The following is a rough transcript which has not been revised by The Jim Rutt Show or Michael Nielsen. Please check with us before using any quotations from this transcript. Thank you.
Jim: Today’s guest is Michael Nielsen, a scientist who has worked on quantum computing, AI and more. You can learn more about his varied interests michaelnotebook.com. Welcome, Michael.
Michael: Hi Jim.
Jim: Yeah, it’s good to have you here today. But today we’re not going to talk about quantum computing, nor AI, but rather one of Michael’s other interests, metascience. If you are interested in quantum computing and our quantum universe, I had a great chat with Seth Lloyd back aways on EP 79. Feel free to check it out. I’m using as a starting point for our conversation today, Michael’s recent essay slash short book with co-author Kanjun Qiu. Is that how you pronounce that?
Michael: Qiu. Kanjun Qiu.
Jim: Kanjun Qiu, okay. Titled, A Vision of Metascience: An Engine of Improvement for the Social Processes of Science. With that out of the way, let’s just hop in. What is metascience?
Michael: It’s a little bit up for grabs at this point, Jim. It’s not quite a field, it’s more of a proto-field than anything else, so I guess you get to define it, and part of the point of the essay for us anyway was just figuring out what we thought the opportunity was. Basically, really, it’s in the subtitle you read, thinking of it as an engine of improvement for science, a way of figuring out what institutions, what social processes might work best to support discovery, is the short version.
Jim: Yeah, as were talking about in the pregame, I think about it as, like any set of cultural and cultural institutions, there’s a series of knobs which may or may not be obvious, and how do you set the knobs? How do you invent new knobs, etc.? And it’s funny that while there’s been philosophy of science for quite a while, which has had some impact on science, the idea of metascience, studying science as an engine, as a process, is not really something that there’s been a whole lot of, which is surprising considering the stakes. I was trying to figure out how much money is spent worldwide on science, and it’s funny, when you do a Google, you mostly get R&D spending, which comes out at about $2.4 trillion. If you take the usual rule of thumb that maybe 15% of that’s actual science, we’re probably talking something on the order of $360 billion maybe. Does that seem like a reasonable number for you?
Michael: It’s pretty reasonable. I think the NSF tries to keep track, and they usually estimate, I think it’s about $400 billion for basic science worldwide, so you’re almost on the money.
Jim: Yeah, not bad for a schwag based on a simple heuristic. And so we’re talking about big stakes here. That’s something on the order of more than half the US defense budget, so we’re talking about some big impact. But more importantly, the unfolding of our society, what we have as a human race depends very significantly on how well we do our science, and it would certainly seem that this bears some serious consideration. Let’s now talk about what… Give an example here for folks, which you did a great job in the essay, putting the meta in metascience. When you talked about it, there’s a very vivid example, I found from Nick Szabo. Now, is he Satoshi? I don’t know. It could be. Probably the best candidate out there. Anyway, you tell the story. In the early renaissance, sailing and all that stuff, this was really very vivid for me to bring the meta into metascience.
Michael: Well, just the idea, so just to give your listeners some context, this isn’t related to science in any way. It’s really a question about to what extent systems change, and in particular, change in the way things are financed can have an impact at the ground level. If you go back to 14th century Genua, that was the time at which maritime insurance was first invented, so basically it used to be the case that if you were financing an expedition, say you were a merchant who wanted to get goods from one part of the Mediterranean to another, your ship runs aground, well, bad luck, you’re ruined. And this ability or this creation of the concept of an insurance premium meant that instead of losing your entire fortune, actually you’d be much more likely to lose only just that relatively small premium.
To see a flourishing of trade, a flourishing of shipping, you might think in advance what you needed was better ships, better sailors, better training for sailors and so on, but actually the invention of this new abstraction, or the idea of risk, the idea of parceling up risk and finding ways of selling it and buying it, actually turned out to be at least as important as any of those. And we use it in the essay just as an example of the way in which systems change can result in massive on-the-ground changes as well.
Jim: And you make the good point that if you go to talk to actual scientists and science administrators, they’ll usually say something like, “Just fund good people doing good work.” In the same way your Genevese shipping captain might say, “Give me a little bit bigger ship with some bigger sails and maybe a stronger rudder, then we’ll do better.” But look at things from a different dimensional perspective and say, “Ah, there’s some things that are outside the system itself which define the system.” That’s why I found that to be such a nice, strong and vivid example.
Michael: I can tell that you’ve actually asked that question, Jim, just from the tone of voice you use is remarkably similar. Yeah, it’s true. One other thing as well that really contributes there to why people are sometimes skeptical, there’s not really any mechanism for institutional change. In fact, there’s a lot of centralized bottlenecks, so you chat with a lot of people, they’ll complain about the National Institutes of Health or something like this, but to get any change, really their only vector of change is actually to try and convince the director the change should be made, and that’s quite a centralized kind of a bottleneck. That’s also a reason for skepticism of metascience, that if you can’t actually make large changes, it all starts to seem a little bit like just sort of abstract theorizing rather than something that can concretely be done.
Jim: And one of the things I liked about the paper, it was a mix of abstract theorizing and concrete things that could be done, but without losing the abstract. At the Santa Fe Institute where I’ve spent a lot of time, we try to do theory, practice, theory. Our side tends to be the theory, but we generally partner with people that are doing experiments in the field or gathering data, because otherwise of course the theorists just like to sit in the… Close their door and just theorize, and the experimentalists just want to experiment. You got to get the two together to really have good generative science, in my mind. Anyway, before we jump into the next topic, a term that you use, which I thought was again quite apropos, was the concept of intellectual dark matter. Why don’t you tell us what you meant by that?
Michael: I just mean that there’s an awful lot which scientists know which project ideas or possibilities which, however, there’s no institutional awareness of these things. Let me give you just two examples to illustrate the point. One is fictional, the other one is actually happening. I’ll start with the fictional one first. One of the program proposals that we make in the piece is for what we call the Century Grant Program. It’s simply basically just to put out a call saying, “Look, does anybody have any ideas for scientific projects which shouldn’t be funded for 3 years or 5 years or 10 years, but no, funded for 100 years? And there’s a few examples of this in the past, probably the most famous is the Mauna Loa Observatory, which is where they found the carbon dioxide is increasing. There’s a few others. The Framingham Heart Study and whatnot, but they’re all done in this very bespoke way.
