ai.bythebay.io: Lyle Ungar, Measuring Psychological Traits using Social Media
Recording: ai.bythebay.io: Lyle Ungar, Measuring Psychological Traits using Social Media
. Good, so I want to talk about no neural nets or deep nets today, although I use them and I like them. And I want to talk about very little about question answering or facts. I want to point out that language is not just about facts and retrieving information and getting uber ordered, but in fact language is how people talk to each other, communicate, connect. It tells a lot about who you are. And I come today, although I'm a computer scientist, with work that's half actually psychology, and I want to argue that AI is in fact a large piece of understanding people and communicating with people, not just about fitting models. So communication is more than facts. As we get more and more better AI systems for chatbots, they're asking something like, can I help you, which could mean many things
It could mean, can I sell you my product today? It could mean, how are you feeling? Are you depressed? Do you need some help? I want to talk today about some of the techniques and research we've been doing to understand these sort of pieces better. So to try and actually ask this. So natural language processing has mostly been about what people say. Find the facts. Is the banana green? Is there a cat in the picture? Where is it? But I think it's also about how people say things, how they feel, what it says about them. And language varies. Men and women on average talk differently. Young and old talk differently
People talk differently to their parents than to their friends. They talk better when they're excited and happy than when they're at home. They talk differently when they're writing formal essays and when they're tweeting. All this varies and all this tells a lot about the person who is using the language. Also, we haven't talked much about data collection here, but a lot of AI work is finding the labels on the data. And my labels are not, did you find the right fact, but who are you, the person providing the data? We go into hospitals and say, do you use Facebook or Twitter? Will you share your Facebook post with us? Can we cross link them with your medical record? We get people to take questionnaires to measure their personality and correlate that with their Facebook. Get information about who the people are, which then drives their decisions, what they're doing. So I want to know what will your language tell me about who you are, what you care about, how you're going to consume health data, whether you're going to be absent from work, are you stressed, are you happy? So the approach at some level is really simple
And all the methods are going to be very simple today. So we get a bunch of volunteers to share some sort of social media. I'm going to talk mostly about language. We collect images as well. Instagram, Facebook, Twitter, take a survey, share some information. How much work have you missed in the last month? How many drinks did you have last night? How many friends have you talked to in the last three days? Get information about people. And we then go through the language. We, our computers, pull out the words
I have a very generous definition of a word. Correlate those words with survey results. And say, what do these words tell us? Build predictive models. You can be fancy and do deep learning and nice models. But today I mostly do really simple statistics to do things. And mostly today is embarrassingly simple. We take a document. Think of it something like a tweet or a Facebook post
And we count up how often each word shows up inside it. And then often we'll do some sort of clustering. How many people have seen LDA, latent duration allocation? Yeah, good number. Okay. So standard preprocessing, grab 100 million tweets, whatever, cluster the words together and find topics. The topics are nice. They end up being sets of words that co-occur that are related. Things you might not think in advance of putting together
I've got chicken, cheese, dinner, yum, soup, made rice, bacon, bread, yummy, fried, eating satellite. Something that tells about them. And then we take these as features that go into predictive models. Mostly fairly simple linear or logistic regression with very careful regularization. But we'll skip all that today. So every document is a mixture of topics. This is a simple sort of illustrative example, but actually real. Take a document the William Randolph Hearst Foundation will give
Each document is a mixture over topics. And each word in the document, okay, what topic did that come from? Then we go through and give arbitrary labels and call the topic children or education. The labels are made up and not so real. But just we get a set of words that tend to co-occur. And every document is now a mixture over topics. And I can look at your documents, your Facebook posts, and say what mixture of topics do you talk about? So that's it. So what are we going to do? First set of results. We have 70,000 volunteers
They shared their Facebook posts. They have to consent to Facebook and to me. Give me permission to download them. These are collected by our colleagues at Cambridge University. They say, are they male? Are they female? They take a standard personality questionnaire. There's a whole journal of personality and social psychology that basically publishes results of this standard personality questionnaire. Lots and lots of papers. Very well validated scale
Psychologists, oh, no, don't do the Myers-Briggs. Do something real. Use the ocean model, the five factor model. So we have questionnaire results and language. I'm going to show you some of those. But before you get there, let's look at something where we all have some intuition. How do men and women talk differently in Facebook? I got to apologize. This all is a little bit cliche before I show these, before I get shot on Women's Day for, but don't worry, I'll show men as well as women
