Michael Nielsen – How science actually progresses

27:27There is almost identical independent creation of that idea between Alfred Wallace and Charles Darwin. So much so that I think Wallace sends his manuscript to Darwin and is like, what do you think of this idea? And Darwin is like, fuck. I don't think that's an exact quote, but I think it's pretty much right. And then so they actually end up presenting their ideas together in the spirit of sort of sportsmanship. And so then, yeah, why was this period in the 1860s or 1850s? What was that the right time for these ideas? For when you come up with different ideas, one is geology. So in 1830s, I think Charles Lyell figures out that there's been millions and billions of years of time that's existed on Earth.

28:04Then paleontology shows you that actually organisms have existed, fossils have existed for that entire time. So life goes back a long time. And in fact, you can even find fossils for intermediate species that show you the tree of life. In fact, between humans and other apes as well, there's intermediate humans. There's the age of colonization, and you have all these voyages. We're going to do this biogeography. And I guess that all must have been necessary because, in fact, there's a huge history of parallel innovation and discovery in history of science. So maybe it is another piece of evidence to actually more had to be in place for a given idea to be discovered.

28:39Because if it's not discovered for a long time and then spontaneously many different people are coming up with it, that shows you that actually the building blocks were in some sense necessary. Yeah, yeah, yeah. I mean, this example of Layel and other geologists, sort of early 1800s, basically having this idea of deep time, that does seem to have been crucial. I know Darwin was very influenced by Layel.

29:10and if you don't have at least sort of tens or hundreds of millions of years, evolution just starts to look like a non-starter. We should be seeing radical change. In order to make it work on sort of a timescale of say 5 ,000 to 10 ,000 years or 6 ,000 years, you would need to be seeing evolution occurring at a massive rate sort of during human lifetimes, And we're just not seeing that. So that does seem to have been a blocker. It's interesting to your question, what other blockers were there? Were there any others? And I don't know. Right. Or, yeah, how much earlier could you in principle have come up with that if you're a much smarter?

29:51Actually, let me just go back, sort of zoom out to your original question. So you're talking about sort of the verification loop in AI. And an example I think that should give you pause there is the big signature success so far is certainly AlphaFold. And of course, AlphaFold really isn't about AI. A massive fraction of the success there is the protein data bank. So it's X-ray diffraction, it's NMR, it's CryoM, and the several billion dollars that was spent obtaining whatever is 180 ,000 protein structures. It's basically the story of we spent many, many decades obtaining protein structure just by going out and looking very hard at the world experimentally, and then we fitted a nice model at the end of it, and that was like a tiny fraction of the entire investment.

30:43But it's definitely not... That's a story of data acquisition. Yeah. Principally, it's not only. I mean, the AI bit is very, very impressive. It's quite remarkable. But it is only a small part of the total story. Off of the world, it's very interesting. And I philosophically wonder what you think of it as scientific theory or scientific explanation. Because if over time, I guess the world has become harder to understand. As I'm saying things, because you're such a careful speaker, I say this phrase and I'm like. Go for it. Will he actually buy that premise? But yeah, we need to fit models to things rather than, at least in some domains, we're trying to fit models to things rather than coming up with underlying principles that explain a broad range of phenomenon.

31:29And so compare, say, the theory of general relativity or any theory which just nets out to some equations versus alpha-fold, which is encoding these different relationships between different things we can't even interpret over 100 million parameters. and are those really the same thing? Because GR can predict things you could have never anticipated or was never meant to do, like why does Mercury's orbit precess? And alpha fold is not going to have that kind of explanatory reach. And I want to get your reaction to that. Yeah, I think it's an incredibly interesting question. I mean, maybe a really pivotal question in the sense of, So if you sort of take a very classic point of view, you want these deep explanatory principles.

32:19You want sort of as few free parameters as you possibly can. You want very simple models, which explain a lot. And alpha-fold doesn't look anything like that. And so you might just sort of say, oh, well, it's nice. It's maybe helpful as a model, but it doesn't have... It's not a scientific explanation. So that's kind of... That's like a conservative point of view. That's sort of answer one to the question. I think answer two is to say something like maybe you shouldn't think about AlphaFold as an explanation in the classic sense, but maybe it contains lots of little explanations inside it. And so maybe part of what you can get out of interpretability work is you can go into AlphaFold and you can start to extract certain things.

33:03Maybe basically by doing sort of archaeology of alpha fold, we can actually understand a great deal more about these principles. You can start to extract it. Oh, that circuit does this interesting thing and we learn this. So I don't know to what extent that's been done with alpha fold. I know it's been done a little bit with some of like the chess models. I believe it's alpha zero. there seem to be some strategies which were certainly borrowed by Magnus Carlsen at least which he seems to have just taken from AlphaZero. I mean I don't think there's any public confirmation of this but there were some experts have noticed that he changed his game quite radically after some sort of some public forensics were released on how AlphaZero worked.

33:49So that's kind of a sort human beings are starting to extract meaning out of these models. And maybe that starts to lead to sort of viewing the models as a potential source of explanations. You need to do more work because they're not very legible upfront, but you can extract them potentially. And I think that's kind of an interesting intermediate situation where they're not explanations, but you can extract interesting explanations out of them. You can use them as kind of a source. And I think the third and the most interesting possibility is no, they're a new type of object in some sense. They should be taken very seriously as explanations, but in the past, we haven't had the ability to really do anything with them.

34:30And now we're going to have sort of new, interesting new sort of actions, which we can do. We can merge them, we can distill them, we can do all these kinds of things. And there's going to be sort of almost a new, it's a big opportunity sort of in the philosophy of science to start to do that. There's sort of like an anticipation of this in some sense, I think in the way, certainly I know some mathematicians and physicists who, I mean, historically, if you had like a 100 page equation, which, and that's the kind of thing that does come up, I mean, there's just nothing you can do if it's 1920. There is nothing you can do.

35:09At that point, you give up on the problem. And now today, with tools like Mathematica, you can just keep going. And so that's an object now. That's a thing that you can work with. And there are examples where people work with these things that formerly were regarded as too complicated. And sometimes they get simple answers out of the end. That's just an intermediate working state. And so I sort of wonder if there's going to be something similar is going to happen in in this particular case where you could take these models and sort of just use them in a little bit the same way people do with Mathematica and take them seriously as they're not explanations in the classic sense, but there'll be something else which interesting operations can be done on.

35:54The thing I worry about is suppose that you, it's 1600 and you're training or 1500 and you're training a model on, this is a weird history where we developed deep learning before we had cosmology. But suppose we live in that world and you're observing how there's the stars, they don't seem to move, the planets have all these weird behaviors. And then you train a model on that and then you do some kind of interp on it and trying to figure out, well, what are the patterns we see here? What you'd see are just these, you just keep, be able to keep building on Ptolemy's model. You'd see like, oh, there's more epicycles we didn't notice.

36:28There's another epicycle. It's the parameters, whatever to whatever, encode epicycle this, parameters, whatever, encode the next epicycle. So if you were just trying to figure out why is the solar system the way it is from observational data, you could just keep adding epicycles upon epicycles, but it really took one mind to integrate it all in and say, here's what makes more sense overall. So, I mean, there, like, you know, I mean, this is sort of to my point that we don't really understand what to do with the models. Like, sort of, we don't have, like, the verbs necessarily yet. But, you know, it is certainly interesting to think about the question, you know, where you start to apply constraints to the models.

37:13You know, it's sort of essentially saying, what's the simplest possible explanation? Or, you know, can you simplify? Can you give me sort of the 90-10 explanation? Can you go further and further and further sort of in boiling it down?" So it might be that indeed they sort of start out by providing a very, very complicated many, many, many parameter model. But you can just force the sort of the case and basically that's scaffolding, which maybe they, you know, is sort of the very early days of their attempt to understand something, but they're forced through that to a much more simple understanding.

37:52Sorry for misunderstanding but it sounds like you're saying maybe there's some sort of regularizer or some sort of distillation you could do of a very complicated model that gets you to a truer, more parsimonious theory. But just take Ptolemy versus Copernicus. So you start off with lots of Ptolemaic epicycles and then you try to distill this model and maybe gets rid of some of the epicycles that are less and less sort of necessary to get the mean squared error of the orbits to match. But at some point it has to do this thing, which is like switch two things. Yeah, yeah, yeah. And locally it actually doesn't make things more accurate.

38:28Yeah, yeah, yeah. It's sort of in a global sense that it's a more progressive theory. Yeah, yeah. And there's some process, which obviously humanity did over its bandwidth, did that regularization or did that swap. But raw gradient descent, it seems like I don't really feel like it would do that. I mean, you think about the example of going from Newtonian gravity to Einstein's general theory of relativity. And these are shockingly different theories. And the question is, what causes that flip? And as nearly as I understand the history, what goes on is Einstein develops special relativity. And pretty much straight away, he understands.