And instead we say, “Well, let’s tip this upside down and say, ‘Actually, there’s probably a ton of such ideas out there in the world, in the minds of individual scientists, but they’re not visible to institutions at all at the moment. They’re a type of intellectual dark matter. Maybe you should just actually essentially build a detecter for them.'” Say, “We’re going to do this kind of funding, what are your ideas?” That’s a theoretical example, and an example that I really like that is just coming into existence right at the moment, some friends of mine, Adam Marblestone and Anastasia Gammick has started what they call focused research organizations. And basically this is… Again, it’s like taking of something that’s done bespoke in the past, something like LIGO or the Large Hadron Collider, and trying to do it at scale.
They’re saying, “Let’s find projects which might cost a few 10s of millions of dollars, let’s say 50 million dollars, where we really, we understand how to build some new instrument or take some new dataset and we just need to assemble the right group of 50 or 100 people to go and do it, maybe over 5 years or 10 years.” And their thesis is that actually scientists, again, this is a type of intellectual dark matter, there’s actually a lot of ideas for these focused research organizations that are out there in the heads of individual scientists, but at the moment they don’t have any way of going to the National Science Foundation or the NIH, or anybody like that, and there is no program to do it, so again, they’re building a detector for this intellectual dark matter, and they’re starting to fund some pretty cool things doing it.
Jim: Yeah, it’s interesting you mentioned that, because one of my own critiques of particularly academic science, is it’s built around the scope of a principal investigator plus some graduate students who turn over every four or five, six years, nowadays a few research scientists thrown in the mix. But the depth of problem you can attack with that model is only so deep and you run into other alternative models, for instance, spend a fair bit of time at the MIT Brain and Cognitive Science Department where I’m on the visiting committee, which is a governance board for it, so I know what they do pretty deeply. Last time I was up there, I went over and talked to the Broad Institute, which is across the street, and they do something quite differently. They have six projects targeted on big, big, big problems. They’ve funded each one with about $100 million and the teams are about 70% professional research scientists. There’s a PI nominally in charge, but it’s the research scientists that really drive it.
And these are multi-year, 10 year or more projects, and they’re attacking a class of question which it would not be feasible to ask if you were a professor in a typical academic department with 5 graduate students, 2 postdocs and 1 research scientist. I think that’s actually very interesting, and you have a little graphic in the paper. Shows the whole space of science and scientific design and the current world of science, a small subset, maybe 20% of the space. And I think things like that are very apropos. The other thing that’s worth talking about is how ad hoc and frozen accidental a lot of our institutional structure is. Think about the invention of the scientific journal. Some guys writing to each other back in Europe in the late 17th century.
And one of my favorites, or let’s say least favorite, but a frozen accident that’s become scarily prominent is the silly thing called the H-index. I’m like, “How did that happen?” I’ve heard all kinds of weird stories how it got invented. I don’t even remember the calculation, but it’s some calculation, a number of papers written, divided by the impact factors of them, and it supposedly indicates who’s hot shit in science, and it’s actually a pretty damn arbitrary figure, and you could even argue that it’s for certain kinds of things that you’re looking for, it may actually be an anti-indicator. But somehow it’s now become this thing that in many disciplines, very prominent on people’s CVs. My H factor is 4.3. Wow, aren’t I cool? Right?
Michael: Yep. Again, you’ve certainly talked to some people about this. They do… What’s one thing that’s scary is actually how much scientists will know about each other’s H indices as well.
Jim: Yeah. Oh, absolutely.
Michael: Not just their own.
Jim: Yeah, absolutely. And it’s a measure with absolutely no theoretical justification. It’s a completely ad hoc. It’s not entirely worthless, or put it this way, there’s some signal there, but signal of what, I don’t think anybody actually knows. It’s quite interesting. And of course, the other one that was a frozen accident is the… And it totally changed, actually, the organization of the field of science, was the invention of impact factors by Eugene Garfield and friends at ISI. In fact, I know a fair bit about them because oddly enough, when I was the CTO of Thompson Reuters, one of our companies was-
Michael: Oh, wow.
Jim: … Web of Science. And I never did get to meet Garfield, even though he was still alive in those days. He came and visited occasionally, but I did work with the business leaders of that business a fair bit and got to understand it, and what a racket that turned into. And again, if Eugene Garfield hadn’t invented this method of calculating impact factors, he and some other folks, the worship of the impact factor would not have completely changed how science is structured.
Michael: It’s funny too, because Garfield himself had quite mixed feelings about that stuff. He wrote several papers where he basically talks about the limits of citation analysis and the limits of ideas like impact factor, but once it was created, it was a little bit of a beast that he couldn’t control. I heard the CEO of, was a springer, springer of [inaudible 00:13:54] at the time, say on stage once how wonderful it was that journals just had this sort of automatic quality branding that they could use to sell, and what an advantage this was from a sales pitch point of view. And of course he was talking about impact factor, which is go to your librarians and say, “Well, this has such and such an impact factor, so therefore it should be in your budget.” It doesn’t make a lot of sense.
Jim: Yeah. Yeah, and there’s also the… Was it Campbell’s Law, or one of these guy’s laws that once something becomes a measure that people behave upon, the measure is no longer much good. Probably whatever it was Garfield found is different than what those factors mean today. Quite an interesting… But again, I think the bigger point, the meta, meta point is that our scientific operating system is a whole series of frozen accidents essentially, that wasn’t thought through, wasn’t optimized. And you gave a nice little funny story early in the paper about imagine some aliens came down and looked at how we did science. Would they say everything that we did was exactly right? And of course the obvious answer was, “Hell no.” And someone could argue that metascience is at least an attempt, and of course it won’t get it right either, but it’s at least an attempt to back off a little bit and say, “All right, how can we tweak the system? What things are counterproductive or useless or both? What things are productive, but could be made better?” Etc. And-
Michael: There’s a really, really funny thing about you talking about the way stuff gets frozen, and you read the early history of NSF, NIH, DARPA, all these organizations, there’s so much that was just made up on the fly, often with the idea of, we’ll fix it later, but then there’s no mechanism for change within a lot of these organizations, so in fact it becomes impossible to fix later. I don’t know. How is the NIH going to fix its panel system? They have, whatever it is, 180 different areas or something now. They need to get this enormous amount of buy-in from some incredible number of people before they could do it, so it just gets very, very hard.