And cliches go both ways. So what are we going to look at? We're going to look at the words that are most predictive of being a female Facebook user. And they'll do the same for males. Now most words are the same, but they aren't all. So what's it look like to be a Facebook, average Facebook user, female? That's cool. Excited and love you. Oh, by the way, some automatic grouping pairs of words like love you or so happy or my husband. So we do statistical correlation to group words
You old guys, you know what that little less than three means? Hard. Yeah, hard. Okay, you're not old enough or male enough. I don't know. Okay, good. Okay. And words like so with three O's, so with four O's, right? None of this normalization, none of this porter stemming, everything's a word, because it all reveals something about who you are. It's sort of obvious
All the coding I've used is that the size of the word, all this throughout the talk today, size is the strength of the correlation and the color is the frequency. So like it on or loves her is pretty rare, but reasonably indicative of being woman or yummy or hubby or herself. Cool. Okay, now before I get roasted for cliches about shopping and mommies, what's it look like to be a male? Let me point out it's not all bad. We talk about beards and engineering and shaving. Yeah, okay, sorry. Hidden amongst the clichéness, there are some actual psychological insights. Women talk about husband or boyfriend
Men talk about my girlfriend and my wife. When you go back and look at the original Facebook posts, women are talking about their husbands and their friends' husbands, statistically. Men are talking only about my wife. I never talk about your wife. Okay, I'm a statistician, you know, I get toward the data. You can interpret it as you want, okay? Cool. Okay, so far so good. So we can also look at the LDA clusters
And again, just to get a flavor of what you get by doing the clustering the data. What are women talking about? Happy birthdays, cute, adorable babies. So excited. I love the O's. And the love and the wonderful families and boyfriends. What are guys talking about? Government, football, fighting, Xbox. Oh, and swearing. A lot of swearing
So we get some intuition. Okay, that's great. But we know about male and female. Can we do something more like say, well, before I get there first? How accurate is this? 92% accuracy, male or female, based on the words. Not perfect, but good enough for a lot of marketing purposes that it's reasonably useful. Okay, let's switch to personality, five-factor model. Start with extroversion. What's it like to be an extrovert? Not bad, actually
I want to point out that some of these can't wait. What's interesting about that particular extroversion signal? How do they spell can't? No time to put that apostrophe in there. I can't wait. And these are controlled for age and sex. And so this is gender balance. You'll see girls and boys and parties and getting and chilling and life's pretty good. Unfortunately, computer scientists tend to be somewhat more toward the introvert side. Okay, again, I'm not sure it's bad
Pokemon and anime. And these are all reflecting the time period. These are about three or four years old now, so they're a little bit dated. Apparently, related, drawing. There's these sort of words of thinking a little bit. A little bit more mentioning of depression, which actually does correlate somewhat with introversion. You know, Doctor Who's not talking about very much, but guess what? It's more on the introvert side than... Yeah, it's sort of fun
And books and reading and pieces there. Cool. So we can do a couple more of these, and we will. Let's do one or two more. So well-adjusted... Oh, so these are neurotic. people, you do not want to be neurotic. This is generally bad
Not again that words are really good, and I'm being very simple about them, but statistically, finding pair of words that show up more often together by chance than you would expect gives you really nice disambiguation. Neurotic people talk about being sick of. They're not sick. They're sick of. And it's really nice to have a little poor man's word sense disambiguation just find pairs of words that go together. If you talk about depression or depressed, you're highly neurotic, but people don't talk about it much on Facebook. That's not something they're going to put on their Facebook accounts. We'll talk later more about depression
Anti-neurotic or well-adjusted Americans... This is what you wanted. What are people talking about? The Lakers. The Lakers, yeah, and the Celtics. So sports. Now it's interesting. Doctors know that if you're depressed, getting out and exercising is good, but I've never found any published literature, and half my colleagues are in this paper or psychologists, where it says watching sports is good for you. Now maybe it's because you're closer to the normative average person
Maybe it's a bonding experience, a group experience. You can get together with your friends. But it is interesting to me that sports and talking about sports on Facebook correlates with being well-adjusted. And lots of them, right? It's volleyball and snowboarding and soccer, things you do, workout, things you don't do, watching the Lakers, right? There's not a lot of Lakers posters in this Facebook set. And then, of course, one that's not surprising, religion. Religious people in America are more agreeable. They're happier. Well, actually, they have more diabetes and obesity, too, but that's a different story
