The Learning Curve Podcast: Paul Kent and Lauren Leek Paul: Welcome to the Learning Curve podcast. I'm Paul Kent and I'm a mathematician, and this is a podcast where I get to have interesting conversations with AI leaders, innovators, and people who are doing cool stuff. Today's guest has a PhD in political economy. She's a postdoctoral research fellow studying polarisation and inequality across the UK. She consults for defence think tanks and she's also somehow finding the time to build a highly-read data science Substack where she uses machine learning to tell stories about the cities we live in here in the UK. So, very warm welcome to you, Lauren. Lauren: Thanks a lot. Thanks a lot for the introduction. And I think we should caveat by saying lots of the people watching this might be used to thinking of AI in terms of large language models, but there's a whole range of techniques and approaches that existed way before ChatGPT came along. Techniques like machine learning and data science, etc. Paul: And the ones I'm really enjoying your output on. You have a PhD in political economy and you've worked in defence research and public opinion analytics. How does that all start? Where did Lauren Leek start her trajectory? Lauren: I have a background initially in economics, where I learned a lot of the statistical methods which I still use nowadays. Then I got very interested in the bigger political economy questions, did a masters in political economy, and then I wanted to continue research in this. The best place to do that for me was in a political science department, actually, where I got introduced to all kinds of people studying political behaviour, doing more sociological approaches and stuff like that. Very quickly I noticed that what I liked the most about it -- I liked all the topics that people were discussing, but I liked approaching this from a methodological angle. How can we research this? How can we measure it? Which data should we use? These were the questions that really intrigued me. I think the rest sort of naturally followed from it. People started contacting me and were like, "Oh, you know how to do this. Can you help me with this research design?" That's how I sort of got into all these different angles and projects. I'm just intrigued by certain curiosities and the ways in which we can answer them. Paul: What I notice is that on your Substack you write to a general audience, but you don't talk down to them. You talk about the mathematical models you're using, machine learning approaches -- that seems very intentional. What's your mission with your Substack? Lauren: I think my mission is twofold. First of all, I do want to show that rigorous research has a place in the public debate. A lot of data journalism is quite descriptive, but I try to go one step further: we have this methodology, can we do more with this? That's on the one side. On the second side, I want to make this accessible to more people. Even the people who have no idea what I'm talking about -- I still try to have a story line there. If they're interested in the methodologies, they can always look it up or go one step further. But I try to explain it in such a way that you don't need to know the technical details to move on and get the larger message. I discuss whatever is on my mind and whatever I'm curious about. I try to find a data angle to that. Paul: You tend to build a tool at the end that people can interact with using local data. You do one on London where you can check your own postcode and see which side of something called a "latte line" you lie on. You do one with pubs where you get to find out how much at risk your pub is of closing down, or losing trade because of how isolated it becomes. And you have these tools and assets that you create that are freely available on your website. It's one of the things I love about data science: it brings all this huge amount of data and produces something at the end that's queryable and relatable. Is that something you always try to do? Lauren: Yeah, exactly. I want it to be relatable for people, and how is it more relatable than when people can literally put in their postcode and see how all the stuff I've been calculating relates directly to them? I go from making things really complex and then in the end I try to shape it all into something that's very relatable and visually attractive for people to play around with or try out themselves. Paul: In your most recent piece, you're looking at the socioeconomic divide in London. It started with you cycling around the city -- you're a big cyclist. And you have a hunch. What is that hunch? You're sat in a coffee shop, I think? Lauren: Yeah, exactly. I have this habit of sometimes, when I have nothing to do, cycling to a random borough in London and going into their local coffee shop to see what people are talking about, what the discussion is, who's there. What I noticed is that depending on where I go, which local neighbourhood I go to, the discussions are completely different. There are completely different people sitting in there. I was basically wondering if I could formalise this a bit more rather than just going on my hunch. Paul: You talk about how when you go into a coffee shop, the coffee is always good. Lauren: That is mostly true. The coffee is nearly always good, but the people getting the coffee are different. And that's an indicator of your local community, but that local community can be very different just one street from the next. Paul: There's this divide that you had a hunch existed and you wanted to see if the data backed it. Where do you go to get this data? Can other people do this, or is it because you've done so much research you know where to look? Lauren: I would say everyone can do this -- you just need to know where to look. A lot of the data for this piece came from census data. The problem with census data is often that because these regional administrative units change, there are a lot of things that change over time. It's quite difficult