I don’t think it’s actually an accident, in some ways, that an organization like DARPA is pretty functional in some ways, and I think it’s, maybe, this is just speculation, maybe because they’re a little bit later than some of the others, so they actually get to look and say, “Well, we don’t like the way they did this, that or the other thing, so we’re going to design our system to be better.” But once the system itself is in place, unless it’s designed to change, it’s pretty hard to change.
Jim: Indeed. It’s interesting, I just read an article recently somewhere about how the baby DARPAs are not doing very well. There’s now an energy DARPA and I think there’s an environment DARPA, there’s three or four or five of them, and this article basically looked into them and said, “None of them are even close to the level of creativity and the impact that the original DARPA had.” And it may well have been that the original DARPA was a fortuitous design and the right people at the right time.
Michael: Hard to know. You have to read those articles, I think, with a little bit of skepticism. Sometimes there’s a political agenda which you don’t necessarily know-
Jim: Yep, of course.
Michael: And there’s also, there’s just the big fact, which is a lot of these, like the energy DARPA that’s… What is it? It’s 13 years old now. Well, when the original DAPA was 13 years old, they wouldn’t have looked too crash hot either. They’d done some things which in retrospect we know were very important, but at the time, that’s 19… 13 years after would’ve been what? 1971, I think. They’d already done the work that led to the creation of the internet, but in 1971 you couldn’t have pointed to anything. Nobody would’ve been excited by that. They would’ve just shrugged their shoulders and gone, “Well, okay. Maybe that’ll be important but don’t see it right now.” [inaudible 00:17:47].
Jim: Yeah. In fact, 1971 you would’ve looked at that and said, “What a simple-minded protocol. Why would anybody use a piece of crap like that?” I actually know the two guys that designed TCP/IP and they had no idea either. They were… Yeah, so anyway. Yeah, that’s a good point too. We’ve been talking a fair bit up in theory space and evolutionary theory, frozen accidents, things like that. Maybe for the audience it would be useful now to bring it down to some tangible examples, and we’ll get back to process and theory again. And you laid out early in the paper a list of examples, and you made it clear that you don’t necessarily vet that these are all the best possible ideas or an exhaustive list or anything else, but let’s pop through a couple of them.
You mentioned the Century Grants, which is the idea of attracting the dark matter of things that could take up to 100 years. Today, nobody bothered proposing that because there is no funding source, and if there was a funding source, maybe we’d get great ideas. But another one, this I thought was very nerdy and I liked it for that reason, which was fund by variance. Maybe you could tell people what that one is.
Michael: Actually, for me anyway, it comes out conversations with venture capitalist friends. One VC friend in particular commented to me that unless at least one of the partners at his firm strongly opposes a bet, they won’t invest. And the reasoning was pretty good, which is, his reasoning was that in venture you make all your money… It’s not enough to be right, everybody else needs to be wrong, and as he explained, the best evidence that other people are going to be wrong is if somebody who he thinks is very smart and respects strongly disagrees with the investment. There’s something to that in science as well. Certainly a lot of the most important discoveries were, particularly early on, very controversial or else in some cases just completely illegible. Things like… Oh, no, actually the meteor impact theory of the dinosaur extinction. Even five years after the paper had been published, most paleontologists still thought this was wrong.
There’s this great New York Times article where they actually went off and polled a whole bunch of the paleontology community, and a lot of them thought the meteor impact theory was crazy. And there’s many, many examples like this where often good ideas are just very polarizing, and rather than trying to get consensus that this is a great idea, sometimes you want a few people to say, “I absolutely love this idea,” and maybe a few other people say, “I absolutely hate this idea,” rather than having a bunch of people say, “Yeah, this is a pretty good idea.” Which is I think the more common situation in the consensus-based model we have at the moment.
Jim: And the beauty, you can actually turn it into an algorithm. You can calculate the variance from the referees, so you could actually have, well, this one… It doesn’t have to even be a judgment call. It could be a straightforward calculation. Another one I thought was very clever, and you presented it in a good punchy fashion, was failure audits. Everybody claims they do high risk audit and then you quoted some, I forget which agency, we do high risk science-
Michael: European Research Council.
Jim: And it’s like, “93% of our stuff works.” Well, guess what, dude? [inaudible 00:20:54]. Talk to us about failure audits and firing program managers who don’t fail enough.
Michael: Yeah, the thing you pointed out, there’s this just… One about the European Research Council is they have this long report on their own activities talking about how high risk everything is, but also it is, it’s like 90% of the stuff works, so I don’t know quite what they’re talking about. But certainly after the fact, if you’re running a funder and after a few years you start to evaluate the success or failure of your programs, I’ve noticed in chatting with individual funders, if you ask them what their big failures are, usually they get really tongue tied. They like to talk about how they support high risk research, but actually ask where are the craters left upon impact, they usually don’t actually have a list at hand.
And that’s started to smell a little bit suspicious to me, so I think just being of honest about it and basically writing down your list of failures, maybe making it public, an anti-portfolio, but then also just evaluating. If you think that you have a high risk program and 50% of the stuff isn’t failing, maybe there’s a bit of a problem. Maybe you’re actually not taking enough risk, and… Ultimately, you can imagine having the program manager or maybe the director of the funder, job under threat if they don’t actually produce what they claim. A way I like to think of it is so many funders will say they want high risk work, but their processes ensure that they will only fund low risk stuff, so they’re not buying what they say that they’re buying, and this would be a way of hopefully correcting that, particularly if it’s done publicly.