39:09I mean, it's a very obvious observation. In special relativity, influences can't propagate faster than the speed of light. And in Newtonian gravity, action is at a distance. In fact, it's straight away in special relativity, you could use Newtonian gravity to do faster than light signaling. You could send information backwards in time. You could do all kinds of crazy stuff. And so it's not a big leap to realize, oh, we have a big problem here. And so that's the forcing function there. You've realized that your old explanation is not sufficient. You need something new. And then you're going to start by doing the simplest possible stuff.

39:51And it just turns out that a lot of that stuff doesn't work very well. And so you're sort of forced. In fact, it is interesting. He's sort of forced to go through these steps of gradually it gets quite more complicated and it's sort of wrong in a variety of ways. And the final theory appears really shockingly simple and beautiful, but it's gone through some somewhat ugly intermediate stages. Yeah. Yeah. So if you're thinking about what does it look like to have AI accelerate science, there's one for maybe well-understood domains where we just want local solutions, like how does this protein fold?

40:31We just train a raw model using gradient descent. Then there's things like coming up with general relativity, where you couldn't really just train on every single observation in the universe and hope that general relativity pops out. And so what would it require? Well, it also certainly wasn't immediately discovered, right? So it was a lot of decades of thought. And I guess you need independent research programs where people start off with these biases, where Einstein is just initially motivated by this thought experiment of, you know, can you distinguish the effect of gravity from just being accelerated upwards?

41:06And then you just need different AI thinkers to start off with these initial biases and see what can germinate out of them. And then the verification loop for that might be quite long, but you just need to keep all those research programs alive at the same time. Yeah, I mean, I think there's like, I mean, this point that you make about sort of keeping all the different research programs alive, like that I think is very important and somehow central. role. I mean, a great example is situations where the same answer has been correct in some circumstances and wrong in other circumstances. So the planet Uranus was not in quite the right spot.

41:45And people very famously predicted the existence of Neptune on this basis. Wonderful, massive success for Newtonian gravity. The planet Mercury is not in quite the right spot. you predict the existence of some other distorting planet. Turns out that doesn't exist. Actually, the reason Mercury is not in the right spot is because you need general relativity. And so you've pursued very similar ideas and has been very successful in one case, and it's been completely and utterly unsuccessful in the other case. And I think, I mean, a priori, you can't tell which of these is the thing to do, and you actually need to do both.

42:23And so, I mean, this is certainly, it's very true in in the history of SOADS that, you know, this kind of diversity where you just have lots of people go off and pursue lots of potentially promising ideas, you just need to support that for a long time. And it's, I mean, it's hard to do that for a variety of reasons. But it does seem to be very, very, very important. So this example of Uranus versus Mercury is very interesting. In one, I think it illustrates sort of the difficulty of falsificationism. Like the orbit of Uranus is in some sense falsifying Newtonian mechanics, but then you make some ancillary prediction that says, oh, the reason this is happening is there must be another planet which is in fact perturbing Uranus's orbit, and I think it's Le Verrier in 1846.

43:18That's right. Point a telescope in the right direction and you find Uranus. Neptune. Oh, it's there. Yeah, yeah. Neptune, yes. But with Mercury, yeah, it's observed that the ellipse which forms this orbit is rotating 43 arc seconds more every century than Newtonian mechanics would imply. So people say that there must be a planet inside Mercury's orbit, they call it Vulcan, and point in telescopes, it's not there. But if you're a proper Newtonian, what you do is say, well, maybe there's some cosmic dust that's occluding this planet. Or maybe the planet is so small we can't see it. Or maybe there's some, let's build an even more powerful telescope, or maybe there's some magnetic field which is sort of occluding our measurements.

43:56And this happens over and over, right? Like, you know, there's just so many stories which are exactly like this. Right. I mean, an example I love from, you know, in the 1990s, some people noticed that the Pioneer spacecraft weren't quite where they were supposed to be. And so, you know, you can get very excited about this. Oh my goodness, general relativity is wrong. We have like in a bit, you know, maybe we're going to discover the next theory of gravity. And today, the accepted explanation is that, no, actually, there's just a slight asymmetry in the spacecraft. It turns out that, you know, the thermal radiation is like slightly larger in one direction than the other, and that's causing a tiny little acceleration towards the sun.

44:39And most of the time when there's these apparent exceptions, it's just something like that's going on. It's very much like the Vulcan, the Mercury Vulcan case. But every once in a while, it's not. And a priori, you can't distinguish these. But I mean, science is just full of these. It's funny too, like the way we tell the history of science, it sounds so simple. Like, oh, you just focus on the right exception and you realize that you need to throw out the old theory. Right. And lo and behold, your Nobel Prize awaits. But in fact, these exceptions are all over the place. And 99.9 % of the time, it just turns out to be some effect like this thermal acceleration in the case of the Pioneer spacecraft.

45:26So, you know, sort of unfortunately, there's a lot of selection bias going into those stories. And the thing is, there's no ex-anti heuristic which tells you which case you're in. And just to spell out why I think this is important, is because some people have this idea that AI is going to make disproportionate progress towards science because it makes disproportionate progress towards domains where there's tight verification loops. And so it's really good at coding because you can run unit tests. And science may be similar because you can run experiments. And I think what that doesn't appreciate, one, is that experiments actually don't, there's an infinite number of theories that are compatible with any given experiment.

46:03And over time, why we glob onto the, well, at least in retrospect, we think is a more correct one is, as we're discussing in this conversation, sort of hard to articulate. Lactatus actually has all kinds of interesting examples in the book about these kinds of hostile verification loops that are extremely long lasting. So one, he talks about his Prout or Prout, I don't know how to pronounce it, but there's this This chemist in 1815, he hypothesizes that all atomic nuclei must have whole number weights, and they're basically all made of hydrogen. And the reason he thinks this is because if you look at the measure rates of all elements, it does seem that almost all of them do have whole number weights.

46:45But then there's some exceptions. Like for example, chlorine comes out at 35.5. And so then there's all these ad hoc theories that people in this school keep coming up with, like, oh, maybe there's chemical impurities. But then there's no chemical reaction you can do which seems to get rid of this. Maybe it's fractions of whole numbers. It's 35.5. It can be halves. But actually if you measure chlorine even closer, it's 35.46. It's actually getting further away from the correct correction. And later on, what is discovered is what you're actually measuring is different isotopes, which cannot be chemically distinguished.

47:17They can only be physically distinguished. But so then you just have 85 years before we realize what an isotope is, where the verification is actually actively hostile against you, against the correct theory, and you just need this remnant to be defending. There's no extantive reason it's the preferred theory. As a community, we should just have people defend, try to integrate new observations, even if they don't seem to fit their school of thought with what they believe. And hopefully, if enough of that happens. Anyways, yeah, I guess the thing I'm trying to articulate is the difficulty with automating science.

47:50Yeah, I mean, the question is, where is the bottleneck at some level? And are we primarily bottlenecked on one thing or one type of thing, or are we bottlenecked on sort of multiple types of thing? So certainly talking to structural biology people, they seem to think that AlphaFold was an enormous advance. It was a shock. So at some level, yes, AI can, it seems certain it can help us speed up science. So it is helping with a certain type of bottleneck. Yeah. That doesn't mean, though, as you're saying, that it's necessarily going to help with all kinds of bottlenecks. And sort of, I suppose, the question you're pointing out is, like, what are the types of bottlenecks that remain and what are the prospects for getting past them?

48:35I think even in the case of coding, like, it's really interesting, you know, talking to programmer friends, at the moment, they're all in this state of shock and high excitement and they're all over the place actually kind of talking to them. you do wonder, where is the bottleneck going to move to? So certainly one thing that a lot of them seem to be bottlenecked on is now having interesting ideas, and in particular having interesting design ideas. So there's not really a verification loop for knowing, oh, that design idea is very interesting. So they're no longer nearly as bottlenecked by their ability to produce code, but they are still bottlenecked by this other thing.

49:16They always were, formally, they weren't bottlenecked on it because, you know, just writing code was, took so much of their time. They could sort of have lots of ideas while they were, you know, they'd take three weeks to implement their prototype and then they would implement the next version. Now they're taking three hours to implement the prototype and they don't have, you know, as good ideas sort of after that from a design point of view. Last year, I predicted that by 2028, AI would be able to prep my taxes about as well as a competent general manager. But we're already getting pretty close.