to put it all together and create a coherent dataset. But in a sense, everyone can do this data work. A lot of the data I use in these Substacks I make available afterwards on my GitHub page. The main thing was the census data, and then I had data from other Substacks -- the restaurant data or the pub data, for instance. These are either publicly accessible or, in the case of Google Maps, I created the scraper myself, which I also shared on my GitHub so that everyone can basically have access to this data. Paul: And you use techniques that you share: PCA, K-means, gradient boosting, SHAP values are the four I picked up on reading it. So PCA -- what was the plan there and why did you need to use it? Lauren: So that's principal component analysis. I had a dataset of around 30 variables or something like that, and these variables all tell a different story and are highly correlated. What PCA does is compress this into the most variation. What explains the most? How can we push this into one score? The first principal component explains the most of the variance in this score. I just took that, and when you look at what underlies it, it's a sort of "poshness gradient" in this case. PCA was a shortcut to compress the data into something that was easier to use afterwards. Paul: So you've got all this data, you don't exactly know how it all explains things, but it correlates -- which means they sort of explain the same things. So you mush it together using PCA so that you've got fewer variables, and then you do what with it? What's the next step? Lauren: I do something called K-means clustering, whereby you define a certain number of clusters. In my case, because I wanted a clear division, I put down two clusters, and then it looks at whether certain clusters are forming in this data. If you define two, it will basically split up the data in the most logical way. I tried three or four clusters too, but I noticed that with two clusters I got the most coherent results. Paul: And so this is an example of unsupervised learning -- you give it the data, you say "I expect there to be two clusters," and you let it work out how to make that happen. And the clusters explain, as you described, those who make the lattes and those who don't: a socioeconomic divide. And it turned out it maps to geography as well. Lauren: Exactly. I didn't put any geography variables in there, but when I looked at where these clusters were, it lined up geographically. You could clearly see that lines were forming through London without me specifying anything. Paul: That's really cool. When you put the geographical data back in and plotted it, was there anything that surprised you geographically? Anything that popped out? Lauren: What I noticed is that a cross formed -- along the Thames, and through the middle of London where there are quite nice neighbourhoods. And then you see, often in the corners of London (if you had to make London into a square), those are the corners that are a bit behind socioeconomically speaking. Paul: And you then used gradient boosting and lastly SHAP values. What was the thinking going through that latter part of your workflow? Lauren: Basically, I wanted a way to explain what was going on: which variables are driving this, and stuff like that. To do that I used a machine learning method called gradient boosting, which builds decision trees -- a common machine learning method -- and does this until all the error term is explained as well as possible. By doing this I got an accuracy of 95%. I could explain, with the variables I had, 95% of the clusters I had formed. Then I could say which variables are driving this. The main driver turned out to be house prices, which makes sense, but it's also an outcome variable. What I really wanted to know is what's behind the house prices. What I then noticed is that education is a big driver. And then I did the same with SHAP values to see whether this holds more systematically. Paul: Really interesting. You started off looking at the data, the data said house prices is the thing that explains this divide, but actually that's more of an outcome. When you remove the value of the house, you revealed that education was the number one factor. When you look at the explanatory values, does that become actionable? If you want to improve your life in the future, education appears to be the biggest predictor of whether you'll be buying the lattes or making the lattes. Lauren: I would say so, but I also want to be clear this is not a causal argument. I don't do a within-subject study where someone gets educated and I track whether that person moves from one side of the latte line to the other -- which would be an interesting follow-up I could do as an academic. But yes, this is pure correlation. On average, people who live inside the latte line are more educated than people who live outside. Paul: Then you build a really interesting argument about the history of it, the political aspects of it, and you pick a particular story around Canary Wharf. Lauren: Yes. So I was looking at whether I could explain the patterns I see. The first story that came up is the wind direction. With the factories, the wind came from the east, so you probably didn't want to live in the east. The west was the better place to live. Then in 1947, the green belt established certain areas around London where they couldn't build, which meant that density became higher inwards and reinforced this division. Then we got the right-to-buy policy in the 1980s, where council houses were sold off to private individuals at a discount. On an individual level that's great, but for an area it removed the best council housing stock. And then the third one -- which for me was the most interesting because it was also the most recent -- was the development of Canary Wharf. That development went outside of the council and outside of any democratic approval. A new body was set up to develop the Docklands. Given that I have data from 2001, 2011, and 2021, I could see that