Transcript#

This transcript was generated automatically and may contain errors.

Welcome to the test set. Here we talk with some of the brightest thinkers and tinkerers in statistical analysis, scientific computing, and machine learning, digging into what makes them tick, plus the insights, experiments, and OMG moments that shape the field.

When I asked Joe Cheng what gets him out of bed in the morning, he told me it alternates between duty, curiosity, and fear. Joe is the CTO of Posit and the creator of Shiny, a dashboarding tool heavily used in the R&P communities.

In this episode, Joe unpacks that answer — the quiet routine of duty days, what it means to be a reluctant leader, and why building AI tools today feels like excitement mixed with existential dread. We're joined by his longtime collaborator and my co-host Hadley Wickham. They're opposites in a lot of ways — Hadley all conviction, Joe all caution — but they've spent a decade riffing off one another. And honestly, I love their dynamic, so I'm so excited for you to listen to Joe Cheng.

All right, Joe, welcome to the test set. I feel like this is a really stacked interview because obviously you're the CTO at Posit, joined by our fabulous co-host Hadley, who's chief scientist at Posit. I'm really excited about, honestly, having worked a bit with the Shiny team, your work on Shiny, but also so much more stuff. You work as CTO at Posit and obviously everything you're doing with AI. So just super excited to have you on.

Yeah, glad to be here.

Maybe even just kick it off, could you tell us a bit about how you and Hadley have interacted over the years? Do you hang out? What's that like?

Yeah. Yeah, I think Hadley and I, we started working together in 2012, when you joined us at our studio at the time. And it's a little embarrassing how funny this is going to be, but one of our first interactions — and me not really knowing what to expect from you, just being the creator of ggplot, and that was about all I knew, and someone who could help us kind of bring some taste, like our data scientist taste, to what we were doing.

But you asked me a question about parsing CSVs. Do you remember this? And I gave you basically the beginning of an explanation. And I said, you want to use a deterministic finite automata, and you can think of it as moving through a few states, and you were basically like, got it. And I was like, okay, come back when you have questions. And you came back with like, you mean like this, and you had fully sort of mapped it out.

And I think that pattern has been pretty consistent with us working together, is that it feels like when I work with Hadley, I just have to say the beginning part of a sentence, and he's like, I got it. And he actually gets it. Whereas I think most people — Like, stop talking, I'm bored. Yeah, yeah, say less, say less, yeah.

Whereas I think especially at the time, a lot of people that I was working with that did not come from a software engineering background, they came from more data. I was just used to, okay, I'm going to have to repeat this a number of times, we're going to have to approach it at multiple levels of abstraction before they kind of fully understand. And with Hadley, it's always been this shorthand.

And that really carried into when Hadley wrote Mastering Shiny, the book that I wanted to write but was not able to. And Hadley, you know, it's easy for him to write books, it's easy for you to write books, but you were missing so much, not just details about Shiny, but the details about why things were a certain way.

And I don't think I could have transferred the information to someone else, because there was so much information to transfer. But I felt like with you, Hadley, it was like, in 15 minutes, I could convey months of work of, you know, really intricate details about the way reactivity works, and why there's sort of like this push pull effects to reactivity. And I really felt like for someone else, each one of those points would have taken like a couple of days to really get them to understand. And with Hadley, it was always just like, okay, I got it.

Push is, it's a very, it's an exciting way to work, I think, when like we're able to transfer that kind of context, like almost as fast as we can, we can talk. And I hope that like, when we work together, I'm able to hold my own enough that, you know, it's a pleasant experience for Hadley also.

No, I think one of my like, biggest regrets of like, Studio Stage Positive is just that we haven't had like more projects to work on together that we could like really crank on. Because that's, yeah, like I love writing Mastering Shiny too, because I'm like, okay, yeah, I can ask these questions, and I get clear, obvious answers back. And that makes sense. And I can like push you on bits where I'm like, I like, I wouldn't have done it this way. I mean, get like, get answers that I'm like, okay, yeah, that's, I didn't think about that.

From the outside, like having seen Mastering Shiny before I knew the context, it was so surprising, like, just to be like, Hadley, what are you doing here? Like Mastering Shiny, but it is so valuable. Yeah, I could see the value of like really picking up something you're like less familiar with and that that transfers to people who are reading it, that you had to kind of really wrap your mind around it and versus a person who built the thing and kind of knows it intimately, figure it out.

Joe is one of the many people I've given my advice to, like writing a book is easy. You just have to write for like an hour a day. And then like after a year, like you'll have a book, but I don't believe like anyone has successfully followed that advice to complete the book. So it's clearly not good advice.

Joe-shaped problems and coming from academia

Hadley, do you think about like, when you think about throwing things at Joe, like what, what types of problems stand out to you as like Joe shaped problems? What would you be like? We got to put Joe on this.

Thinking back to the early days, like coming from like academia to Posit, it was just amazing to feel like my ideas were like respected. I was like, Oh, like, like these people just like believe what I am doing is like useful and important, just so like freeing and then be able to be able to like interact with them on that like technical level.

And to never, there's definitely this like kind of, I sometimes feel myself falling back into this when I like visit universities, but there's kind of like academic one-upsmanship where you're like, you know, everyone's a little insecure in their positions. So it's about like making sure you're kind of like, you know, making yourself feel good and other people feel not so good.

Because like one of the things I felt when I was in academia, it was like so hard to ask questions about things that I did not know, that it felt like every statistician should know. And there was like never that at RStudio. Like I could ask Joe and JJ what seemed like the dumbest questions. And they were just like endlessly patient with me. And it didn't like, it didn't feel like they were like, Oh, what an idiot. He doesn't know this. They're just like, yeah. Like people are fairly good at some things and don't know about other things. It's just, yeah. All in stuff together. That was, that was like, so that was just an incredibly like freeing experience for me.

I could see that coming out of like those like classic academic brown bags where you're getting like hammered at the end.