49:47As I shared before, I use Mercury both for my business and my personal banking. So I recently gave an LLM access to my transaction history across both accounts through Mercury's MCP. I asked it to go through all my 2025 transactions and flag any personal expenses that seemed like they should actually be charged to the business. And this worked shockingly well. Mercury's MCP exposes a bunch of detailed information. Things like notes and memos and any JPEGs of receipts and PDF attachments. So my LLM had plenty of context to work with. One of my favorite examples happened with a charge to Bay Paddle.

50:21If you looked at the vendor alone, you would have had to assume that it's a personal expense. But the LLM looked at the receipt and the attached note in Mercury and realized this was actually a team bonding exercise from our last in-person retreat. So a legitimate business expense. I imagine it will be a while before traditional banks have MCP. Functionality like this is why I use Mercury. Go to mercury.com to learn more. Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column N.A., members FDIC. You have a very interesting take.

50:56I think it was a footnote when I know where your ass is and I couldn't find it again, which was that it's very possible that if we met aliens, that they would have a totally different technological stack than us. And that contradicts, I guess, a common-sense assumption I had that I never questioned, which is that science is this thing you do relatively early on in the history of civilization where you get to a point and you have a couple hundred years of just cranking through the basics, understanding how the universe works, et cetera, and you've got it. You've got science. And then basically everybody would converge on the same quote-unquote science.

51:29And so I found that a very interesting idea, and I want you to say more about it. Yeah. Yeah, I mean, I think probably the idea there that I'm at least somewhat attached to is the idea that sort of the tech tree or the science and tech tree is probably much larger than we realize. I mean, we're sort of in this funny situation. People will sometimes talk about a theory of everything as a potential goal for physics and then there's this presumption somehow that physics is done once you get there. And of course, this is not true at all. If you think about computer science, computer science basically got started in the 1930s when Turing and Church and so on just laid down what the theory of everything was.

52:17They just said, you know, here's how computation works. And then we've spent 90-odd years since then just exploring consequences of that and gradually building up more and more interesting ideas. And those ideas are, to some extent, you can just regard as technology, but to some extent, insofar as they're sort of discovered principles inside that theory of computation, I think they're best regarded as science and in some cases very fundamental science. Ideas like public key cryptography are, I mean, they're just incredibly deep, very non-obvious ideas, which in some sense lay hidden already sort of in the 1930s.

52:54And so my expectation is that there will be different ways of exploring this tech tree. And we're still relatively low down. We're still at the point where we're just understanding these basic fundamental theories and we haven't yet explored them. A thing which I think is quite fun is if you look at just the phases of matter. When I was in school, we'd get taught that there are three phases of matter or sometimes four phases of matter or five phases of matter, depending a little bit on what you included. And then as an adult, as a physicist, you start to realize, oh, we've been adding to this list.

53:33We've got sort of superconductors and superfluids and maybe different types of superconductors and Bose-Einstein condensates and the quantum hall systems and fractional quantum hall systems and, and, and, and, and, and. And it's starting to turn out, it looks like actually there's a lot of phases of matter to discover. And we're going to discover a lot more of them. And in fact, we're going to be able to start to design them in some sense. I mean, we'll still be subject to the laws of physics, but there is this sort of tremendous freedom in there. And this looks to me like, oh, we're down at sort of the bottom of the tech tree.

54:06We've barely gotten started there. And I expect that to be the case sort of broadly. Certainly in terms of, I think programming is a very natural place to look. The idea that we've discovered all the deep ideas in programming just seems to be sort of obviously ludicrous. We keep discovering sort of what seems like deep, new, fundamental ideas. and I mean we're very limited, we're basically slightly jumped up chimpanzees, so we don't, you know, we're slow and it's taking us time, but what do we look like sort of another million years in the future in terms of all of the different ideas which people have had around how to manipulate computers, how to manipulate information.

54:58I think we're likely to discover that actually there are a lot of very deep ideas still to be discovered. So who was it? I think it was Knuth in the preface to the art of computer programming. So something like, he started this book back in the 60s and he talked to a mathematician. He was a bit contemptuous and said, look, computer science isn't really a thing yet. Come back to me when there's a thousand deep theorems. And Knuth remarks, and he's writing this now decades later, the preface, there clearly are a thousand deep theorems now. And that means, like, it's really interesting to sort of think of it.

55:37Like, what's the long-term future? As you get higher and higher up in the tech tree, like, choices about which direction we go and sort of how we choose to explore, you know, I think it's potentially the case that, you know, different civilizations or different choices mean that we end up in different parts of that tree. And in particular, just things, I mean, sort of very basic things about, you know, we're very visual creatures, certain other animals are much more orally based. Does that bias sort of the types of thoughts that you have, and then you extend it to sort of much more exotic kinds of civilizations where maybe just sort of their biases in terms of how they perceive and how they manipulate the world are maybe quite different than ours.

56:29And that might make some significant changes in terms of how they do that exploration of the tech tree. It's all speculation, obviously. No, this is such an interesting take. I want to better understand it. So one way to understand it is that there might be some things which are so fundamental and have such a wide collision area against reality that they're inevitably going to discover like general activity. Numbers, numbers. Yeah, yeah. Like of all of the intelligences in the Milky Way galaxy, maybe that number is one. Actually, arguably, we've already increased the number. But, you know, of all of those, what fraction of the concept of counting?

57:11And, you know, it does seem very natural. What fraction have discovered, you know, the idea of some kind of, you know, decimal place system? Interesting question. Like, and maybe we're missing something really simple and obvious that's actually way better than that. What fraction got there immediately? What fraction sort of had to go through some other intermediate state? What fraction use linear representations versus, say, I don't know, a two-dimensional or a three-dimensional representation? I think the answers to these questions are just not at all obvious. It's a lot of design freedom. On theoretical computer science, this is going to be extremely naive and arrogant.

57:53But I took Scott Aronson's class on complexity theory, and that was by far the worst student he's ever had. But what I remember is like there was this period that you were the pioneers of where we figured out here's the class of problems that quantum computers can solve and how it relates to problems that quantum computers can solve. It's like groundbreaking. Oh, crazy. This works. And then since then it's been this – literally it's called Complexity Zoo, this website, which lists out here's all the complexity classes. And if you have this complexity class with this kind of oracle, it's sort of equivalent to this other class.

58:30and that it feels like we're building out that taxonomy. And so there's a couple of ways to understand what you're saying. One, maybe you just disagree with me that this is actually what's happened with this field. Another is that while that might happen to any one field, the amount of fields, who would have thought in 1880 the computer science, other than Babbage or something, the computer science was going to be a thing in the first place. So the amount of field, we're underestimating how many more fields there could be. Yeah, yeah, for sure. Or maybe you think both or maybe a third secret thing, but I'd be curious.

58:56I mean, a very common argument here is sort of the low-hanging fruit argument, the argument that says, oh, there should be diminishing returns. And in fact, empirically, we see this, right? The amount of scientists in the world is just exponentially increased. And I mean, I think it's worth thinking about, like, why do you expect diminishing returns? And how well does that argument actually apply in practice? an analogy I like is actually thinking about going to some event, going to a wedding or whatever and you go to the dessert buffet and they've put out 30 desserts. And of course, naturally what people do, the best desserts go first.

59:40I mean, we don't quite have a well-ordered preference there, so maybe there's some difference, but human beings are fairly similar. So the best desserts will go first. And this is an argument, you know, for why you expect diminishing returns in a lot of different fields. If it's relatively easy to see what's available and people have similar preferences, then the best stuff goes first. And, you know, it just gets sort of worse and worse after that. And sort of a very static snapshot in time of scientific progress. Maybe there's some truth to that. But if somebody is standing behind the dessert table and is replenishing, restocking the desserts and keeps adding new ones in, it may turn out that a little bit later, much better desserts appear.

1:00:29And so you're going to go and eat those instead. And scientific progress has a little bit of that flavor. We go through these sort of funny time periods. Computer science is a great example where computer science basically arose as sort of a side effect of some pretty abstruse questions in the philosophy of mathematics and logic. And so you've got these people trying to attack these rather esoteric questions that seem quite high up in some sense in sort of exploration, quite esoteric. And they discover this fundamental new field and all of a sudden there's an explosion there. So sort of the diminishing returns argument just didn't apply there.

1:01:12We just weren't able to see what was there. And this has been the case over and over and over again. Sort of new fields arrive and all of a sudden, boom, it's actually easy to make progress again. Young people flood in because you can be 21 and make major breakthroughs rather than having to spend 25 years mastering everything that's been done before. It's obviously very attractive. and I don't understand, I'm not sure anybody understands very well sort of the dynamics of that, like how to think about why the structure of knowledge is that way, that these new fields keep opening up. But it does seem empirically at least to be the case.

