The following is a rough transcript which has not been revised by Vanishing Gradients or Jim Savage. Please check with us before using any quotations from this transcript. Thank you. hugo bowne-anderson Hey there, Jim, and welcome to the show. Jim Savage G'day! How are you? hugo bowne-anderson I am very well yourself? So we've been talking data for years, and we've never done so in public. So I'm super, super excited for that. Some of our best data conversations have been over dinner and in bars connected to laundromats and such things. And I actually hope later on to talk about the role of conviviality and sharing food, in business and in data science, as well. But before we get into all of that, I'd love to hear about your trajectory. You're director of data science at Schmidt futures now, and I'd love to hear kind of where you started and your path to where you are now, Jim Savage sure thing, so I'll give you a fairly short version. I started out as an economics undergrad at Latrobe University in Australia, it's a fairly small university, and got really lucky in having a couple of, you know, pretty great mentors, and, you know, honors advisors and those sort of things, who really pushed me to develop these economic modeling skills, I was able to land a job at the Australian treasury, when I was fresh out of undergrad, and got placed in this this really funny team. It was the macro econometric modeling unit. And my job for the first year out of school was translating these large macroeconomic models from one language, it's called tsp into another weird language called EViews. And I'd have to like, just line up two editors and translate the code. So I'm not sure if that science was really a thing back then. But I kept on, you know, opening these file, hugo bowne-anderson I had several early projects where I was translating code from MATLAB to Python, as well. And I felt like it was almost..David Graeber has a book called bullshit jobs. Duct Taping, essentially, which is a job that comes from software originally, but being the Kluge in many ways, or the blue tack, but it's important, Jim Savage it's a great experience, I had this this guy, Michael Cooper, it says, Treasury who sat over my shoulder for, you know, I don't know, first three months of my job, and just like, walked me through how to code. I never coded anything before my life. I'd done models, you could solve with paper and pen on a chalkboard or something, you know, you've ever been on shortbow. But it was such a great experience and learning to actually take some model and fill it with data. But there was this. Yeah. At the top of all the code, there was this, you know, the copyright and who wrote it, and when they wrote it, and I think there was this guy whose code I was translating Anthony Goldbloom, I'm not sure if you've heard of him, he started a company called, hugo bowne-anderson Kaggle. He's a fellow Australian. Jim Savage Yeah. And it's about that time that he, I think, was at the reserve bank. And people were talking about this guy having left to do this startup thing in, in data science. And I'd never really heard the term. And it was kind of interesting to me. And so that's kind of how I heard about it. But it took me years and years before I thought of myself as being a data scientist, I left the Treasury after burning out completely on the carbon tax work that we did in 2011. moved to Mexico, and I was going to enjoy early, early retirement, I think I was 26 or something. And so moved to Mexico took salsa lessons and spent eight hours a day, just going through I think I was like, Andy Gelman, Jennifer Hills book, and learning , reading a paper a day and kind of coating them up and just just playing for one coating at the time, all and that was in R and, and, you know, eventually ran out of money, moved back to Australia, and landed this really cool job at the Grattan Institute, which is a think tank, mostly economic policy, based in Melbourne, but it's on the University of Melbourne campus. And it was just a really fun experience. So it was fun, because like it was essentially the first data scientist they've hired. They didn't call us that back then. But they had you know, we had a speechwriter, who had previously been the prime minister, speech writer on staff we had like PR people who really understood how do you take research and put it on the front page of a newspaper? We had these program directors, who knew how do you spin a proposal so that you can get into a minister's office and sell to them? You know, how do you actually have influence with something you don't have an organization like the Treasury Department, where your influence is by law, you don't have to earn influence? When you're a small think tank over you know, 25, 30 people, you need to go and work out who are the people, you're trying to influence? What influences other people and are you doing that? And so that that was a really great kind of Less than a somewhere in there because one of our sponsors, sponsors for the federal government, University of Melbourne, national australia bank, I think it was Google, and a few other orgs. Because of Google's sponsor, they sent us over this announcement that there was a program called the Eric and Wendy Schmidt data science for social good fellowship at the University of Chicago. And so I applied, and I got it. And so it took, you know, I think I had a baby at that stage was like 11 months, and a wife who didn't really want to go to Chicago for three months. And so we took her went to Chicago and had had a great time where I met all these like fantastically sharp, American data scientists. hugo bowne-anderson What time of year did you go? Did you take your family to Chicago? Please don't say December, Jim Savage no, no, this is summer. And you don't want to go there in February. So we had this, this fantastic experience, it was a wonderful experience, they currently have applications open. And if you release this podcast well enough, people should definitely apply. hugo bowne-anderson I'll include a link in the show notes as well. Jim Savage Great. So we went back to Melbourne and just like, really felt that my place was in the United States, I almost immediately kind of started trying to make plans to move back. So I started interviewing with a couple of American firms and found this is wonderful little startup, it was like two guys. And they had this idea that because of mobile payments in East Africa, being used to pay for solar panels in these kind of police to purchase agreements. So how they work is you get these little solar panels, and you stick them on your, on your pretty small house. And you can unlock the panel by texting money. And so you take some, some shillings to the provider, they send you back a code, you punch the code into solar panel, it unlocks for seven days or a month or something. And it's his way of providing credit for solar panels in areas where there is no electricity that had this interesting aspect. Because those companies that are selling these solar panels doing so on credit, it's very capital intensive, they need to import the solar panels into Kenya or Uganda or Tanzania or wherever else they up. And then they only receive the money over some time. So if they need to grow, if they want to provide more electricity to more people, they would need capital. And so the insight of Daniel Goldfarb, who was the Monaco founders, was that he could use this secondary source of information from the mobile payment providers to verify that the financials that we were getting from the solar companies were indeed correct, not falsified. So you could you could go and get the ground truth data, look at what the payments had been taken to their financial records, and say, Oh, this, this company is legit. You can also build a whole bunch of models on top of that data to work out, hey, you've originated, you've loaned that 1000 solar panels, how much are you going to receive by when, and you could use that information to create financial products and make loans that were way way cheaper for originators of solar panels. It turns out that solar panels is a pretty good industry, it grows very quickly. But the model that that the team discovered, really scaled beyond that. And so they started to go into motorcycle taxis and agriculture finance and small business financing. So things that never I don't think they're in non countries throughout Southeast Asia and Africa. With I think I've hit $160 million out the door. Now they're real up. hugo bowne-anderson And this is called Lendable, right? Jim Savage They're also hiring an econometrician right now. So if you know any econometricians let me know. So a fantastic company. That was his experience being a real life data scientist. And then about a bit more than three years ago, someone reached out to me, they said, we've, we saw in your CV impact investing, Eric and Wendy Schmidt, data science for social good. This is new philanthropy out outfit, they called Schmidt futures, you might want to talk to him. And I was refugees now three years. And it's the best fit job I've ever had. hugo bowne-anderson And I'm really excited to chat about your work there a bit later. There are there are several things that came up that you mentioned that I think um, be great to sink our teeth into soon. The first thing that we'll get to soon is the role of economics in data science and the tools that have been employed for decades in economics, that I think are woefully missing from data science, including as you know, something like the sophisticated techniques of causal inference used in economics you mentioned Andy Gelman's and Jennifer Hill's book which brings of course causal inference to mind it also brings Bayes to mind. You mentioned R and I think tools such as R kind of opened up The world of data science so, so incredibly, something when we're talking about data for good that came up when talking about, let's say, American companies working in East Africa is how we, what our responsibility is to different stakeholders, and especially cultures that we may not have enough understanding around. Another thing I'll want to dive into, I think, as you mentioned, we've got the ground truth data. And I think ground truth is something that I am personally inherently skeptical of. And it's, it's worth talking a bit more about what happens in the data collection process, and how that can approximate ground truth. And who decides what ground ground truth is, that's just like, all of these things, I think are going to be great, great, um, proverbial meat, or let's say it's 2021 2022. So plant based meat to sink our tea. It's beyond beyond data science. But before we get into all of that, I'm interested in why data and what I mean by this is a bit about me, like, on and off, I'm relatively bullish and bearish, and skeptical and cynical and excited about data science and I understand, like, you know, we should be taking, like, at least 30 day moving averages of how we feel about the industry we work in, you know, in multi dimensional space. But I do ask myself a question more and more like, Why? Why are we doing everything with data? What type of value can data bring? So maybe, just from your personal experience, why do you why do you work with data as opposed to working in other other functions in organizations? Jim Savage Yeah, by accident, I have many mottos. One of them is like, never ask a model, always ask a modeler. So I think about data science as a fantastic training ground for the next generation of leadership. So if you think of the people who bubbled up into senior leadership in the 80s, or the 90s, it was like lawyers and MBAs and economists, they end up running, having their hands on the levers of power. And then through the late 90s, through to now, it's engineers, it's it's people with CS degrees, dropouts from CS programs, who end up with a hands on levers of power, I think what we'll probably see is more people coming up from a data science track. And so I think the real value of data science is, it forces you or enables you to test assumptions very quickly, the person with access to data can have an idea, and quickly whip up a model or whip out a very easy piece of analysis and test their idea and get immediate feedback, rather than having to go on hunches or theory, or anything else. And I think that that is the real value of data science in like a pretty macro sense. hugo bowne-anderson I like the idea of data, data skills bubbling up to the top, I suppose there's a question, do we want people who are very sophisticated data scientists doing that? Or is is the play to have people with significant domain