Hello, OddLobots listeners. I'm Joe Weisenthal.
Tracy, we were down at South by Southwest recently.
Yeah, that's right. I didn't get to have any barbecue in Austin, but we did have some really good Mexican food. We did have really good Mexican food and we recorded a really fun episode of our podcast.
That's right. More importantly, we recorded a good episode.
That's right. So, check it out listeners. We did a live episode of the podcast talking about sort of the prospect of mass white collar displacement due to AI as well as the politics and how politicians should be thinking about this and how voters are thinking about this. Our guests were David Shore, founder of Blue Rose Research, and Byrne Hobart, who writes the excellent DIFF newsletter and a general partner at Anomaly Fund. Take a listen.
Thanks everyone for coming out on a cold Sunday morning.
Yeah. Hello, and welcome to a live recording of the OddLots podcast. This is definitely going to be the most uplifting of all the conversations that have happened this weekend, right?
I think so. I think it really will be. We are... Everyone here is going to leave feeling really good about the future.
All right. Well, let's kick it off. So, we have two great guests. We are going to be speaking with David Shore. He is the founder of Blue Rose Research political consultancy pollster knows a lot about AI and we have Byrne Hobart, the founder of the DIFF, a great newsletter that everyone should read and a general partner at Anomaly Funds. And we're going to talk about all things AI and job loss potential and the politics of it and so forth. And so, David and Byrne, thank you so much for joining us on stage here.
So, let's just start on this question. David, you're pretty like this is happening now. The economy is going to look radically different maybe even a year, 18 months from now and you're trying to wake politicians up. This is happening right now. But like tell us what's happening right now or what's about to happen.
Yeah, you know, I'm not an AI expert and I don't want to claim I'm one, but you know what I'll say is that I use AI a lot. I use Claude Code a lot. I think in December I spent maybe like 15% of my pre-tax income on Claude Code overage fees. And I think there's a real disconnect among my friends between, you know, people who use Claude Code and people who don't where I could really feel the extent to which I can do so much more now versus a month ago and a month before that. Like the scale at which these things are getting better and the speed at which things are getting better is really jarring. And I think you might not notice that if you're just using ChatGPT like you know ChatGPT is like a little better than it was a year ago. But again it's very hard to make predictions about the future but I do think that if things continue to improve on this scale then things could get really weird really quickly.
And Byrne, one of the reasons we wanted to talk to you is because you're sort of at the intersection of technology and finance. If I look at the valuations of big tech at the moment and a bunch of the AI companies, they basically suggest that they're going to take over the entire economy. When you see those numbers, what do they suggest to you about the future of labor?
So, I think right now it's definitely true that you can think of this coherent category of this is an AI company or this is a non-AI company, but they're incorporating AI and they have these distribution advantages, so they'll probably do well at that. But I think that it's a mistake to think that this is a durable discreet category in the same way that you could refer to a lot of companies as electricity companies. And if it were 1925 and you're trying to figure out which stocks to buy and you're just really really bullish on electricity, then you would really really care that RCA is exposed to electricity and I don't know, US Steel is mostly not. But eventually every company becomes an electricity company just in the sense of you would run a very different business if the lights did not turn on. And so it gets subsumed by the rest of the economy. And you see that with software too where there just a lot more companies that have software developers and are writing software for internal use. And they're not software companies per se. You know, they're restaurants or like tractor companies or whatever, but they have that element. And so I think that's part of what you see just early in the roll out of any general purpose technology is that you have a lot of really narrow specific bets and then over time the impact gets so widely distributed that it's very hard to trace and you have to kind of go back and look at the history and look at things like okay the rise of the car leads to the rise of the suburb but it also leads to the rise of the grocery store because or like the supermarket where you can have a much larger selection and that means lower labor cost per unit sold and that means lower cost overall and that works if people are not walking to get their groceries. It doesn't work if people are walking and it's like a daily, you know, stop on the way home from work or something. So, we will probably see a lot of those weird kinds of outcomes. Like, I think if you're thinking about the risk to a given career, I think for most of the careers people worry about, the mean compensation goes up, the median compensation of people who are in that industry right now, the median compensation they get from being in that industry probably goes down where a lot of people get washed out. But this has happened before. Like the spreadsheet did not eliminate investment banker or accountant as a job. It actually made it a more lucrative job. But it also made it a more measurable one where you just can't slack off the way that you perhaps used to be able to somewhat slack off in some of the white collar professions. It's just a lot easier. Like the expectation for output is so high.
I'm on Twitter all day and a lot of people are tweeting and I'm like it seems like a lot of people are slacking off these days. I'm like how do you have time? I have time because I'm like a journalist and I sit, you know, I create words professionally.
Some would argue that you do not in fact have time to be tweeting all day.
That's true. But David, so okay, there's been a lot of progress. No question. There are new harnesses for AI models that increase all of our capabilities and so forth. That's different than we all know that's true. That's different than like job wipeout in a significant way. What is it about to you that you think yes there is progress but progress on the scale of this is an imminent thing that we have to be talking about that could really reshape white collar labor?