If they actually publish the results of their failure audit, that actually starts to build credibility. Scientists know that when a funder says that they want to support high risk work, they know that usually that’s not true. And so of course they’re very cautious about… Why waste your time submitting a genuinely high-risk proposal when you don’t expect it to be actually funded? If a funder is however credibly saying, “Look, in the past we really have done this,” then it starts to become a little bit more interesting if you’re on the application side as well.
Jim: Let me add one improvement to your proposal there, which is that of course, being humans, being gaming the system… In fact with the rule, they’re going to intentionally fund a certain number of bozo projects of incompetent buffoons, so-
Michael: That’s right.
Jim: … make it an actual failure audit, and then in my capital work I, when I’m trying to assess a team, look at their successes, their track record, I look for noble failures as opposed to incompetent failures or bozo failures. And so I would add to your idea that there needs to be something like an identification of noble and worthy failures, and to distinguish them just from bozos and just bad ideas and-
Michael: Yeah, that’s a great term. That’s a great term.
Jim: Yeah, I think it’s a really good one. And then the last we’ll go on to, there’s some other ones in your list, but one that I personally feel very strongly about is you propose an open source initiative. And we talked about the institutional lack of depth in the classic PI plus a few graduate students and a postdoc or two, and it really shows in the software they create. In fact I have a term for it, which is graduate student ware, and it sort of works for them, but then when they try to put it out for other people, there aren’t… Been a few that have been brilliant, there certainly have been a few.
Some of my friends at George Mason have put out the MASON Agent-Based modeling system, for instance, which is really an incredible set of tools. University of Chicago with their star logo, so occasionally graduate student ware turns into really good software, but usually it doesn’t. And the idea of putting more money into professionalizing scientific software, which is the way I read Open Source Institute, strikes me as a great idea. Maybe say a little bit more about that.
Michael: Yeah, it’s partially putting more money into it. There’s actually… There’s a few different initiatives which are putting more money into it, but there’s still not… The basic unit of work is still the paper, even for the people doing that, and you can just imagine switching to an economy where the basic unit of work is in fact the software package, and that would change a lot of things. You think about the way certain art institutes, they’re not about… You don’t get a job at those art institutes by showing the 50 wonderful high impact papers you wrote. You get it by actually showing the portfolio of artistic work that you do. If you take seriously the idea that software is actually an important expression of research ideas, then that ought to be the fundamental thing sometimes. You don’t have these stupid fig leaf papers that are being written, which then makes people actually… They waste all their time writing these papers. It becomes actually the important thing rather than focusing on the real thing. A really good example actually is, which some listeners might be familiar with, is Jupiter Notebook.
Jim: Yeah. Yeah.
Michael: Which is this incredibly important piece of software now, just this basic operating environment for lots of data science, lots of AI work, and tons of other things. God, the amount of scientific work that’s done there is just incredible. That started by Fernando Perez at Berkeley, but I think Fernando had a lot of trouble really just caused by the fact that fundamentally the university and his academic colleagues wanted him to be writing papers about it, not making the software better, and so because he is very persistent and very hardworking, it all worked out, but… I don’t know. That’s bad news, that you want the thing itself, Jupiter, to be… The quality of that to actually be the objective evaluation, so the proposal for the Open Source Institute is really just to start to set up a research institute or research institutes where the basic unit of evaluation is software, it’s not the paper.
Jim: And I’d suggest that an additional boost that will come from that is it will draw different kinds of people in. It’ll bring in really good software engineers who also are interested and passionate about scientists but actually aren’t research science type people. And such people exist, I know for sure that they exist, and that would upgrade the whole ecosystem by having an honorable home where they get great status for creating, same way somebody gets a paper in nature, someone gets a-
Michael: Absolutely, absolutely.
Jim: … paper that’s being broadly used and broadly cited, for instance, would be a wonderful thing. Last one before we move on from your list and then people who are interested reading in the paper, there’s more in this first list, and this is one that I hadn’t thought of but I thought it’s really interesting, is at-the-bench fellowships, tell us about that.
Michael: Actually, just… James Phillips pointed this out to me, that at the… What is it? The Laboratory of Molecular Bioscience, which is this incredibly famous molecular biology place in Cambridge. Most work, certainly in most American universities, is done by grad students. The old joke is that PIs are often machines for turning coffee into grant applications, they’re not actually doing the work. And if the work is very simple and doesn’t require a lot of training, that maybe makes sense, but if there’s a lot of returns to increased expertise.
Then actually you sometimes want the people who’ve been doing it for 20 years to actually be at the bench, so we just propose essentially a fellowship to support senior scientists who want to go back and just do their own thing, maybe in collaboration with two or three other people, which was the Bell Labs model, it’s the LMB model, and there’s a couple of other famous places where that was true. Somebody like Sydney Brenner, classic at the LMB, did his noble prize-winning work with two or three other people rather than running a lab with 30 grad students where he didn’t really know the details of anything that was going on. And I think that kind of deep expertise, it just enables you to do things that are impossible in any other way.
Jim: And I was lucky enough to tour Bell Labs in it’s still glory days and it was amazing. These really famous guys actually did work. What a concept, right?
Michael: Yeah. Yeah, for sure.
Jim: Yeah, and the other place where that still does exist is in some of the business labs. For instance, Microsoft Research very much works that way. Small teams of people working for sometimes years on really important projects, and so it does exist, but not much in academia, and it would be great to be able to bring that back.
Michael: Jim, a place where it does in academia, which is I think really interesting, is mathematics. Not everywhere in mathematics, but a lot of mathematicians, they still spend 30, 50, sometimes 80 hours a week actually just doing the work, and I think mathematics is in really great shape, partially as a result of that. But the top mathematicians just are astoundingly good.
Jim: Sit there, make new proofs. I know very little about the academic side of mathematics, but that does make sense. Now let’s move back to the conceptual. One of the things that you guys talk about is a way of thinking about science funders as detectors and predictors. Talk about that model a little bit.
