OpenAI Chief Economist Ronnie Chatterji: measuring the greatest technology of our time

Ronnie Chatterji is the Chief Economist at OpenAI, where he leads research on how AI reshapes growth, jobs, and the economy. He is also an economics professor at Duke. He was previously the White House CHIPS coordinator implementing the $52B CHIPS Act, and before that the chief economist at the US Department of Commerce. In this episode of Summation, Ronnie and Auren discuss: The three things people actually do with ChatGPTThe 10x jump in people handing their work to AI Why the next big ...

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  I think about, wow, I really want my kids to become great writers and I'd love for them to learn how to write. And then, once they can write really well, then interact with AI rather than having AI shape their voice. A lot of parents are concerned about that.

• 0:10Auren Hoffman?

  My guest today is Ronnie Chatterjee. Ronnie is the chief economist of OpenAI, where he leads research in how AI reshapes growth, jobs, in the economy.

• 0:20Speaker 00

  How is what they're working on today going to affect what you're seeing downstream tomorrow? That, to me, is the biggest edge I have working inside OpenAI, more so than like playing around with a product where honestly, like, there's people who are much better at that than me, but being able to understand a little bit what's coming on the research agenda, that's very helpful to someone downstream who's doing economics.

• 0:41Auren Hoffman?

  Hello, fellow data nerds. My guest today is Ronnie Chatterjee. Ronnie is the chief economist at OpenAI, where he leads research in how AI reshapes growth, jobs, in the economy. He's also an economics professor at Duke. He was previously the White House CHIPS Coordinator implementing the $52 billion CHIPS Act. And before that, he was chief economist at the U.S. Department of Commerce. Ronnie, welcome to Summation.

• 1:05Speaker 00

  All right, it's awesome to be here. Thank you for having me.

• 1:08Auren Hoffman?

  I'm super excited. Now, you possibly have maybe one of the best data sets in the world of how people actually use AI. And we read all about it, but what is actually going on? Like, how do people actually use all these hundreds of millions of people who are using AI every day?

• 1:25Speaker 00

  Well, first, I'll say it is a privilege to be able to do this kind of work. And what gets me excited every morning to do the job, I love my team, I love the organization. I also love the data. I love being able to say things that are evidence-based about how we're using AI. There's so much interest in AI right now as a social, economic, and political phenomenon. And for me, on the economics side, I'm able to use these data sets that have consumer data, enterprise data, and data on new tools, like agentic tools, let's say, like Codex, and be able to actually say what we are doing. So it all started last year in 2025. I had been at the firm for about a year, and we released a paper called "How People Use ChatGPT." And I just want to dwell on that for one second. As a son of a professor, it's always a dream to write a paper with such a simple title, right?

• 2:10Speaker 00

  Nothing technical in it. It's exactly what it says it is. How do people use ChatGPT or how people use ChatGPT? And what we did is we took the consumer data set at that time, first study of its kind, and we basically analyzed patterns using classifiers. And so what we do is we use LLMs to classify the discussions. That means I don't have to read anything and it's privacy protected. And we organize it into groups. And basically, I'll tell you two key results that really got to me, and happy to take other questions about it. One is, there's really three modalities, three intents when people are interacting with ChatGPT circa 2024, 2025. And that is asking, trying to get some information about something, doing, trying to get ChatGPT to do something for you, or expressing, talking about feelings or topics that are neither asking or doing.

• 2:52Speaker 00

  First, asking is definitely the number one at that point in 2025. And we'll talk about how that might have changed.

• 2:58Auren Hoffman?

  I happen to be in Nashville. What is a good sushi place in Nashville or something?

• 3:02Speaker 00

  That's exactly right. Or, I want to read my kids a great bedtime story about baseball because the World Series is happening. How do I do that? The second thing is doing, right? The agentic tools have become much more useful, or agentic task. The expressing being third was kind of a big surprise. I think a lot of the popular commentary on ChatGPT at that time had really focused on that category. And it's not that it's not important. I mean, we have north of 900 million weekly active users by now. So it's still a large portion, but it's third in terms of the ranking, and by a lot. And I just think this is like a really simple, but I think profound reason to do research. We got to figure out what the right answers are. If we want to guide thought leadership, enterprise decisions, policymaking, we need to know what the answers are. So for me, the asking, doing, expressing.

• 3:42Auren Hoffman?

  When you're asking a question like, okay, I'm going to have a tough conversation with my boss or something, help me walk through it. Is that expressing or is that actually, is that asking? Like, where does that fall?

• 3:53Speaker 00

  That would be expressing, but I do agree, like, this is where the classifier system, we usually try to have some subclassifiers and run a lot of validation tests, because that one could be both. They'll ask you for some advice about what to do, but also it's maybe expressing how you feel about your boss or how you feel about that situation. So, you know, a lot of these things can be both, and we have a methodology in the paper to try to separate between them. That kind of measurement error probably would still keep expressing, you know, even if accounted for at the third place position. So that was one big thing. I think the other big thing was, you know, what are the number one things people are asking about? I think writing and sort of seeking information. So your Nashville example, perfect, or my bedtime story example, but also writing. And I wanted to talk a little bit about that just because, you know, my parents are immigrants to this country, and I always think about them, and I also think about my kids.

• 4:37Speaker 00

  I think about, wow, I really want my kids to become great writers and I'd love for them to learn how to write. And then, once they can write really well, then interact with AI rather than having AI shape their voice. A lot of parents are concerned about that. For my parents and people who come from other countries with variations in language skills, that writing support can really help. And I'm seeing that already in my correspondence with people who are speaking other languages besides English. They'll write me an email or a text message that looks great. And I imagine a lot of it's because it's AI assisted. So we have to think about both of those margins, like people who maybe now have a seat at the table who didn't have it before with their writing, but also how to develop those skills in young kids like you and I have.

• 5:13Auren Hoffman?

  One of my friend's parents, they're originally from Russia, and they live in New Jersey. And, you know, they work a normal job, but like, they found it to be super helpful and just like helping them express things from like a Russian standpoint and stuff like, so they get the right email to their boss, et cetera.

• 5:29Speaker 00

  Very important. And, you know, those are kind of translational gaps that would hold people behind in the workforce. You know, there's lots of research that says, you know, if you can't communicate or you can't, you know, ask for that raise and negotiate the salary. So I think, you know, these are the margins of AI on both sides. You know, it creates new challenges, it creates new opportunities, and that's why I find my research and the work at OpenAI so interesting.

• 5:46Auren Hoffman?

  That was done kind of late 2024, 2025, now where, let's say, six, seven quarters later, like, I'm sure a lot of this changed. And the doing stuff is now, now we've got Codex, now we've got all these things where the doing stuff has. I feel like in December, the doing part really shifted for me. So I would say before December, I was much more in asking mode. And then, post-December, you know, we had like, we had the Claude Code, Claude Co-Work come out. Then we had, you know, we had the OpenClaw, then we had Codex, right? So we had this kind of Cambrian explosion of these like doing things that came out after that.