Curiosity, duty, and fear

Joe I know you probably gave me one of the most intriguing answers to a question I asked before this, which was what gets you out of bed in the morning? And you said a mix of curiosity, duty and or fear. I, I was going to ask you and have you repeat that, but I'm actually so intrigued. I thought I'd just lay it out there and, and try to get your take because that I've been just haunted by this statement. And I'm so curious about that.

So, um, I think on my best days, it's curiosity on the best days. It's like, I have this burning idea and I can't wait to see if it works. Those days are not often. But when they happen, not only are they so fun and so exciting, but often those are the days where there's like a big breakthrough, you know, um, and once in a while, if lightning really strikes, you know, the early days of Shiny, it was like that every day. Uh, the, the early days of working on, uh, LLM APIs was like that every day. Um, and that's, those are just like highlights.

I think like 90% of days it's duty that it gets me out of bed. It's just this feeling like the company's counting on me. My team's counting on me. My family's counting on me. Our users are counting on me and, uh, it's just time to get up and make the doughnuts, you know? Um, and there is something that I think is very noble about that. There's something that I think there's something very satisfying about just quietly doing your job when there's no glamor in it. Um, it just, that just feels like that's most days and that's okay.

There's something that I think there's something very satisfying about just quietly doing your job when there's no glamor in it.

Um, and then there's the fear days where I'm, I'm driven by a sense of anxiety or dread about a got to stay ahead of our competitors, got like the, or the fear of like being embarrassed on stage, you know, or the fear that we're, that I'm personally falling behind in adopting some new technology or understanding some new hot concepts. Uh, those are the worst, but fear days, I think they can easily lead to working really hard in unproductive directions. Um, or sometimes you can counterproductive directions because like the impulse to do something is so strong as like a way to like cathartic cathartically bat away the fear.

Um, and unfortunately I feel like there's been way, a disproportionate amount of fear for the last year just because AI is just, it's, it's just moving the playing field so much and so often and so quickly. And uh, and I, it's, I don't think it's, it's just me, it's just us at Posit. It's the whole industry. I think, um, it, it's yeah, it's just like a time of fear, like excitement, yes, but also like this just existential dread and worry that underlies everything.

Shiny's origin and the fear of missing out

I find that framing so interesting and there's so much to unpack there, but I, it does also kind of remind me of your like 2022 RStudio Conf keynote, where you talk a little bit about, um, creating Shiny and some of that environment. Like as I understand it, you mentioned like being on the verge of quitting essentially before kind of like hitting that, um, key idea.

So I almost wonder, like, I'm really interested to unpack a lot of what you said about fear in the context of AI and in today, but, but to get there, I'm almost, I'm curious if you could bring us back to like 2012 and some of what you mentioned at the keynote, like that, like mix of curiosity, duty, and fear in 2020, 2012 was Shiny.

It's funny. I didn't make a connection to that moment, but you're exactly right. Like that's exactly a perfect illustration of what I'm talking about, where I think, um, I started working at RStudio with JJ in 20, uh, 2009. So by 2012, it had been about three years and, um, at the time three years felt like a long time to me. Uh, but also I just, I felt like I've been working with JJ for a number of years. I think the world of JJ, I really, I mean, it was, the whole point was to work with him. That was the whole reason for joining RStudio was to keep working with JJ, but I had this feeling like, is this all that I get to be?

You know, um, a lot of my friends were, uh, at that point starting companies or exiting companies, you know, um, like getting successful exits. Others were, um, I had a friend who was like the head of project, uh, or product management for Android. You know what I mean? Like just people who were like my classmates and peers that I was like, if I don't start a company, who am I? Like, like, can I even say that I've been successful? And I was really consumed by that.

Like I think what I would now call FOMO, you know, um, but at the time I just felt like I'm just, am I okay with this? Like I'm just going to be a failure, not a failure, but like my, my world is so small compared to, um, you know, people I considered peers.

So I didn't have an idea, uh, like I didn't know what kind of company I wanted to start. I just knew like, I need to be a founder otherwise, um, I won't consider myself successful. And um, yeah, like JJ was super supportive about it and he was like, okay, I mean, like I think you should have an idea first, but like whatever you decide, I'll, I'll support you in that. And um, yeah, it was, I, I, I committed to him, like I will be with you and focus only

on our studio until the last day of the USAR conference that year. And I woke up on the last day, um, you know, basically the last day before I was going to start a several week soul searching process that was going to hopefully turn into, I have a killer idea and I leave and, you know, start that company. Um, and that was the day I thought of Shiny.

And I think I described it as like a switch being flipped. Like it really was like a night and day difference before and after that, um, that like thunderbolt of, of, um, Shiny struck me and like that I had thought that this was about me really wanting to start a company really, whatever, but it just, I don't know that, that a technical idea could so completely erase that feeling, um, of needing to prove myself, needing to go do this, needing to measure up to my peers. Um, I never would have, would have predicted that.

So that, I mean, it really was like a different mode. Like my brain entered like, okay, let's build, you know, um, and who cares, who cares what your title is while you're building it? Who cares, uh, you know, what the, what people are going to, whether people are going to credit you for this or whether you get to call yourself a founder. Um, it was just like, let's build, like, this is going to be awesome if it works. Like how could we not, how could we not just focus all of our energy on, on doing this?

Um, that was like a huge lesson for me that like the thing that you think is the problem, the thing you, the thing that you think is causing you unrest, like maybe it's not the thing. But the biggest takeaway is like, I feel like it would have been miserable starting a company to start a company. It would have been miserable trying to be a successful founder so that I could measure up to, um, by the way, people that I will never catch. Like I could not have kept caught had I started my own thing. I truly believe that, uh, who are just born to do, you know, big things.

And I think I was born to build little medium size, little to medium sized things. Um, and that in, when the right idea strikes, that's more than enough for me, you know.

And I think I was born to build little medium size, little to medium sized things. Um, and that in, when the right idea strikes, that's more than enough for me, you know.