Michael: I really wanted to provide an abstract way of thinking about what funders do, partially just to stimulate people to think of other ideas. And so all we’re saying is the obvious point that what you’re doing as a funder, really as any kind of institution, is you’re trying to build a detector, so you find signals of promise out there somewhere in the world in the scientific community, or outside the scientific community, which you then try and amplify. But you also, you’re making predictions fundamentally about the future. You can’t know in advance how things are going to come out. If you did, it’s not really research, but there are many different inference methods that you can use potentially, and just wanted to draw attention. Yeah, mentioned that very simple example before, moving from a consensus-based model of funding to a variance-based model, but there’s a lot of other ways you could do that.
You could start to package up multiple different grants where, in fact, maybe you require something like they’re actually working on opposite hypotheses. You’re using one to de-risk the other, or the whole package to do so kind of de-risking. That’s another kind of change in inference method, and there’s a lot of ways one can potentially do that kind of thing, so just wanted to draw a little bit of attention to those two variables, so to speak. Knobs, in your previous way of thinking. The knobs that can be turned or dials that can be turned in slightly different ways, rather than just using the existing system, which is basically, it’s consensus-based peer review pretty much everybody does.
Jim: Yeah, and you gave some… Again, to try to bring it back to the tangible for our audience, they sometimes get a little tired of our pinheaded yammering about theory, you gave some good examples. What I really liked was on the detector side. This would really force people to turn their detector sensors up to full blast, is the idea of endowed professorships by 25. When I moved from business to science governance, one of the things that struck me was how long it takes these days for people to get established in their career and to become a tenured professor now. Really hard to do it much earlier than 41 or 42. Holy goalie. Back at my entrepreneurial days, I burned through about seven companies in that time, and so 41 or 42 before you’re solidly arranged, and yet we look back at the history of science and when people did what, and most of the best discoveries were at a much earlier age, so talk a little bit about this seemingly insane idea of endowed professorships for 25-year-olds.
Michael: I don’t think it’s particularly insane at all. It’s just giving… Exactly to your point, there’s no good a priority reason to wait until somebody’s 35, 40, 45 to give them tenure. It’s I think certainly true that you’ll make some mistakes if you’re trying to find 23, 24-year-olds, but you’re also going to prevent a lot of people from leaving who get a little bit bored. A lot of my friends in Silicon Valley, they wanted to have research careers, and yet they figured out when they were 19, 22, 23 that they were going to have to wait until they were 40, they could go off and start a company, so they left.
And that’s in some cases just a tremendous… It’s a tremendous pity for research. Maybe it’s a… I don’t know. It’s hard to know whether it’s good for the world or not. Maybe they should have just started companies instead, but I don’t think it’s such a crazy idea to give people their independence at that age and see what they can do. Maybe they just do systematically something a little bit different. Einstein was 26 in his miracle year when he made, depending on how you count, maybe four or five Nobel Prize-worthy discoveries. Not bad.
Jim: Yeah, at least four. At least four. Not bad for a [inaudible 00:34:07].
Michael: At least four.
Jim: At least four. For a guy who wasn’t even in academia at the time.
Michael: Couldn’t get an academic job, yeah. Indeed. Doing it as a hobby on the side.
Jim: On the side. He did at night, or actually while he was not doing his day job, apparently, but of course it would require some development of some new skills in the funders to essentially detect these people. I don’t think that’s an impossible task. I think there are people that could do that.
Michael: I don’t think so. I don’t think it’s really that difficult at all. You’d certainly be making some guesses, you’re not going to be based on somebody’s 15 tear track record at that age. But you’re always guessing anyway. Gosh, there’s a fair number of people who, tenured professors, in fact in some cases there’s a few Nobel Prize winners who did their Nobel Prize-winning work as graduate students and then pretty much did nothing else ever after. And of course they got the tenured job because everybody knew that they’d done this amazing thing, but past results are not a guarantee of future performance, as they say, so I’d certainly like to see the experiment done. In fact, it has been done a little bit, but never fully that tenure by 25 model.
Jim: Yeah, I think that’s a very interesting one. Now, on the predictor type activities, this was, again, this is something that was completely fresh idea to me and hadn’t ever actually even focused on the fact that it’s a real thing, but it obviously is, and that’s what you call the illicit, the secret thesis. Tell us about that. That’s really clever.
Michael: Yeah. Yeah, this is… I don’t think it’s happened to me, but I’ve heard a number of scientists complain sometimes literally directly to the funder that they don’t want to tell them about their project or they don’t want to tell them… They know something special that they don’t necessarily want to disclose, because it’s going to… If it’s disclosed in the grant application, it will be known then by the competitor who does the review, and if your idea is good enough, sometimes that’s the last thing in the world you want to disclose. I know I’ve sometimes had that where I’ll think I’ll have some special technique or something for a problem, something a little bit cagey in the grant application, you want to claim that you’ve got it but you don’t actually really want to give the details. And this is crazy, because of course this is the most important thing.
Honestly, if you would just give the five line version of the grant application, you would go in the five lines in sane world, but instead you’re in this situation where you don’t want to disclose it to your competitors, and so you leave it out or you wink at it a little bit. And so we just make this proposal that says, “Look, there should a short extra box on some grant applications where you could just disclose any extra information that you certainly don’t want to be shared with referees or whatnot, which the program manager can at their discretion decide actually to include in the evaluation.” Now, sometimes people are just going to say, “Oh, there’s nothing.” But other times I expect it would actually be the decisive element. Again, we’ll see.
Jim: Yeah, that’s like, yeah, let’s apply Riemannian non-euclidean geometry to space time and see what we get here. Something like that.
Michael: It worked out all right. Actually, it’s a great example because people will sometimes say, “Oh, David Hilbert actually discovered the field equations of general relativity before Einstein.” Maybe he did it, but what those people don’t know is that Einstein had actually given lectures at Goettingen with Hilbert in the audience where he made exactly this point. And it’s funny you say it, it’s actually an example where Hilbert leaned quite heavily on having learned the secret thesis in advance. Anyway.
Jim: Yeah, that’s very, very cool. It’s a good example why this would be a good institutional improvement. Now, as we’re moving on in the paper, you talk about metascience as an imaginative design practice. Talk about that a little bit, then the kind of people who you think might be good at it, and who might not be good at it.