• 6:22Speaker 00

  Aron, you and I are in the same group chats, apparently, right? Because it was the same for me, right? So here's how I think about it. In 2025, we were in chatbot world, at least for most of us. And we know we had sort of early adopters thinking about it in different ways. So when we wrote that paper, you know, that was the right paper to write at the right time. And for hundreds of millions of people around the world, people who are signing up for AI services every day, that's still the right, you know, frame. But during that period, late last year, I think, which you referred to, that's really when things started to change. We started to move from sort of chatbot world. And I won't skip a step, reasoning models came in between that and really improved things too. And then into more agentic AI. We kind of felt it. People like you and I were messaging, we were talking to our friends and saying, this is that moment. And so we decided to write a second paper in the series, and it's really about agentic AI now.

• 7:07Speaker 00

  So we have this amazing data set on Codex, and that Codex generally has grown incredibly fast. Like, I think Tebow was on X the other day saying, you know, beyond 7 million, up to 8 million in terms of users. It's also about how people, what people use it for. So we wrote this paper on Codex, we released it a couple weeks ago. Again, sort of the first paper to comprehensively look at the agentic AI work from the perspective, and this was kind of interesting. We did it for consumers, we did it for enterprise, but we also did it for people inside OpenAI. And this was a big methodological step and something I'm really proud of the team. Our lead author on this, Drew Johnston, really was instrumental on this. But if you think about that insight, he said, look, let's measure how people in OpenAI are using it, how they're making that transition.

• 7:47Auren Hoffman?

  They're going to be this pretty far down on the most advanced users, you would assume.

• 7:51Speaker 00

  You know, we live in the future, I feel like, at OpenAI. And if we can show people what that looks like, it both provides some inspiration and guidance for what the firms of the future might look like. And so a couple of things from that. One is, at least on the individual, the consumer account, you're seeing the number of people delegating a full workday. Okay, a full workday, eight-hour task to Codex. That proportion grows 10x during the early months of 2026, from around January to May. Insane growth in terms of people literally delegating a full workday to Codex. Inside OpenAI, in that period in 2025, when you were seeing that Cambrian explosion, we saw basically our research and engineering groups pivot almost entirely to Codex and agentic AI. January to April of this year, everyone else caught up.

• 8:36Speaker 00

  Finance, right? Econ, research, people org. So if you, we put this chart on X the other day, again, basically show that the ramp for the less technical groups, and I include my own group in that definition, which is kind of funny to say, that group grew significantly in those four months, early in 2026. And so if, in your own organization, you should ask that question. If my, if my technical folks are already fully agentic, how long will it take for everyone else to catch up? And I think four months is a pretty rapid clip. That's what we saw undocumented in that paper. Still getting a lot of great feedback, so if people are reading that paper and have ideas, please let us know. But those are the two key results. I think the 10x increase in delegating a full workday on the consumer side, and then the big shift within OpenAI on the non-technical side, where you see now we're fully saturated on Codex and these other groups as well, including my own.

• 9:21Auren Hoffman?

  I was, the other day, I went to Toronto, quick trip, it was like a 24-hour trip. And so I just had to Codex on my local, on my work machine. Hey, I do this thing, you got 16 hours to accomplish this task. Here's the task that I want you to do. Just go crazy, do as much as you can in that 16 hours. Went to Toronto, came back, literally checked in 16 hours later, and it was awesome. It was just awesome to see like how much I was able to do. There was a few things, obviously, couldn't do everything, but it got like, it got a lot of work done during that time. I had my weekly tokens, I was like, I gotta, I gotta make sure I use these, right? Like, you know, I'm an economist too in a way. I'm like, let's use the tokens, get them out of the way.

• 10:01Speaker 00

  I mean, and Aron, I think what you're seeing at the individual level, what you're seeing in companies, what you'll see in governments is, we're moving from, you know, asking for an answer from a chatbot to really being able to delegate workflows. That's really important. And when you delegate, you can both do things simultaneously. I'm sure you, I mean, you've written about it, it's hundreds of agents doing sort of scouting for venture deals, right? So parallelism is really, really important and a new way that we're able to work. All of us are managers now. If we're not economists, we're definitely managers. And the last piece is with skills, you can actually routinize these things and say, this is what I want to happen and have it a repeatable process. So being able to parallelize and routinize, that combination is going to really change work. And I think we're still grappling inside organizations how that's going to change the way we work.

• 10:43Speaker 00

  It really is more like electricity or general purpose technology, where we're going to have to learn how to work around it given its new capabilities. That's kind of how I'm thinking about it.

• 10:50Auren Hoffman?

  When your team is internally is working with your own agents, how do you guys like give these agents feedback? Because like, what we're, in our organizations, we're all giving, they're all giving feedback to these agents. We're almost treating them like a remote employee. Or we're saying, okay, here's, here's some, here's some feedback, here's some whatever. And of course, just like any remote employee, if the humans give you conflicting feedback, like they get confused, right, as a, which would normally happen, right? I might say, hey, I want, I always want the presentation like this. And someone else says, oh, actually, you should always put the presentation like this. And then like, there's a conflicting kind of thing. And they have to say, oh, for Orin, do it like this way, but for the other person, do it like this. So this whole like nuance that goes, but how do you guys do that?

• 11:35Auren Hoffman?

  How do you guys like collaborate together on an agent?

• 11:37Speaker 00

  I think there's a couple of key things. One is, I do think more context is better. This is the part of managing agents that's very similar to managing humans. I think I'm, I'm working on some ideas around what's going to be similar, what's going to be different. As you know, I mean, economics background, but I've taught in a business school for a long time. So I'm really interested in how we manage agents. One way that's not going to change in management is you got to give clear instructions and context. The more you give, the better it does. And you know that from talking to Codex every day, the sense that you basically, when you give a specific interest in context, I said, okay, I'm going to do an interview with Aron Hoffman. I sort of know him pretty well already. He's going to ask me some different questions. How should I be thinking about things that maybe I want him to know or his audience to know? If I explain who I'm talking to, the fact that I know him, it's already building in that conversational tone. So it preps me differently than if I'm talking to someone I've never met.

• 12:22Speaker 00

  And also, they know your audience, right? So they could figure out, okay, it's Aron, it's Summation, and they understand what your audience is likely interested in. That kind of thing helps Codex become a much better sort of prep engine for me. And that is how our team works a lot. Detailed markdown files, you know, there are people on my team, and you look at some of the authors on the Codex paper, there's some of the power users, they're really thinking about markdown files that they share and have sort of full instructions and information that they're always adding to.

• 12:46Auren Hoffman?

  To me, this has been still like the janky thing, like within the organization. If it's like personal, it's great. I got my own thing, it's quite, like, but like internally, we've got like this like janky GitHub thing that we're kind of all committing to and stuff like that. Like, how do you guys like have, because you're a little bit in the future?

• 13:03Speaker 00

  We share skills. So the way I think about it is if Drew Johnson or Alex Richman on my team develop a skill, they can give it to me and I can use that skill and I can access that. That's, that's the key.

• 13:14Auren Hoffman?

  For us to collaborate on the same skill at this, like, we have to like put in GitHub, we have to kind of like, you know, it's kind of this really janky thing. There isn't like, at least, at least you guys might live in the future where there's like this great collaboration way, like the single player mode's amazing for whether it's Codex or Claude Code or these other types of things. The multiplayer mode still is like, it still feels like it's still being worked out in a way where you're trying to collaborate on a team or small team or other types of things.