It's interesting to hear too, like, if I understand you're saying that at the time, like a lot of your framing was like, I need to start a company and be successful or do a big thing. But looking back, it sounds like you're saying like, you really appreciate the level of support that you had at the time to kind of find something and to work on it that actually like you're realizing like the position you're in was pretty good and actually like a good fit for you. Yes. I think there are a couple of reasons. Like the first was, yes, I didn't, I didn't have to do all the other work that comes with starting a company.

Um, the second is like having JJ there, I have like a built in co-founder and I know, I mean, Hadley knows this about me. Well, like I, I have very high levels of self-doubt, uh, which is often helpful, you know, like it's often helpful because it, it stops me from getting too lost in my own blind spots most of the time. But to be a true, especially solo founder, you need conviction, uh, and you need it all the time. Um, that's very difficult for me.

And then the third was like, I didn't want to build this and not have people use it. And we had our studio, like it was working, you know, we had a community that was building and I, it was more important to me that this thing launched to a receptive audience than it was for me to, I don't know, have a hundred percent of the equity or whatever it was, you know, um, that, that might've been competing with that priority. So yeah, I, I never, I didn't even really give it that much thought. It was like, as soon as it was, oh, this is the idea. It was like, yeah, I'm, I'm obviously doing it here.

Shiny's impact and the move to Python

I mean, Shiny is like vastly more successful than most startups. Like, even if you just like looked at the impact of like Shiny, um, Bama, that's, that's like pretty, pretty incredible. No, it happened like pretty quickly, like it was like that, like people like that was like, when I remember when I first heard about Joe first heard about Shiny, I was like, oh, a web development framework for, ah, wow. That's something I really need. Uh, but then it like, it turns out actually that is what I need. Like people need a way to like in their ideas and the little interactive apps.

And it was like such a killer, like it just made like data scientists look so impressive to people because they're like, oh, I just thought you were this like data scientist and now you're an app developer. Like I remember like people in the early days who were like, you know, you Shiny saved me $200,000 because I prototype, I could prototype this app in a weekend instead of hiring a developer to do it for me. Like it was, yeah, pretty crazy.

I guess jumping forward, it seems like 2022 was another really kind of big point where you were really focused on like Shiny for Python and our, our studio, I think at that same cover, I think you announced in that keynote, uh, our studio changing its name to Posit. Uh, Hadley and JJ announced it. Yeah. That morning. Yeah. That's right. Yeah. Yeah. What was 2022 like then? Like thinking about what, what do Python users need and kind of that, how was that for you?

Yeah. Um, it was, yeah, a really fascinating couple of years there. It's a complicated story because we've done Python stuff for a while, right? Like we did, um, we did reticulate and there was some Python support in our studio, but I think it was, it was clearly like, this is an interop story, not a, we're not like actually targeting, um, you know, making Python users lives better, uh, until like 2021, 2022. And, um, we had talked about it many times. I think like at a lot of company work weeks, you know, the leadership team would get together and look at each other and be like, Python, Python, anyone, you know, and, um, and it always felt for years, we kept saying, no, we can't afford to dilute our focus. There's so much to do for R that we can't afford to dilute our focus. Um, and also like, what kind of message is this going to send to the community? Like is this, does this really make sense for us to do?

And, um, what I recall, you correct me if I'm wrong, Hadley, is that when we started talking about the idea of Posit being around for the very, very long term, right? This is not just a company we want to be around for 10 years. We want it to potentially be around for a hundred years. Uh, that was when it went from, should we do stuff with Python to when do we do stuff with Python? Because it just became clear if we truly believe that we want to be around for a hundred years, we can't be a one language company. That just like, it doesn't make sense. It won't even be Python. It was not going to be the terminal node. This is just, this is just a journey that we're on and we're going to get through many languages. So, you know, may as well, it's just a question of when do we start, you know, n equals one becoming n equals two.

Because it just became clear if we truly believe that we want to be around for a hundred years, we can't be a one language company.

Um, so I think in that sense, it went from a very difficult decision to a very, very easy decision. And then it was just a question of how do we do this? Like how do we kind of reach Python users where they are? And to be honest, like it is a challenge for us. It continues to be a challenge in a way that is very different than the challenge with R. I think R was a difficult, um, it was difficult because R was not seen as a language that was for a bunch of the things that we thought it should be for. Um, so maybe we had, we had to overcome skepticism that R was even for this. And we had to demonstrate like, no, R is amazing for this. And over time when those battles, um, you know, one heart and mind at a time, Python is almost the exact opposite problem where like Python is acknowledged to be good at so many different things.

And in any corner of the Python universe, there's six things vying to be the solution for that thing. Every once in a while, one thing just completely dominates like UV recently. But I think it's like much more normal for there just to be, you know, a lot of, a lot of different options and, and, um, like Python users just have to make sense out of many, many different, um, you know, competing ideas and competing viewpoints.

So I think for us, the challenge was and is how do we take, um, we don't want to be too literal in how we transfer the lessons from R over to Python. Python is a different language. It has different idioms and different norms. So we don't, we don't want to slavishly copy, um, you know, what, what we have been doing in our world, but at the same time, we do have a point of view, right? Um, like we may not, um, we, we may not literally want Shiny to work exactly the same way that it does in R, but we believe that this reactive model is useful for a large class of applications that are relevant to data scientists. You know, um, like we feel that there's something really fundamentally good about the way literate programming worked with RMDs and we want to, um, you know, make that available to QMDs, but with a totally different architecture, um, and yeah, just like working in a totally different way with the Jupyter kernels versus with, you know, in our execution engine. So, um, you can stop me at any time, cause I could go on about this for a while.

But I think one thing that I do want to point out that I feel like, um, it is, uh, I think Jonathan McPherson said during, um, last year's PositConf keynote that he gave that the tools we build shape our thinking or the tools that we use shape our thinking. And there is, it has been interesting to realize once you hold these two ecosystems up right next to each other, how much our users thinking is shaped by what the tools can do and how much that's true for Python as well. And also true of the limitations of these tools, um, like not to pick on any particular.