1:01:54Despite the fact that that is the case. Yeah. Take deep learning, right? Obviously, this is an example of a new field where the 21-year-olds can make progress and it's relatively new 15 years or so when it sort of gets back into high gear. But already we're in a stage where you need billions or tens of billions or hundreds of billions of dollars to keep making progress at the frontier. and there's a couple ways to understand that. One is that it actually is harder than the kinds of things the ancients had to do or is more intensive at least. Second is it might not have been, but because our civilizational resources are so large, the amount of people is so large, the amount of money is so large, that we can basically make the kind of progress that would have taken the ancients forever to make almost immediately.

1:02:45We notice something is productive, immediately dump in all the resources. those. But it's also weird that there's not that many of them. Like, I feel like deep learning is notable because it is one big exception to the fact that it's hard to think of other examples. I think that's a consequence of sort of, you know, the architecture of attention, right? Like, at any given time, there's always a sort of a most successful thing. You know, maybe if deep learning wasn't a thing, maybe you'd be talking about CRISPR. Maybe you'd be talking about, you know, whatever it is, maybe, you know, maybe we wouldn't think about solving sort of the protein structure prediction problem as a really a success of AI.

1:03:23Maybe we would have figured out how to doing it with sort of curve fitting, like, you know, more broadly construed. And we'd just be like, oh, wow, like we took a lot of computing resources. But protein structure prediction might, you know, be an enormously important thing. So there is always sort of a biggest thing. and I think what you're pointing out is more a consequence of the way in which attention gets centralized. It's basically fashion is sort of what I'm saying. It's not just fashion, but there is some dynamic there. There's a very interesting and important implication of this idea that the branching is so wide and so contingent and so path dependent that different civilizations would stumble on entirely different technology sects.

1:04:08There's a very interesting implication that there will be gains from trade into the far, far future, which might actually be one of the most important facts about the far future in terms of how civilizations are set up, how they can coordinate, how they interface with... Like, there's not this, like, go forth and exploit. It's actually, there are humongous gains to trade from adjacent colonies or whatever. Yeah, sort of. There's a question of, like, what's actually hard? You know, if it's a question of, if it's just the ideas, well, those spread relatively quickly. It's relatively easy to share ideas.

1:04:44If it's something more, it's almost sort of a Dan Wang kind of an idea where it's actually sort of, there's some notion of capacity. You need all the right text, you need all of the right manufacturing capacity and so on. And so civilization A has very different kind of manufacturing capacity and it's just not so easy to build in civilization B, even if civilization B is kind of ahead, then I think that that becomes true. There is actually comparative advantage, which is really worth, I mean, it's going to provide massive benefits to trade in both directions. Eventually you're going to expect some diffusion of innovation.

1:05:22It is funny to think about what the barriers are there. A fun thought experiment I like to think about is sort of GitHub, but for aliens. So, you know, somebody presents you with all of the code from some alien civilization. And I mean, I don't even know what code means there, but the sort of their specification of algorithms. And it's so interesting, like it would have many interesting new ideas in there. And it would take forever for human beings to dig through and to try and extract all of those. Because one reason, I mean, the origin of this for me was actually thinking about proteins in nature.

1:06:04We've been gifted just this incredible variety of machines, which we don't understand really at all. And we just have to go and sort of try and understand them on a one-by-one basis. We're still understanding hemoglobin and insulin and things like this. And no doubt, there's hundreds of millions of proteins known. So it is a little bit like that. We've been gifted by biology just this immense library of machines, no doubt containing an enormous number of very interesting ideas. And we're just at the very, very, very beginning of understanding it. So actually, I mean, that's I suppose kind of your point actually is I need to relabel your argument slightly, but you sort of think of that as a gift from an alien civilization, which obviously it isn't, but you think of it that way.

1:06:56And it's like, oh my goodness, like there's so much in there and we're going to study it. And goodness knows how long we could continue to study it. There's tens of thousands of papers about the hemoglobin and things like that, and we still don't understand them. And yet we're getting so much out of it. Just think about But insulin alone, it's such an important thing. That's an incredibly useful intuition problem that you have on Earth. I had Nick Lane on where he had this theory about how life emerged. But like whatever theory you have, basically something like DNA, four billion years, and you have an alien civilization come here and be like, There's all these interesting things to learn about material science, about, you name it, right?

1:07:43Like about - Think about kinesin walking along. Like, I mean, and we know almost nothing about these proteins. And yet the tiny few facts we do know are just incredible. The ribosome. Yeah. You know, another example. I mean, this miraculous sort of device, little factory. And all seeded by just like there's this particular chemistry on Earth with nucleic acids and carbon-based life forms that that chemistry gives rise to all of these interesting things which an alien civilization would find very interesting. And so that seed which must be one among trillions of possible seeds of, I mean, just of general intellectual ideas leads to all this fecundity.

1:08:25That's a very interesting introduction poem. I want to meditate on this gains for trade thing because I feel like, I think there's something actually very interesting about this idea that if you have this vision of how technology progresses and how it might be different in different civilizations, it has important implications about how different civilizations might interact with each other. Like the fact that there are going to be these huge gains from trade. It makes friendliness much more rewarding. Yes. Right? Yeah. That's a very important observation. Yeah. I hadn't thought about that at all.

1:08:54That is a very interesting observation. Yeah. It is funny. I mean, comparative advantage is something that people love to invoke. And it's a very beautiful idea, obviously. There are limits to it. Like, you know, it's a special limited model. We don't, you know, chimpanzees can do interesting things. We don't trade with them. And I think it's sort of interesting to think about the reasons why. And part of it is just power, I think. Once there's a sufficiently large power imbalance, very often, not always, but very often groups of people seem to sort of shift into this other mode where they just seek to dominate.

1:09:47and maybe that's something special about human beings but maybe it's also sort of a more general sort of a thing. They're no longer, they give up, you need all these special things to be true before groups will trade and it's not necessarily obvious. I think the big thing going on here is one, transaction costs and two, comparative advantage does not tell you that the terms on which the trade happens are above subsistence for any given one producer. So people often bring this up in the context of, well, humans will be employed even in a post-AGI world because of a great advantage. There's like five different ways that argument breaks down, but the easiest ways to understand are, why don't we have horses all around on the roads because there's some comparative advantage between cars and horses?

1:10:37Good example. Well, there's huge transaction costs to building roads that are compatible with horses and cars at the same time. In a similar way, AI is sort of thinking at 1 ,000 times the speed and can sort of shoot their latent states at each other are going to find it way more costly than the benefit in just terms of interacting with you to have a human being in the supply chain. and second that just because horses have a comparative advantage mathematically does not mean that it is worth paying 100k a year or whatever it costs to sustain a horse in San Francisco. That subsistence is going to be worth the benefit you get out of the horse.

1:11:20I do think it's interesting just the sheer fact that my expectation and my intuition obviously differs a great deal from yours on this, is that most parts of the tech tree are never going to be explored. There's just too many interesting ways of combining things. There's too many sort of deep ideas waiting to be discovered. And not only we, but nobody ever is going to discover most of them. So choices about how to do the exploration actually matter quite a bit. Interesting. It's something I really dislike about sort of technological determinist arguments. I'm willing to buy it sort of low enough down when progress is relatively simple.

1:12:03But higher up, you start to get to shape the way in which you do the exploration. And it's interesting, people, we are starting to shape it in interesting ways. I mean, there's various technologies that have been essentially banned. You think about DDT, you think about chlorofluorocarbons, you think about restrictions on the use of nuclear weapons, the Nuclear Non-Proliferation Treaty. And those kinds of things, they weren't done before the fact, but they're starting to get pretty close in some cases where we just sort of preemptively decide we're not going to go down that path. So that starts to look like a set of institutions where we are actually influencing sort of how we explore the tech tree.

1:12:51Yeah. On where you would see these gains from trade, obviously it would be, you'd see the most where it's pure information that can be sent back and forth because the information has this quality where it is expensive to produce, but cheap to verify and cheap to send. And so it'll be interesting how much of future productivity or whatever can be distilled down to information. Right now, it's kind of hard to do because you can't really transfer... If China is really good at manufacturing something, well, there's this process knowledge that's in the heads of 100 million people involved in the manufacturing sector in China.

1:13:22But in the future, it might be easier if AIs are doing... I mean, the question about sort of to what extent does our fabrication get sort of very uniform and get really commoditized, like 3D printers have been the next big thing for at least 20 years now. Why do they still not work all that well? Why are they still not actually at the center of manufacturing and sort of what comes after that? It is funny to look at, say, the ribosome by contrast. It really is at the center of biology in a whole lot of really interesting ways. and whether or not that's the future of manufacturing is something very simple, sort of where everything goes sort of as throughput through, I don't know, maybe it's a bioreactor or something like that.