expertise, getting data skills, and then taking that with them. Jim Savage think I've met a huge number of people who have the domain expertise, who just like, one day, go and become data scientists, you you definitely made some. I think there are more curious data scientists who then go and get domain expertise in a topic. I think they're definitely like, economist or political scientists, or physicists or whatever, who are like, they go and study something. And that what used to be the track that you get instead of science. Now there are that as long as programs enable that. But you know, I think that we can, that we should be expecting that a scientist develop domain expertise, because they are curious people right. Now, now, which is not to say that, that basic analysis skills should be universal, everyone on your team should be able to execute a simple SQL query, rather than being blocked by a data scientist who's like on the bus or something, because those are just day to day skills for almost everyone now, but when it comes to who needs to be able to do the end to end data science stuff? I think its data scientists, and that they should be doing it because it's like specialization is is important for everything. hugo bowne-anderson There's also a good question, like how much domain experts should know. So I think one example is if someone in HR is using machine learning in a hiring flow. Now, there's a question of whether this should happen at all. Because of all the bias we've we've seen, I mean, we've seen Amazon scrap their, their tools because of the amount of bias in those but in the case that someone in HR is using this, it's arguable that they definitely need to know about confusion matrices, false positive precision, recall these these types things or maybe we develop, I mean, I joked that confusion matrix, it turns out that that's the best name for this this matrix, but I do think we don't necessarily need them to know about about gradient descent, but they do need to know about the aspects of data science, ML and AI that can impact what they do. So an analogy I like to use is, when you drive a car or a bus, you don't need to know how the internal combustion engine works. But you need to know if you drive at a particular speed into a wall, you're gonna kill everyone. Yeah, does that make sense? Jim Savage You know, I think people learn this by case studies. The best three courses like academic courses have ever taken, both undergrad and grad level have been have had the flavor of their history that goes through, hey, here's some models here, public policy decisions that were made with them. Here's how the models were wrong. Here's how science reacted and improve the model. And I think we can do the same thing with data science, we ought to be like, showing, okay, so, you know, Amazon were using screening algorithms to like, you know, select people, and then they found it was like, really bad. So they moved on. And then they did this, like, there's some response. So I think that like teaching people those methods, you know, stories stick in people's minds, and we should be using stories to kind of like, illustrate all the problems. The other, the other side of it is, the problems are the interesting thing. I think data scientists are a set of people who are specially engaged by challenging problems. And we can talk about this a little bit later. But I'm really interested in how we can really improve ethics and data science, by turning the ethical question into the central research question. hugo bowne-anderson I'm really excited about that. The other thing that comes to mind, when we're talking about these skills bubbling up, and you know, the people who control the levers of power, I think as, as you put it, this concept of executive data science that I've heard you heard you talk about now, there are many ways to slice data science, one way to slice data science is in to descriptive analytics, which can be exploratory as well and include insights and dashboards and that type of stuff predictive, which includes but isn't limited to machine learning. And then prescriptive analytics, which is decision theory telling you how to make decisions, which I'm really excited to come back later, to discuss with you, Jim. But the way one way you slice that space, which I really like, is into kind of tool building and tools for automation. What we want to automate away, then, is something which augments humans in a number of ways of figuring out how to get machine and human intelligence working besides each other. And then above these is what you call executive data science. And I think this ties into our conversation about what leaders... the types of skills they need to have. So maybe you can tell us a bit about what you even mean by executive data science? Jim Savage Sure, yeah, I think it's worth like, really describing the first two as well. So these, at the bottom you've got is like, what is data science as a product. And there are many, many roles, which involve people building some model or something with the fundamental aim of improving recommendation system or making a news feed better, or whatever it is, they've got some model and research and they they'll tweak the model and see whether it, you know, makes people stay online longer, or whatever it is. And there's like an entire industry of that people get paid very, very well to do this, or, you know, too often, it's probably about surfacing really high quality ads to people, which is something I care deeply about, then I think, you know, you've got a lot of tools. And I think a lot of the dashboard world sits in like augmenting people. So dashboards, search, those are things it's like, how can you... decision support, and then there is this, you know, executive data science thing, which doesn't really have a name yet, so I'm just gonna call it exectuive`` data science. And to understand what it is you kind of have to have an appreciation of what what do executives do? You know, and the list is pretty long. It's very difficult to write down an executive job description, but they do things like you know, they allocate resources, they create vision and strategy, they go out and sell the organization to try to attract top talent. They establish internal incentives and hold people accountable to those incentives. They make judgment calls and break ties and adjudicate like interpersonal squabbles, and those sorts of things. I think about the sorts of things that people do in the automation space as almost being like, on off, it's a little level change. If you get, you know, you go what you build the recommendation system, and then like, it just has some level shift on what the firm is doing, what the whatever you doing, it just boosts things and you can generate huge amount of value that way. Enormous amount of value. But it's not like what it's not sort of value that differentiates the The Wealth of Nations like Why are some countries rich and some countries poor by some companies worth a trillion dollars and others fizzle out after a couple of years? I think those sort of things are much more influenced by strategy, by who you attract, what sort of networks you're able to get yourself into, and so on, so on and so forth. And that's why, you know, I think about the trajectory, what's changing the growth rate of the good that your organization is doing. And so I think there are a bunch of ways of using data science for those things is really about like executive as a researcher, or having executive offices that do research into what is the competitive landscape? What are the products are out there? How am I, my competitors going to respond to this? And what available data that we have to really test those assumptions? So at the moment, this happens, because you've got like, really smart kind of McKinsey type people who move into those sorts of roles that are very analytical, they know how to boil down like the most important principle components and think through strategically. But I just think that the data scientists toolkit rather than management consultant toolkit is the one that is probably likely to benefit from those so likely to do well in those sorts of roles in the long run. Now, I think that the big challenge we've got here is that the sharpest data scientists all get soaked into the modeling world. And they kind of poopoo dashboards, they poopoo that sort of descriptive stuff, because it's not hard enough, it's not engaging. And yet it's the most important thing, almost any, any senior executive will tell you that like much rather keep the dashboards and keep the the recommendation systems hugo bowne-anderson Absolutely. Well, there's an impedance mismatch between what people liked doing and a good out of what creates value. And I really want to have that conversation now. So to set the stage, I'll include a link in the show notes, you gave a talk several years ago with the decidedly unclickbaity title of productizing structural models at what was then called Data eng con. And now I think it's called Data Council. And I just actually want to give a shout out to Pete Soderling, who you introduced me to several years ago, and has been one of the most inspiring community builders, as far as I'm concerned in the data space. If you haven't heard of Data Council, check them out. If you have heard of them, check them out even more. Pete creates really wonderful spaces for people to connect. And I've actually met many of my colleagues and close friends through data Council and data eng as well. If you're listening, open invitation to come and join me on the podcast as well. And if you're not listening, I'll find you. So Jim, you actually gave a Class A really nice, like Harvard Business Review two by two in this talk, right? Have I honestly loved it? Of what I just want to check my notes to make sure I get this right. Yeah, the on the x axis was the additional impact of success. So what type of additional value data science activities can create, and on the y axis is the probability of success. So it was framed around prioritizing effort, essentially, what people are good at versus what value they create. So maybe we can talk about both of these? I don't know if you'd like to take them, you know, turn by turn. We've already talked about some things that create value. yeah, and value is a hard one. That was that I generating app. It's been like all over Twitter last couple of weeks. Yep. wambo is one vs. I honestly can't remember the name. Jim Savage it is. So some really smart people have made this tick. And it, it's a lot, it's a lot of people. And maybe it's kind of fun entertainment, maybe they're pitching something really great in the future. But you know, I don't consider things like that, to have an enormous impact in the long run. But so hugo bowne-anderson I will stop you there. And I think we can probably agree to disagree on this. But I do think, um, discussing the value of art is actually inherently subjective. And, you know, these things are multi causal, and they're lagging indicators as well. So we don't have a, you know, if you told me Picasso wasn't wasn't valuable, I'd say, Well, you know, look at the amount of money people pay for his artwork, right? Jim Savage I think with like a data scientist toolkit, helping get monetary policy right, helping change regulation that might keep people poor, helping improve diagnosis of new disease, like I think there are lots of things that have genuinely... there's a lot of apple pie out there. And those things are the things that we ought to be focusing on especially those things. hugo bowne-anderson It's a false dichotomy. My we carry value to diagnosis and this this type of stuff, whereas I think it's just far more an open open space and without art would die spiritual deaths. So that's my only statement there. Jim Savage And you get places where people have had died spiritual deaths, because they kind of sterile in terms of Yeah, yeah. And I live in one of the world's great, you know, artistic cities and enjoy that a lot. So I totally, totally by that this is very valuable stuff to do. I have made choices in my own career to work on topics that I think have are going to have very large, positive externalities, and kind of layout tools to those sorts of topics, hugo bowne-anderson where is the mismatch between what creates value, and what people are good at, or what people can do and enjoy? Jim Savage but I use this kind of example. So Wambo is like, definitely an example where like, it's actually creating a lot of a lot. And we know this, because people like running it, they share it as like. And so it's probably one of these