There are certain Python packages say for, you know, building data oriented web apps that someone coming from Shiny for R would look at that and be like, that'll never work. How do you handle X, Y, and Z critical scenarios? So it, it, it is a really interesting challenge for us to reconcile these, these two value systems and, um, and kind of learn by doing and, and see what resonates with people and, and see what doesn't, but it, it, it has been surprising over and over again as, as we've engaged with the Python community.

Yeah. Yeah. It's so interesting. I also love if you throw, um, like dbt's community, the like SQL dbt community in the mix, now you have like, uh, three, a three piece stew. Yeah. And then let's layer AI on top of that too. I mean, yeah, exactly. And it does feel like SQL is going to be our N equals three, I think. A hundred percent. Yeah, but not yet. But like that is a direction we're moving across so many different projects. Like if like every, every dataset just has to use SQL.

Managing Shiny's growth and backwards compatibility

I'm curious too, between like 2012 and 2022, like managing the work, I'd, I'd imagine the way the work rolled out was really different in 2022 compared to what you were and how you're working in 2012. I'm curious if you could say a little bit about that aspect of things. Yeah. Yeah, totally. Uh, 2012 was like a fever dream of coding. Um, I mean, I look back now on some of the things that Winston and I did in those early days. And, uh, like, I don't know if I could do them today. Oh, with Claude, I could totally do it today.

But, um, you know, I remembered HTTP UV, which is the, the web server we had to write for R, uh, that was like a lot of C++ code. It wraps the lib UV library and then, um, imports, uh, some HTTP parsing code that I think Ryan Dahl live coded. And I like watched his YouTube video and then grabbed the script that resulted. Anyway, we had to build a web server package for R and I remembered that taking six months. And then years later I went back to do some sort of archeology and it turns out that we had done it in like a month or something like just way, way, way shorter than I ever could have thought.

And, um, over the years that became harder and harder to do with Shiny as, um, well, first of all, it went from, I wonder if this will work to like a lot of people are counting on us. And, uh, anytime we had a backwards compatibility breaking change, it like people let us hear it loud and clear. Like we do not appreciate this. Can we talk about that for a second? Cause I think that's another interesting difference between you and me because like when people complain to me, I'm like, well, like this is the right way to do it. So I'm sorry, but suck it up. And you're like, Oh no, I irrevocably harmed these people's ability to do their work.

Yeah. In a, in a lot of ways, Hadley, it's amazing that we're friends. Like there's so many things that I feel like we're so, our personalities are so different. And I think that's one of them is that like when you are convinced that something is right, like you're very unafraid of being like crystal clear about like, you know, you have that opinion. That's okay. I get to decide. And it's going to be like this, which is, which is not always a good thing to be clear like that. Like, like sometimes I excite and that sometimes that means like it's stock holding the wrong opinion for too long. Cause I'm not listening to people. But also, I mean, I, I was going to caveat that by saying you don't hold onto ideas too hard. Like if someone does prevent you present you with what you consider real evidence that you might be wrong, like you're very quick to change your mind.

It's just that like in the absence of that evidence, you're not going to be shy about like, well, I understand that you feel that way, but I'm going to go with this cause I believe that it's right.

And, um, and I think my sort of politeness drive and like, let's all get a long drive. Um, makes it much historically it's been much more difficult for me to be bluntly like, okay, agree to disagree. We're going to do it my way. Uh, so like a lot of shiny tickets would kind of Peter out, you know, like just maybe if I ignore this, it'll go away.

And I like to think that I really have learned from you, Hadley, in this, in the value of just say it, you know, just say it. It's going to be fine. People will actually appreciate it. Like people, most people do not get mad. Some people get very mad, but most people don't get mad about it. And oh my gosh, like double and triple that if these are people who work for you. I mean, like it is a moral obligation to be clear. And not the light. That has been a painful lesson for me to learn over the years.

if these are people who work for you. I mean, like it is a moral obligation to be clear.

But it's just so interesting. I think that like it feels like you were born this way, Hadley, you know. And I was born like in a very different attitude about this.

Which I don't think I actually was. Yeah, like I don't know how I got to this point of certitude, but. When do you think you hit that point, Hadley? How did you, when did you crest Certitude Mountain, you know? I don't know. I kind of feel like a lot of that like I got when I got to RStudio. Maybe it was like suddenly like, oh, these people I respect now like trust my opinions. Like maybe I do know what I'm talking about. I just assumed that, you know, there was some old crusty professor who was berating you. And in that moment you were like, never again. We'll never bend the knee to a lesser idea or whatever.

I have realized like I am primarily driven by like anti-role models. Like there are not many people I can point to where it'd be like, I want to be like them. There are lots of people I can point to. I'm like, I want to be the opposite of that. Interesting. I love it's like the machine learning model where it's like classifying binary classifier, but it's in reverse. So you just flip it. You're like, not that, not that, not that.

Reluctant leadership

Yeah, that's wild. I think this also gets to Joe. When I asked you before this, how you'd describe your leadership style, you just said reluctant. Maybe you could unpack that for us a little bit.

Oh, funny. Did I leave that in there? I forgot. Maybe I made it up. No, no. I definitely wrote it. I just thought I was like, just kidding.

Okay. Yeah, I am for sure a reluctant leader. I ironically went to school for management and I thought that I wanted my career to be managing technology projects, managing engineers, maybe someday being a CIO. And it wasn't until I graduated, graduated MIT, only having studied management. And like, right after I graduated, I was like, I actually programming is my main thing. And I had just left the institution that would have been like the best place to study that.

I was able to, like in my first full-time job within a few months, I was managing a team of, I had two reports and it didn't take me long to be like, Oh no, like this, this is not for me. Like, I'm not the right shape to do this. This feels hard and bad. And yeah, I had to let someone kind of counsel someone out within the first few months. I was like a 22 year old. I mean, I just, everything about that job, I was like, I would so much rather be coding and it would be better for the company if I was coding. So ever since then I've really resisted being a lead or a manager of any kind, but circumstances keep compelling me to step into that role.