during this development the share of working-class people around Canary Wharf changed. The very local people didn't change so much, but it was the two-to-four kilometre radius around it where working-class people were basically pushed away. You can see a stronger decline of the working-class share than the London average, which is evidence for the argument that a lot of the geographical and spatial distributions we see are driven by specific policies. Paul: And then you propose some restorative action, which I think embodies your approach to this. You believe strongly in justice -- that flows through a lot of what you do. Why is that so important, do you think? Lauren: I try to point out the problems we have currently, without only screaming that this is wrong. I want to also offer a solution. This is how we can go forward. There are certain policies we can reverse, certain actions we can take as individuals, or pressure we can put on policymakers to go a certain way. Paul: Really interesting. In your workflow, do you use generative AI? And if so, where and how? Lauren: I actually do use a lot of generative AI. It's also interesting to say that when I was doing my PhD, ChatGPT came out and started becoming more established. Back then I used ChatGPT as a tool to classify sentences in central bank communication. Now I use it in every stage of what I'm doing. First, in the brainstorming stage -- normally I go to some friends and ask what they think of an idea. I now also ask LLMs as a sort of sense check. Then in the second phase I have my idea, I start writing my own notes, start forming a story out of this, start gathering the data myself -- because unfortunately a lot of the data gathering is not so good yet with LLMs. But once I have everything in a central folder on my laptop I start: okay, what can you do with this? Often I have an idea -- you could do this machine learning model or this simple classification model -- and then I prompt the AI to write a pipeline for me, which saves me a lot of time especially in the data cleaning stuff. I'm also always very wary of the results it gives me, because often in these machine learning pipelines it's the small parameters we choose that matter a lot for the outcome. Especially if you use LLMs as a classification tool, the hyperparameters such as the temperature, or things like how many examples you add in a batch you throw to an LLM -- stuff like that really matters downstream. So I try to use the code the AI gave me, but then be very conscious about the decisions it made for me, question it on those decisions, and try different settings. It's always me interacting with the AI the way I would normally interact with a coworker or a research assistant. Paul: What excites me about that is: if I could do while doing my PhD what I can do now, I could experiment a hundred times faster. I spent so much time as a mathematician writing code that broke, then I'd fix it, and sometimes I'd set it running an experiment for a few days and then what came out was rubbish because I made a mistake. That loop now is so exciting. Research is going to speed up. You can now declaratively get an experiment going essentially. You do some final checks and you're good to go. Stuff that took me weeks, maybe months -- you can now spin that up so quickly. Can you talk a bit more about what tools you're using to get it done? Are you using VS Code, coding in Python? If you could suggest a framework to somebody who just wanted to play around with some data and ask some questions, what would you suggest? Lauren: My setup is: I open VS Code, which is really the go-to place for me. I have different programming languages I all access via VS Code. I put all my data in a folder called "data", I create a folder called "code" and in it I start off a Python file. Then I go to my AI of choice -- I currently use GitHub Copilot, which has access to Claude, ChatGPT, even Grok if you want, Gemini, etc. Then I start playing around with the agent modes. You give it a prompt and it will basically write the code for you. You can accept or decline certain things and you can always go back and override it. Then I have a code folder, a data folder, and I create an output folder where a lot of these visuals and in-between steps go -- CSV files produced in between that I can check to see what's going on. What I end up with is a very clear, clean pipeline, and if I want to go further I can always build on this basic structure. Paul: So we've just talked about using generative AI to build non-generative AI -- to do classical machine learning and data science. And I think that's a really powerful loop I'm seeing emerging in businesses. A lot of businesses have rich data that describes their operations, but problems they want to solve with it -- dashboards that describe their sales, what's going on in their factory, user data about some tool they're using. Our clients in particular are really starting to build these dashboards and get useful metrics. And the mistake you can make is to use generative AI in the loop, rather than using generative AI to build a data science pipeline that, once built, is deterministic. The problem with these models is that first of all there could be secret updates or constant changes, which also changes what comes out. And even if there are no changes or updates made to these models, because of their probabilistic nature you could get different answers at different times. If you want a regular data science pipeline, that's not what you want. You want a deterministic answer. You want it to always give one clear answer and you don't want a black box. What's good about these often simplistic machine learning pipelines is that you know exactly what decisions are made on the way, and you're in control of pushing those decisions and making sure the data that goes in gets used to the best extent. Paul: In your latest article, you talk about going to check out the latte line in Sydney and cycling around in Australia. What's on the scope