• 13:41Speaker 00

  And that's the key to making it work in organizations, I think. My view is that, you know, to make this work in organizations, you need sort of a shared memory, right? The idea of like, Aron and I might be the CEO and the COO. We want to bring in the legal counsel. Maybe you have to talk to her and I'm not in that conversation. You guys have a shared memory because it's about me. But other times, the three of us have to get together and the memory has to be different. Like, these are all important things to think about in terms of working in organizations. Governance, permissions, like, you know, what do we let our agents do? All these are key in enterprise. This actually also goes to why, as the capabilities are going up and to the right, we haven't always seen the impact in the enterprise commiserate with how quickly the capabilities are growing. That's the biggest question I get. If capabilities are advancing, how come we're not seeing more in the organization? And my point is, you know, you need to get a bunch of things right on the organizational

• 14:25Speaker 00

  structure, on the governance, right, on the processes, and then we will start to see a lot of benefits. You see this in your individual work, right, where you don't have to ask anyone, you just run 16 hours of Codex while you fly back and forth to Toronto. In most organizations, there's a lot more to it, and I think that's what is being navigated now. You know, I obviously, you know, have my own sort of vision of it, but I feel like we're moving faster than we have with previous technologies. But people are so excited about what they're seeing with AI in their personal use, they're expecting us to move faster on the enterprise. It's one of the few technologies where the consumer is experiencing the revolution, maybe at the same time the enterprise is. And that creates, I think, an impatient fly, maybe faster.

• 15:01Auren Hoffman?

  Because you work at a regulated bank or something like that, or some, like, they just have to move slower because there's a lot of scary things and you're not sure about, so that stuff might leak. Whereas if you're just at home, you can just go crazy, yeah.

• 15:13Speaker 00

  That is, there's something interesting there about how public perceptions around AI are forming, how attitudes towards enterprise adoption. I think with things like electricity, for example, or the semiconductor, the integrated circuit, I bet the, the sort of the implementations in the industry, people were seeing the magic there first, maybe far before.

• 15:29Auren Hoffman?

  Yeah, because they're so expensive.

• 15:30Speaker 00

  Right? And this is something different because we've kind of democratized intelligence more quickly than people maybe imagined. So to me, that's, this is the kind of stuff that keeps economics interesting, you know, the usage data part, but also, you know, who's seeing what at what time? Who's getting exposed to the capabilities and what the impact is on enterprise? Super interesting stuff for our team.

• 15:46Auren Hoffman?

  How far in the future does like your group get, go when, because like, I'm, I'm, I get the latest release from OpenAI, right? So I'm, I'm, let's say at day zero, or, you know, day, day plus one or something like that. Are you day minus 30, day minus 60? Like, what do you get to play with that? Because, because obviously, you don't want to be playing with things that are breaking all the time. So what do you, what do you get to play with in your org that's like a little bit further or further? Are you a quarter ahead of us, a month ahead of us?

• 16:14Speaker 00

  It probably depends on the particular sort of product or technology. I'll say two things, you know, the, you know, being an employee, you can see some things before and it helps you think about, oh my gosh, when, when everyone else gets this, what might be the economic impact? I will say though, maybe more importantly, one of the cool things about being in the AI lab, and I talked to my friends in academia a lot about this, you know, there are pros and cons, you get to interact with people building the technology. You know, as excited as I am about all this stuff, I'm not the one building it, right? But if you talk to the people who are building it and you get a sense for where their minds are and where they, they are living in the future in a whole different way, then you bring it back to your field of specialty. In my case, economics and business. And you think about, how is what they're working on today going to affect what you're seeing downstream tomorrow? That to me is the biggest edge I have working inside OpenAI, more so than like playing

• 16:58Speaker 00

  around with a product where honestly, like, there's people who are much better at that than me, but being able to understand a little bit what's coming on the research agenda, that's very helpful to someone downstream who's doing economics.

• 17:07Auren Hoffman?

  Why is the perception of like favorability, unfavorability of AI so much lower in the US than most other developed countries? Like, why are, you know, whether it's Asian countries, even many European countries are like way, have a way higher favorability rating than the US?

• 17:25Speaker 00

  I've traveled probably to about 13 countries and experienced different attitudes across all of them. And so as a social scientist, I don't have one answer, but I can tell you some things I've picked up that might be helpful in thinking through this. I mean, one is, when you go to East Asia, I think there's a lot, because of demographic challenges, less immigration than we have in the West, there's more of a sense of thinking about complementing and extending workers. You know, I think a lot about this where it's like, look, if your, if your biggest challenges are, you wanted to devote more of your economy to care for people, let's say the elderly or kids, or you need more workers in particular industries to be globally competitive, you're seeing AI as an accelerator in that instance because of the realities in your country. And in South Korea and Japan, they're much more comfortable.

• 18:07Speaker 00

  And this is because of previous generations of robotics and bringing technology into society. If you see Seoul and Tokyo, different story. So for there, I think it's, it's, it's national differences in demographic realities, immigration flows and kind of attitudes towards technologies kind of flow from there. I think a lot of emerging markets, you know, I've traveled to India and Indonesia and talked to folks there, or any frontier kind of market, they see this as sort of a leapfrog opportunity. I mean, when we teach, when we teach these technology trends at business schools, we talk about the mobile phone and how a lot of places in Africa skipped the generation and, you know, skipped the landline, went right to the mobile phone. And I think that leapfrogging attitude and story is very resonant with places in emerging markets where they say, look, we see the technology and geopolitical situation as being up for grabs. This is a great opportunity for us to use these tools.

• 18:52Speaker 00

  And to your point, you know, a user in India is using Codex and getting the same kind of power that I'm getting from it. So think about that. How many other goods can you say that of, right? So that person has the power in their hands that they haven't had before, and I think it's changing attitudes in the emerging markets. And I think that might be one of the reasons that you see some differing attitudes. I do think also at the same time, you know, how AI is being used in society will determine how people think about it. If people are seeing it in applications to healthcare, solving problems that are real in their lives, I think that makes a big difference between if its phrase is some sort of like esoteric, technical thing. And so I think it just depends on who you're talking to. It's also very broad. I find this when I talk to audiences, some people want me to go really, really deep on some technical idea. Other people just want to talk about the political, social and ethical issues attendant to AI.

• 19:37Speaker 00

  It depends on who you're talking to and it's hard to do all that in one presentation. I think people get what the message maybe that they're looking for from the different speakers, and sometimes that might affect their attitudes as well.

• 19:46Auren Hoffman?

  The debate around data centers, to me, is a very interesting debate in the US. Do you have the, well, let's say some, one side maybe more on the left, which is more typical, they were like the same people who are against pipelines or against fracking or against warehouses, right? That's kind of like, you expect that in a way, like there's some group of people are going to be against those types of things. But there's also been a lot of people on the right. Who to me was also like a little bit less and maybe they're worried about the trucker job or some other type of thing. So it's been this like odd coalition of folks sometimes who who've been kind of who've who've who are who have like big concerns about some of these like new technology things that are happening.