Maybe the church dinners, but... So how well do we do? We look at the big five, openness, conscientiousness, extroversion, agreeable neuroticism. In blue, I have the accuracy, more as higher correlation of our models. In red, we had a bunch of people, friends, Facebook friends fill out the same questionnaires. And we're about as equal as your friends are in estimating your personality. That's cool. We do it with language. Other people have done this with Facebook likes
It's about as equally accurate, get a flavor. What movies and shows and stuff you like tells a lot, again, whether you're gay or straight, whether you're extroverted, you like different stuff. So it's cool. How is this being used? Since there's some entrepreneurial table. IBM has a product out that does a, not nearly as accurate as mine, but a fine version of categorizing personality. They're selling pieces there. There has been a lot of news recently of starting to price everything like car insurance. If you're conscientious, you're going to drive more carefully than if you're a little sloppier
That's part of being conscientious there. There's been a lot of press before Cambridge Analytica. They were claiming for a while that they won the election for Trump. That they do targeted personality based advertising. I don't know. Two days ago, they announced we didn't do any cypher graphics with the Trump campaign. So now they're saying they didn't do it. So it's gone back and forth
But they've gotten a huge amount of press for a lead. Well, they do have, I've talked to them. They do have a database of basically every Republican in the US and have pulled out huge amounts of text and information about them for targeting marketing. I personally don't think that targeting by personality is nearly as effective as 20 other ways of targeting people. Like how unhappy are you with the current political climate? It's a good way to target. So I'm not sure psychographics are the way to go. But they've certainly got a lot of press where they're either both claiming and denying being involved. Interesting to watch
So that's personality. We also go to the hospital and ask people if they're feeling depressed and check their clinical record for depression. It'd be really nice if people say you're pregnant, probably not me, you're at risk for being depressed. A lot of postpartum depression, even prepartum depression. Could people volunteer to share their Facebook and let us monitor it for signs of depression and then notify you, your doctor, your husband if you have one, your partner, can we get something that shares that? Can we understand depression? When you look at the questionnaires used to assess depression, there are a couple of different factors. I'm going to show just two of them today for a time. One is a set of factors which are labeled low mood. I often feel blue
I'm often down in the dumps. And one which you would view as low self-worth. I don't feel comfortable with myself. I'm not pleased with myself. I dislike myself. So we have these questionnaires from a bunch of people, thousands, and we have their Facebook posts. We can say, okay, what language correlates with low mood? Alone. If you say the word depressed or depression, you're in a low mood, but people don't use those a lot in Facebook
If you say I'm alone or anymore, it's a very negative sort of term actually. People don't love anymore. Someone. Those sort of terms are indicative of low mood. Another factor of depression, the low self-worth. Biggest word? Why. If you're posting why on Facebook, it's not because you're a scientist trying to understand the nature of the world statistically. You may be different
It's actually apparently you're not sure of things. You have low self-worth. You're a little bit of a loan up there, but much more don't and would and probably and trying. Often you see words you weren't expecting that after the fact, oh yeah, okay, I see what's going on. These same sort of methods are used a lot for sentiment analysis and products. But I think it's interesting to look not just at what do you think about this product, but how do you feel? What do you care about? What mood are you in? We can then build these models. We've got models personally based on 70,000 labeled people and then go to new people and not give them questionnaires. I personally hate filling out questionnaires
On the other hand, I'm willing to share my data to anyone who's going to offer me something useful, like my doctor. So we go to the hospital and say, great, share your Facebook with us, please, and Instagram and Twitter and whatever. And we will then estimate the personality and estimate the words and see how they correlate with anything. And I'll show you just one result looking at depression. People with a clinical diagnosis of depression, these are LDA topics, tend to talk more and the, this is the law of the fine print, don't worry about, but it says compared to only looking at age and sex and race, all of which are predictive, right? Women get depressed a lot more than men. But controlling for all those statistics, it's still the case that people who talk more about headaches, somatic symptoms, stomach hurts, people talking, and these are people in the hospital, we're recruiting them in the ER, talking about doctors more, likely to be depressed. People talking about somebody and help and please, alone, which we saw before. Get this flavor of what are the depressed people looking like in our hospital sample
And it's something that doctors should be looking for, but mostly don't. It's a really common problem. It leads to lots of problems with treatment because depressed people are less likely to follow through, lower compliance adherence. We'd like to know about who they are. It's ridiculous that everybody walking to the hospital is given the exact same discharge instructions. They hand me a piece of paper and I go, I'm a computer scientist. Don't give me paper. There are people who are almost illiterate, people who are PhDs