1:14:07So you send the information and then you grow stuff or you have some 3D printer that actually works. And if they're good enough, then actually it does become much more a pure information problem and some of this process knowledge becomes much less important. Jane Street has a lot of compute, but GPUs are very expensive. And so even optimizations that have a relatively small effect on GPU utilization are still extremely valuable. Two of Jane Street's ML engineers, Corwin and Sylvain, walked through some of their optimization workflows at GTC. You're not bottlenecked on the network being too slow.

1:14:42You're bottlenecked on waiting for a different rank in your training, not having completed the work. They talked about how Jane Street profiles traces and diagnoses bottlenecks, and then how they solve them using techniques like CUDA graphs and CUDA streams and custom kernels. With these sorts of optimizations, Corwin and Sylvain were able to get their training steps down from 400 milliseconds to 375 milliseconds each. This 25 millisecond difference might sound small, but given the size of Jane Street's fleet, that improvement could free up thousands of B200s. Jane Street open-sourced all the relevant code.

1:15:13If you want to check it out, I've linked the GitHub repo and the talk in the description below. And if you find this stuff exciting, Jane Street is hiring researchers and engineers. Go to jadestreet.com slash thwarkesh to learn more. I can ask a very clumsily phrased question. So there's these deep principles that we've discovered a couple of. One is this idea that, hey, if there's a symmetry across a dimension, it corresponds to a conserved quantity. It's a very deep idea. There's another, which we've written a lot about, written a textbook about, in fact, about there's weight. There's ways to understand this thing of what kinds of things you can compute, what kinds of physical systems you can understand with other physical systems, what a universal computer looks like, et cetera.

1:15:58And is your view that if you go down to this level of idea of Noether Serum or the Church-Turing principle, that there's an infinite number of extremely deep such principles? Because I feel like what makes them special is that they themselves encompass so many different possible ways the world could be, but no, the world has to be compatible with actually a couple of these very deep principles. I don't know. I mean, you know, I just, all I have here is speculation and sort of instinct. My instinct is we keep finding very fundamental new things. It was very, I mean, for me anyway, quite formative to understand, as I say, you know, I gave the example before, there's these wonderful ideas of church and Turing and these other people, ideas about universal programmable devices.

1:16:45This is, and then you understand later, oh, this also contains within it the ideas of public key cryptography. And then you understand later, oh, that also contains within it the ideas. People refer to it as cryptocurrency or whatever, but there's a very deep set of ideas there about the ability to collectively maintain an agreed upon ledger, which is built upon this. And there's probably many deep ideas to sort of... Right. It actually took whatever, it's taken many years really to figure out the right canonical form of those. And so just this fact that you keep finding what seem like deep new fundamental primitives, I find very, for me, that has been a very important intuition bump.

1:17:30And it's across, I mean, I've given that particular example, but I think you see that same pattern in a lot of different areas. What is your interpretation then of this empirical phenomenon where ideas like whatever input you consider into the scientific process or technological process, economists have studied this a million and a hundred ways. It just seems to require, even at actually a very consistent rate, X percent more researchers per year. So there's this famous paper from a couple of years ago by Nicholas Bloom and others where they say, how many people are working in the semiconductor industry and how does it increase over time through the history of Moore's law?

1:18:06And I think they find like Moore's law means computing increases 40 % a year or transistor density increases 40 % a year. But to keep that going, the amount of scientists has increased 9 % a year. Something like that, yeah. And they go through industry after industry with this observation. And so is your view that there are these deep ideas, but they keep getting harder to find or that no, there's another way to think about what's happening with these empirical observations? I mean, so first of all, all of their examples are narrow, right? They pick a particular thing and then they look at some particular metric.

1:18:42Nowhere in that shows up, like GPUs don't show up there, right? Like in the sense of, oh, all of a sudden you get this ability to parallelize. And that's really interesting.

1:18:55So there's sort of a lot of external consequences that are just alighted from basically, they have these simple quantitative measures. They look at it in agricultural productivity. They look at it in a whole lot of different ways. But you do have to focus narrowly. And I suppose I'm certainly interested, as I say, in this fact that just new types of progress keep becoming possible. But there is still, I think even there, there does seem to be some phenomenon of diminishing returns.

1:19:33is that intrinsic? Is that something about the structure of the world? What is it? Well, one thing which hasn't changed that much is sort of the individual minds which are doing this kind of work. And maybe those should be sort of being improved as well, or some sort of feedback process going on there. And maybe that changes the nature of things. I I suppose I look at scientific progress up into, let's say, 1700, something like that. And it was very slow and also it was very irregular. You had the Ionians back sort of five centuries before Christ doing these quite remarkable things. And so much knowledge would get lost and then it would be rediscovered and then it would be lost again.

1:20:21And you'd have to say that progress was very slow. And there, it's partially just bound up with the fact that there were some very good ideas that we just didn't have. Even once you've had the ideas, then you need to build institutions around them. You actually need to solve a whole lot of different problems about training, about allocation of capital, about all these kinds of things, even just about basic sort of security for researchers so they're not worried about the Inquisition or things like that. So there's all these kind of complicated problems. you solve all those complicated problems and then all of a sudden, boom, there's a massive sort of burst of scientific progress.

1:20:56If you're not changing it, if there's some kind of stagnation there, if you're not changing those external sort of circumstances, yes, like you may start to get sort of diminishing returns again. But that doesn't mean there's anything intrinsic about the situation. Maybe something just external needs to change again. Obviously, a lot of people think AI is potentially going to be a driver. I mean, it certainly will at some level. In fact, to the extent you can think of a lot of modern scientific instrumentation as really, I mean, at some level, kind of robots. What is the James Webb Space Telescope?

1:21:33Well, it's unconventional maybe to describe it as a robot, but it's not completely unreasonable either. it is an example of a highly automated, very sophisticated system with electronically mediated sensors and actuators where machine learning, in fact, is being used to process the data. So in that sense, we're already starting to sort of see that transition. We've been seeing it for decades. I have this smoke a joint and take a puff thought, which... I think we've had a few. Yeah, yeah. Well, I think we're going to do that part of the conversation, You can help me get my foot out of my mouth and figure out a more concrete way to think about it.

1:22:13So to your point that AI, there's a natural revolution, the Enlightenment, and now there's AI, and each might be a different pace or a different way in which science happens. If you think about the pace of how fast such transitions have been happening, You can draw over the long span of human history, this hyperbolic of the rate of growth is increasing. So, yeah, 100 ,000 years ago, you have the Stone Age. You go back even much further, how long probably it's been around, it would be like, let's say millions of years and 100 ,000 years ago, the Stone Age. Then 10 ,000 years ago, the Agricultural Revolution.

1:22:52Then 300 years ago, the Industrial Revolution, each marked by this increase in the rate of exponential growth. and then people think it's going to happen again with AI. But that would happen potentially even faster. It would not have occurred to somebody at the beginning of the Industrial Revolution that the next demarcation in this trend will be artificial intelligence. And so if things are getting faster and it's hard to anticipate what the next transition will be, I guess we just think of this singularity between now and AI and that's really what distinguishes the past from the future. but just applying the same heuristic that many people in the past should have had.

1:23:35Maybe the intelligence age is also quite short and the next thing after that is we don't even have the ontology to describe what it is but the future will not think of the past as like there was pre-intelligent AI and post-AI. No, that seems... I mean, obviously we can't prove this but it certainly seems quite plausible. I mean, part of the issue, of course, is just the substrate we have available to conceive seems all wrong. You can't speculate with a bunch of chimpanzees about what it would be like to have language.

1:24:14Just to sort of pick a major transition in the past, the transition itself is the thing. Right. And it seems likely if we're talking about taking a puff kind of thoughts, I'm certainly amused by the idea that there's going to be some transition involving artificial general intelligence using classical computers. But actually there'll be an interesting transition with quantum computers as well. They're probably capable of sort of a strictly larger class of potentially interesting computations. So maybe actually the character of sort of AQGI or whatever it should be called is actually qualitatively different.

1:25:01So maybe there's sort of a brief period between those two things. I mean, as I say, you know, this is just speculation, but it's certainly amusing. Is there a reason to think that? Because from what I understand, there's been, for decades, people like you have put pretty tight bounds on the kinds of things quantum computers can do. And so it'll speed up search somewhat. It will do, and the kinds of things that extremely speeds up, like Schurl's algorithm, it seems like it, again, maybe this is to your point that we can predict in advance what's down the tech tree, but at least from now here, it seems like you break encryption, but what else are you using Schurl's algorithm to do?

1:25:36Yeah, I mean, we've only been thinking about it for 30 years or whatever. It's 40 or so years, not for very long. And we sort of haven't in some sense thought that hard about it as a civilization. So, you know, does it turn out that it's very narrow? Maybe. Does it turn out that it's very broad? That's also like a really radical expansion. That seems distinctly possible. Like, keep in mind as well, we've been doing it without the benefit of having the devices. Right. Like, that's a pretty big bottleneck to have. If you're thinking about computer science in the 1700s and you're like, okay, do and and or.