marginal cases, there are other cases of like, a lot of effort that don't create much delight. And those are probably in the kind of low, low additional impact category. And there are these other areas where we probably have the tools to do it well. So there's, like, you know, I think, predictive analytics, for example, or A/B testing online, we've got tools to do this. So that if we're working in a in a deterministic system, predictive models can now basically solve any deterministic system. I mean, they they're playing computer games, they're playing Go and all that stuff, once we've got an open system. hugo bowne-anderson a deterministic system where you may need to know the rules as well, right? Jim Savage But they're pretty good at working out what the rules are. So long as we've done some scoring function yet what what we're not as good at is when we're working with an open system. So that is when the information that the model is using only captures a part of what's going on. And so this is like, you know, forecasting monetary policy, most of the kind of like stuff on websites, and like recommendation systems, it's not, we can change the odds a little bit. But we're not like making perfect predictions or building technologies that will win every single game. So I think that's something that we're fairly good at. And the answer, if you want to improve those sorts of models is to try to collect more data, which has like, possibly positive, possibly negative implications, depending on what's being done with the model. It also has some really big questions. You mentioned ground truth before, the thing that kind of motivates me, at the moment is this question of measurement error? And who does measurement error fall onto? I think there are many, many people. Everyone has a story. We know on Twitter, when somebody there are these threads where it's like, oh, what has been some some time that you've been underestimated? And it's a very emotive topic, people, everyone has this story about when they were underestimated... this grudge they hold for years and years and years. And models will do this all time if we train them on data that has like a measurement error that differs across groups, and that will systematically hurt people. And so I think we have to be pretty careful there-- itss not deterministic systems. And then there are these counterfactual systems like we're, that's causal inference. I think, some data scientists are really, really good at causal inference that -- hugo bowne-anderson I think that's what's a minority. And that's something I want to get to. Jim Savage So there's causal inference, and then there's like structural causal inference, like what happens when I go and change the rules of the game, what happens when I introduce a new product to a market, what happens when I, you know, change strategy in an organization and how its engaging with its competitors. And those sort of rules that when you change, sorry, those sort of bits analysis, when you change the rules of the game, there are people who have those tools, they don't necessarily call themselves data scientists. But I think that is an area where we can have a huge positive impact with toolkit that we've got, if only we like, add a few more tools. hugo bowne-anderson This is super important. And let's be very explicit now. And probably circle back to it later. But the way I want to frame this is that, as you said previously, to me and elsewhere, if you throw a neural net at something at a system in equilibrium, and we'll come back to what that means, because I think that's actually a term that not enough data scientists have a robust understanding of, and then we perturb the system out of equilibrium and try to use that neural net to make inferences, it'll make precisely the wrong inferences, right? Jim Savage Yeah, I have a really close friend back home, who came to me once and said, I've got this data on sales of some product. And my predictive model is doing extremely well for every store we've got, what the price was on every day, how many how many of the products that are sold, and we can fit this model and it's got amazing predictive performance. And it says that when I increase the price of this good, the sales will go up like you've discovered my big problems in econ know that, and why would that be so? Well competitive markets, right, competitive markets, where what you observe that P that p and q, price and quantity are determined in equilibrium. And, you know, the heuristic here would be the store manager know is when something is popular when it's hot, and they'll increase the price. And people also know when it's hot, and they will go and buy more of it. And the price is actually responding. So never reason from a price changes, as economists like to say. And so the whole point of structural modeling is that you interpret the data through the lens of a model that has some concept of equilibrium. And that enables some very powerful types of analysis. For example, what happens if we introduce a new product? What happens if we impose a tax on something? What happens if we add a new store, to the mix and so on? That's like a very powerful set of tools, especially in the field of industrial organization in economics, that we, as data scientists should be using much more. hugo bowne-anderson Absolutely. And it's funny, we, we should leverage the knowledge that all data scientists have in their specific domain. So I don't know about equilibrium from economics. I know from my STEM background in chemistry and biophysics, essentially, right, and cell biology. Jim Savage ecologists also! equilibrium, it's got, like, predator prey models, and there's going to be some equilibrium in that system, and you perturb it, and you never, it's never thinking about things as predictive models, the animals worth worth using. And I, like Kaggle has has been fantastic in its as a training ground for many data scientists. And they've, I think it's had a huge impact on the quality of predictive models. It is it is meant a lot of people consider a predictive model as being the kind of best thing to want to build. And because, you know, science does say that we should be able to make predictions from science from something, but it's prediction in a counterfactual that actually matters for decisions. I want to know what will happen if I'd take decision one or decision two. And you cannot infer that when all the people in your dataset took a decision I did, because it was the right one for them. hugo bowne-anderson exactly. And I'm glad you mentioned Kaggle in that light, because this is something that I hope to come back to time and time again, in this this podcast is not necessarily only Kaggle, but leaderboard style competitions, they have lifted up a whole variety of people to be capable of doing incredible things along with kind of the API's and tools such as scikit-learn and XGboost, you know, and all the PyData and R ecosystem as well. I do think when something has such a massive groundswell, it's worth interrogating what has worked, and what hasn't. So leaderboards, competitions, don't teach you best practices about collecting data, or dealing with data drift, or the data engineering necessary to or experiment trackers and this, this type of stuff, they also I think, um, have taught us to think that, you know, a 10 to the minus six improvement in accuracy, if we, you know, use GPUs or TPUs instead is maybe a good thing. And that may not be a good thing. So these things are worth interrogating. When something a methodology gets this much usage, let's say, I do want to move on. Before that, I just want to recap that the things we're good at down to the things that we're bad at, because this is really important, we're good at prediction, in no one systems where we know or can learn the rules, we're pretty good at predicting in noisy systems, when there's a robust signal, we're not great at causal inference, something I'll throw in here, which I want to get back to, we're not great at uncertainty in decision making. And then we're not great at structural modeling and kind of systems involving lots of different agents, particularly when we're away from equilibrium. Now, some of these things that we are good at add value, like image classification, for medical diagnosis, right. So X ray classification is one. Now, I think when we're talking about businesses, predictive analytics can add value, it definitely does at the top in tech. But there's a question with like small to medium sized businesses where the machine learning does add value over analytics. And I'm saying this because I've seen so many companies try to adopt a machine learning strategy, where it just they don't have the scale that it will actually have a positive ROI. Essentially, I've seen machine learning teams be hired a great cost even a few machine learning engineers and data scientists. It's incredibly expensive for a company. And then I've seen those functions be gutted. So that's something we're thinking thinking more about. Jim Savage I would say, for most startups, the first data scientist, you don't spend all their time trying to scrape data on who the competitors are, who the customers are, what conferences they're going to how they're going to meet them what the networks of people are. I think just like understanding that kind of competitive intelligence stuff I think is way way more important for most startups, than does this subject line work better than that subject line, which that's incredible for building predictive predictive models for versus is hugo bowne-anderson great. And the other thing I mean, you know, people have heard me talk before No, I'd probably babble on about this far too much. But like, the times I've seen startups hire their first data scientist who becomes a data engineer for 12 to 18 months, maybe include that in the job. JD, right. But having someone who can can can do the data engineering stuff, and as excited about it as well, I think is important. So where I want to get to now is, in no uncertain terms, I do think we have a methodology crisis in the industry, I actually think we're kind of in the midst of a series of interlocking crises that I want to kind of elucidate more as throughout this, this podcast, there's a credibility crisis, there's a tooling crisis, there's a lack of standardization crisis, there's a data culture crisis, I think we're probably entering some sort of labor market crisis as well. And maybe I'm being hyperbolic here. But I actually, I think it's worth bringing these things to the forefront. why I'm so excited, among many other reasons of talking to you is I think part of the methodology crisis is that we don't have enough widespread tools around causal inference, decision theory, structural modeling, I want to talk about whether these techniques are necessary or could be helpful for robust data science and how we can spread these tools, more tools and techniques, whether there are tools for them as well. Our mutual friend and colleague Adam Kelleher here will probably argue that it isn't a tooling problem, we actually have have the tools that may be actually a cultural, cultural and norm- based problem and labor market problem as well. But the way I want to set the stage here, is I think, we've been living in a hallucination. What I mean by that is, I don't you may recall, Chris Anderson's piece from 2008 in WIRED called the end of theory, the data deluge makes the scientific method obsolete. Okay. Now, the argument here was in the age of big data, we don't need theory. And and, of course, Chris was being provocative. But the argument was, in the era of big data, we have so much that we don't need to understand things anymore. All we need is to look at the data and it will tell us what to do. Now, I want to set the stage historically, as well. What I mean by that is that this is an argument that actually goes back at least as far as Francis Galton, and Pearson, who essentially developed it, I mean, gotten developed co relation, which we now call correlation, right? And regression to the mean, and these types of things. But what they did was essentially, part of their project was to subsume causality into correlation, saying correlation is enough to describe causality in the world, I'd recommend people, if they're interested in this, look at Judea Pearl's book of y, where he goes into this, I just want to make clear that Judea Pearl has a significant bias, I'm sure he admit that as well. In terms of telling this story, there's also a book called Bernoulli fallacy, which we may get back to by Aubrey Aubrey Clayton, which goes into kind of the historical