They're giving the same, you know, can we profile communications more with who we're talking to? So I want AI to recognize that communications are at people who are in moods, who have certain personalities, depression states, understand that. What else can we do? Well, I think increasingly as more and more of your communications are available to the world or your company, how many people use Slack? Somebody's in-house? Yes, I'm in Silicon Valley. Okay, cool. Almost north. So you can read the stream and see how someone's doing. Are they engaged? Are they disengaged? What about stress? What are people talking about who are stressed? What about stress? Headaches? Feel like crap. Body hurts. Stupid people
Okay, it's annoying. I hope this is not you. Pain. Depression, not talked about very much, but again correlates with stress. Anxiety. Me and I, interestingly, in retrospect, obvious. If you're depressed, you'll talk more about I than about you. If you're stressed, you'll talk more about it
Didn't know that, but now I do. And tired of, again, not to say tired, tired of my head. And people who are low stress, what do they talk about? I like it. Breakfast. If you're posting out bacon and breakfast on Facebook, you're very chill. You're waking up. You're alarm clock there. You're talking about yesterday and tomorrow
You're so excited. You can be nervous and stoked. You're feeling accomplished. It's not that the stressed people are sitting, unstressed people are sitting on beaches drinking margaritas. They're still engaged. In fact, there's a very nice theory which we're working on. I have the results. But I think we'll pan out that there's good stress and bad stress
People believed in the 50s there was a type A personality and stress caused heart attack. Not true. If you're stressed because you're working on that project that you really wanted to get done and you have a deadline for tomorrow, or if you're CEO of your startup and you're working insane hours, no bad health effects. If you're stressed because your boss is saying you need to do that, but she won't give you the right resources, it's a stupid project, and you're just working late at night on something that has no point, more cardiovascular disease, more suffering. So I think there's good and bad stress, and I'd like to be able to measure it, quantify it, know about people, know how to help address it. Hey, you're looking a little bad stressed. Are we not, is that boss not empowering his group enough? Not giving them enough autonomy? What can we do to do that? And measurement's the first piece there. Cool
I want to shift from modeling people to modeling communities. One can profile whole groups, whole sets of people, as well as profiling individual ones. So instead of now taking Facebook and individuals, I will take tweets, it's easy to grab 10 billion tweets. You can map maybe a quarter of them to the US county they come from, if they're from the US. A tiny number, a few percentage actually have geocode, latitude and longitude, but a lot of them the profile has some address. If it's from the US and says San Francisco, it's probably from San Francisco. Some of them say, my backyard, or the universe, or a state of hell. Okay
So three quarters I throw out. But okay, I'm down to, you know, only a couple billion left. Fine. So it's a reasonable measurement piece. So I've got a lot of tweets mapped to the US county. And then the government, namely the census department, does a lot of surveys about who you are. And the centers for disease control do a lot of surveys about what people are suffering from, from diseases. Actually, mostly don't
We know exactly what people died from fairly well. What you're actually suffering from is pretty badly measured. That's a different story. So we can now take by each county, the language used in Twitter, the demographics, which we'll control for, because it matters. Health varies a lot with age and sex and race. And health outcomes, mortality, or happiness outcomes. So I'm going to show you some of those results. Just one quick slide
On a scale of one to five, how satisfied are you with your life? What LDA topics? Same topics. We do 2000 topics for Facebook. In fact, I use the Facebook topics here on Twitter. Okay, I'm lazy. But it's okay. What are counties that are satisfied with life talking about? Personal training, suggestions, meetings. Okay, the red one is ones that are unhappy. The bored and tired
Where does money show up in terms of happiness? Haiti, donation. Sure you see when this is collected from. Donation. Happy counties talk more about donations. In fact, if you look at the poor counties, there's actually no correlation. Among the wealthy counties, the ones that talk about donations are happier. And the ones that talk less about donations are less happy on average. Profile of a community
And that's consistent with individual level results. Psychologists do these great experiments. Here's 10 bucks. Go buy yourself a nice little present. Come back in two hours and how happy are you? Here's 10 bucks. Go buy one of your friends a nice little present. Come back in three hours. Ask me how, ask how happy you are
Who's happier? Buy the 10 bucks present for your friends, not the 10 bucks one for yourself. So this correlates individual level models, group community level models to understand. Communities have personalities too, as do people. Different way of understanding them, different use of data. They're trying to look at those. So let's look at cardiovascular disease and a particular mortality from arteriosclerotic heart disease. Dying of hardening of the arteries. Pretty common cause of death in the US