1:26:13Yeah, yeah, yeah. What are you going to do? You can't anticipate Bitcoin. You can't anticipate deep learning. I mean, maybe you could if you were sufficiently bright, but it is a pretty hard situation. Right. What is your inside view? Having been in and contributing to quantum information, quantum computing back in the 90s and 2000s, what is your telling of the history? what was the bottleneck? What was the key transition that made it a real field? And how do you rank sort of the contributions for Feynman, to Deutsch, to everybody else that came along? Yeah. So, I mean, let's just focus on sort of the question about sort of what actually changed.

1:26:57So why was quantum computing not a thing in the 1950s? Like it could have been. Yeah. You know, somebody like, I don't know, John von Neumann, good example, absolutely pioneering computation, also wrote a very important book about quantum mechanics and was deeply interested in quantum mechanics. Like he could have invented quantum computing at that time. And I think there were quite a number of people who potentially could have. So why do we have these papers by people like Feynman and Deutsch in the 80s? And those are, I think, fairly regarded as the foundation of the field. There are some partial anticipations a little bit earlier, but they were nowhere near as comprehensive and nowhere near as deep.

1:27:40And, well, you should ask David. You can't ask Feynman, unfortunately, but he'll know much better than I do. A couple of things that I think are interesting. One is that, of course, computation became far more salient, sort of late 70s, early 80s. It just became a thing which many more people were interested in, partially for very banal reasons. You could go and buy a PC. You could buy an Apple II. You could buy a Commodore 64. You could buy all these kinds of things. It became apparent to people that these were very powerful devices, very interesting to think about. At the same time, in the quantum case, That was also the time of the ball trap and the ability to trap single ions and so on.

1:28:24And up to that point, we hadn't really had the ability to manipulate single quantum states. So you kind of got these two separate things that just for historically contingent reasons had both sort of matured around sort of, let's say, 1980 or so. And somebody like von Diamond could have had the idea earlier, But it is, I think, quite an interesting story about Richard Feynman. He went and got one of the first PCs, which is around 1980, 1981. And he was apparently just so excited with this device. He actually tripped and hurt himself quite badly, sort of carrying his brand new computing device.

1:29:15That's a very historically contingent sort of a coincidence. But having somebody who's very, very talented and understanding of quantum mechanics, also just very excited about these new machines. It's not so surprising, perhaps, that he's thinking then, what similar story could you have told 10 years earlier? Like there is just no, the conditions don't exist for it. So I think that's, I mean, it's quite a banal story. One of the things we were going to discuss was this idea you had about the market for follow-ups. And I think this is actually the perfect story to discuss it for because you wrote the textbook by the field, right?

1:30:00Mike and Ike is the definitive textbook on quantum information. And so you presumably came in after Deutsch, but you identified, in the 90s, somehow identified it as the thing that is worth following up on and building on. And instead of talking about it more abstractly, I'd love to actually just hear the story of, like the first-hand story of how did you know that this is a thing to, of all the things that were happening, physics and computing, et cetera, that I want to think about this problem. Sure, sure. So, yeah, Reed Weinman writes this great paper in 1982. David Deutsch writes an absolutely fantastic paper in 1985, sort of sketching out a lot of the fundamental ideas of quantum computing.

1:30:43So I'm 11 in 1985. I'm not thinking about this. I'm playing soccer and doing whatever. But in 1992, I took a class on quantum mechanics that was really terrific, given by Jared Milburn. And I just went and asked Jared one day after the fifth lecture or something, I said, do you have anything, sort of papers or whatever that you could give me? And he said, come by my office in a couple of days' time. And I did. And he presented me with a giant stack of papers, which included the Deutsch paper and included the Feynman paper and included a whole bunch of other sort of very fundamental papers about about quantum computing and quantum information at a time when essentially nobody in the world was working on it.

1:31:30He was. He'd actually, I think he wrote the very first paper that proposed, I mean, sort of a practical approach to quantum computing. It wasn't very practical, but it was actually in a real system. And so in some sense, I'm benefiting from the taste of this other person. But as soon as I read the papers or take a look at the papers, these are exciting papers. They're asking very fundamental questions and you're sort of like, oh, I can make progress here. These are things that one could potentially work on. Deutsch has this sort of conjecture that basically there should be, I don't know what the right term for it is thesis or what you would call it, that a universal model quantum Turing machine should be capable of efficiently simulating any system, any physical system at all.

1:32:28This is a very provocative idea. I think in that paper, he more or less claims that he's proved it. I'm not sure that necessarily everybody would agree with that. There's questions about whether or not you can say simulate quantum field theory effectively. And that kind of question is I think very interesting and very exciting there. It's obviously a fundamental question about the universe. You know, here's some wonderful ideas in there about sort of quantum algorithms and where they come from and what they mean and what they relate to the meaning of the function and questions like this, which is still not agreed upon amongst physicists.

1:33:15So yeah, there's just some sense of, oh, I am in contact with something which is A, deeply important, and B, we as a civilization don't have this. And so of course, you start to focus your attention a little bit there. I'm not sure I got the answer to the question that... Maybe I misunderstood the question. Yeah, yeah. Let me think of how it would phrase it. Maybe I'll explain the motivation first. Yeah, yeah. So in a previous conversation we were discussing how could you have known in the 1940s that Shannon's theorems and Shannon's way of thinking about communication channel is a deep idea that goes beyond the problems with pulse code modulation that Bell Labs was trying to solve at the time and it applies to everything from quantum mechanics to genetics to computer science, obviously.

1:34:09And one of the, I think, an idea you stated that we didn't get a chance to talk about yet was this idea, well, Shannon published this paper, there's all these other papers, but there's a market of follow-ups where people gravitate to and build upon Shannon's work and how do they realize that that's the thing to do and how does that process happen? And so I guess you gave your local answer, you read these papers and you immediately realized, okay, there's work to be done here. There's low hanging fruit. There's some deep provocative idea that I need to better understand and I could, you know, tractably make progress on.

1:34:44Yeah. I mean, so, you know, to some extent you're sort of saying, okay, I wanted to get into this game of contributing to humanity's sort of understanding of the universe. And you are applying sort of this low-hanging fruit algorithm. You're like, relative to my particular set of interests and abilities, where should I pick up my shovel and start digging? And there it was like, oh, this looks like quite a good place to start digging. and different people, of course, chose very differently. It was a very unusual choice at the time. It was 1992. Very few people were thinking about that. Yeah. Fast forwarding a bit.

1:35:29So you've been – I don't know how you think about your work on the open science movement now, but did it work? Like what would have – what did success there look like? What is it that that movement is trying to accomplish? Yeah, I mean, this set of ideas about open science, I mean, it's interesting. You didn't stop and define open science there, which I think 20 years ago you would have had to do. People recognize the phrase. People have some set of associations with it. Most often they have a relatively simple set of associations. It means maybe something about making scientific papers open access.

1:36:08Very often they have some set of notions about maybe it means also making code openly available. Maybe it means making data openly available. But already those are, I think, very large successes of the open science movement, which is to make those salient issues. Those are issues on which people have opinions. And then there are relatively common arguments. An argument like, so this is sort of the meme version, publicly funded science should be open science. that's a distillation of a set of ideas which you might be able to contest. But if you can get people actually sort of thinking about it and engaged with that kind of argument, that's a very fundamental kind of an issue to be considering in the whole political economy of science.

1:37:01If you go back, say, three centuries, there was a very similar kind of an argument prosecuted, which is the question, do we publicly disclose our scientific results or not? So if you look at people like Galileo and Kepler and so on, the extent to which they publicly disclosed, like it was done in a very odd kind of a way. Sometimes they did bizarre things where famously they published some of their results as anagrams. So basically, they'd find some discovery, they would write down the result in sort of a sentence like his, the discovery of the, I'm trying to think of an example. I think the moons of Mars, I think was one such example.

1:37:52I'm getting it wrong, was it Hooke's Law? Anyway, it doesn't matter. The point was they'd write it down, but then they'd scramble it, publish that. And then if somebody else later made the same discovery, they would unscramble the anagram and say, oh, I actually did it first. This is not an ideal way. This is not an ideal foundation for a discovery system. And then it took a very long time, over a century, I think, to obtain more or less the modern ideals in which what you do is you disclose the knowledge in the form of a paper. there is then an expectation of attribution and so there's a kind of reputation economy which gets built and so basically oh such and such did this work so they deserve the credit for that and that's then the basis for their careers so this is sort of the underlying political economy of science and that made a lot of sense when what you've got is a printing press and the ability to to do scientific journals then you transition to this modern situation where in fact you can start to share a lot more you can start to share your code you can start to share your data you can start to share in progress ideas, but there's no direct credit associated to those.