aspects of thinking about correlation more than causality. It's also worth mentioning that both Galton and Pearson were your Genesis, if I recall correctly, Peterson was the first Galton chair of eugenics at Cambridge, I think, but I may have got the institution wrong. And I'll fact check that after, after the fact. This is actually an I'm not just just saying this in order to appeal to the Twitterati, or something along those lines. My what I really want to do is make clear that the statistical methods that have been developed and that we still use, especially in frequent ism, actually were developed to sciencify, the eugenics movement. And for that reason, these things are worth interrogating as well. Now, all that having been said, there's an incredibly rich history there. My real point is that we don't do causality or causal inference correctly. And there are huge historical forces at play that mean that we don't, so how important is causal inference? And how do we get out of this hole? Man? Jim Savage Yeah, look, I think it's, if not like, the most important problem, then one of them that we as data scientists should be concerned about, in my world, it is what everyone's concerned about. So so when you say, when I keep on hearing this, I got that it's not science, causal inference. I'm like I ever everyone's seen as being called inference these days. I mean, this guy is Nobel Prize to Joshua Angrist. for their work on natural experiments are using observational data and that for experiments to generate causal inferences about policy shifts, hugo bowne-anderson beautiful and let's just give a shout out to Angrist's book on mastering metrics, which if anyone's interested in learning some more causal inference stuff. That's what that's a beautiful book to get you started. Jim Savage Yeah, and just yesterday, Nobel Prize outfit, Nobel Foundation that has dropped the Nobel lecture from all three, they're really really worth watching. They're they're a generalist or like an undergrad level, Introduction to very serious topics. And Josh in particular, I work with Josh on the board of Avella, which is a impact focused startup is very, very committed to exploring education and inequality in education and how we can do things about how can we use these sort of methods to improve education policy. And so I think a lot of the listeners to your podcast would be pretty engaged by, by his Nobel lecture on, for example, do fancy schools actually do much turns out, that's all selection, that treatment. hugo bowne-anderson I will include a link to those in the show notes. What I'm mostly hearing is that you and I have pretty serious selection bias in who we talk to. tell me about the importance of causal causal inference. Jim Savage okay. So causal inference is like, one of the tools that I just think we should all be learning we should all be aware of, you know, my take on this is that most data scientists, hopefully will not stay data scientists forever. I hope many of them go on become entrepreneurs, go and become executives go and become generalists. And take with them this mindset of curiosity, and a recognition that research is really important, and an ability tests options quickly, one of those really big tools is causal inference. And it's almost knowing that it exists is enough for many people. And knowing the sorts of problems that good causal inference can uncover in your data. And uncover in your, your own beliefs is enough. So there's this case with, you know, the famous research by many people associated with Angrist, including Parag Pathak. So they've, they've done all this work, looking at the effects of, you know, elite... these exam schools. And I live in New York City, and all the time on NPR, or in the New York Times, you'll read about the crisis in admissions to some of these fancy schools. So, Stuyvesant, you know, had like seven African American kids in a huge class last year. And because they say, we've got these emission thresholds, and if you're on the other side of it, we'll give you an offer. And if not, then sorry. And not enoughpeople are applying from minority groups and not getting in. And it's like, a big issue. This is such a salient political issue, we spent so much kind of cognitive effort just talking about thinking about it, arguing about it. So these schools have better be really good, they'd better add a huge ton of value to justify the disappointment that we have at people missing out. It turns out, when you consider people who just get into those schools and just go to the next best, there is no additional impact on where they go to college or their test scores. It's all that these schools are really good, yes, that the average kid who goes to these schools is very, very smart. And those friendships, not because this fell off. But in terms of a lot of the observable outcomes that we really care about, like, are we leveling kids up on tests or getting them into college and different rates? The answer is no. And in Chicago, there's some follow up studies there that shows that these super salient exam schools might actually be drawing kids away from really high performing charter system, and sending them to these schools that aren't boosting them as much. So you go and do your exam, you had an offer at a high performing charter school, that would have really helped you, you get the offer to a fancy school where yes, your average classmate's smarter, but it doesn't actually level you off very much. You go to that school, and you do worse than you would have otherwise. And so I think, knowing like you go into the causal inference, or go and learn about causal inference, you go through all these case studies. And you kind of know what to start looking for when you're doing your own analyses. And I think that's the important bit. Same with decision theory. So decision theory, but decision making under uncertainty is a set of tools we've got that really say, well, let's say I don't know, whether an effect is zero, or five, it could be anywhere in there. But I know that if I take the medicine, and the effect is zero it's going to hurt me. Think of this like chemotherapy. And if I take the medicine, and the effect is five, I get a lot better. And I've got some uncertainty about whether it is in zero or five, then what you can do is basically assess how you will feel what is your cost benefit look like at all those points along the uncertainty and take the weighted average according to how uncertain you are each of those points. And you average over those that's called, you know, integrating your loss function over your posterior, abilfoyp, and it tells you what action you should take. It's, you know, there's this great Deirdre McCloskey, she has this wonderful economist, historian and just interesting person, she's got this line in her wonderful essay on the cult of statistical significance, which says, you know, imagine you're, you're playing, you hear someone in the distance yelling. But you know, at the, you can't be sure that she isn't, you know, hearing kelp, kelp over like a heated game of Scrabble? Are you sure that you wouldn't go in and save that person or at least go and check it out? It's like, we know intuitively, that we have this loss function, we would like it is really bad if someone needs help. And we wouldn't run to the rescue, even if they're just saying kelp, because they're playing Scrabble. So we know that we would do that. And not having that baked into our, our use of estimates from models is kind of important. hugo bowne-anderson And I think this is actually worth elucidating and spelling out. With even more clarity, I've actually got her quotation from you in front of me. So I'm gonna read this out, if someone called Help help in a faint voice in the midst of lots of noise, so that at the 1% level of significance, and we're going to come back to what that means. It could be said that she's saying kelp kelp, which arose because she was perhaps in a heated game of scramble, you wouldn't go to rescue her. So what we have there is some sort of, we have a likelihood or probability on one axis, then kind of impact on the other. And we want to, we don't only want to look at the likelihood, if there's a small chance that something really full on is happening. Maybe we we want to go there right now, the reason I want to talk about the 1% level of significance is because what she's essentially calling out there is statistical significance testing, along with baking statistics into the decision making processes. Okay. So this is actually a message I sent you on signal. The other day that I'm going to kind of paraphrase, it looks like a simple example. But it's nuanced because it involves hypothesis testing. And most people, including myself, and a lot of working scientists can't easily intuitively recall what the 1% level of significance means. Okay, I asked you, could we change this example to it's 90% probable that she's saying kelp kelp, but that that actually misses the point. Okay. So what it actually means is, the 1% level of significance is the probability of hearing this hearing Help Help, as sorry, hearing what you hear, assuming that she called Help Help is 1%. So it's a low probability that you'll be embarrassed by a false alarm, essentially, right. But the fact that she's saying maybe saying Help Help is enough to know that perhaps you should actually run and do that. So when you convolve, the loss function with the posterior, which in this case is help help or kelp kelp kind of turns out that you should go and check things out. Is that if I butchered that, or is that pretty reasonable? Jim Savage That's pretty reasonable. as reasonable as any description of uncertainty gets. I think this is there these heuristics? Again, I don't, I don't literally mean you need to go and write out your loss function all the time. I don't literally mean you need to calculate the posterior all the time. But my hobbyhorse is saying you need to you always be integrating your loss function over your posterior. Because it's an exceptional heuristic you need to think about how are you uncertain? And how would you feel if you made this choice at all those points of uncertainty that you have? Because if you're not thinking like that, so and I think like this, this actually motivates a lot of the precaution that was reasonable to take, especially at the beginning of the pandemic, like we had no idea where the R was, like 1.5 or seven, we had no idea where that because people were in hospital for so long, whether the fatality rate was half a percent or 10%. And it's like, in that moment, if you remember, February in 2020, when Yes, about 1% of people had died, but all these people who hadn't been released from hospital yet, and we're like, oh, we have no idea what case fatality was, what should we have done? Well, we should have locked everything down. And if we would have done if we had Ebola. Now once you start to resolve the uncertainty, then you're like, Oh, now we can do the cost benefit analysis, which I think is part of the economy that we should open up which are the types of activities that are fairly low risk, let's like be as reasonable as we can in opening things up. But we got that so wrong, because we don't have this mindset of ablfoyp. And it's because like, you need to think okay, it might have been 10 It could have been 10 are might have been 10 and case fatality rate might have been 10. And that that would have been awful. We would have like decimated huge amount of population, and shutting it down shutting everything down for a few weeks while we work it out would have been exactly the right thing to do. Now we've got lucky that wasn't, but it could have been for a moment. And so and you want to always evaluate the quality of decision based on what information you had then, like, you know, you don't ever say he was really smart, he bought a lottery ticket, and he won. You're like, no, the expected value was was less than $1. Like, that's a weird thing to do. hugo bowne-anderson Exactly. And, and so actually, at the time, I wrote an essay for rally radar, which I'll link to in the show notes called decision making under uncertainty, I think something along those lines and the two key points you made, the latter point you made is the first one I made in the essay, which is you judge the quality of a decision not on an outcome. That's a total logical fallacy, actually, the second is you want to consider likelihood and impact. So once again, it's always be integrating your loss function of your posterior. What we're doing here is projecting that down on onto one dimension, and I want to come back to kind of kind of