Michael: Yeah. I gave this… Your request, this example of the invention of insurance back in Genua in the 14th century, and that didn’t rely on making better ships or better sailors or any of these kinds of things, or better judgements about what expeditions to do. Instead, you introduced this new kind of abstraction that was the idea of risk, the idea of insurance. All these things were extra things in the world, so that’s an active imagination to come up with those kinds of ideas, and something like, in a very modest way, something like the anti-portfolio or the Century Grant Program or things like this. You’re introducing new abstractions for intellectual dark matter, which hopefully can be used to drive systemic changes in what kind of work is being done, so we’re just pointing out that really it is an imaginative thing.
You need to invent these new abstractions, fundamentally. It’s not just something where it’s going to be handed to you on a platter. I should say, I don’t think that those ideas are necessarily quite as imaginative as maritime insurance must have been in its day, but that’s the kind of thing that we’re trying to gesture at, but it’s not just… The reason we point this out and make a big deal out of it is because it’s not the way that scientists tend to think about the world. They’re much more interested, typically, I say as a former theoretical physicist, they’re much more interested typically in taking extant phenomena and trying to get to the bottom of them, maybe pushing things a little bit, like questions like what happens if you increase the pressure or the temperature or you drop the temperature?
Those kinds of questions are very interesting, but the idea of trying to create entirely new abstractions, design abstractions is not something that you necessarily get a whole lot of training in. It’s more the kind of thing which programmers do. It’s certainly the kind of thing which designers do, and we think that that’s what ultimately a lot of metascientists will do. They’ll be the person that thinks up clever new ways of insurance, or things like that. It’s a very long answer, but it’s hard to-
Jim: It’s a good answer, an important answer. And particularly you pointed out, hey, people that drive cars don’t necessarily know how to design cars, so-
Michael: No, that’s exactly right. Yeah.
Jim: And actually I took that, I stopped and paused for a minute when I read that part of the paper and I said, “Hmm. I wonder who might be good at it.” And I came up with this very provisional answer, which is, if one were to create a new discipline of metascience, perhaps its home ought to be in the business school.
Michael: I think it’s pretty close to what a lot of finance people do. There’s certainly people thinking about how to find new signals of promise and then how to find good ways of doing predictive reasoning, and they invent new instruments all the time, so they’re designing these new abstractions. I think it’s pretty close to what finance professionals do in some ways, at least.
Jim: And not just finance, but I also think business administration. They think of institutions and structures and signaling and information flows and decision-making. My alma mater is actually the Sloan School of Business at MIT and it was a pretty hardnosed place.
Michael: Oh, really? Yeah.
Jim: And I could imagine… It’s a bit of a stretch, it would require some creative new generative thinking, but I could imagine a metascience department there.
Michael: It’s funny actually that you’re saying that. One of the people… Well, actually, several of the people I most admire, particularly Pierre Azoulay, who you mention later on, that’s where he is. Exactly there.
Jim: Interesting. Yeah, it’s just a thought that we toss out there.
Michael: Yeah.
Jim: Right, let’s move on to the next part of your paper, part two, the decentralized improvement of the social process of science. Bottlenecks inhibiting decentralized improvement. Talk a little bit about some of these barriers that are out there.
Michael: If you ever [inaudible 00:42:20] I take up a new job, Jim, I think you should maybe go into audiobook reading or something like that. You’ve got a nice sonorous tone. Yeah, this is just the point that existing institutions are not… They’re not designed to change. It’s very much the situation of… You think about Blockbuster video. They didn’t turn into Netflix and there was no institutional way by which they could change in that kind of way if, certainly if people inside Blockbuster had been interested in doing that kind of thing, and I don’t think their performance reviews would’ve been too good. We run down… I won’t try and give you a list of all the bottlenecks, but this basic point that a huge amount of change is just bottle-necked through these enormous centralized institutions.
You have three or four funders responsible for a massive fraction of all of the funding. The NIH alone is almost half of the funding for basic science in North America. How do they change? Well, you can’t convince the director to make the changes, or they only have very limited bandwidth. Maybe they can make one or two big changes a year. I think that’s actually probably not even possible, but it’s a heck of a bottleneck. I don’t know. I’m getting a bit lost here, to be honest.
Jim: Yeah. Yeah, it seems like that those are certainly true, and you talk about that there have been experiments at smaller scales or institutions that are quite different. Even our own Santa Fe Institute is quite different, right?
Michael: Yeah, yeah. Yeah, yeah.
Jim: We pride ourselves on Think Different, as Steve Jobs might have said, and we’ve been fairly successful, but we’re tiny, or 12 full-time faculty people. Maybe the secret is why we call it meta, metascience at the top. Maybe a dictate from the Congress that 10% of funding will be used to start new funders. Startup funders. You talk about that a little bit. The idea of startup funders.
Michael: Yeah, that… Actually, not just startup but also funders with a sunset clause, where forcibly retired. If you want to get new institutional ideas, maybe turning things over occasionally is a good idea, so you don’t just have the same old, same old happening year after year, or you’re not stuck with the existing institutional processes. I don’t know. There’s actually, there’s a lot in your question. It’s pretty hard to had to get a through line on it.
Jim: Yeah. Well, let’s see. Who [inaudible 00:44:53] could we go with on that? One of the things you talked about is, again, some of the other, the frozen accidents, the Shanghai Rankings for instance.
Michael: Yeah. Oh, yeah. Yeah.
Jim: Yeah, so essentially zippo in the way of turnover. They make the Soviet Union Parliament look dynamic in comparison.
Michael: And this is just a funny thing I guess I just happened to notice a few years ago, which was there’s a few world rankings of research universities. The oldest is the Shanghai Rankings, the oldest I know of anyway. And if you look at the top 10 in that rankings, they basically haven’t changed since they started, since 2003. I think there has been one organization has moved out of the top 10 being replaced by another, but otherwise there’s just been this tiny little bits of shuffling around in the top 10. And when I tell this to scientists, they will say, “Well, of course you would expect Harvard is going to stay in there all the time. That’s where all the best people are.” And it’s kind taking for granted the idea that the best people would never ever move. Now maybe in an academic context, that seems to be true, there is this incredible stasis. Of course, in lots of other parts of the world, things are very dynamic and they move around a lot.