1:39:00It's not at all obvious how much reputation should be associated to them. That's all constructed socially. And so making it a live issue is, I think, a very important thing to have done. And that's, I view anyway, is one of the main positive outcomes of work on open science. Shall we give you a really practical sort of example to illustrate the problem? For a long time in physics, there was a preprint culture in which people would upload preprints to the preprint archive. And in biology, this didn't happen. There was no preprint culture. That's changing now. But for a long time, this was the case.

1:39:46And I used to sort of amuse myself by asking physicists and biologists why this was the case. And what I would hear sometimes from biologists was they would say, well, biology is so much more competitive than physics that we need to protect our priority. And so we can't possibly upload to the archive. We have to just publish in journals. And then I would sometimes hear from physicists, physics is so much more competitive than biology that we need to establish our priority by uploading as rapidly as possible to the preprint archive. We can't possibly wait to do it with the journals. And I think this emphasizes the extent to which this kind of attribution economy is just something we construct.

1:40:30It's just something which we do by sort of agreement. And so any attempt to sort of change that economy results then in a different system by which we construct knowledge. And so there is sort of this very fundamental set of problems around the political economy of science. You know, sort of we've got this collective project and how we mediate it depends upon the economy we have around ideas. One of the sort of things you've emphasized as a part of this project of open science is collective science or groups of people making progress on a problem where no individual understands all the logical and explanatory levels necessary to make a leap or a connection.

1:41:19Outside of mathematics, what is the best example of such a discovery? I mean, I'm not sure I have a well ordering of them to give you a best, but I mean, yeah. An example that I think is very interesting is the LHC, where it's just this immensely complicated object. I actually, years ago, I snuck into an accelerator physics conference. I didn't know anything at all about accelerator physics, but I was just kind of curious to see what they were talking about. And this particular group of people were experts on numerical methods, in particular on inverse methods. And so it basically turns out, you know, inside these accelerators, you have these cascades.

1:42:01So a particle, you know, will be massively accelerated. Maybe it'll be collided. And then you'll get a shower of particles which decays and decays and decays. And there's just this incredible sort of, you know, consequential shower, which is ultimately what you see at the detector. data, and then you have to retroactively figure out what produced it. And so there's these very, very complicated sort of inverse problems that need to be solved. You've got this final data, but you need to figure out what produced it, and that's how you look for sort of signatures of these. And what many of these people were was they were incredibly deep experts on simulation methods for sort of following particle tracks.

1:42:43And like this was really deep and difficult stuff. And I'm like, wow, you could spend a lifetime just learning sort of how to do this and how to solve some of these inverse problems. And you would know nothing about, or you would know very little about quantum field theory. You would know very little about detector physics. You would know very little about vacuum physics, all these other things that are absolutely at work. Very little about data processing, very little about all these things that are absolutely essential to understanding say the Higgs boson. And I don't think it's possible for one person to understand everything in depth.

1:43:19Lots of people understand broadly a lot of these ideas, but they don't understand sort of everything in the depth that is actually utilized. That's why there's these papers with well over a thousand authors. And those people can talk to one another at a high level, but they don't understand each other's specialties in that much depth. I mean, things like, as I say, detective physics, vacuum physics, these kinds of solving of inverse problems. Like, this stuff is incredibly different from each other. And, you know, to understand it in real detail is serious work. How do you think about prolificness versus depth?

1:44:00Where, I don't know, maybe Darwin's an example of somebody who's like just gestating on something for many decades. there's other examples where Einstein during the year comes with special relativity is just doing a bunch of different things. Pais talks about how they were all relevant to the eventual buildup. Yeah I mean you know it's something I stress about a lot sometimes I feel like I'm you know too slow. Actually it's funny that I mean the Darwin example is really interesting like you know prolific at what like I mean I god knows how many letters he wrote it must have been an enormous number.

1:44:36So you're certainly very active. There's two types of work that tends to be involved in any kind of creative project. There's routine stuff. And there, you just want to avoid procrastination. You just want to like, how do I get good at this? Or how do I outsource it? And how do I do it as rapidly as possible? And just avoid getting into a situation where you're prolonging it. And then there's high variance stuff where you actually, you need to be willing to, you know, take a lot of time. You need to be willing to go to the different places and talk to the different people where in any given instance, most of it's just not, it's not going to be an input.

1:45:20And somehow sort of balancing those two things. I think a lot of people are very good at doing one or the other, but it's hard to, you know, it's almost like a personality trait, sort of, you know, which one you prefer. And people tend to end up doing a lot of one and not enough of the other. So I certainly, you know, sort of try and balance those two things. I mean, it's such an interesting example. I mean, 1905 is just this extraordinary year. Like you can delete special relativity entirely, and it's an extraordinary year. You can delete special relativity, and you can delete the photoelectric effect for which he won the Nobel prize and it's still an extraordinary year, like plausibly a multi-noble prize winning year.

1:46:03So what's he doing? Yeah, I mean, maybe the answer is just he's smarter than the rest of us. And there's a lot of luck as well.

1:46:16But certainly for myself anyway, like trying to identify those things that are routine that I should get good at and then just try and do as quickly as possible. That's yielded a certain amount of returns. But also being willing to bet a little bit more on myself on sort of the variance side has also been very, very, very helpful. That's really hard because intrinsically you're putting yourself in situations where you don't know what the outcome is going to be. And so if you're very driven to be productive and whatever, and actually mostly it's not working over there. You're like, let's reduce this.

1:46:53Like it doesn't feel right. When I worked in San Francisco, actually a practice I used to have each day was instead of taking the 15 minute walk to work, I would take the more beautiful 30 minute walk to work, partially just because it was beautiful, but partially also as just a reminder to like, like that there are real benefits to not being efficient. But it's not an answer to your question. I mean, really, I think all I'm saying is I struggle a lot with the question. I mean, there are these, Dean Keith Simington, I forgot his exact name. Yeah, I know who you mean. Has this famous equal odds rule where he says the probability that any given thing you release, any paper, book, whatever, will be extremely important for a given person through their lifetime is not that different and really determines in what era they are the most productive is how much they're publishing.

1:47:47Any given thing has equal odds of being extremely important. Maybe just think of some of the most successful creatives or scientists that are just doing a lot, like Shakespeare is just publishing a lot. And, of course, there's kind of examples. You know, Gödel publishing almost nothing. But broadly speaking, you need a very good reason to be avoiding it, basically to not do that. It's funny, I mean, I've talked to, I've met a lot of people over the years who you talk to, they're clearly brilliant and they're just obsessed that they are going to work on the great project that, you know, makes them famous and they never do anything.

1:48:30And that seems connected, like it's a type of aversiveness. I think very often they just don't want public judgment. Something that I would love to see, there's an awful lot of biographies and memoirs and histories of people who achieve a lot. I wish there was a very large number of biographies of people who are fantastically talented who just missed.

1:48:58I've known people who won gold medals at IMOs and things like that, who then tried to become mathematicians and failed. What happened? What was the reason? And I suspect in many cases that's actually more informative than anything else. You have this essay that I was reading before this interview about how you think about what is the work you're doing. And writer doesn't seem like, as you say, was Charles Darwin a writer? What exactly is that label? I'm a podcaster, right? And in a way, obviously, our work is very different. But I also think a lot about what is this work and how do I get better at it?

1:49:43And in particular, how I can make sure there's some compounding between the different people I talk to on the podcast, where I worry that instead of this kind of compounding, there's actually I build up some understanding that's somewhat superficial about a topic and then depreciates and I move on to the next topic and it sort of depreciates. and so I think there's this question there's a lot of podcasters in the world who will interview way more experts than I have or have and I don't think they're much the wiser or more knowledgeable as a result so it's clearly possible to mess this up and I wonder if you have thoughts or takes or advice on how one actually learns in a deeper way from this kind of work yeah I mean it's sort of an incredibly complicated and rich question.

1:50:34I mean, it does seem like the question is, how do you make it a higher growth context? How do you make it a more demanding context? And you can do that in relatively small ways, but that might have a yield compounding returns, or you can do something that is maybe more radical. Maybe it means actually starting a parallel project in which you do something that is actually quite a bit different. There is something I think really interesting about how being very demanding can simply change your response to something. Something that I would sometimes do with students and sometimes with myself, was really aimed more at myself, was they would say some week, oh, I'm going to try and do this work over the coming week.

1:51:17And then the next week would come by and they hadn't solved the problem or whatever. And you're sort of like, if a million dollars had been at stake, would you have put the same effort in? And the answer is no, sort of invariably. like they've tried, but they haven't really tried. I think that's a very familiar feeling for all of us. Often you could do a lot more if you had just the right sort of demanding taskmaster standing by you and saying, look, you're barely operating here. And so I do sort of wonder a little bit about like, what's the demanding taskmaster? What What can they ask you that is going to make your preparation way more intense?