Just because like, to be honest, every time it's sort of been like, well, it's not ideal, but Joe's going to have to do this. You know, like nobody looks at me and thinks like, Oh, let's bring in our top notch management talent, Joe, you know? It's always more like, well, it seems like it can't be anyone else. So it's, we're going to have to settle for like having Joe lead this. That was the case with Shiny.

I think that is a very negative characterization. I can bring, I can bring a lot of evidence. I think that, you know, I can say that because I think we've tried, we've tried many times, like, or at the very least, at the very least people are like, let's, let's give Joe a break and have someone else lead. But there's always, there always seem to be reasons why that doesn't work out. I think in the early days of Shiny, there was just, I had such a clear vision in my head and it probably took like four or five years before just the set of things that I like in my head were already built, but they needed to be like actually built. Like that would have been a hard environment for someone else to step in to a place of leadership when like, I created this project, like I needed to review all the PRs. I hired all the people, like it was like the more I played that role, the more it ingrained me into that role and made it hard for someone else to come in and, and take my place.

And I think the same thing happened with the AI stuff that I'm working on now. I'm leading a team doing sort of focusing on AI within Posit. I mean, every, every team is working on AI in some way, but there is a team that is specifically like, this is what we eat and breathe in and dream and sleep about AI. And yeah, I think once again, it was like, who else could run this, you know, given the shape of, of the charter of this team and what needs to be done. So I'm always like looking for ways to, especially the CTO title, I'd really prefer not to be CTO. I've made that pretty clear to leadership and it's just like, this is kind of how it has to be for the foreseeable future.

So like in my, in my fantasies, I imagine, you know, someday having my, and my title go back to engineer and having like someone that I really, you know, trust and enjoy working with be the manager that I report to, and just, you know, being able to focus on the part of the job that, that I, I most enjoy. But I feel like we're not close to that being the right thing for the company, you know? So until that, until that happens, or I just hit my limit and I'm like, I can't do this anymore. It kind of has to be what it is.

But I will say I have, once again, I think like Hadley and I have talked about this a lot. And one thing that has made it much easier is like learning from Hadley, like Hadley manages teams in a way that works for him.

I was going to say, if you wanted to characterize my leadership style, it would be benign neglect. You're giving people autonomy. You're letting them thrive. Is that?

Like, I don't know. I think like Joe, Joe is a better leader than me, but he also cares about it more than I do. So I'm just like, I'm not, I'm like, whatever, someone's going to do it. Yeah. Better than malignant neglect, I guess. Yeah, exactly. Better than malign neglect.

But I think Hadley, what makes it work is that the people that work for you, that is actually what they prefer and enjoy. And it's shown in the tenure that we have on your team, on a lot of our teams, like it's working, you know, like this is the way that they want to be managed. And it would be really, it would feel really bad to be managed by a more traditional, you know, scrum, you know, agile, whatever.

The universal adapter trap

And I think being willing to be clear eyed about what do I need as the manager, like, what do I need this team to be in order for this to work for me? And how can I select people for this team that that's going to work for them as well, instead of I think what I thought before, which is, if I can imagine this person being a big contributor on my team, then I will become whatever shape I need to, like, I will be the universal adapter for all of the spiky shaped individuals on my team. And it is my job to fill in all the gaps to do whatever emotional work with them needs to be done for them to be effective, to fill in whatever technical gaps they might be deficient at for them to be effective.

And I, you know, I did that for like a number of years. And that was, you know, a recipe for burnout kind of predictably. And also, I think didn't work great for the people that I was doing this feeling like I was doing this in service of, you know, so this latest iteration of both the Shiny team and the AI core team.

I'm really surprised. I thought you were saying that that worked for you. So I at least just want to dwell there for a bit, because I, I thought you were saying, like, when I gap fill for a person who's like really effective, I feel like I'm really doing a good job. But it sounds like that you mentioned, like, burned you out a bit and kind of didn't work for them. Like what? What went wrong?

I think it, it is rarely as cut and dry as like, this person just needs this, you know, they're great at everything except for X. If I can provide X, they're like incredibly effective. What it looks like more of the time is I have this person that is a middling amount of effective. How many different things can I try to, you know, to compensate for that or to, to align them in a way that they can be maximally effective. And it is not a feedback loop with super tight intervals. It is not a feedback loop with super clear signals.

And instead, like, it feels, I mean, it feels a lot like parenting, maybe. It feels like I'm doing a lot of work to really understand someone, like, and not just what they're saying is the problem, but what is the problem behind the problem? You know, sometimes like, it feels like almost like we have to do therapy to get to the bottom of like, what is standing in your way, or why are you feeling so angsty about this or that? And that is really exhausting. It's really exhausting work when you're doing it for five, six, seven people. And if you finish that work and you don't see a big change, or you finish that work, and people don't recognize you as doing that work. And in fact, like that work not being legible to the other managers in the organization who really kind of just see that you're doing things in a weird way. And yeah, I've definitely come in for some criticism from the rest of the organization.

So yeah, I think at this point in my career, I'm much more excited about finding people with compatible shapes to me. And one thing that's helped is that I'm more clear on what a compatible shape looks like to me and the kinds of people that I think can be very effective under my leadership, or even as a collaborator, and the kind of people who like, it's going to be a struggle. And if the stakes are high enough, I'm happy to say, sorry, you're going to have to change seats or you're going to have to get up so someone else can sit there, because it's what I will need for this to be a sustainable kind of an operation. Boy, this got very frank.

So yeah, I think at this point in my career, I'm much more excited about finding people with compatible shapes to me.

No, I honestly think it's so helpful. And even you framed it as burnout, but maybe in another way, it's also kind of like being kind to the person. Like if they won't have a great time and in that spot, like, am I kind of like artificially propping them up versus giving them like frank feedback? So I often like think of gap filling is like really a really powerful strategy. So it's something that I'll definitely think about.

Yeah, I mean, I think it can be really powerful, but it can also be like disenabling in the sense of like, if you were just going to like, if someone is not doing something that they like should learn how to do, and you just do it for them, like that, you know, that there's lots of reasons to do that in the short term, and it can be really powerful and useful, but in the long term, like that doesn't necessarily like serve you or serve them.