We know what counties. On the left is my estimate of the east coast where people are dying or not dying based on CDC reports. On the right, a model fit from our data. But I'm not interested in predicting and I know how many people died, as CDC told me. What can we learn about it? So here are a bunch of different pieces of prediction. Looking at correlations. We start with information about demographics, males and females. Blacks have a higher rate of mortality from heart disease than whites
Obesity is bad. Hypertension is bad. Diabetes and smoking are really bad. The most important single factor, if you can choose something to be healthy in the US. Be educated. Wealthy is not bad. Educated is even better. So those are highly predictive
More educated counties are healthier. Put those all together. You're almost as good as just using Twitter. Now, Twitter doesn't know who has diabetes and who has hypertension and who's smoking and who's black. Well, they do sort of. I can predict race pretty well from Twitter. On average, people differ. But on average, males and females and blacks and whites tweet differently
And what I'm getting is a significant boost just using Twitter to predict dying of heart disease over all these standard factors that epidemiologists use. That's sort of cool. But I'm interested in insights. Prediction is not useful. What can we see? What sort of words? And there are a bunch that reflect socioeconomic status. So I've cherry picked a few that are more psychological. Higher status occupations. Counties that talk about skills and conferences and meetings
Oh, we're pretty good. What places have higher mortality? Hate, bull, past, despise, and being tired and bored and exhausted. Now, this is a little bit weird. The first set of data I showed you, are you depressed? Show me your Facebook. The people tweeting here are not the people dying of heart disease, right? The people tweeting are 20 or 30 year olds, 40 year olds. The people dying are mostly 60 and 70. It's a different set of people tweeting than dying. So the community has a personality
You're tweeting about bull and you're tweeting about anger and I'm dying of heart disease. Right? It's reflective of the community that you're a piece of. So it's an interesting piece that it's showing what the personality of the community is and how that reflects overall health. As we move toward more interventions, I've talked to a lot of companies that are interested in improving, happily, the welfare of their employees. How do we measure that? Historically, companies have hired Gallup or Qualtrics or these companies to go in and do polls and ask people lots of questions once every year, which then for reasons obscure to me takes three months until they actually give the results back to the managers. The processing is so expensive. Okay. How sad? But now one can actually get people to share
And I should point out the obvious, if you're working for a company, they own all of your slack information and all your email that comes to the corporate account and all the text on the corporate computer. Okay. This is privacy issues. You want to be a little bit not creepy about this, but to get some profiling of what's happening. How are people feeling? What is the culture and subculture and different subgroups within the community? I think it's really nice to be able to measure community well-being and particularly as one does more interventions. How are these working? How are they not? How did that project affect it? How did that new hire shift pieces? So what am I arguing? I'm arguing that language tells a lot about who you are. It can reveal sort of obvious things like age. And a lot of the words indicative are ones you'd never think of
The used more by old people and males. The usage of the word the is radically declining over the 20th and 21st century. You count the number of thes, the percentage of words in presidential addresses. It's half what it was under Obama's, half of what it was 100 years earlier. Going way down. Okay. It makes sense with women and young people. Who are the fastest language change agents? Young people and women
Men and old people lag behind in change. Okay. Interesting pieces. Use of the word I. Depressed people use the word I more than non-depressed. We saw a couple of examples there. Stress you get more. But lots of these words, not just the common words that psychologists have found, but lots of much rarer ones
We can do personality, do emotion, anger, happiness, stress, burnout, life satisfaction I showed you. I didn't show you a bunch of results on political orientation. Not surprisingly, Democrats and Republicans talk differently, but even more importantly, very politically engaged versus non-engaged. Different pieces there. There's a bunch of interesting work coming out suggesting that when you look at what drives elections, it's not the obvious things of economy, which everybody quotes and does make a difference. It's life satisfaction. How happy, how satisfied am I? And even more, how do I think things will be for me in five years? So if you look at Brexit, Trump, these surprising results, the pollsters were not very good at capturing how people felt. Right? Who would you vote for? They know very accurately
They don't know as accurately, will you vote or not? And they really don't know, rather surprisingly, given the amount of work that goes into targeted advertising and selling candidates, how satisfied are you with how things are going? How optimistic are you about how your life will be in five years? Politicians tend to not think as much about that as maybe they should. We're measuring empathy. There's a big movement that says that you should train doctors to be empathetic, care about your patients, feel their pain. Oh, this is a terrible idea. You're treating dying cancer patients all the time. You don't want to feel their pain. You want to be compassionate. You want to care about them, but you don't want to be flooded with their pain