• 20:28Speaker 00

  One thing I think is really interesting is that this is a technology where, you know, obviously, software is at the key, but hardware, I mean, to your point, is really, really important and infrastructure. So it requires us, you know, not just to think about sort of what's happening on the software side and the revolution there, but also hardware and infrastructure at the same time, which, you know, obviously has longer lead times, more capital investments. That's that's what's affecting the business models of the labs, too. They're in a fundamentally different business. And so I feel like these revolutions happening at once are why you're seeing lots of different patterns of how people are engaging with AI and a scrambled politics that doesn't look like what we've seen before. I think that's actually a harbinger of things to come, where things aren't going to look like the playbook of the 1990s anymore. This is a different kind of technology and different kind of discussion. I, for one, I mean, obviously, these things are challenging for all parties involved, but it is good to

• 21:12Speaker 00

  have people expressing their voices on these things. I mean, this is the process that should work. So the staff of these different groups are organizing and sometimes organizing together for the first time, you know, it's a sign that this is new and we haven't done this before. So I feel like this is how I look at it. And we'll probably be having a different conversation six months from now or a year from now than we are now in terms of how the impacts of AI are perceived.

• 21:30Auren Hoffman?

  December was like this different step function. It was a huge step function. Since December, you know, you and I are talking now in July 2026. So since December, it's been like, to me, it's been like this progression that's happened, but has not yet, it hasn't been like the new step function. At some point, I expect maybe even this year, there will be another step function. What would that look like? Like, how should we be thinking about that from our lives?

• 21:56Speaker 00

  Actually, let me let me slightly disagree with you here, but then agree with agree with something you said before, because I think you've been really focused on adoption as much as capabilities. And I will say, I think the next step function is actually adoption. If I think about the way I'm using Codex now, and this is literally because I had a really great trainer at OpenAI who sat us down, a group of about seven people, small training, role-specific advice about how to use Codex, gave us a harness to use along with it and said, and walked us through. And I, look, it depends how technical you are. This is the kind of person who was coming around to our computers, making sure all the plugins work, making sure that, you know, because you know how these things are. You could be in a big room and you're, you know, you're, you don't have the right version or there's something conflicting on your computer. People get slowed down. I, for me, had a really effective training that met me where I was, and that accelerated my usage

• 22:40Speaker 00

  so much and allowed me to skip a bunch of steps. And after that, the way I use it is now totally different than it was before. And I feel like adoption vis a vis capabilities improvement is something to really watch. You're going to see tremendous capabilities improvements. You've been seeing that measured by the evals. But I think the unlock, that step change you're going to see is actually more likely people actually adopting agentic use cases for things that make sense in their role. And then you're going to see things get very, very interesting. I also think that adoption needs to be looked at, not just do I use it once a week or not, but the intensity of use. And that's something we measure a lot. In economics, you talk about the extensive margin versus the intensive margin. I think we're past this idea of just thinking about breadth and also thinking about depth. So those are the two things I think that you might see from this step function. I'm sure there will be so many interesting things getting cooked up on the research side.

• 23:25Auren Hoffman?

  We don't have to make any new gains in capability, and we still have so many gains in just adoption, essentially.

• 23:33Speaker 00

  Every time I use it for something new and someone in research, I'll say, gosh, you know, when did you, when did this, when was this possible? And they usually say, oh, about six months ago, Ronnie, you're just catching up with us, you know? So I do think some people know earlier, some people understand it in some people's workflow. Now I'm doing things where I realize, oh, gosh, like, I never needed to do that before because I wasn't using Codex, and now I need that, and Codex can do that, too. And so I think this, this is the change in the next six months. Because I do agree, I don't know what that November, December conversation is going to be this time around in our group chat, but I hope that it is something around adoption and more people using agenda tools.

• 24:03Auren Hoffman?

  Well, I found in organizations that have been this, that it is very contagious. And there's kind of like two ways the contagion happened. One is like you're all in the, you're all like right next to each other, working, and you just kind of like you're talking about it. So people see what each other do. The other way is you're using some sort of group chat like Slack, and people are adding the agent in the public channel, you're in the marketing channel or whatever, and they're asking the agent to do something for them. And then everyone else starts to catch on, oh, I can do that. And then, like, at least in our channels, the agent is responding to at least half of the Slack threads anyway, like chiming in, jumping in. You're like, oh, okay, like, I understand, like, here's

• 24:44Auren Hoffman?

  what's going on. So of those two things that have really driven the adoption for us.

• 24:49Speaker 00

  I find it very similar, and I, you know, I hope it's true in all organizations, it's true at OpenAI. It's a collaborative sort of encouragement, you know, it's not as if, okay, someone's saying you have to do this or whatever, you know. What I find is it starts with leadership. So, you know, Sarah Fryer, my manager, the CFO, she really made it clear to all her direct reports and the teams underneath us that, you know, using Codex to accomplish our task is really important. And she understood we all had different roles and different ways of pursuing that. But she really encouraged us, and she provided that training that I talked about. She did it with us, which is really important. So I think sometimes we do forget the human element of this, like having my manager recognize this is really important, having us do a training altogether, her included, really important. And Sarah, having it something that when we do our one-on-ones, she asked me about, again, tailored and customized, I'm not, you know, in a different role in

• 25:33Speaker 00

  the finance org, I'm the chief economist, so it's a different job, but trying to explain how we're using it and why it works well for that, really, really interesting. So I feel like that's one thing that people often forget, the human leadership in making this work. And also saying, look, this applies to everybody, from the most senior person in the company, to the chief economist, to everybody else, technical, less technical, that's really key. And then the other thing is seeing my colleagues, particularly my team members on econ research, do things that when I was a graduate student would have taken me like three weeks. They come back to me and they're like, hey, I just cleaned this dataset afternoon. And I just said, oh, my gosh. So it inspires me, although they joke, it means I give more work.

• 26:06Auren Hoffman?

  Cleaning data is like, it's like for me, which I, because I spent, I spent much of my career just like munging through data and stuff. And just now that I could do it in sometimes a few minutes, it just, it blows my mind, but it also makes me so sad that I spent so much of my career doing that.

• 26:22Speaker 00

  Somewhere Travis May, our common friend, he's shedding a tear, right? Because you guys did amazing stuff, right? Although he's probably building the next thing, I'm sure. But with with folks like us, we did the grunt work to match the data and many people, more than me, I remember my dissertation, just big data matching exercises. Now those things can be done a lot quicker, but now we have time to ask more interesting questions. So for me, it's been a really interesting thing. When I see these advances in the methodologies and the data collection and processing part, I think this should free up more time for us to ask more interesting questions. And, you know, in AI research, they call it taste. In social science research, I also call it taste, which is just like, what kinds of questions do you ask? And you have more flexibility to do that when the execution part is easier.

• 26:59Auren Hoffman?

  How is this showing up in GDP? Like, how, where is this being affected? How is this happening? I mean, there's obviously like the data center side and the selling, you know, the OpenAI revenue and, but how is like the productivity, how is all these other things like happening in the broader economy? And how do we even measure those things?