Yeah, I think like, I don't know, like took me, I also had like, I don't know, it took me a long time to get to the point where I can like give people like direct feedback where they're like, you're like disappointing me, like you're not like this is not working the way it should do or it needs to work. That like that is like,

that is really tough to deliver, but it's also better than the alternative because typically people who are not doing well, like they know it, they can get that sense. You might not be saying it directly, but they know you're unhappy. And if they don't know like what is wrong, like they can't respond. Yeah. I'm curious your thoughts on today now where you mentioned you're leading an AI team. What, yeah. How do you think of things today?

Excitement and existential dread in AI work

Yeah. Yeah. The first thing that struck me actually, it was, it was one of your questions you asked me earlier, Michael, before this recording, it really made me think about the fact that a lot of the earlier moments in my career, the exciting moments were just exciting. And this particular moment is exciting mixed with existential dread.

And like that, that combination is so weird. Like it is such a strange sensation to build something that you're both incredibly excited about. And also should we even be, should we even be doing this? Like, is this the right thing long-term for the industry, for people who are going to use this for the planet? And yeah, I have felt more conflicted about this work over the last couple of years than like at any point in my career where the sort of push and pull of AI has like literally kept me awake at night wondering, you know, should I be doing this? Should I abstain from being involved in this technology at all?

And this particular moment is exciting mixed with existential dread.

So that's the first thing I think that is really striking about this moment is how much sort of ethical soul-searching I feel like a lot of us have had to do.

And for me, I've gotten to a place, we can talk about it if you want to, I've gotten to a place where I don't lose sleep about this. I'm quite convicted that the work that we are doing is not positive and it is important that we continue to do it. But the other, there are some other things that are really weird about this moment. This, like the technology, like LLMs as a sort of core technology, the fact that they're both so incredibly capable and until like very, very recently, I'd say so, so unintuitive, like it's so unintuitive what they're going to succeed at, what they're going to fail at, that it's almost impossible to come to any conclusion about whether something's going to be useful until after you've built it, after you've built it. Like it's hard to reason from first principles, whether you should harness them this way or that way.

And then at the same time, because the tools that you build so fast, that's actually a viable strategy. Like we don't ponder on paper, we ponder with prototypes now, you know. And it's created this whole new motion of just, you know, my team and the teams that I'm in close contact with are just throwing off massive numbers of ideas every week. And it's no longer a generation problem, like generating features, generating code. It's like a curation problem. It's a conviction problem, like developing conviction about what you think is right for the user. That is now the big struggle and ideally the key differentiator for us, because collectively we have so much experience in data science and we have like really strong opinions about the right and wrong ways to get answers. If correctness is important to you and reproducibility is important to you, you know, we have a lot of opinions about that. And I think a lot of people, a lot of competitors or ostensibly competitors have different priorities, which gives us sort of a unique viewpoint to build around.

But it's still, like, it doesn't feel like any other motion in my career, you know, where it's the conviction and it's the editing. It's not like it takes you X amount of time to come up with the idea and then 10X that amount of time to execute. It's really trying to think of like, what do we actually think is the right way to harness these things? And then the building of it is relatively straightforward. And then on top of that, it's changing all the time, right? Like every model, the lessons we're learning about building harnesses collectively as an industry, it's changing so fast that it adds another level of complication to try and to form conviction around things. Because you have to have conviction that it's going to stay constant for six months or hopefully more than that. So a really weird time.

The ethics of working on AI

Is the thing that was keeping you up at night asking like, what would be useful for people? Or was it something? No, it was, is it ethical to work on these things? Societally. Is it ethical? Yeah.

Is it like considering the negative side effects of this technology existing, not to mention the negative potential side effects that especially two years ago were quite extreme, like people were, well, I guess some people still do, but like, it was much more in the discourse that these things could kill us all or some approximation of dooming the human race. And it certainly like, what effect is this going to have on knowledge work? What, what effect is like this amount of change this fast going to have on our economy and society and the copyright, you know, the copyright infringement, I guess that these models were mostly built on being an ethical problem.

And then also, I think on top of that, in the particular domain of data science, the ability of these things to so quickly and easily give you answers that you don't know if they're correct. And feeling like, I don't know, we could build tools that are like handing out loaded guns to the general populace. And then just like having a warning sticker on it, like warning, this is a loaded gun. So we're cool, right? Like we did our job, we warned people. So it's really on them. But like, if you know that people are going to, you know, shoot their feet off or whatever with these guns, is it really ethical for you to release that like into the marketplace? So those were the kinds of questions I think that like, I really legitimately struggled with.

Should I go to, you know, to Reef, our CEO and tell him like, I want to keep working here, but I don't want to work on AI anymore. I'm conscientious objector, or maybe even like, if this company is going to continue to develop AI, then, or develop solutions around AI, then I need to work somewhere else. Which is ironic, because I was literally the one leading the charge at the time, you know. So it was not a statement about whether I should continue to be neutral. It's like, I'm literally evangelizing this technology within the company and to our customers while feeling intensely conflicted about whether it was ethical to even have an open AI account or whatever.

And I don't mean to say like, I knew that it was unethical. It was that I was really like riding that line of, I can't decide if this is like net very positive, net very negative, like catch me on a different day and I'll have a different answer. I think it's like so hard to kind of balance because it feels like the good things are like so good and the bad things are so bad. They're both like really large numbers. And what do they cancel out to?

And like, I think our, at least your mindset now is like, they need out to something that's like, you know, mild to moderately positive. Like that doesn't mean there's still not like lots of really bad things, but maybe mostly kind of balanced out by the good things. And like we have the opportunity to kind of at least try and accentuate the positives and make the negatives worse for like some, make the negatives less worse for some group of people. That's right. That's right. Yeah.

And I think for me, it's not so much like today's trade-offs, like the harms and benefits that have been realized today, I do agree are mild to moderately positive. It's the potential harms and the potential positives that really, like, it's like the error bars are like this, or the error bars are very wide, I guess, going forward.