elucidating quite what all those those terms mean. But before that, depending on your risk appetite, you can look at a I suppose a matrix, right, a risk matrix of impact on one axis, and likelihood and probability on the other axis and see the different possibilities and based on your appetite for risk with your risk friendly or risk averse, like, of course, like where you put your 401k will vary if you're 20, as opposed to 60. Right. So maybe you don't want to project it down at that point. I want to come back to the chemotherapy case you mentioned, because you could imagine, and this is a slide from a talk he gave at NYU, or maybe one or two years ago at the conference Jared Lander organizes. So fantastic conference also great, amazing, amazing community builder. And yeah, super, super lovely. You had a slide there, which had the effect size of a chemotherapy drug, which... it looks super effective. I was like, oh, yeah, I'll take this. But then when you look at the cost of taking it, which can include, you know, harm to your body, and all of these things that we know about chemotherapy, it actually when you ablfoyp essentially, when you integrate of your cost function, what you see in the particular case you gave is that it doesn't outweigh the benefits don't outweigh the cost. So essentially, that's what we're, we're computing there. Correct? Jim Savage Exactly. So there are many, many cases where you might have some significant effect, fairly precisely estimated, but not perfectly precisely. And you look at that effect. And it's like. hugo bowne-anderson ,And it'snot just statistically significant, it's a practically significant effect as well. Jim Savage If I take, if I take this drug, I expect to live four months longer. The thing about like that, if I take this drug, it gives me I expect to live four months longer. But that four months isn't deterministic, it's not like I will definitely be four months longer if I take it might be zero months, it might be eight months. And there is some uncertainty about that. Now the drug also make you lose your hair, you vomit for three days, you have to go to a 12 times that you're in immense pain. Now, if you if it doesn't do anything, if you're not dying at the same time you would have anyway, but you went through all this pain, you're worse off. Yeah, if you go through all this pain and you live eight months, then you're better off. But it's that uncertainty. And you need to think about that uncertainty. And like there are great cases of doctors prescribing things that they when they get to, like terminal stage of, for example, cancer, they don't actually prescribe themselves the same drugs, I've been prescribing because they understand the uncertainty and understand the cost. And they're like, Ah, I just want to go home and spend time with my family and take morphine. And so like this is a really important thing that you do. Even if your loss function at that three month point of expected increase in life is negative like is a look at the expected point. It doesn't mean that your expected loss is negative, it could be quite a large loss. hugo bowne-anderson So the other thing that I want to discuss explicitly now is you've done something very clever, Jim. And what you've done is you've snuck in Bayesian inference without telling anyone. Right, so we're talking about a posterior distribution, which essentially is what you know about your parameters with uncertainty after you've collected your data and also based on any prior information you may have what you've also done and a la Deirdre McCloskey is compared it with frequent ism, in a way and in a situation in which the Bayesian announcer is the right answer. And the frequentist answer is the wrong answer to the question, this is not always the case. Okay. And I'm skeptical of people who pure Bayesians because you can only ever probably be a Bayesian as far as I'm concerned. Sorry, not sorry. And but what I do want to say is the reason one of the reasons this mistake has been made in this case and is in a lot of others comes down to something I referred to earlier, there's a book by Aubrey Clayton called Bernoulli's fallacy, okay. And this is an incredible book because what it points out is a fallacy that and to stand up to Bernoulli, as well, the man and the entire family and to title a book based around Bernoulli's fallacy is such a wonderful act. And it's also so true. The fallacy here that Bernoulli made when doing the balls from the urns and the flipping the coins, and all of that was that Bernoulli, when he wanted to talk about the probability of the model, given the data, was actually talking about the probability of the data given the model, okay? Now, these things can coincide in several cases, and frequentists can get the form of given the ladder, but a lot of the time, this is the frequentist error. So maybe you can speak a bit to the importance of Bayesian inference, Bayesian thinking, Bayes theorem, all of all of these things, where what we really want is, um, I suppose something you would probably call like, well-calibrated uncertainty about the distribution of your unknowns, given the observed data. Jim Savage So I guess it's maybe my spicy take here, that it really doesn't matter. Like I spent years and years and years of my life, becoming a fairly competent Bayesian statistician. And I don't really care. The reason I think it's worth learning that stuff is because it teaches you a certain way of thinking, you build models in a particular way, you have to think about what is the generative model, you have to think about, like, what is my data like? And you have to like, rather than just say, Hey, I'm going to try and predict the outcome, or I'm just gonna, like, run some regression, interpret some coefficients, be like, How did my data come about? What is that process? And think through it that way? I think that's the important thing about Bayes is because in order to write down your likelihood function, you do need to consider like, what was the generative model. hugo bowne-anderson And you need to make your explore your assumptions explicit, right? Jim Savage Yeah. And it's that the becomes thoroughly addictive. You just think about how to build models that way. I don't, I'm not like completely sold. Like if, if I write a model, and it fits well with maximum likelihood. And my uncertainty is isn't too far different, but it runs 1000 times faster. I'm like, Oh, this used the maximum likelihood, like, I'm not literally doing ablfoyp on everything single thing. hugo bowne-anderson Because Because maximum likelihood is Bayesian inference with a uniform prior. Jim Savage Yeah. Yeah. So and sometimes sometimes, maximum likelihood will not work. You've got like, a bunch of hierarchical models where the likelihood function is degenerate. And you're, you just don't do it very well. So like, sometimes you just need to use Bayesian methods to fit models. But it's not because you're a Bayesian, it's just because like, those methods are what you use to get the correct estimates for your model. Yeah, and I'm not even like dogmatic about like, I need to write out loss function, I need a posterior and I need to plug those things. Like, I mean, these things seriously, like you do need to think about, like, what is your loss function, you do need to think about, like how uncertain I am and what the model is. Because that is how you understand the system and understand the consequence of decisions in the system. I don't think that you literally need to be fitting Bayesian models and running a loss function and doing like, integrate, like, that's fun. It's like a little bit super analytical. I think it builds habits of thought. Especially, there's hygiene in distinguishing the probability of an event. And your, how you feel if that event happens, distinguishing those two things like you don't wanna be the person who's like, I don't think Donald Trump's gonna win, because you don't want him to win. That's like a foolish sort of, you want to model your own biases. Exactly. Yeah. And so gives you the toolkit to think about those two things separately, yep. I do not think you actually literally need to be fitting models or literally need to be building them or doing anything, once you have the learning them is the point that is the human capital development process for a topic decision making that will hopefully benefit you later on. hugo bowne-anderson Great. And I do think I would kind of want to zoom in on generative modeling before we move on to the work you're doing now, because I think specifying several levels of approximation, increasing levels of approximation the process that generates the data you have also comes back to this idea of ground truth right. It also allows you to specify your data collection process, right. So you can put in the biases you have along there, whether it be selection bias or survivorship bias, or any of these things right, or Berkson's paradox, or whatever it may be. So to wrap up this section on generative modeling and budgeting So, what tools if someone were interested in in doing this type of stuff? What tools would you suggest they use? Jim Savage So I think learning base Stan, and doing things in R, is just like, the fewest lines of code to most expressive, I mean, pythons are wonderful language. It uses like, so many lines of weird code to anything like the volleys class or whatever, to just like, you know, generate some, some fake numbers. Well. hugo bowne-anderson Yeah, good. But you're right as well. But if you are a Python, a still interested in Pydata, do check out pi MC r because you don't have to do a bunch of OOP to do that. And you can write several lines of code in order to specify your model system, you know, context managers and a few like idiomatic stuff. But if you'd like working in Python, that works, but I definitely think Stan is incredible. And said, code for you and a lot of other people, right? Jim Savage Yeah. Because it allows you to just like describe your model in a shorthand, it looks a lot like how you're describing on a chalkboard. And if you're used to writing down models like that, it's very, very cool. Yeah. Now, I would always say, before you ever look at the data, you write down what you think the model is, and you simulate fake data from that model. hugo bowne-anderson Great, because otherwise, you're already overfitting to something, right? Jim Savage Yeah. And it also tells you like, hey, so and this is process where you say, here's a prior distribution I think my parameters might live in, I'm going to draw some fake numbers for those priors. I'm going to use that to simulate some fake data. And then you look at it and you're like, there's nothing Why the data that I observe, like, Well, my priors must be completely wrong. And so I go back, and I'd trim my priors. Now learn how our model works. And often that that takes like, months for a sophisticated model, you'll learn Hey, what are unreasonable assumptions about its model. And you do that before you go and fit the model. And it's that process of learning how the model works and what's weird in it. There's just like this, this amazing trip, it's like, you touch the stars, you're you're like, really see God, at that point, because you fully understand some big complex model. I thoroughly encourage people to take that approach. You just learn so much myself. hugo bowne-anderson I'm glad you mentioned seeing God. Because before we move on the OG risk matrix, as far as I'm concerned, is Pascal's Wager. So so there you go. And if you want to go back, anyone listening, and check out some Pascal's research, how he was thinking about probability, how him and Fermat were thinking about probability at the time, really, really beautiful, incredible work. But it's the 21st century now. And we've moved past all of that, it seems, I'm excited to jump in and hear about what you're working on now at Schmidt Futures. Jim, tell us a bit about what you're up to. Jim Savage Yeah, so my role at Schmidt Futures is pretty weird. I get to be first boots on the ground for programs, a program might take couple of years to spin up and try to build programs, they're gonna have some outsized impact. Schmidt Futures is a philanthropic initiative that was co founded by Eric and Wendy Schmidt. Eric used to be CEO of Google. And, you know, I think there's this insight from his experience at Google that finding people who are really exceptional and supporting them and challenging them can yield outsized impact, specifically, not being too prescriptive in what they do, how they get there, but making sure that you're providing some combination of support, and