• 27:18Speaker 00

  I think it starts with this, that you're right, that the capex channel, right, the investments in sort of hardware and infrastructure did drive in the US some increase in GDP, especially in 2025. And I'm sure it'll be true in 26 as well, if the data holds up. I think the second piece, though, that is not measured in GDP or traditional statistics is what economists call consumer surplus. You know, it creates value for me. So you think about those examples that we talked about before, you're in Nashville, you're asking where to go, what restaurant to go to, what bar to go to to see great music. I'm thinking about what story to read my kids. I'm saving time fixing stuff around the house. Those things actually create a lot of consumer benefits, but they're not priced, and many of us are using the free versions. You might ask, okay, well, that's definitely value. Where does that show up in GDP? The answer is it doesn't. And consumer

• 28:03Speaker 00

  surplus is something economists have thought of different ways to measure. There's a good method by Eric Brynjolfsson in the Digital Economy Lab called GDPB, which can be applied here. And we've done a little bit of the work on this. But it's probably in the hundreds of billions of dollars in terms of things that ChatGPT, for example, is creating for consumers. They're not priced in the same way. So I think, like, we should first start with, it's creating a lot of value, and that's not controversial. What is challenging is when does it show up in the productivity statistics, which is very important as well. I think for that, I mean, compared to previous technological booms, right? Usually you'll have a lag of several years before that happens. I think we're on our way to sort of a more quicker turn, given how transformative AI is in the adoption part of it. Not surprising to me, it hasn't showed up in terms of driving productivity and going up to GDP yet. You are seeing to see the scattered anecdotes. St. Louis Fed just released some good data

• 28:47Speaker 00

  on this with some, some stories from sort of firms in their districts. That's the kind of stuff that needs to aggregate up to get to where you see it at the macro level. I will say that right now, you see a lot of heterogeneous use of firms, right? So you see the real power users in the real power firms. You might be, are in the power user in a power firm. They're at the 99 percentile of every distribution of token intensity of the kind of workflows they're doing. There's a big split between those people in the 50th percentile, both within the firm and across firms. And that looks a lot more like the IT revolution in the late 90s and early 2000s, where a set of firms became super IT intensive. They got the managers who could complement the IT. They saw big productivity increases compared to firms that didn't. I think it's going to be more of a story of who's using this to get productivity and some who are less so or less able to. That's probably

• 29:32Speaker 00

  the thing to watch in the productivity statistics over the next couple of years.

• 29:35Auren Hoffman?

  In some ways, if everyone gets all the benefit all happening at once, like, how does it, like, in some ways, it just kind of increases the bar that's happening. This is one thing, like, I don't know if I'm getting my taxes done from my accountant, like, there's a, there's like a, you know, that maybe they just dropped the price because it's, instead of charging me X, they charged me X over two because they can do it in X over three times. And then that, that's all I need. I just need my taxes done correctly from them. But then there's other things where like, I just, I want more. I'm like, my, my bar for what I want goes up. And then they have to still spend the same X amount of time on it and they still have to now they're just like delivering me more stuff and there's maybe like a consumer surplus, but like the same dollars go to what they do.

• 30:20Speaker 00

  This is why economics is so interesting. There are two points you've made here at least. One is there's not a fixed set of output that one can make. So if you're writing code or doing tax returns or graphic designs, we shouldn't assume that the level as of July 2026 is the level that is sort of the only one that's possible. When you lower the price of something and demand is elastic, right? People might buy more of it. That could actually create some more jobs in those areas too, as people are starting to provide more of that, right? I think about, we released a paper on AI and jobs and we find, you know, let's say anywhere from like, you know, 14% or so of jobs in the United States, we think have this kind of quality that when sort of the price of producing something goes down, you might actually see the mix expand. So that's one. The second thing, which is more

• 31:03Speaker 00

  subtle, is the evolution of our tastes and wants. You know, now that I can talk to Codex and do things that I couldn't do before, I'm actually looking for some different sort of other parts of the workflow to complement that I never even knew. And I might be able to buy things or think about buying things I didn't have before. It might create job descriptions that never existed before. And so these are the kinds of things where the economy really grows on. And research by David Autor and at MIT and other people has shown that like a good proportion of jobs are things we didn't even have names for, like, you know, when they started counting this stuff. So this is what gives us optimism. It is, of course, hard to know exactly what they are. So I'm sympathetic. I think about the new grads all the time who are coming out of school or people starting school this year. It's hard to know exactly what those are going to be, but that has been the story of technological change in the economy for a long time, in the United States in particular.

• 31:46Auren Hoffman?

  What do you think is the most important skill they should have in this new world that we're in?

• 31:51Speaker 00

  Yeah. Well, I, it's funny. I don't even think about it for a random one. I think about it for my own kids and kids, their friends and things like that. And but of course, for anyone listening, my view is that technical skills are still going to be important. I think that, you know, it's so funny. We sort of like over-calibrated in some ways to STEM. I think you remember like 10 years ago, everyone was saying that we should all study STEM. Countries were introducing mandatory coding classes for four-year-olds or whatever. And now everyone's saying, oh, like, you don't need to learn any of that stuff. I think we swung the pendulum too fast both times, right? I think those hard skills are really important. Like, if you look at the typical computer scientists or data scientists at OpenAI, they understand the way Codex works much better than I do, right? As an ECAT major. And I think there's something to that. Eventually, the operating system or the interface becomes so good that, you know, people can use it similarly.

• 32:35Speaker 00

  But there's always a lot of knowledge that complements the tech stack that's really, really important. And that, by the way, I also think learning things like computer science, the logic that goes into it, the formalism that goes into it, the math that goes into it, it's not just for computer science. The theory, it's used in so many other ways, as a way to think and view the world. And it's the same reason if you're interested in philosophy or social sociology or economics, I'd encourage you to study that. So one is learning how to think, learning hard skills, sort of STEM skills, still really important. I also, though, do think the human skills, like, that are going to be real compliments to AI, I think you probably see this in your org too. You might have hundreds of agents running and sourcing investments. The question, Orin, is who's going to make the decision over what to invest? And ultimately, if the investment goes south, who's responsible? And I think at the end of the day, right, that's going to be you and the team. And

• 33:18Speaker 00

  so making decisions, delegating the work, and really inspiring the other people in your team, humans and agents, that's going to come to people like you and that's what we're training people for. So I feel like those human skills are going to be really, really important, even in a world where machine intelligence is ascendant. And so that's what I would ask folks to focus on. I also think flexibility is important. Like, I think there was never a sure path or a sure thing to any particular destination point. It's probably going to be more uncertain in the future because of all these technological changes. And so flexibility and resilience, really difficult to develop, but are important. And I have a lot of empathy for people who are going through that now in their careers and developing those skills. And we just got to do the best we can by them and try to do both from an educational standpoint. I think about this in my professor role, from a policy standpoint, from a business leader standpoint, to give people the support

• 34:03Speaker 00

  they need. But that's my hopeful take on it.

• 34:05Auren Hoffman?

  At least in the last six months, when I've been like evaluating young talent, the most talented people have one very key thing in common, which is they have all taught themselves something relatively complicated from a YouTube video. It could have been like teaching themselves how to play guitar, which is, you know, I've not, I don't know how to play guitar, but I imagine it's a super complicated thing to learn. And they didn't do it because their parents told them. They didn't do it because their teacher told them to do it. They did it because they, they had their own passion to go do it. They went to YouTube or wherever they went to. Maybe they went to ChatGPT or wherever. They went there and they went kind of step by step with this kind of like very uncertain

• 34:48Auren Hoffman?

  learning curve and taught themselves how to do something complicated. I think if you can do that one thing, you're kind of set for the rest of your life.