Yeah. Yeah. I think the other thing that makes it so hard to think about is it does feel like if we as a company don't buy into it, then like the company is not going to exist. And like how that, like, I think that just makes that it's so hard to tease apart that kind of like pure ethics from like, okay, well, I, you know, I like my job. I'd like to continue having this job. I like the people I work with. I like, they want them to keep having more jobs. And like, at what point are you willing to pull the fire alarm and say, like, like we have to stop this train, even an extent to that, like by not doing anything with AI, like, like that's the end of the business. And, you know, five Latinos like that just add all of this, like additional, like emotional weight to it. Yeah.

And I think like for my ethical, like values and framework, like I did not really, I tried to remove that effect because like, I don't like everyone can look at this a different way, but for me, if I thought that it was the wrong thing and it was net harm for us to do this, but I didn't want people that are close to me to lose their jobs, that would be like, these principles are inconvenient. So therefore I'm not going to follow them.

But the one that really kind of, I couldn't shake and ultimately won out over the others was that us abstaining or me, me abstaining, even if I could get the whole company to abstain, that would not slow down the development or adoption of AI one bit. Maybe if everybody in like in a coordinated way, every company in the world decided to abstain, then yes, but, but us taking this action would have no effect.

But the good that we can do is in the specific area of data science and not just data science, but like reproducible data science. And we actually have a differentiated viewpoint there and we have outsized influence there. And that, because of that, because we have like, we have no agency to slow down AI. We do have agency to help people avoid some of the harms in getting incorrect answers. That sort of extra agency there makes it to me, like we have also a special moral and ethical obligation then to, you know, to serve that role for our community.

But the good that we can do is in the specific area of data science and not just data science, but like reproducible data science. And we actually have a differentiated viewpoint there and we have outsized influence there.

And that ultimately, like once I landed on that conclusion, I was like, okay, this debate is over for me personally, you know, and I, I worry about other things now, but I don't, but I don't, I don't worry about, you know, should I be abstaining from this, you know, this technology altogether?

What's valuable for open source in the AI era

Do you think about like for a company like Posit that, especially like our open source teams that have traditionally contributed tools, like you put in like blood, sweat and tears for three months, three to six months or a year, and you come out with something you think is really useful, like a really useful tool. How do you think about that today? Like what's our output today to open source? Do you think it's still tools that people can use? Like what, what do you think is, is useful in 2026?

Interesting. Yeah, that's, it's really interesting how you phrase that. I do think people still are hungry for tools. But maybe this is what you're implying. I think they're maybe even hungrier for perspective, perspective on like how to think about AI, how to, how to use the tools that are out there, how to integrate the kinds of analysis that they've done in the past with these, with these new tools.

I'd say like, this was probably even more so a year and a half ago where I just felt like my meeting, my calendar was nonstop meeting with customers who wanted to talk about AI strategy. I think we're in the sort of experimenting phase now where a lot of those customers have deployed, you know, software agents, they've deployed our agents, they are building their own agents and people are trying different things and learning for themselves what's working. So we're seeing a lot of, you know, at conferences, people are sharing, okay, well, here's what people are sharing. Okay. Well, here's our, here's the agent that we're built, that we built in. Here's what worked well and what didn't. And for some other companies, like here's generation five of the agents that we're working on. And here are the lessons that we've learned over like the last couple of years.

So as that's happening, I think I'm seeing a little less, please show us the way. And they're more coming back to us now and saying, okay, we have opinions now, and we would really like your tools to do this or that, or integrate with this or that thing that we already really like. Where is the MCP for X, you know, or please deliver me a skill that helps me do Y with your tool. I don't know, does that resonate with you guys as well? I'm curious how you guys would answer that question.

Yeah. People spent like blood, sweat and tears. You're saying they got an open source, bringing a tool to like market or bring it to the world of the community. But how, how useful is that in 2026? Like, why should I type it into my cloud code instead of you type it into your cloud code? I think there's some of that.

So yeah, I think that's true. If we know what the problem is and have a good sense of the solution, probably it's like just a package to do something like that is less valuable. What I think is more valuable is identifying that problem.

And I think kind of like we touched on this a little bit with like Shiny, like before, like there's a lot of things that like, but the Shiny added to R that people were like, like, why would anyone want this? And then when you have them, then you're like, oh, actually, this is really nice.

And I, like, I feel like the same way with tidy data, like the principles in hindsight seems so obvious. Like you put your variables in columns and you put your observations in rows. Like, you know, it wasn't really necessary for me to take five years to like, you know, figure that out. Um, so I think there's like, there's still value in that and figuring out like what, like, like naming the problems and being like, okay, you're facing problem X. And then that gives people something that like hang ideas in the head off.

Like the software, yes, is probably still valuable, but there's also like, would you rather do this or would you rather have a team of like professionals who do this day in and day out? Um, but you know, that's our job. And then we also see like all of these problems across all these different companies and that uses all over the world and that generality helped us create better solutions. But yeah, I have a lot of, I think there's a lot of questions now about like, what's the open source? What's the value proposition of open source? It's like fundamentally different now than it was a year ago in ways I don't think we have, you know, we really appreciate it all.

Yeah. I do, I do wonder if that also hits back on Joe's earlier point too, about, I, I think quoting, um, maybe John McPherson, but, uh, of like, if, if tools shape our thoughts, Tools Shape Our Thoughts.

future, we could imagine that's going to be beta for everyone, but you would have to blend the thinking and UI, and maybe that's going to pay off, and maybe it's not. Yeah, it also reminds me too that you said, I think in a previous podcast, you mentioned like working on Roxygen 2, I think, issues. And so just for context, like Roxygen 2 is a tool used to generate documentation for R packages. It's probably not the sexiest or funnest tool to work on, but it sort of like undergirds everybody's or a lot of packages. And maybe that's relevant too, is it's like, do we want a world where everybody just cloud codes their own Roxygen 2 versus like is reasonably happy with the one Roxygen 2 out there? But maybe like the scenario is like you don't even use Roxygen 2, right? You just, you have cloud code updating some comment and updating like a beautiful, you know, interactive documentation website at the same time. I, my suspicion is they were like inevitably tripped away from each other and become super confusing, but like maybe that's not the case. Like maybe that's something that's now like more feasible than it was before.