I also, I just for comparison looked at the largest companies in the NASDAQ. Of course, the NASDAQ in 2003 and the NASDAQ in 2022 looked… They’re not completely different, but there’s a lot of difference. Facebook, which is one of the top 10 companies now, didn’t exist in 2003, and a bunch of the other companies which are now top 10 companies were certainly a lot smaller then, and in some cases weren’t public yet. That’s a situation where you got this tremendous dynamism and things winking in and out, and so you get more institutional innovation as compared to the very static kind of a situation in academia, where it’s the same institutions year after year. And they have a whole lot of incentive to change at the level of how they’re organized. They’re certainly changing the science they do, because they’ve got mechanisms to do that, but they’re not changing the way the institution itself is organized in dramatic ways.
Jim: Yeah, as we talked about earlier, the depth of problem with the PI graduate student model is [inaudible 00:47:03].
Michael: Yeah, exactly.
Jim: … if you don’t change that, you’re picking fruit from six foot trees, right? That’s unfair, my scientist friends, but there is something to that. It’s not designed really to attack the big problems. You have to have other institutional structures to do that. Well, let’s move on now to some examples of where metascience type ideas have been coming into existence a little bit. You talked quite a bit about the rather famous Open Science Collaboration in social psychology, and in fact, one of my earlier podcasts was with Brian Nosek, who was the leader of that project back in EP 12, so why don’t you tell us a little about Open Science Collaboration, the context of where the discipline had gotten to, and what it might mean for improving science?
Michael: Yeah. Well, this is an incredible example, in my opinion, where you’re really actually starting to see massive change in processes inside a pretty large chunk of science, so that… The basic context is, for 40 or 50 years, people doing work in psychology have been pretty skeptical of some of the basic statistical methods that they use. They’re very susceptible to things like P-hacking and the file drawer effect and all these other things. And this has been known for a very, very long time. It was pretty hard to change institutions to do anything about it, and so what the Center for Open Science has figured out, I think very cleverly, a way of getting some effective institutional change.
How that happens is a long story, but that’s, I think the top level thing is they found a way of doing it and they found a way of doing it, very importantly, from my point of view, as an outsider. They didn’t need to actually be running all of the institutions. They didn’t need to get everybody’s consensus in advance. In fact, it was rather unpopular with a lot of people in advance, but they were able to get the changes really rolling anyway, so it’s example of decentralized change.
Jim: Yeah, it was quite interesting. Brian was a tenured professor at UVA, in one of the very top psychology departments in the country, and he couldn’t get it to happen at UVA. I think he probably would’ve liked to have, less personal risk, and he ended up taking quite a considerable risk in jumping out into the void and setting up the center for the open science as an independent organization. And as you point out, he might have… Maybe in retrospect that was a good thing, because it was not popular amongst his colleagues, at least a lot of them initially. Quite interesting.
Michael: Now, he’s got called, was it a methodological terrorist in, I think in the New York Times.
Jim: Well, sometimes you gots to do what you gots to do. One of the things that they have created, which I find very interesting, and it’s starting to spread, more and more journals every year, is the idea of preregistration. I think that’s… It seems like a little subtle thing, but it’s really quite profound. Maybe you could tell our audience what preregistration is and how it differs, and how it undermines some of the flaws that you talked about, like P-hacking and desk drawer outcomes, etc.
Michael: It does sound like a very simple obvious thing. You describe what experiment you’re going to do in advance and what analysis you’re going to do in advance, and in particular what questions you’re trying to answer in advance. You register that on our website. You actually potentially, with registered reports, can get that refereed by a journal who decide, without seeing your results, whether or not the paper is going to be published, and then you go off and you actually take the data. Now, that sounds crazy. In fact, it sounds completely unimportant, why would you do this?
And of course the reason why is it prevents people from futzing around too much with the data. You take data intending to answer one question, turns out you got a completely boring answer, it’s not the thing you wanted, so instead you run 67 different analyses trying to answer different questions, and lo and behold you discover some completely different effect inside the data and you publish that. And the problem with doing that of course, is that if you try enough different analyses, you will find something, but mostly what you’ll actually find is just coincidental effects inside the-
Jim: Yeah, spurious correlations, as we like to say.
Michael: Spurious correlations. What is it? There’s a correlation I think between unemployment in Greece and the rise of Facebook. Were these two things causing one another or are they caused simply by people asking enough different questions that you start to see this? And you know the answer to that question. And by doing this kind of preregistration, you rule out the possibility of doing that kind of data dredging after the fact. And they’ve done a great job of actually getting quite a few people to adopt this practice, and importantly, quite a few journals to adopt this practice.
Jim: Yeah, that’s, to my mind, the real secret. If you get the journals to agree to commit to publish the result, of course that means they also have to agree to publish negative findings, which is quite rare in the journal space.
Michael: That’s right. That’s right. Yeah. And then what you really want ultimately, of course, is for your peers to actually take it much more seriously if it’s been done that way, so not just to pay a lot of attention and give tenure to people who’ve obtained sexy results using pretty low quality methodology, but actually to people who’ve looked for the really high quality evidence. And that’s also starting to happen. Actually, one of the things, when I first heard about this, I thought, “It’s not that important.” And one of the things that changed my mind about it was starting to hear, talking to individual… It was individual psychologists and individual neuroscientists actually who would say things like, “I don’t believe my own papers before 2016.”
And that’s a shocking statement. They were discounting a large fraction of their prior professional work, and it was just because they’d… They’d written those papers in good faith, they thought they were doing outstanding scientific work, and now they don’t believe the statistical methods that they were using. And that was pretty shocking to me, and it starts to, I think point at the cultural shift that’s happening there.
Jim: Yeah, and it is interesting and quite profound. I read that quote, I go, “Whoa, that’s a sign of a real cognitive change. A change in how people are framing their whole careers.” Which really doesn’t happen too often. Who was it that said, was it… One of the 19th century German guys. Science progresses one funeral at a time.