for Lauren Leek? What's she going to be Substacking about next? Lauren: Basically I want to do whatever excites my curiosity at that moment. I still want to push through the analysis I've been doing on pubs and see whether in Australia there's indeed a different business model going on. I already hint at the fact that a lot of this is actually driven by poker machines and betting, which is a completely different business model. I don't know if it's better or worse than pubs closing in the UK. It's something I really want to explore. But more generally I will be looking at things on the intersection of public policy -- very policy-driven decisions which you can turn around, places where there's very interesting data. Something I've been carrying through is also this urbanism: what's happening to the cities, what's happening to communities in these cities, how do people live together. That's something I will keep pursuing. Paul: During your PhD you looked at communication with central banks. Can you talk to me a little bit about where that came from, what your goals were, and what you actually did? Lauren: Central banks make monetary policy decisions but they do a lot more than that. Their main goal is to keep inflation to a certain level, but their main tool for this -- other than interest rates -- is communication. The things they say on a daily basis really shape how the economy reacts. We know that central bankers talk very cryptically. They often say things between the lines. They try to be as boring as possible basically, just to make sure the message gets transferred to the market as clearly as possible. A lot of people go through the words these central bankers used manually, try to code it up, try to give it certain classifications. This took ages. When I started my PhD, the main way to speed this up was unsupervised text methods, but it was just not specific enough to read between the lines. This is where ChatGPT came in. At the beginning we started experimenting with: what's the underlying meaning of this sentence? Where is the central banker going with this? And we saw it was really good at detecting certain pressures that were forcing central bankers to say certain things. It could really go that one step further that a lot of the unsupervised methods didn't go to. It's very difficult to get a deterministic answer, to get the same answer consistently. So we had to do a lot of experimenting. We tried different models. We even fine-tuned an adapter in a model ourselves to get to better answers. In the end, we had all this data, we could run this at scale for every central bank in the world from 1997 until now. We got our classifications of the pressures on monetary policy that were going on, and then we wanted to make it accessible to the public. So we created a website, centralbankalk.eu, which allows you to access a dashboard and play around with a lot of this, to look up on a sentence classification level what pressure an LLM could detect in certain sentences. I think this was my first taste of making something quite technical and niche very accessible to the broader public. People could easily see that, say, the Dutch central bank was very much under pressure to say certain things -- and you don't have to know all the technical LLM models and all the machine learning behind it to get that. For me, this PhD project was the first step in bringing something quite technical and difficult to understand to a broader audience. Paul: And is it still up? Does it still work? Lauren: It's still up, it still works. My thesis is also on there, so if anyone is interested in checking that out. Paul: Okay. So we're going to start a business together -- me and you, Lauren. Lauren and Paul Inc. we'll call it. We're going to solve a problem with AI. What problem are we solving together? Lauren: Whoa. I mean, if I knew that, I guess we would have a very successful startup going on. A lot of my decisions are made by many of the apps I have -- Google Maps, the Times -- not only where I eat, but also the way I drive, stuff like that. And these platforms in general have so much data about me. I think one of the things we can solve as independent data scientists is seeing how these algorithms and the data they have about us really influence our daily lives. Paul: Okay, let's flesh that out a little bit. These large companies have loads of data on us, but they're also nudging us to change our behaviours. As exemplified by the restaurant search thing with Google Maps -- the choices that you have are somewhat chosen for us. Where else do you see that happening? What other pockets of data do you think exist that we don't really know about, but that are shaping our everyday? Lauren: I think it's something I'm also still thinking about, because once you're conscious of it you start seeing it everywhere. I think even the answers your LLM gives you -- I've been hearing stories that advertisements are already being trialled, and whether that feeds back into the answers they give, whether there's some kind of paid promotion behind this, would be something to keep an eye on. I would say also just the general Google results is something you keep an eye on. Even the Airbnb recommendation -- what gets put first, what gets put second. Same with Booking.com. All these platforms that people rely on to make decisions and people often see as very credible, but there is an interest behind it. Often there is a lot of money put into this to make sure that certain recommendations come up closer to the top and certain recommendations do better. Being aware of this and figuring this out in the hotel business, the restaurant business, even the AI business -- the generative AI you use -- I think is a promising avenue to pay attention to. Paul: Right. Lauren and Paul Incorporated is starting this week. Well, Lauren, lovely to speak to you. Thank you so much for being here on the Learning Curve and I'll speak to you again soon. Lauren: Yeah, thanks a lot for having me and yeah, speak soon.