challenging them. So it's like, let's go find really, really smart people who are going to do the thing that they're good at, and help them get there. The important thing about this is that we don't have a very strong like, cause prioritization lens. And a lot of people in philanthropy. They say, like, we work on x, we work on Y, the Schmidt Futures model is more the fit between the person and what they're doing matters so much that really let people who are exceptional decide what they should be working on. And we just provide the sort of support that we can. hugo bowne-anderson My question and maybe this is where you're going is how you use your skills. What is director of data science for these types of initiatives even look like? Jim Savage Yeah, so let me walk you through two inititiative this as well. I spent last couple of years on, and it was one that I'm starting now. So bit over two years ago, were like, Wouldn't it be great to get if I'm if I'm perfectly honest, my old boss Dylan, he was like one of the first hires at what became Lyft and he's a really nice very well-intentioned guy. So he dropped out of college, went helps out lift. He went back to college after a couple of years and and you know, that mindset of it. How To Get A Job at lyft? Well, I was neighbors or at school friends or something, I don't know what was, was like, fella, right? And he saw in his friend who started lyft, like, oh, people like me can start an almost really impactful companies. And so I was like, Oh, I can, I can do that. And then as an Australian that was like, something really weird to me. I always thought like these people who started big companies are like something else. They get like, they're geniuses or whatever. hugo bowne-anderson We'd call them tall poppies in Australia, I think. Yeah. Jim Savage Yep. Yeah, they would like if you're an Australian, and someone talks, says, Hey, here's this cool thing I want to do. You would politely insult yourself as a way of telling that person that they've self deprecation is one of those superpowers. Yeah, it's sick. I don't know if I'd go that far. But it's absolutely poisonous. I think it's as, as a mindset. It's something that Americans are really good at. So Americans, like they don't do that at all. They're like, oh, cool, you're crazy idea is wonderful. hugo bowne-anderson let's go to the moon. No, seriously, let's go to the moon, Jim. Jim Savage So it's this question of like, hey, how could you create a program that would do two things. And it's to find a large number of really talented young people who you can hopefully convince at a key juncture in their lives just before they get to college that they should be having a career dedicated to serve as follows. And so how can we do that, at the same time as leveling up, people who miss out, you're always going to have people who miss out on a program. And so this program called RISE was created RISE is, as one of the world's largest opportunities in terms of like, support, each of the 100 global winners every year, is eligible for over half a million dollars and support through their lives. It's also an operative prescriptive program. So for example, the Rhodes Scholarship, you go to Oxford for two years, if you win RISE, you get sort of support that you need. And you would work out with the rise team, like what sort of support you actually do need. And if you're going to be a social entrepreneur, that's very different to if you're going to be a member of parliament but we would hope that we can provide the sorts of support to help you be your best self. Now, that's extremely expensive. And so it'd be a shame if we only benefited winners. And so the entire point of RISE is can we attract a large number of applicants and make an application that both gets a good read on their skills, but also helps them start to change your mindset. And so we did it by working with a celebrated psychologist at Penn Wharton, Angela Duckworth, she helped us put together this kind of application process, where people would to apply, they're going to do something, they work out a project that will take them roughly eight weeks, we kind of adjust for how much time they're likely to have. If they're looking after siblings, whatnot, then we can expect them to set to do the same sort of work as someone who's got all the resources in the world. And they know what they're good at. And let's say let's assess people at what they're good at. And you work on a project for eight weeks, and we provide structure and support to you doing that. So there, we have discord channel where all the kids can actually meet the other dweebs like them. We've got like these online coaching sessions, we've got like all these other kind of like, resources that people get, and all these interesting mindset interventions that have been demonstrated in quite large studies to help people kind of think about themselves differently. And so it was just like, it is fantastic program. And applications are open until December 22. I'm not sure if this will be released before them. But if you don't, if you're 1017 year olds, they should definitely apply. Now my role in that was helping to like really set early parameters. So like, who should we be looking for? What questions do yield information about people? What sort of information should we be collecting? We did this incredible study that, and I'm allowed to call it incredible because like I wasn't, it's not my study, hopefully inspired a bit of it. But but really, like the team working on it is just a really, really amazing team. The basic idea is, hey, if you, let's go to design an application process. Let's start with a population for whom we know there are outliers in our quote, unquote, ground truth sense. Let's then send them down a couple of different application paths. And check to see whether that those application paths are able to identify these people who we already knew were outliers. And so what we did, we recruited 16 classrooms around the world in gifted and talented schools or programs and took a class in each of them. Were the two teacher had been working with the kids for at least a year and the kids all knew each other pretty well. We then sat down with all those teachers and said, Hey, can you like, score your kids? And we've built this rubric, can you score your kids on, you know, intelligence, empathy, integrity, perseverance, one called spike, so we had like some calling, and boostability. And also creativity, which can code for intelligence in women in some cultures. And we then asked all the kids in those classes, hey, can you nominate the top three most quote unquote, like empathetic kids or top three kids with the highest integrity or whatever it is? Now, we knew that the people were kind of talking about same constructs. Because when the kids when one child was ranked highly by one person for integrity, they're much more likely to be ranked highly by like kids in a classroom. When you looked at what the classmates were saying about who the outliers were those outliers but much more likely to be identified as high scoring by the teacher as well. And so it probably talked about something that at least is shared in people's minds. It doesn't necessarily mean that it was. hugo bowne-anderson this is really interesting effect. If we're trying to talk about ground truth, the fact that we're getting ideas from different people, and a sense of shared reality, I think, is actually incredibly important if we're to use the term ground truth, which I'm uncomfortable with, but I'm okay for the purposes of this conversation. But the fact of sourcing it which essentially will know there's an argument, we need to look at the details, but this will essentially remove certain types of measurement errors, systematic bias, these these types of stuff as well. Right. So crowdsourcing, have you come across the jingle jangle fallacy? No, I haven't. I'd love to hear about a new fallacy. Wow. Jim Savage It's like a metric. hugo bowne-anderson That's the Hugo fallacy. Sorry, gone. Jim Savage I forget the which one's the jingle, which one's the jangle fallacy. But it's basically that you might have two similar sounding terms that actually described very different constructs. Or you might have different terms, for example, empathy, integrity, we know, I describe that to you, as you and you can just if I ask you, Hey, what is integrity? Then you've got an answer. And then I asked you was empathy, and you've got an answer. And those things are not the same. And then you go to the data, and you look at people who've been graded on empathy and integrity. And they're not two independent factors. They're like a single, single correlated, very, very, very correlated thing. And in fact, what we found was that although people were agreeing on who the outliers were, they weren't agreeing on like, who literally has high stability or who has high empathy, because there are really only two or three dimensions. So they roughly correlate to are you a pretty good student? And do you have some kind of grit street smarts type stuff? Okay, that makes sense. Hugo is was waving his eyebrows at me for those not watching video. Maybe I'm waving my eyebrows here. Yeah, so. So those two factors explained most of the variation that we saw. It's not graduates that are it's like some some, hopefully, fairly validated scores. hugo bowne-anderson Can I just ask, there is one other thing, though, when, and we'll come back to this when talking about data for good. I just wonder like terms like integrity and empathy. Maybe they have pretty like Western connotations to how they're interpreted perhaps, like, if if you go and do something like this in Nigeria, for example, and we're using terms that we we validate, and that we think like, What are the dangers associated with that? And how do you include stakeholders and feel free to say, let's come back to this later. But I, I think that's important. Jim Savage Yeah. So we did this study in Brazil, Kenya, UK, Hong Kong, not sure if we could defeat that Vietnamese class or possibly US and Mexico. And what we found was this effect of there being very few factors that describe us the things that people agreeing on who was exemplary, more or less held in all the countries. So we allowed the parameteres to vary by classroom. But they didn't vary by very much. hugo bowne-anderson my question is does asking the same questions in different countries even make sense? Or is there some sort of colonial bias there? Jim Savage Yeah. And we actually found I remember, there was one school, I think it was in Zambia. And the teachers rankings were all the same. So in theory, integrity, it's like, ah, and that basically, it's just like, a large class 45 kids that probably don't know them very well. So I'm going to look at the Gradebook, and score. And so it's like, well, I don't know if that that is quite as, as good because like, there's think about education differently. And comparing that to like a school In, I think the other school where the school in New York, it was Midtown, it's like a very kind of hippie dippie sort of school, where the teachers get to know their students very well. And like, I just don't know if we're measuring the same thing. Yeah. Okay. Which, which doesn't mean that we shouldn't try just it's like, you know. hugo bowne-anderson It's very worthy of consideration. And I think it comes back to uncertainty around decision making these types of...when we know that it may not work the same way. Or we may be putting our own biases under different cultures, then when decisions are made, incorporating as much from the different stakeholders as possible and make and co-evolving and c0-making those decisions. I think, Well, the Jim Savage An interesting thing that we learned from this is that we then interviewed, as many of those kids, I think it ended up being 110, or 112, of the kids with panel interviews that were structured, so they were interviews, they all use the same questions that could gauge us across different students and how they related to each other. They we recorded all the interviewers scores on every attribute, individually, so the interviewers would all do the interview, and then they would record their score. And then they would come together at the end and say, oh, so what do we think and that would kind of negotiate a group score, a panel score. And what we basically learned from that was that they didn't do especially well, at pulling out the outliers. And when we broke--when we really decomposed that error. So who were they making the mistakes for? We found that it was the kids in the so we also had biographical data on the kids