I have lots of beautiful word slides I won't show you that capture how that feeling varies. These all tie together. So what are we seeing? A move, I think, as one looks at this whole sort of chatbot conversational agents. The AI world has worried a lot about what people are saying. What product can I sell you? What attributes do you like about this product or dislike? What's in this picture? What's the right label for it? Lots of talking about what. I think that's great. And I think most of us will stay employed using AI to solve what problems. I like taking my Uber or Lyft and having it show up at the right time
I like being able to talk to my phone. This is all great. But I'm pitching for something more. What I want is for AI also to realize that communications are not just about conveying information. When you're having a drink with someone, hopefully an hour up there and check, it's partly networking learning facts, but it's partly how do I feel about it? Is this someone that I trust? Can we connect? Can we work together? And AI has been really fairly pathetic about looking at people and saying, is that someone I could work for? Is that a good personality fit between me and that company? Me and that boss? What can I pull together? So I'm making a pitch here that now the data are available, not that hard to get about the language people are using, the images that they're posting, their photographs. There's also great work, but it's much harder. I don't do it on listening to people's speech and looking at their motions to do personality assessment. But I think this is part of AI is understanding people and not just understanding facts
With that, I want to thank you. And I want to thank the very large team of people that helped me make this happen. Thank you. Hi. Love the work. This is a wonderful talk and especially thinking about the how, the question of how people are speaking is very, very insightful. I want to ask you a question about how do you control for the fact that a lot of people have personas in social media? You know, like the, there are different communities. You said before you speak differently to your friends than you do to your parents
And knowing that these different silos of my social network, I might be speaking differently. Like, how do you control for that in your research? Great question. There's sort of two pieces, one of which is not fatal at all. One is that everybody tries to make themselves look a little bit better. You tend to try and look happier on Facebook or Twitter. And that piece doesn't bother me too much because the statistical correlations are still there. The other problem is much worse. Communities use social media really differently
My community tweets mostly about business and about the conferences we're at and the latest product releases we have. Down the street from me, across the road in Camden, are a bunch of 19-year-old black women. They're tweeting the way that I text to their closest friends. They're saying, hey, let's hook up at the bar tonight. So they're using the media really differently. And I think it is something that's important. I think something that's been rather overlooked in most of the research is the context. Which you're socially talking about one kind of context for understanding a question
But the broader context is, of course, you're different at home and at work and different communities is different. So I think there's a bunch more work to be looked at how does usage vary across communities. I think there's also a real danger that as you build these models, you get something that works really well for the mainstream people. These are whites in their 20s and 30s in big cities. And then you go on to try and apply it to 15-year-old Hispanics in the countryside. And all of a sudden your model is crap. So I think this sort of transfer learning of how do you work across them is a widely open problem. And I don't have a good way except to agree with you, yeah, I got to watch out for that
Thank you. We'll do one more question right here. Thanks so much for your talk. There's a way in which marketing is designed to change behavior at scale. I wonder if your project gives any insight into influencing well-being at scale. Yes, so we would like to influence well-being at scale. My project is called the World Well-Being Project. You can see all these slides at www.p.org
So I do want to influence well-being. And the first one I'm going to influence well-being actually is really stupidly, that which is measured can be changed. And governments have good measurements of the economy. Governments have crappy measures of well-being. So my first attempt is incredibly stupid. I want every person in government to have a little well-being meter. And it's not just a lot of happiness. It's quality of relationships
It's feeling of meaning and accomplishment in life. Am I contributing to my community? So I want subtle meanings of well-being. But I want to give measurements of well-being. Then we can start to ask interventions. Okay, if we make free yoga training available or meditation available to everyone in the community, does that help? Then we can start to measure it. So I do want interventions at scale. I want targeted marketing of well-being. I want to sell health and happiness and satisfaction with life
I want to sell go for a walk with your friends. No money in that, but it's a good idea. I guarantee you, if you remember one thing, exercise is good for you. Being with friends is good for you. Okay, it's two things. So, but my argument is measurement is the first way to start driving change. Thank you. So exercise, well, a group walk is a nice day, right? Yep
No, great. Thank you, again. Great. Thank you guys. Thank you.