now. If you took the money from NIH, And you gave it to computer scientists, physicists, engineers, but, but you gave it to them to do things which might be useful in biomedicine, right?
So like AI for biomedicine or some better like imaging capability or some better way to, you know, colony x rays or something, you know, I used to joke to these biomedical guys, like we would get more bang from that buck if we gave the money to these other types of scientists, but said they have to do something which is relevant for biomedicine.
Because it looks to me like every time there's a big breakthrough in biomedicine, it's because of some alien technology, like some aliens developed the laser and gave it to you. You don't actually know how a laser works, people at NIH, but, but people outside invented this laser and gave it to you and then now all of a sudden you can do a bunch of stuff, right?
Or these other guys figured out how to sequence genes, genomes, and like we gave that to you. So my joke was we should just give the 50 billion to these other scientists that are outside of NIH, not NIH funded people, and you'd actually get more biomedicine progress out of that than those guys produce.
Of course, they, the people, would get extremely mad. If somebody were sitting next to me and I just said that and that guy was an NIH funded researcher, they would, they would just blow their top. Yeah, I think the
Host of the Information Theory Podcast: reactions to the piece were pretty, pretty angry as well, like you could see on Twitter.
Yeah, I
Steve Hsu: believe it.
Host of the Information Theory Podcast: Alright, so, I have a bunch of random questions as we're wrapping up. Yeah, go ahead.
You know, we're recording this November 2024, and about a week agoDonald Trump was reelected as president. Uh, so what was your experience finding out about Trump's election? What happened?
Steve Hsu: I was in China on a trip, mostly a scientific trip.
I gave scientific talks on physics, like at the Chinese Academy of Sciences and other universities. That particular day I was climbing on the side of a mountain at almost 5, 000 meters, it's called Jade mountain in Yunnan province, which is in the far Southwest. It's kind of near the Tibetan plateau.
And I was climbing up and I had access to the internet because of Huawei 5g. Through my phone and then I have Google fly is my telecoms provider in the U. S. And that is roaming in China. So you just get unfettered access through that while you're in China. No firewall or anything like that. And so I was able to follow the election real time.
There's a 13 hour time difference. So you know, Mhm. Midnight on election night on the East coast, just like one in the afternoon when we're doing our climb and I'm following the election. I'm like, Oh my God, Trump is going to win in a landslide. I was kind of expecting him to win, but I didn't think I thought it was going to be close.
I thought it might be a nail biter. I was kind of glad that I wasn't in the U S because I didn't want to follow it that closely. It looked like he's going to win by a landslide. So I actually was screaming out like. You know, you know, MAGA, Trump, Trump is gonna win, you know, it's a landslide and these Chinese people found it, you know, they were like looking at me kind of strangely, but actually, I think one other person fist bump to that might have been another like tourist or foreigner or something, or maybe just some Chinese guy who understood.
But anyway, it was a peak little pun intended here. It was a peak experience. to followDonald Trump's, some people say a third triumph in a row while climbing a mountain at 5, 000 meters.
Host of the Information Theory Podcast: So what made you, what made you so happy about it?
Steve Hsu: What made me happy was thatI almost entered the Trump administration.
You're breaking news here because I don't think I've ever said this publicly. So in 2016, when Peter Thiel was playing a big role in the transition, because he had supported Trump during that election, he had brought an entire binder full of Names of talented people who are willing to serve in the White House.
so I almost joined the Trump administration in 2016. I don't want to say too much more about it. It was at a very senior level that would have required Senate confirmation. but as a result of that, I got an inside view. Of our intelligence services spying on as well as Peter Thiel and other people in that circle.
I think this is kind of well documented now, though, like it will never be fully directly addressed like in the New York Times article, but they've kind of admitted that they did get a FISA warrant and they did monitor everyone within two hops in the network. Two, two, two hops in the social network around Trump.
So that would have included a lot of people like Peter Thiel and myself. So we were being spied on potentially, by the U S intelligence services, which did everything they could to undermine Trump. And so he did fight the deep state. Uh, during his first term. I just think it's incontrovertible. Now, plenty of people who are like on the left or believe what the New York Times prints is generally true, just won't accept what I'm saying to you.
But if you care, if you actually care about what I'm saying, like you can just brush it off. But if you actually care and you do your own research, you will easily convince yourself that stuff like this happened. There's an inspector general report within the part, I think the Department of Justice on specifically on like the FBI activities.
Yeah. uh, I think the inspector general's name is Horowitz. You can go look this up. Of course, like, this is the kind of thing that like, it's a multi hundred page report and it's completely buried by the legacy media. But if you want to go read it, you can see there's lots of interesting stuff in there.
So deep the state tried to undermine Trump, unbelievable amounts of lawfare, just crazy things meant to take him out of the race. And I'm totally okay. Like I have very close friends who hate Trump, who don't want to vote for Trump, would never vote for Trump, who think he's a disaster. We're still friends.
It's not, not a big deal. But if you say it's okay for us to bend the law because he's so bad that we can do these things which are for which there's no precedent, but we could use these tools to try to lawfare or whatever to try to attack him and take him out of the race, we can, you know, totally co-op the media sensor information like about the Biden laptop.