• 34:55Speaker 00

  I think about this as parents. I can't help but go back to this. Like, maybe our job as parents, the best we can do is, you know, rather than like, you know, quoting the graduate, the Dustin Hoffman scene where they say, plastics, you know, we can't do that, right? We don't know what that is. But what we can do is say, you know, what is it that you want to learn how to do? And here are some tools that you can go and chase it. And, you know, it could be applying, you know, advanced statistics, understanding their favorite basketball player's performance. It could be like building a computer. It could be figuring out a way to get into outer space. It could be a way to solve a social problem in their community, whatever it is, right? But using tools, AI and otherwise, to kind of be able to work through, learning how to do that, it's really empowering. And that sense of accomplishment, I think you're going to want to feel that again and again and again. And it cultivates a sense of agency that I can control my own environment

• 35:40Speaker 00

  and make a difference. And I have the tools at my disposal to do that. I think that's our hope, right? In terms of how we're trying to build these things. What I think about every day, actually, in terms of the evolution of AI.

• 35:48Auren Hoffman?

  A lot of AI leaders are leaning into UBI, this universal basic income kind of thing. How do you think about that? Like, in some ways, like, yeah, it's good to have a social safety net, in some ways, but also like, okay, does it sap the agency out of people? Or does it, like, how do you think about that? But not just from an economist, but from a philosophical standpoint?

• 36:12Speaker 00

  My views are informed a lot by my government service. I mean, I think the one thing I've had, I think I work with people who are much more talented than me, but I had like a couple of different kinds of experiences, you know, as a professor, you know, writing articles, going deep on things in industry here, and also working in government, as you mentioned, the chief economist of commerce and working at the White House. I worked in those government roles during two, you know, very seismic events in our economic history, you know, the great financial crisis and then again during COVID. And so I thought a lot about, like, what do, what do people need when there's economic disruption, lots of change? I think, again, because of those experiences, one thing I lean on is we actually have some of those levers in government now. They're called automatic stabilizers. So when the

• 36:52Speaker 00

  economy goes sort of down and we see an increase in unemployment rate, right? We often will trigger other sort of things to ratchet up, like more increased food stamps, unemployment, and eat, and things that are done legislatively, things that happen automatically too. So I would say if you look at the fiscal monetary response to the last several crises, you'll see the foundation of a toolkit to deal with economic disruption. That was built, you know, in the, in the wake of the Great Depression and the lessons that economists and other policymakers had. So while I love building new things, I love thinking about new big things, I actually think focusing on improving the safety net, stabilizers, using our best forecasting to figure out what could happen and improving those capabilities in government, that's really important. It's not as sexy, but fixing like the state unemployment system,

• 37:36Speaker 00

  some of which run on really old software platforms, can actually be a huge benefit to getting people the help they need quickly. It's, it's not as revolutionary as new policy idea. And, and OpenAI has released some policy ideas, like an industrial blueprint, but it's one way to start. And then you think about the bigger ideas. And I think it does have to be very creative across government. There's not going to be one silver bullet here in how you do it. But I would focus on the automatic stabilizers, think about some new policy ideas, getting kind of bipartisan support on top of that. But my, my, in terms of immediate things we want to work on, I think the levers that we have could be sharpened and improved.

• 38:08Auren Hoffman?

  We talked a little bit about it, but how also you guys, you using your, your, your team is using AI?

• 38:13Speaker 00

  I think I'm saving my team time. They're going to laugh when they listen to this because they think I create work for them. But I, here's what I do. I love doing conversations like this and I do a lot of public speaking and you know this, I know this. It takes prep to do these things, to think about, hey, what have I done? And in terms of my research, what do I know? What do I know that's interesting to other people? And so often, if you're someone like me or you, you'll have a prep document with all the different things that you've done and some suggested things to say. And right before the discussion, and usually you're back to back to back, you are trying to cram that information into your head and it's never good enough. And you have to drag some poor person on your team into it who has to sit there and listen to you, you know, for each engagement and help you. I use Codex for that. And as I was telling you, you know, the team still made the prep document. Eventually, I think we'll, we'll be able to outsource that too

• 38:57Speaker 00

  to Codex. But instead of reviewing it with them the half an hour before we start the show, I just have Codex read it to me and with me. And I say, hey, you know, what if I mentioned this fact about how quickly non-technical parts of the org at OpenAI adopted, you know, agented tools? How would that sound? And it says, well, actually, you know, the asking doing point you made in the ChatGPT paper is a really nice contrast because you had asking ahead of doing and now it's changing. That's something that my team might have come up with, but I don't have to waste their time. And then what I say is, look, how much is this stuff is things I've said before that people, you know, might find interesting, but maybe it's boring, boring to me if I said it before? What's something new I've learned in the last two weeks that maybe I put down in one of those markdown files? And aha, I've had about AI and science or AI and another part of my life. I'm going to make sure I bring that into the conversation with Orin. So, you know, I have more fun,

• 39:41Speaker 00

  he gets something new. That has been the most sort of recent interesting use for me. And the voice to voice, as I talked about, I like talking to Codex and hearing it talk back. The rhythm works for me. I can also type and do some other things at the same time. That's my newest use case. And for me, something really resonates with me about the voice. I realize I'm not necessarily average or typical there. The, the talking and the talking back, that was aha for me in ChatGPT and it's an aha for me now with the agented tools as well.

• 40:05Auren Hoffman?

  So one of my friends, his, his head of sales is always just calling him to work something out. And it could take sometimes 30 minutes and the head of sales is like ruminating. And all my friend is, is he's the CEO, he goes, uh huh, uh huh, uh huh. He never says anything. And like, at the end, like, the head of sales just like had figured it all out on his own. It's like, it's like, I'm not sure if I, like, if I could about AI, just kind of like, it's like, you just need to have somebody there to just like help you think it through.

• 40:32Speaker 00

  Yeah. And I'm hoping to save them time. So now what are they doing? Instead of spending a half hour with me on the call reviewing a document they've already written, they're doing something else for the team, you know? And I think that, to me, could be a big creativity unlock for, for a team like ours that has like economic research, external obligations, internal obligations. We're trying to balance the, the production function of the content as well as the delivery. That's really important for our process.

• 40:53Auren Hoffman?

  How do you use AI to either be a better father or a better husband? Like, in the family? Like, how are you using it?

• 41:00Speaker 00

  I definitely try to fix things around the house. I think it's debatable whether I become a better husband as a result. I'm definitely volunteering more to fix stuff. You know, if our kid, it was about six months ago, but I had written about it internally. It's basically like, you know, her bike was broken and, uh, you know, I'm usually like the last person people ask to try to fix like the derailleur on a bike, right? It's just, I don't even know how to spell it, much less, you know, find it, fix it. I, you know, I used the, the video and I watched and chat kind of patiently guided me through it. And it was funny because like I kept making the same mistake over and over again and kept saying the same thing, like look for that little notch. And finally, after half an hour, I was like, oh, that one? And it's like, yeah, but it's very patient with me. And I was able to fix it. Now meanwhile, my family went on a bike ride, came back, but at least, at least it was fixed and my wife, you know, maybe had some renewed confidence in me.