1:52:05The most helpful thing, honestly, is for some subjects, it is very clear how I prep. Like I'm doing an upcoming episode on chip design with the founder of a company that is chip design. And he wrote a textbook on chip design. And yesterday I went over to his office and we brainstormed five sort of roofline analysis I can do. And if I understand that, I have some good understanding. The problem is with almost every other field, there's not like you, I don't know, when I interviewed Ilya three, four years ago, it's like implement the transformer. And if you implement it, like you have some nugget of understanding you've clamped down.

1:52:44And with other fields, it's just like, I vaguely understand this. It's not clamped. I vaguely understand this. I vaguely LLM'd about this. I LLM'd about this. But there's no forcing function that you do this exercise. And if you do it, you will understand. So, I mean, really what you're sort of saying is you can do a good job at podcasting without actually attaining this. Exactly. And that's the problem from your point of view. Yeah. You want to sort of change your job description so that you are internalizing these chunks. Right, right. And just getting this kind of integration each time. time.

1:53:15And it seems to me like what that means is you actually want to change the structure of the work output at some level. I mean, lots of people think... There's this terrible idea that people have that they should be in flow all of the time. And of course, as far as I can tell, high performers just don't believe this at all. They're in flow some of the time. You certainly see this with athletes. When they're actually out there playing basketball or tennis or whatever. Ideally, they are in flow much of the time. But when they're training, they're not. They're stuck a lot of the time or they're doing things badly.

1:53:54And I suppose I wonder what that looks like for you. That I would be extremely satisfied with. The problem is I just like, I don't know what the equivalent of do the 64 lapses for almost a... And so this is a thing you can change by choosing guests where there is a legible curriculum. And so maybe it's a mistake for not having done that. Or also, like, there's no real way to prep for Terrence Tao or something. And, like, there's no curriculum that's, like, a plausible one. I think there's one failure mode. So there's many failure modes. But one is if you could do one dynamic I'm worried about, a long-term dynamic, is that you can have a good podcast and there's a local maximum.

1:54:34But for no particular guest or topic are you going deep enough that you've – I think my model of learning is if you don't really understand the deeper mechanism, you're just mapping inputs and outputs of a black box. Yeah, yeah. And that just fades incredibly fast or is not worth it in the first place. And you kind of just move on and it's over. Yeah. And you kind of need to build the intermediate connection.

1:55:00And it's unclear. I think actually AI in a weird way is really easy for that reason because there is a clear thing you can do. Just implement it, right? And then you understand it. We're almost, if I applied that criteria elsewhere, do I just not do history episodes? Exactly, Ada Palmer. Like, what, you know, wonderful to talk to, incredibly interesting, but for you personally, like, what changed? Right. Yeah, there's some things I learned. I think I could have done, if I had maybe allocated more time, especially after the interview, to like, let's write up 2 ,000 words on everything I learned and how it connects to other things I know and something.

1:55:36And maybe that's the thing worth doing is spreading out the episodes more and spending more time afterwards consolidating. But yeah, I think I would pay basically infinite amounts of money if there's somebody who's really good at coming up with, here's the curriculum and here's the practice problems you need to do and here's the exercising you do after the interview to clamp what you have learned. Have you tried doing that with somebody? It's hard to find something. I mean, I haven't tried super hard, but it seems like... Isn't it tough to find somebody who could do that for every single kind of discipline?

1:56:07Maybe I should just hire different ones for different topics. Maybe, or there's something about like, I mean, what problem, you know, are you solving sort of for each episode? And I mean, as far as I can tell, like, that's the only way I really understand anything is that, you know, I get interested in something. At first, I don't even have a problem, but there's just some sense of there's some contribution to make here and gradually you home in and there's a problem. And then you, I mean, funnily enough, I mean, spending time stuck is incredibly important. and I sort of, you know, that used to just be annoying.

1:56:39Now it seems like, oh, this is actually maybe even the most important part of the whole process. But that very hard oneness of it means that, you know, I internalize it afterwards. I often find actually, if I, you know, I've written sometimes 10 ,000 word essays in, you know, a couple of days and I've written them in, you know, three months or six months. I feel like I didn't learn very much from the ones that only took a couple of days. Interesting. Whereas, you know, some of the ones that took three months, I'll be, you know, 15 years later, I'll still remember. Yeah, can you describe outside of physics how you learn of the ones that took three months?

1:57:26I mean, by far the most, you know, the common things, there's always some creative artifact. Sometimes it's a class. Sometimes it's engagement with a group of people who, you know, there's some collective creative artifact that you're working on together. I mean, you might not even be aware of it, but you're acting as an input to their creative ends in some way. And sometimes it's just, you know, it's an essay or a book or whatever. it's one of the reasons why I often quite enjoy doing podcasts I mean particularly I said yes to come here partially because I know you ask unusually demanding questions and so it's sort of that's an attempt to get this sort of perspective from a different it's a different kind of enforcing function so you're trying to pick sort of the most demanding creative context for this interview I went through like three lectures of the Susk and Sesha Raltzuri book The problem is that there's almost no practice problems in it.

1:58:27And so I hired a physicist friend who's going to like, I haven't done it yet, but it's like every lecture I want like a bunch of practice problems go to them. And I'm planning on being appropriately humbled. How do you make it as jugular as possible, right? Like the higher you can raise the stakes, the better. I mean, the interview is in some sense high stakes, but also it doesn't necessarily test deep understanding. Yeah, but I don't think the interview is that high stakes, right? You're not writing a book about special relativity, and you're not trying to write a book that replaces the current, whatever the existing standard textbook is.

1:59:02That's a really high... A phrase that I find particularly difficult, and it's a funny one. People will talk about going deep on a subject. And it turns out different people have different ideas of what this means. Some people means they read a couple of blog posts. For some people, it means they read a book about it. Some people, it means they wrote a book about it. And I think what your standard is, the standard you hold yourself to, determines a lot about your ability to integrate knowledge in this way. I don't know what your experience has been, but I found that I'm getting, in some sense, it will move much faster on some things to the help of AI, but I don't know if I'm learning better.

1:59:51And I think it's probably because the hardest thing, the thing that is most demanding, is so aversive that you try to take any excuse you can to get out of it. And just having back and forth conversations where you gloss over. It's entertaining, but not necessarily anything else. Yeah, so it's such an easy way to get out of the thing. In fact, it makes it easier because instead of doing some intermediate thinking, there's always a next question you can ask a chatbot. Yeah, and it's somewhat valuable. Like it's not, I mean, that's part of the seductiveness, of course. Like it's not actually useless.

2:00:26But yeah, it can sort of substitute for actually doing the thing that maybe you should be doing. It's interesting that, like the extent to which, to what extent should you be outsourcing that kind of stuff? And to what extent, it's really, there's some sort of interesting judgment call about, you actually, there is a whole bunch of routine work that you want done. And in fact, it's low value for you. So you may as well get, if you can get a chatbot to do it, you may as well. So somebody interviewed the pioneering computer scientist Alan Kay years ago, and he was asked what he thought about basically Linux.

2:01:07And if I remember his answer correctly, he basically said, look, you know, it doesn't have anything to do with computer science. It's just a great big ball of mud. There's a few interesting ideas in there which are worth understanding. But mostly, all you're learning is stuff about Linux. Like you're not actually learning anything which is transferable. I thought there was like a very, like that there's a certain kind of seductiveness to some things where, you know, it's sort of a Rube Goldberg machine. You can just sort of learn about all the bits and it feels kind of entertaining. But if you step back and think about the question, you know, what am I actually doing here?

2:01:48It might not actually be meeting your objectives. Maybe you want to become a, you know, a sysadmin and learning Linux is a great use of your time. There's no harm in that at all. But if your objective is to understand the fundamentals of computing, it's much less clear that that's a good use of your time. I thought that was certainly an answer I've thought a lot about where you actually need to, for a certain type of mind, there is a seductiveness in just learning systems and confusing that with understanding. Yeah. Okay. I'll keep you updated on how to discuss. Yeah. I owe you a text within a month of some revamped learning system.

2:02:28I'll be really curious if you, I mean, it's also true, right? Like tiny incremental improvements in this. I mean, they're just worth so much. It's sort of the main input into the podcast. It's great that the bookshelves are fancy and I've got a Blackboard or whatever. But really, like the thing that makes the podcast better is if I can improve the learning, I do. So it's, yes, it's worth every morsel of improvement. All right, thanks for the therapy session. Great notes on that. Thanks, Michael. All right, thanks, Kutakesh.