Yeah, it's fair. I, yeah. And I did say that about Roxygen 2, but I do also vibe code a lot of web apps. So maybe these two things. Yeah, I guess. Yeah. I mean, I feel some of this where I feel like I am writing like more base R code than I ever did before, because I'm like, yeah, the syntax is kind of ugly, but like, I don't have to write that. I don't have to wrestle with it. And like, when I'm trying to make a package like lower dependency, I'm like, well, like Org can figure it out. It can implement this algorithm from scratch in a way that's, you know, I'm like, I wouldn't do this because I don't think I'm, I don't trust myself, but I trust Core enough to do it as long as it's lower than a bunch of tests. So yeah. Like what about the cost of dependencies versus writing your own code? It's like that trade-off has fundamentally changed. Like it's going to take us years to understand the consequences.

Like what about the cost of dependencies versus writing your own code? It's like that trade-off has fundamentally changed. Like it's going to take us years to understand the consequences.

I love people heard it here first. Hadley loves writing base R and is doing it all the time. And trust agents more than he trusts himself. That's our soundbite.

The origin of the name Shiny

Yeah. No, Joe, this has been so, honestly, so inspiring. I think just seeing you're like thinking about like leadership and a lot of these like moments in time, I think your willingness to, I think just like lay it out there, honestly, is so helpful. I figured we'd end with something. What do I want to say? Maybe like a weird fact, but what one I figured I dig back is as I understand it, Shiny is named after a particular sci-fi series. Is that right? Yeah. It was a slang word in Firefly by Joss Whedon. And they would just, they would just use the word shiny to mean like good. Like I'm not sure the way they, they dropped their made up jargon. They just did it in such a casual way that like, it felt like this is actually like how they speak. They're like, oh, that's shiny. I can't even do it. Like I still like even trying to do it, I overemphasize a little, but they, they would just say it in such a casual way that when I was watching it, I, I thought to myself, oh, that'd be like a shiny would be a fun name for some kind of open source project. And then not long after I came up with the idea of Shiny and had to create a repo for it. And it suggested the GitHub forum said like having trouble thinking of a name, consider it was shiny something Octocat. And I was like, oh man, it's gotta be shiny then. And that it stuck. Do you watch a lot of sci-fi or what?

No. And actually, I mean, I tried multiple times to watch Firefly just because I felt like I don't know, it had such high nerd cred and it didn't land for me. The first couple of times I tried to watch it. And then it was this particular time I was sick in bed with like a stomach flu or something. And that time I got hooked and I was like, okay, like I watched this thing end to end. And the more I re you know, I've rewatched over the years, the more I've appreciated it. But it's funny because like, I think in the beginning, the early days of Shiny, when people heard this, they're like, oh, you are also like a diehard Joss Whedon, you know, like super fan of Firefly. And I was kind of like, sure, okay. But really, it was just I that name, I thought was catchy. But yeah, now, of course, the name has taken on such rich meaning for me. And, you know, with rewatches, I certainly like, like it more. So I'm not like a hardcore sci-fi guy. But that that series now obviously means a lot to me. Yeah, I love it was like aspirational. You like aspired to watch Firefly. And then through effort and grit, you like got there. Eventually. Yeah. You're like totally working. What I mean, what TV are you getting into? Are you like a Love Island fan? Are you?

Yeah, it's interesting. Like, I really have. I have a hard time getting into shows often. And then when something kind of gets hooks into me, it really sticks. So the last show that I was like, really, really into was Severance. And then now, I'm a little late to it. But Pluribus. Is that how you say it? Pluribus? I guess maybe I like Apple original series. But yeah, I think I'm like three episodes in. And I'm like, wow, this is not what I thought it was. This is like, very entertaining. It's kind of like The Last of Us, but the zombies are incredibly friendly and trying to help. At least so far, no spoilers. But geez, well, yeah, so much appreciate you coming on. It's honestly so interesting to hear like you, you too, having interacted for so long and like, Bill, really interesting things that Posit and I think to have these, but especially like Shiny, yeah, being like so load bearing and industries like the pharmaceutical industry and research there. It's, it's been so interesting to hear about. Yeah, your journey there. And like into reluctant leadership and, and all that. So really appreciate you coming on. Any, any final words for I don't know the people actually, just last last thought about this.

Finding a trusted peer outside your reporting chain

And I mean, I feel like it's turning into like the Hadley and Joe buddy show. But it just occurs to me as, as we're talking about this, that it is interesting, like Hadley and I, we don't, neither of us reports to the other. We're, we're, it's like we're in each other's orbits enough to know pretty well what's going on with the other person. But there is no, like, I feel like there's no competition at all between us. There's no like worrying about if I say the wrong thing in front of Hadley, what are there going to be like consequences. And that is incredibly valuable. Like that's a very helpful person to have, especially when you have the right amount of complimentary and similar like commonalities. I like, I do highly recommend trying to find that, you know, like, you know, wherever you are and whatever you do, like to find someone who is not in your reporting chain who, you know, if you can really trust and, and whose opinion you respect and go out of your way to, to exchange ideas and, and support each other.

I like, I do highly recommend trying to find that, you know, like, you know, wherever you are and whatever you do, like to find someone who is not in your reporting chain who, you know, if you can really trust and, and whose opinion you respect and go out of your way to, to exchange ideas and, and support each other.

Yeah. No, Joe, thanks. Thanks so much for coming on. It's been so great. I honestly, yeah. So inspired to your approach to Shiny and leadership. So really appreciate having you on the test set. Thanks. Thanks. Appreciate you having me. And I love the podcast and I've, I've learned so much from listening to other, other guests. So thank you. Yeah. All right. See ya.

The test set is a production of Posit PBC, an open source and enterprise tooling, data science software company. This episode was produced in collaboration with creative studio Adji. For more episodes, visit the test set.co or find us on your favorite podcast platform.