That's all okay. I disagree. So I feel like Trump overcame so much to win this third election. That's why I felt exhilarated. Not so much cause I'm like the biggest Trump fan in the world. I see a lot of faults in Trump and I saw how dysfunctional his first term was. Peter Thiel actually kind of got disenchanted with Trump and just kind of after a while stopped, you know, really playing an active role in the administration.
So, there's lots of negatives.
Host of the Information Theory Podcast: How did that relationship, how did that relationship break up? 'cause you know, from the beginning it seemed like he was, he was going to play a pretty big role then, kind of silently. He just disappeared from the scene.
Steve Hsu: I, I don't, shouldn't comment too much about this, but one thing that.
You know, if you just ask in Silicon Valley circles, a lot of people just say like, look, Trump is too mercurial. And if, if you're a Silicon Valley guy and you're used to dealing with like, in a way Silicon Valley is like a super high trust environment. Like people are competing against each other. But once we decide we're on the same team, like we're in the same startup or we're in the same cap table of this startup, there's a way, like we, we collaborate, we push things forward.
And we act kind of professionally with each other. I'm not sure any of that exists in the Trump world where like he could just pass you aside or, you know, like, so I don't want to make specific comments about it, but I think Lots of super capable people who tried to work with Trump with the best intentions came away, maybe with a negative opinion.
Okay. And it could happen again. Okay. This time around, it's Elon. Who has a big influence on Trump. Or, and vice versa, but let's see how long that lasts as well. Elon now is in a different situation because Thiel's business interests weren't that tightly bound up with his political influence with Trump.
Whereas I think Elon stands to gain and lose a lot depending on how much influence he has with Trump. So I expect to see him stick it out a lot longer than maybe Thiel did. Also, also like Elon, Elon's much more comfortable in the spotlight. I don't think, I don't think Peter was, Thiel is that interested in being in the spotlight.
Host of the Information Theory Podcast: Are you going to have a role in the next Trump administration?
Steve Hsu: I, no comment. So I, I, I tweeted something asking for people to send resumes if they wanted to contribute specifically, I said, restoring meritocracy and competitiveness to us institutions, especially like the scientific funding agencies and things like that, that we've been discussing.
And I've been overwhelmed with resumes. I've gotten hundreds of. People are sending me stuff. I'm not an agent of the Chinese government. I'm not sending this information to China. I'm only sending it to my contacts on the Trump team. Believe it or not, I was accused of that. The great thing about Twitter is that all, all segments of society are participating.
So, whether I would participate or not, I, no, no comment.
Host of the Information Theory Podcast: So, yeah, I guess that brings me to the final question. The 20th century was great. Time for America. There was a lot of scientific progress. We had these wonderful institutions. We hoovered up all the best scientists from all over the world, you know, from Europe, from the former Soviet unionplaces like India and China.
And, you know, you've been kind of sounding the alarm bells recently.
We're basically losing our, our, our edge, right. Both in science and technology. and you know, for someone who's as cosmopolitan as you, you seem to care deeply about this. A particular country's future for you. What, what is at stake? Like what is it that makes America exceptional and, you know, what makes it worth fighting for?
Steve Hsu: Well, I, I feel very blessed because I grew up, you know, I had an idyllic childhood in Iowa. I had great friends. I grew up in a very high trust, nurturing environment in the Midwest. And so I hope for the best for America and even the fact that like my parents who are immigrants from outside the U. S.
could come here and be accepted and have close friends and live a wonderful life in America, like, I would like to see that continue. I would like to see the best of America continue into the future. I'm worried that that won't be the case. We could easily be in for a hard landing with the huge debt that we've accumulated as a country.
Uh, we don't seem to be able to produce things very efficiently in this country anymore. Uh, with some exceptions, obviously. So I am worried about the overall competitiveness of America. I'm worried about the erosion of meritocracy in American society. Lots of those things. I do think, though, that for the class of people that are so talented, that you know, they can, they can go wherever they want and work wherever they want in the world.
I think America will still be an attractive place for that class of person, at least for another 10, 20, 30 years. I mean, unless there's some, unless, you know, our politicians screw things up so much that they, you know, cause a dollar crisis or debt crisis, like the treasury has an auction, they can't sell their bill.
you know, unless we hit some really, sharp inflection point. I think America has a long way to go where it's still like a super desirable place for elites to be for the average American. I feel like the average American may be less well off than they were, you know, when I was growing up. Now, someone like Tyler would say, Oh no, they're actually, they've got lots more, lots more money or incomes are much higher or whatever.
But when I grew up, it was unusual for mom to be working. You had intact families, at least in Iowa, and usually only dad had to work, mom could spend all her time nurturing the kids, cooking dinner, whatever, not saying that's the best thing, the best use for her talent, but it was a very different situation.
And people did not seem, their existence didn't seem as precarious as the way I perceive, you know, the 50th percentile, 35th percentile American to be living. So I'm worried about those people. Those are the people that elected Donald Trump. I'm not, you know, I'm not naive. Maybe Trump doesn't care about them.
Maybe he just is using them to gain power. But I think someone should be concerned about those people and be fighting for those people.
Host of the Information Theory Podcast: Well, Steve, the fact that people like you are out there and potentially staffing the next administration gives me a lot of hope. I really enjoyed this conversation today.
Steve Hsu: I've really enjoyed the conversation. And I hope, hope your podcast is a big success.