Michael: Planck.
Jim: Planck. That’s right. It was Planck?
Michael: Yeah.
Jim: But when you don’t have to wait for that, that’s a sign that you’ve really done something important. Another thing that you guys mentioned, I know Center for Open Science has been working on this as well, is to strongly encourage people to publish their data and methods. Again, makes it a lot harder to do spurious correlations and other forms of P-hacking if the data’s there, and combine that with preregistration, and it really does a good job of say, cleaning up the method of science quite a bit.
Michael: Yeah, yeah. Actually, I know… It’s always a little terrifying to publish your data or code. It’s funny that the. I don’t know. The final few hours before hitting essentially send, I know I personally find it a little bit of panic. It’s like, “Did I…” Up to that point I was sure, but then when you’re going to make it available to everybody else, all of a sudden, am I sure I’m sure?”
Jim: Yeah, [inaudible 00:54:26].
Michael: That’s the right kind of pressure to have.
Jim: Yeah, it is. I know one thing, I always… It makes me nervous, is when I publish software to go along with something, and I am a talented one-man software writer, but I do so by horrible bad practices, no documentation, crazed structures, and I go, “Anyone who’s going to look at this code carefully is going to think I’m a deranged individual.” And eventually I just said, “I don’t care. It works.” And they can walk through it, but I would really not want to spend my time prettifying my code just in support of a paper, as an example. Perhaps a perverse pressure that some people at least might feel.
Let’s moving on to the next, maybe we’ll call this the final one, we’re coming up on our time here pretty close, which is that you take the example of Nosek and Center of Open Science and generalize that to the concept of a metascience entrepreneur. And you lay out a process by which metascience entrepreneurs, and basically anyone that wants to practice metascience, might think about. Maybe talk about the idea of a metascience entrepreneur, and if you still remember it, the little flow chart diagram that you had in the paper.
Michael: Sure. The top level takeaway is, for one thing, there is no name for this. There is no name for what people like Nosek are doing, this idea of trying to create scalable improvements in the social processes of science. We just felt that this needed a name, and actually it’s resonated a surprising amount. I’m not actually fond of the term metascience entrepreneur, but there definitely needs to be a name for this concept, that kind of person. Partially just for very banal reasons, you want a community of practice, you want people to be able to meet each other, you want people to be able to mentor each other, those kinds of things.
And then I think if you look at the actual changes which have been wrought in science in the last 20, 30, 50 years, so things like the rise in pre-prints, the rise in open access, the rise in open data, these changed methods in psychology, which are now spreading to other areas, they’ve almost always been driven by single individuals or very small groups of people. Typically, very often acting outside an academic context, coming in and just working very, very hard to drive change in some way or another. We just wanted to distill out that and to try and find a few of the key patterns of metascience entrepreneurship to act as a bit of a template for other people.
Jim: Yeah, cool. Yeah, one of the examples you give, which has made again a quite substantial impact in some disciplines less than others, is arXiv.
Michael: Yeah. Yeah, so that’s Paul Ginsberg, who was, I guess a staff scientist at Los Alamos National Laboratory who just realized that it’d be pretty good for initially high energy physicists to be able to publish their pre-prints to… Actually, I think initially it was email, but it quickly moved to the web, which was brand new at the time, 1991, 1992, and he wasn’t exactly supported in this endeavor all that well by his peers at the lab at the time. In fact, he got a negative performance review and left, but he’s changed science dramatically. Certainly he’s changed physics incredibly.
That’s where you publish your paper first now, once you’re done, months in advance of the journal publication, or years. And it spread to a lot of other… Spread to a lot of other fields. Again, for the money invested… I don’t know what the budget is now. It’s got to be on the order of a million dollars a year. They’re supporting the key publishing infrastructure for, I don’t know what percentage of science, 10% or something like that. It’s just an incredible return on the dollar, and it’s not just that return, it’s also the change in working practices. The fact that it is possible for everybody to see in near real-time the advances that other people are making is a really big change, and it’s just wrought by a very small group of people.
Jim: And an interesting after the fact impact is essentially all machine learning, and closely related, artificial intelligence technology, is posted on arXiv, 100s of papers a day, right?
Michael: Yeah.
Jim: And I think one could arguably say that that field would’ve looked quite different if it had emerged, let’s call it the post-Hinton, 2008, 2009 version of deep learning and AI. The field would’ve organized itself differently and probably progressed much more slowly if they weren’t able to accept arXiv. Does that seem reasonable to you?
Michael: Seems pretty plausible to me. It’s funny, there’s like a… It gives and it takes away. The give is it enables everything to go much faster. The take away is I talk to my machine learning friends and they’re drowning under this tidal wave of papers. They can’t even read the abstracts, nevermind the papers.
Jim: Yeah, yeah. And what’s evolved are curators on Twitter who read all or part of the flow and then pull forth the ones that they think are worthy, and some of the people are pretty good. Carlos Perez is a really, really good… He doesn’t spend as much of his time as he used to. He used to seem like it’s full-time job, reading the flow into arXiv ML, and pulling out four or five a day, and he said, “This is worth looking at.”
Michael: Yeah, yeah. Actually, and other people have tools on top of it. I know Andrej Karpathy has this arXiv-sanity thing, which is partially he created to deal with… That’s a tool, it’s kind of an overlay on top of arXiv, created to deal with it. But I don’t think anybody’s fully got it under… Nobody seems all that happy with the way things are at the moment, but [inaudible 01:00:17].
Jim: Metascience entrepreneur, come on, somebody step forward-
Michael: Exactly. Exactly.
Jim: … and let’s add an extra level on top of this. Well, I think we’re about up to our time here. Michael, Michael Nielsen on metascience. A really interesting discussion today, and now, where’s the name of your website? It was michaelnotebook.com, I think it was, right?
Michael: That’s right. Yep. Yep, [inaudible 01:00:40].
Jim: Place to learn more about Michael and his work, and oh, by the way, if you’re interested in quantum computing, quantum mechanics and AI, he writes on those topics as well, so thank you Michael, for a really good show today.
Michael: Thanks so much, Jim. I really enjoyed this.