• 41:43Speaker 00

  It turns out being a dad, I, I, yeah, she still doesn't want me to touch, touch some of this stuff around the house, but with the kids, I, I've really enjoyed like making creative exercises. If I see something they're interested in, it goes back to this point before about agency and having them chase their, their dreams and curiosity. My tendency is always like, oh, let's build a game around that or let's build, um, that's cool. Let's build an interactive story around that. And it doesn't always work, you know? I, I think as a father, you evolve, like sometimes your kids actually just want to lounge with you on the couch and chill. Other times I want to go get ice cream, like, you know, technology doesn't solve that. But, but I do think if I can help kind of meet them where they are with a fun exercise and maybe voice will be the unlock for them, or it could be, you know, typing or whatever it is, or sports or some sort of topic. I'm just trying to look for those things and probe and, and learn a lot from my friends who, you know, who have kids the same age and kind of see what they're doing.

• 42:28Speaker 00

  That's how I've been thinking about it.

• 42:29Auren Hoffman?

  Now, in the economics world, there's a lot of talk about the population crisis and I have this crazy theory that AI could lead to a population boom. How crazy is that theory?

• 42:42Speaker 00

  It, it's not crazy. I think there's, there's countervailing forces that we'd have to consider. Like if you think about as societies get richer, they typically have fewer kids, right? There's been some like, some, you know, obviously like longstanding trends there. But what I find interesting about AI is that if it does create lots of prosperity and more leisure time, we, there's an open question of how you spend that leisure time. For some people, it's obviously, you know, pursuing their interests or, you know, doing things that they don't have time to do at work. For a lot of people, I believe that that's spending time with your family and having more children could be conducive to that, right? You think about if, if how many people would say, look, if I had more time and less stress about work, I'd actually have more kids. And that's a mechanism and a function that could go in your direction.

• 43:22Auren Hoffman?

  Going from zero to one is much harder, but going like for people, if AI could potentially help you go from one to two, from two to three, from three to four, it just like that one little notch like would just massively change the, the population curve.

• 43:35Speaker 00

  So I don't think it's crazy to consider. I think it's something we'll be studying in the years to come, but you do have these macro trends about as societies get wealthier, you know, they have had fewer kids. So you have to think about it against that baseline, but, uh, that's, that's where I think you, you're going to watch this and it's very interesting to think through.

• 43:47Auren Hoffman?

  Self-driving cars seems to me something where that could help get more kids. Cause a lot of like what you do as a parent is like schlep your kids around and do these. So if you could, if it could make it a little bit easier, like sometimes it's like, well, how do we have more than two kids? Cause it's like one parent has to drive here, the other parent has to go drive there. Like how would, how do we do like, so just like logistically, you can see like, oh, well now we can have a third kid or something, right? There starts to be this other like things.

• 44:14Speaker 00

  And when you streamline logistics in general, chores around the house, organization, management, it, it might, it might help. And I think that as you move towards robotics one day and, and more intelligent robots, think about the kind of household labor that can be done with robots, they start to think about the burden of, you know, child rearing in a different way. You arguably will then develop even more refined tastes and have more expectations, but still, right? At the end of the day, some of these technological shifts can make a big difference. I think population is something important to be watching.

• 44:40Auren Hoffman?

  Number of kids per married couple in the U.S. has basically been flat for 30 years. The real population decline has come from people getting married at lower rates. Could AI help for people, you can imagine it could, it could both be a negative because it's like too many other great things and I'm, I'm going to delay or not get married. Or you could see it like, oh, it could be that it could actually help you find your soulmate better or you're more likely to get married. Like which way do you think, or do you think like the forces just countervail or something?

• 45:11Speaker 00

  It depends on the forces and which way they go. Like I think people had same debates about the internet, right? Some people thought the internet might increase loneliness because it would give people entertainment, you know, separate from hanging out with other people. Other people said, wait a minute, now the internet allows you to connect to people you never would have met before. And I think both are probably true for different kinds of people. And so, um, lest I give any date, lest I give any dating advice, my wife would, you know, immediately criticize me as having no real knowledge about this. But I would say you, you could see the forces going both ways and, but I, but I do think this, this is what's so interesting and where I think universities, think tanks, researchers can play a role. We're to be able to study the impact of AI on these, on these trends you're talking about. And that's going to be really important as a foundation for social science over the next 10 years to understand what these things are and, and where it's creating new opportunities and new challenges. I, I'm looking forward. It's not the kind of work we're doing inside the econ group, but it's kind of the kind

• 45:55Speaker 00

  of work I'm sure other people are doing and should do.

• 45:57Auren Hoffman?

  How do we get more data available to more people about the economy, right? You have this unique data set that is internal open AI that very few people have access to. You've got maybe someone like Raj Chetty has access to the IRS data that very, very few people have access to, which is just like insane longitudinal study of hundreds of millions of people, right? Which is incredible data, but the, there's the, they're the unique data sets that are, that are really gated for sometimes for good reason. How do we get more data sets available to more people so that we can do more research?

• 46:36Speaker 00

  When I took this job, it's something I thought a lot about, you know? I got a lot of advice from folks when I took the job and I asked a lot of people who've been successful in their careers and business and government and academia and philanthropy. And I said, you know, what's your advice? I, you know, this seems like an opportunity of a lifetime. Don't know how long I'll have a chance to do it. Why do I maximize it? And a few people kind of honed in on this idea is there's got to be a set of things and maybe it's one thing that, that can happen or not happen based on whether you were there or not. You know, there's lots of things that chief economists, open AI would have done, no matter who it was, right? But there's some things that, you know, your particular interests, your view of the world really matter. And when I came in, I said to myself, we need to release more data for people to use. And the team is obviously very excited about it and, and team members, you know, past and present were really integral in this.

• 47:19Speaker 00

  But as I think about this, releasing that data, it's, it's called signals. We can maybe provide it in the show notes. We have signals for consumer and signals for enterprise. And that was a real important work stream for my team. So as a leader, if you believe on something, you believe in something, you have to prioritize it. Prioritizing means, you know, going to talk to folks in legal, in the data science team, folks who think about privacy, people who think about investor relations, all those kinds of folks and say, here's why this is conducive to our mission at OpenAI, which is to build AGI to benefit all of humanity. Humanity will benefit if they have more information is my simple view and releasing data like signals, which allows us people to see all the trends we're talking about. They can download it, they can use codex to run their own regressions. It's really cool. I wish I had this as a graduate student. So to me, I think the way we do that is get more people in organizations who think that this is a priority.

• 48:04Speaker 00

  In the past, as a person on the outside, working with a lot of companies, I always met some fellow travelers who were really interested in data, but it was never their number one. They were always like, look, I'm trying to get this other business thing done. I'm fine to give you some data for your research project, but by the time we got everything worked out with the lawyers, the person was gone to the next position. For me, this was an OKR, an objective, a key result, and we achieved it. And I'm going to keep trying to push out as much data as we can. Of course it's challenging and there's trade-offs and, and things I have to navigate. I can't do everything I'd want, nor should I be allowed to, but that's one thing we prioritize and I'm proud that signals and B2B signals are out there.

• 48:34Auren Hoffman?

  It's hard because if you make a little mistake, it could set them back. Like if you remember, remember this is like 20 years ago, the AOL searches that got released and it was like, I don't know, someone wrote like, how do I kill my spouse in it or so? You know, you start to like, there was all this weird stuff in there that kind of like derailed it there and.

• 48:51Speaker 00

  Every time we release something, I hold my breath and I, you know, for me, it's did we make a mistake, you know, will someone find? In the end, you also rely on your users and people who are using the data. We have a really dedicated use group of economists and statisticians who are using our data. I'm hearing from them. But you always take a deep breath. You're worried, did I make a mistake or did I not do it right or whatever? And it takes, it takes a lot of guts to do it and so far it's paid off. And I've gotten a lot of support from the organization and, uh, I'm going to keep doing as much as I can on that front.