who are being interviewed. And there were, we looked at one dimension, in particular, that really stood out to us, that was household income. And the kids in the richest 20% of households 23re systematically for almost all the attributes, being given scores that were higher than average by the interviewers relative to what the peers and teachers scored. And in the kind of lower middle class income quintiles two and three quite a large underestimate. When we went to the interviewers, and told them that they were like, Okay, tell me who told me, and I'll tell you about the interview. And we would give them like, well, that person didn't answer the questions. They actually didn't interview very well. And it is really, really interesting, because, you know, I think this is we talk about a lot, I don't think we've got sort of data that we're completely convinced that it's, you know, the whole picture yet. But that sort of polish is really important answering interview answering a question properly. So just to clarify, it was the wealthier children who did better. nice structure was smart. Oh, that was like a star, you know, Situation, Task action result, sort of interview responses that were very, like, great. And the interviewers systematically gave them higher scores than the peers and teachers, right. And the kids in lower middle income were systematically penalized. The kids at the bottom, were on par. Okay, which is really interesting. I think, like, go through my career up if it pretty motivated by like, fairness. I am kind of, I get quite angry at systems that are very unfair. hugo bowne-anderson Well, you're one of my friends who's like a man of meritocracy, which I have my own challenges with, this is probably isn't the place for it. But that's a conversation over many, many more beers in a Brooklyn bar one day. this is very important to figure out. Jim Savage Yeah, that's really, really important information, it means no, the model I've got in my mind of identifying people for support. If we think about, for example, applications for universities, or whatnot, it's known in Australia, where we're both from, that if you get 95, out of 100 as your ATAR school, and then go into a university and do a degree. Let's take all the kids who got 95. And then let's track, how many of them dropout. And how many of them get F's and how many of them get A's and whatnot. There are some great data that shows if you went to like a good private school, you get a higher dropout rate, and a high rate of failure and all that sort of stuff than the kids who went to the selective state schools and like I said schools have so and the regular state schools, the regular government schools that aren't as resourced. The kids who get 95 of those schools that go into same courses have lower rates of dropout and lower rates of failure and what's going on, was measurement error is always mentioned error. So we're trying to use a level threshold. This is like some test score or some assessment in an interview as a way of judging trajectory. And trajectory is a function of how much input you've been given. If people have made a huge set of investments in you and you meet the threshold and I'm comparing you against someone who's had no investment. And they almost beat the threshold, or who should I bet on? Was the kid with no investment, they got there with almost nothing. If you've had like a huge set of investments in you, and you just met the threshold, like, that's your ROI, this really motivates like, what sort of information should we be using? When we are like trying to assess people for this as opportunities, we want to look at what sort of investments have been made. And what's the option is people had for growth. And also what's how they score and interviews and you're all those things. Now, the other side of that... the flip side, which is really, really important, is investments matter. Like, if you have early childhood, a really great early childhood education, you're way more likely to do good things the world, it's a huge source of inequality that some people don't receive great early childhood education. And if we not adjusting, like, so, although we want to be adjusting for exposure to opportunity, the exposure to opportunity also matters because you have got more building blocks. And so this is like really challenging ethical decision of like, which those things truly matter to someone's potential, which says things is simply training for the test. hugo bowne-anderson I think this dovetails really nicely in what I'd like to and I think you'd like to discuss next, the idea of challenging ethical decisions, and what data for good actually is and how we can responsibly approach it. So as we've discussed, you've done a huge amount of work in the data for good space, as has your wife Sue, as well, across a lot of different different dimensions. And maybe another conversation, I'm very interested in thinking through what having a partner who works in a similar space and who you know, you talk about these things with. But I'd like to have a conversation about responsible data for good and to kind of set the seed, this is something I'll include in the show notes. Rachael Thomas recently wrote a blog post on her and Jeremy's fast.ai blog called doing data science for social good responsibly. And I'll actually set the stage by reading a tweet that she wrote a couple of years ago, which included in this post... Rachel wrote, data for Good is an imprecise term that says little about who we serve, the tools used, or the goals, being more precise, can help us be more accountable and have greater positive impact. And she was actually...she's shared a screenshot and was referencing a presentation by Sarah Hooker at the Data Institute in San Francisco, at their lunch seminar. So perhaps we could use that as a launching pad for thinking about accountability and transparency, thinking about including the most impacted people in the conversations, recognizing their values may be different from those of nonprofits or academic stakeholders. And also, in these types of things, what it means to take data privacy seriously. Now, I've just done the worst thing an interviewer can possibly do, which is throw out maybe five incredibly meaty issues. But I'd love for you to pick what resonates with you and maybe we can have a conversation around these ideas. And I just want to stay, of course, this is something we could talk about for five hours in itself. And I'd love to do that and downstream at some point. But at the moment, I think it's important to do this after the conversation we've had. Jim Savage And I encourage the listener to go and read the piece, it's, of course, very important. And there isn't that Hugo has included it. I've read it, I think three or four times now. And I have a very mixed response to it. And I'm not sure whether it's because it reminds me of some less than friendly discussions I have with people, or whether it's, you know, whether I'm interpreting it wrong. It is, of course, highlighting some incredibly important issues. Like, if you are not taking data privacy incredibly seriously, then you're probably not doing data science for social good, if you're not like building if you don't have like the end user in mind and protecting them, then you're not doing data science for good. The challenge, I think, here is that this is a very, very, very important point here, which is that they're especially when we're working in development contexts, or weather groups who might miss out on finance, or whatever you need to if they're not represented in your decision making in your user research in your rollout plan and in your research, you are doing a bad job. I've seen it done really well very infrequently. I should say one things I have kind of disagree with in the article are the pieces of data science for good work that do nothing and they might be like greenwashing or whatever. I think there's that I have a lot of time for young people trying to do good things. I've done those. I think most people have worked on projects that didn't really do very much. And that's okay. And I don't think we ever should tell them they're wasting their time. Because often that's like human capital development, often a network development, like it's important stuff. hugo bowne-anderson I agree with those examples. I think that is slightly straw manny and in the sense that those examples, sure, but I think examples of greenwashing and ethics washing are when big tech companies hire an ethics owner right, and set them impossible OKRs, or KPIs where not a lot is actually done. And that's endemic in, in tech in in general. So I think you're absolutely right. But I think the concern is incredibly important. Jim Savage Yeah. And I think the the right person to impose... so the way I've seen it done best is when the person who is leading the project is like, wouldn't ie be interesting to understand what sort of model can we build, and who is hurt by it? And they treat that as a primary research objective. And they are curious, because they're not just trying to demonstrate success, we do have a culture of showing how my AUC increases and aren't I smart, I think it's much more important to see the points where your model breaks or where your analysis is wrong, because it makes for not only a much better quality piece of analysis or product level building, but also, it's more interesting, more fun. hugo bowne-anderson And it's also important to think who your key stakeholders are. So I'll give an example from another industry, which I think is uncontroversial, which is civil engineering, okay. So bridges and buildings need to be built, there are codes and all of these things, which checklists as well, which need to be built so they don't fall down. If they fall down, people die. Civil engineers are incredibly culpable there, right. Now, let's take a tech company, any tech company that does make some money from online advertising, let's say, in this particular case, I don't want to pick on Facebook or, or whatever. Having said, I don't say I don't want to pick on Facebook. But the idea that who are the stakeholders for said company, a lot of the time, the real stakeholders are the customers who are buying the ads, and the users, user-consumers who use social network or search or streaming or whatever it is, or like, looking at photos or like I don't know who that would be these things, what happens to them are probably considered at best negative externalities, as opposed to actual stakeholders. So is it important to consider everyone who's impacted as actual stakeholders and be responsible to them? Jim Savage Yeah, it goes back to the causality conversation. And Shira Mitchell's on a whole bunch of work on how do you think about both causal inference and fairness, and algorithmic fairness and those things as one because if you're not thinking about causality, and in context of how your systems might hurt people, then you're not doing it right. hugo bowne-anderson Yeah. And check out Shira Mitchell. She's at Blue Rose. And I'm actually going to have her on the podcast actually spoke with a couple of days ago. Amazing. She's great. And I need to thank you once again for introducing me to Shira at a lovely dinner party you hosted in Park Slope couple of years ago. Excellent. Okay. I think there are several prompts. hugo bowne-anderson How should the general populace think about accountability and transparency in these situations? So why should we trust Schmidt futures or Chan Zuckerberg Initiative or anything along those lines to be to be doing like, you say, We're doing data for good, but how what type of conversation between the public and in democracy should happen between philanthropy which has a lot of capital behind it, and and the democratic process and the citizenry? Jim Savage it's really important, I don't think there's really a because like, it's not just philanthropy. It's not just big tech, you're giving away your data to like your restaurant, and one of those restaurants like leaked a database of credit cards. And that's like, it just happens all the time. I think the most important thing is that people like become aware of what sort of things they what is the standard End User Agreement? What is this? What does GDPR give me because everyone's using GDPR as a standard for their own data collections now, and all the other types of protection, education has a whole bunch of protections with those. So I think you need to educate yourself and tell people about it. And there are lots of great videos on all this stuff, I just don't think it is likely, because no one is incentivized to create the sorts of accountability mechanisms that we have with government with big technology companies with flashy with small businesses like to I'm still here, sorry, the thing is, I don't think we'll ever see accountability mechanisms that are not driven by people asking for data. hugo bowne-anderson Yeah. And also, I do want to clarify that I don't necessarily think... I framed it as how to include people in the democratic process, I do think we're seeing deep fundamental flaws in the way democratic processes are implemented. Now, in the past couple of years, we have had what is arguably a total failure of leadership classes in the west and other places, but you know, where you and I are, as well. So wondering whether, you know, regulation is important. Establishing norms in an industry is also important. But whether how much we want governments to know about our data, as opposed to corporations, it isn't even clear what this what this trade off is, right? Jim Savage I've got quite a few friends in the political data world. And one things that you know, about or everyone knows about in the US is that you've got these voter files, and they're all merged with credit card, like your credit card company has sold your data, your you know, if you went to a big, tall and wide, or whatever those stores are for, for like unusually shaped men, if you went to other stores, like they have sold, your data is in some database somewhere. They know how much dog food you buy, hugo bowne-anderson well, I shop at a place called high and mighty. And I used to buy shoes at a place called Bigfoot. So there you go, because I'm size 17. Jim Savage in the US, then those companies would have sold your data to a Data Broker, who would have merged onto a big file that has like 8000 columns, and that is merged with the voter file, and it's sold to anyone who will pay. And they use that to train machine learning models to work out who to target. This is a well known thing, all types of politics uses these sorts of methods. One of the really interesting things that GDPR is doing in Europe, is because the transferring of personal data is like far more challenging unless you give informed consent. And so in the political world, they can't, they don't have access to that granular data. And so the kind of low level targeting is not easy for them to do. And it's like, actually changed the dynamics of the political operations there, which is really fascinating. And so that is, you know, we're already seeing GDPR being because it's like, the, is a huge set of policies that I think anyone understands it fully, but because it's so huge, everyone needs everyone's GDPR compliant, even in countries where they don't have to be GDPR compliant. And that's probably going to mean that we will see an increase in ..., I'm confident that that actually sets the tone and we'll probably see an increase in and privacy. To be honest, like I the other side, like I really like some of the things that we've gotten from a lack of privacy, like we can't pretend that it hasn't had some benefits as well. You know, I if I jump on Instagram, the advertising is so good. I just enrolled my son in after school program that was perfectly targeted to me. And is that kind of sick? Yeah, but I do benefit from it somehow. And, like, hugo bowne-anderson I mean, what we need to talk about is like, a high multi dimensional cost benefit analysis and not reductive projections down to yay or nay. And I do think, you know, how many of these conversations happen on so that's one of the reasons I, I'm starting podcasting again, man, I'm sick of bloody the 180 character reduction to yes or no to good or bad to burn or raise, you know, and that's one of the reasons I wanted to discuss this with you as, as well, because of the work you've been doing. And speaking about case guides, Jim Savage So not sure if we're going to get the opportunity to talk about it in detail. But I think there is this one thing that that piece you said, threw me made me think and it is, you know that the most scarce resource in the world is entrepreneurial energy. If you see, you know, yourself, when you have owned a project and you love it, you are like 100 times as productive as when you are disempowered or poorly incentivized or, or whatever, like, entrepreneurial energy is worth its weight in gold. hugo bowne-anderson Yeah, I don't want to get too Nassim Taleb-y, but he's like entrepreneurs over VCs. That's for another conversation as well. [a] Jim Savage so at the same time, people have feelings. And there's a thing that I've seen time and time again, which is like you get someone who's got ton of energy and working really hard. And then someone makes a bad faith, comment about the work, or sometimes even in good faith comment about work, and it just knocks it out. And, and can really decrease the amount of good, I think we need to find constructive ways of making sure that people do high quality data science for social good. Now, in particular, I think there are terms like, at once, it is critically important that if you if you're doing a large data science project, that you have actively sourced an advisory panel who's going to be is it red team, what's it called, like the team who like, challenges you to sign off on your plan to make sure they're representing people, you should be talking to your end customers, and customers and end users to really understand like, you should be on the ground and doing all that stuff, to really make sure you're solving the problem, like know those things worth taking incredibly seriously. At the same time, there are people out there who know that they can win an argument in a meeting by saying you're not taking x group seriously, they're not represented at the table. And it can be a bad faith, challenge of your work. And it can, and it can snuff out that entrepreneurial spark that that is often driving good. So I just want to be really, really careful that we don't be overly critical of people who are trying to do good and said we support them, right and help them ask the right questions rather than accusing them of racism or sexism, things people because it's like hugo bowne-anderson to have good faith conversations. And that's what I'm interested in here, you and I agree on a bunch of stuff, we disagree on a bunch of stuff, but there's a certain base level of good faithness and respect for another human here. I do also want to acknowledge though you and I are both white males from a certain class in Australia who've been given a bunch, have lived in New York City, worked in certain industries, that that type of stuff, and I do want to make a commitment to I don't even know who the listenership is going to be of this this podcast yet. But I do want to make a good faith commitment to have as many people from as many diverse backgrounds on this show to talk about these issues, because I think these conversations are incredibly important. You and I have posed more questions than we've answered, we've probably missed a lot. I would ask listeners that if you know, you disagree with things we've said or think we've missed things and are frustrated, let us know. But please do it in a good faith manner as well, particularly on social media, because we do want to engage, and well I definitely do and I get the distinct impression you do as well, Jim Jim Savage Yeah, you know, I think this comes, you know, this guy, you and I were introduced by a common friend, what, six years ago or so. And this project I'm working on at the moment, you know, what we're trying to do is, you know, what we learned with RISE was that, you know, it is possible to find some really exceptional people who might not get discovered and change their careers and change the change the incentives for a lot of people's they might let go in, consider a career in service of others. The program that I'm building up at the moment, which is it kind of zooms forward 20 years at what do you do with those people who are roughly around our age who go? But I think we're within a few years which other late 30s? hugo bowne-anderson Sure, mid to late 30s. Jim Savage So I'm 35. So how do you intervene to make sure that exceptionally capable, big hearted people are working? Actually having impact? I think there are many, many, many people in the world, there might be an impactful oriented person, just like in the wrong role, or at the wrong firm or employer. Remember, even the wrong team, the right employer, maybe there is maybe they're like, optimizings click through rate for dog food which is like, fine, it's benign, it's okay. But that person could be doing something much, much better. And so what we're building is a program to really help with those top executives, operations, people will take people, data people, product people, and so on, to make sure that they understand what an impactful career looks like. And they can get into that role, especially these people who were like, six months, two years out from that they know they've got something that they know that very capable, and they pondering what that really high impact switch is. And so, so if if we want to really tackle who's got access to those super big levers where they can do a huge amount of good. It's about redistributing networks. And I think networks are important. Because, like, people hire through networks, they become friends through networks and whatnot, because we can't describe what we want in a job ad, a CV and an interview cannot capture. hugo bowne-anderson We still develop band level groups within large scale societies in a variety of ways. You've made a key point, I think there are all these conversations around equality of opportunity in America in particular, and how the hell like what an ideal, how do you have equality of opportunity, when there's so much network nepotism for one and so much intergenerational wealth transfer on another? I mean, it's all well and good to say equality of opportunity, but how do you operationalize it or deploy it when the system isn't only against it, but it's a totally different frame of living in a lot of ways. Jim Savage Something a lot of it is about, like describing the rules of the game. And people don't know the rules the game. And it works against them, and simply giving people like, if you learnt that, hey, the people who've all made it, they met their sponsor by we're going to a political campaign. And it's like, Hey, do it. That's like, a really cool, put message that like, that's well known in some communities, and not at all known in others. hugo bowne-anderson Yeah, like my armchair anarchist might come out slightly now. But I mean, a lot of people in positions of power have literally vested interests, to not let the rules of the game be known. I want to wrap up now, I'd like to say it does sound like incredible work you're doing I would love to hear more in future about how philanthropy can be kept accountable. In particular, it sounds like good work but I'd love to know more about, like, who decides what's done? And then who decides who decides? As well, I think these are really important questions in terms of transparency and accountability. But thank you for having just a candid conversation around these things. And I do want to say to our listeners, please reach out with thoughts, inspirations, critiques, I'm sure we've missed out a huge amount, and perhaps said things that are incorrect as well. And I'd love for those to be brought to our attention. So Jim, we'll probably have a listenership who are very interested in data, if they made it this this far through the conversation they're clearly very interested in data. I'm wondering if you have a call to action, something that you think you'd like to see people do more in, in the space or what people can can do? Jim Savage Yeah. So I mean, the thing I tell all young data scientists is to spend some of that time that you're devoting to6 tools, and instead devoted to reading research assigned to understand research problems, and how really smart researchers have tackled difficult research problems. I think that's like, if you're developing your career as a data scientist, that is like, a superpower. hugo bowne-anderson Right? And are there any, would it be domain expertise stuff? Are there any research papers that you think everyone should read? Jim Savage They're 100. Just thought with mostly harmless econometrics or mastering metrics, think great, go go read those books, and you'll, your life will be better. hugo bowne-anderson Fantastic. And we'll include links to those in the show notes as well. Cool. Jim, it has been an absolute pleasure! Transcribed by https://otter.ai [a]up to here