• 49:15Auren Hoffman?

  Going back to your government hat on, how do we get the, the government has a lot of great data, like the IRS data is a great example. Okay, yes, it is super, super sensitive, but there must be a way that we could sanitize it and, you know, do some things on top of it so that we could give it to a much broader set of researchers where obviously there's no personally identifiable information as well, or, you know, the Medicare data, you just think of all these like amazing data sets that the government has access to. There must be a way to have your cake and eat it too. We can protect privacy completely and let, have great research on it.

• 49:51Speaker 00

  And we can give, you know, we can allow people to go through trainings on how to handle the data and have access. You know, it doesn't just have to be one researcher, it could be a group and we can make sure that they're, you know, they're qualified to do it. I think if you talk to both multilateral organizations like the IMF, the World Bank, you talk to the Department of Labor, you talk to the Census Bureau, all the different agencies, fiscal agencies in the U.S. government, a lot of appetite for this. And, uh, and I think AI is making it even more pertinent. There's a recognition from the Federal Reserve system to the government, to sort of local community groups, you need more real-time data on AI usage. And that's why we've released the data, but we, we're going to need that from more labs and have it be reconcilable across labs. And again, challenging, but, but something that can be done and something I care a lot about.

• 50:29Auren Hoffman?

  What do academic economists get wrong most about the economy?

• 50:35Speaker 00

  If you think about how technology's diffuse across society, you know, there's all these models of how technologies, well, new technologies, innovations will lead to economic growth. What's often misunderstood or underappreciated is what's called the last mile. And what I think about this is like, organizations don't just seamlessly adopt technology. There's a human element, there's organization, a legal infrastructure that matter. As a result, the rate of adoption may be different than in economic models that we assume. And it could be heterogeneous by different kinds of firms. It could even matter based on the idiosyncrasies of the leader, the CEO and which career preferences are. So missing those kind of frictions sort of make economic models a little less effective maybe in predicting and forecasting what's going to happen because a lot of these things, right, are often assumed to have adjustment costs that are very low or the frictions are abstracted away in the models.

• 51:19Speaker 00

  A lot of economists are working on these kinds of things, but as a person who kind of made his career in a business school and studied a lot of economics, but then applied it to sort of strategy, organization, entrepreneurship, I think a lot of those organizational frictions are often overlooked in much of the economist profession. Not all, I mean, there's fantastic or econ people who are, who are doing that work. Connecting that to the macro and our models of growth and forecasting, that's a key challenge, something we need to think more about. So that, that's the place I think we can, we can do a lot more in economics.

• 51:46Auren Hoffman?

  Of all the things that would be classically in the College of Letters and Sciences. Why is econo-- it seems like economics has for the last at least 20, 30 years has been the most high status thing, like more than history, more than philosophy, more than political science, more than-- why, why is that? Is it just because like it just because we live in like a practical world and it's a practical kind of discipline?

• 52:13Speaker 00

  I found a lot of value in particularly for me personally, my political science classes, my history classes, actually kind of made me more excited about economics because it gave me the context to ask questions that maybe weren't being asked in economics. I think the questions that I was introduced to early in economics were relatively narrow compared to what I saw in the history and poli sci classes. Later I got more exposure to sociology. So I find, I kind of feel like those disciplines give me fantastic toolkits to ask questions. I think economics has the appeal that for a long time when I was growing up, people thought of it as like a practical major to get a job in finance or consulting. Although, you know, if you know those firms, I know you're following them quite a bit, they hire people from all kinds of backgrounds, but economics was seen as one. Two, more quantitative sort of work earlier on. The other fields are very quantitative now and

• 52:54Speaker 00

  have people using advanced statistical techniques for sure, but economics probably adopted that early on and it had more math in it that gives it sort of the advantages of maybe seeming more technical, if not practical. And I think those are the things that kind of, that kind of elevated economics. And I think living in a market society, right, where market is at the center of, of, of the economy, that also makes economics a different kind of field in terms of how students think it's staying. I think teaching matters. I think the job market matters and I think the problems you want to solve in the world matter. And so that's why I think over time, like things like computer science have actually overtaken at some schools, economics majors. And I think, you know, that could be part of a longer term trend. We'll see what happens now with the AI revolution.

• 53:34Auren Hoffman?

  All right. This has been awesome. Last question we ask all of our guests. What conventional wisdom or advice do you think is generally bad advice?

• 53:41Speaker 00

  People often say something like, success leaves clues, meaning you should follow the paths of other successful people, ask them what made them successful. This isn't a lot of like self-help books and other areas. And I actually think that that is relatively bad advice because what I find when I talk to really successful people is that their path, if you actually wrote it down based on their CV, is both incomplete. There's a lot of stuff that's not included there, a lot of struggle, a lot of adversity, a lot of questioning. And two, the path that they took, whenever they took it, very context dependent, much more so than most of us are willing to admit. Very rare to find someone self-aware enough to say, you know, the way I did it wouldn't work for you anymore. It wouldn't work for anybody. And I feel like we should use our heroes, our, our career heroes, our, our occupational heroes for inspiration, but not step-by-step advice.

• 54:25Speaker 00

  Their career paths probably, if your field is changing fast, are not going to apply to you exactly. And I think charting your own path in terms of figuring out what problems do I want to solve. It's really hard to shift from, I want to be that person, right? Famous entrepreneur, politician, you know, sort of enterprise leader, sports star, and say, here are the problems I want to work on. And here's how I'm going to design my career to have a shot to be at the table where those problems are solved, not reach a particular position or sort of mimic someone else's career. Success leaves clues, but it's not in the step-by-step. It's often in the habits, right, to identify interesting problems to solve. That's how I think about it.

• 55:01Auren Hoffman?

  There's probably some things like that are timelessly like, you know, that are probably good advice for centuries or something too, right?

• 55:09Speaker 00

  I would say focusing on the problems you want to solve and it puts you in the contemporary context, that can help a lot. Because I do think I find people that say, okay, well, my, my advisor did his PhD at this age and then got tenure at this age, that kind of thing. Or someone found their first company by this stage of life and like, you know, the economic environment matters. It determines when the best time to find a company, found a company is, for example. Or, you know, someone had seven jobs before they were 40 versus one. Those things depend on the dynamism of the economy. So these are the kinds of things. I think going step-by-step, success will leave clues in, in some ways that are a little more abstract than the prosaic blow-by-blow on the CV.

• 55:42Auren Hoffman?

  Thank you, Ronnie, for joining us on Summation. We've been friends for a long time and I knew this would be great. So this has been, was awesome. By the way, I follow you on X, at RonnyJetterty. I obviously follow you on LinkedIn as well. So this is awesome. Thank you so much.

• 55:56Speaker 00

  Thanks for having me, Orin. It's been a pleasure.

• 55:57Auren Hoffman?

  One more thing before we go. I read a blog called Summation. It's the same as this podcast. The blog is about non-obvious idea sharing on business, talent, data, longevity, and random contrarian takes. If you like the conversations on this show, you'll probably like the blog. It's free. New content comes out twice a month. You could subscribe at orin.substack.com. That's orin.substack.com.