The analogy that I think about a lot is COVID because the thing about AI progress is just that it's really in many ways exponential where the amount of time that an AI can operate autonomously without a human really has been doubling every 11 or 12 days or so for the past 6 years. And you know, when I think about COVID, this was a thing that nobody saw coming. And then it happened really fast. And then I think the political system was very reactive. And I think that a lot of quietly, a lot of Democrats wish that they had handled things a little bit differently. But what's cool is that unlike COVID, you know, we really can see this coming. All of the warning signs are blinking. And you know the other point I want to make here is you know I personally think that there's a lot of potential for large-scale job loss particularly white collar but you know there are a lot of truck drivers there are a lot of Uber drivers, you know, all kinds of people could lose their jobs and this could all happen very quickly. I think the more important point though is that the American people see this and are really quite worried. You know, when you ask people, how likely do you think it is that in the next 5 years there might be large scale job loss because of AI? 70% of the population says that it's either very likely or somewhat likely. And so that's really the main point I'd make there is the reason politically why politicians should act now rather than waiting until there's a problem is you know one the American people are already worried about this and two once it's already happening it will be too late for our political system to respond and so I think it's important to try to get ahead of it.
Definitely want to get more into the politics but Byrne you brought up something that like seems to be kind of standard in these conversations which is we've been here before right? We all well not literally went through the industrial revolution but that's something that happened and we went through the internet boom and the economy and society adapted more or less but when you ask people now what the alternative professions are for white collar workers we haven't been able to really get like a slate of possibilities. I know it's hard to imagine the future but you know if you're an insurance broker and you have been for 20 years what are you going to be doing in the new economy like what are the new jobs that are coming down the line.
Yeah, that is actually a tough question. Like one of them is just like temporarily I think there are a lot of jobs that are basically either humans who are required to be in the loop for regulatory reasons like doctors are incredibly rapid adopters of AI tools and the supply of doctors is sort of artificially constrained and so if you decrease the percentage of their time that they spend on admin tasks and you decrease the rate of mistakes that they make you basically get the equivalent of manufacturing more doctors and in cases like healthcare there is effectively unlimited demand. Like there's yet to be an economy where people don't spend more of their marginal dollar on health.
So, we're all going to be healthcare workers.
I think that's actually like that sector probably will grow and you know that's it's kind of glum to say okay some white collar workers are going to kind of move downscale in terms of status where they will probably have jobs that sound less cool. But if overall output is high enough and if the returns on capital are high enough that people like the economy starts shifting more incremental production into just building data centers that does mean that if you are the complement to a data center like you are a necessary component of this entire supply chain your bargaining power is a lot stronger because there's just a lot more capital that you're adjacent to. If you are completely substitutable by the data centers then you're in a tough spot. We always find that models have these really spiky abilities. Like they're superhuman in some respects and in other respects they are, you know, in terms of things like math ability. Like they're beyond the point where I could reliably distinguish between two models and say, 'Well, this one's really good at math and this okay at math.' But in other domains, they do just kind of fall flat because they don't have this comprehensive world model. And the reason for that is that they're trained on text and text actually skews towards areas that are uncertain and open to debate. So I use the term the maybe sphere which is like if you imagine there's like this bedrock of facts that are so obvious that nobody ever bothers to write them down. And then there's this infinite space of questions that are so weird that it's almost certain they don't have an answer. There's like this little narrow layer like an atmosphere where it's worth asking a question and you might get an answer. So they have a really good world model for the parts of the world that we're either not sure about or that we've codified in textbooks and then a really bad world model for a lot of the obvious stuff. And so it would be kind of weird to say like my job right now is to say extremely obvious things to this superhuman intelligence. I don't even know what job title historically that might correspond to like sort of a servant for a brilliant person who's also like an absent-minded professor. But yeah I think we'll have a lot more manservant for absent-minded professor professions in the future. Manservant for absent-minded professor. It sounds fine.
David, you have a firm and you hire people. Has the nature of your hiring changed? Are you hiring for different types of roles than you would have a few years ago or something like that given the change in technology?
Absolutely. You know, I think just to give a simple example like we used to have lots of copy editors to write polling questions or to write messages and you know the reality is that now AIs are generally better than people at doing that. Not uniformly but we have a lot.
Are they good at writing polling? Like can AI you want to find interesting polling questions right not the obvious stuff is AI good at finding non-obvious polling questions that are non-correlated to things so that you can actually get signal from them?
Well you know I don't want to talk too specifically about that but I will say that there's just tons of copy editing tasks that I think frankly...
Why do you want to talk specifically about that means this is the question I should be asking.
No no no but you know look there's a lot of stuff like translation there's a lot of stuff like copy editing I think the big shift for us is that we're focusing a lot more on person-centric jobs and you know now you can do a lot more engineering than before. So I think there's definitely been a big job shift you know in terms of how we've been hiring. Absolutely.
Maybe we can do a little bit of content creation naval gazing here since I imagine this is probably of interest to a lot of people at South by Southwest as well. But Byrne you write a newsletter on a daily basis. Joe and I do as well. How are you using AI just in your sort of day-to-day?
So, I use it a ton for research and one of the specific use cases is asking questions like, does this joke land or hey, I'm making a statement, you know, a kind of narrow technical statement about a domain that I'm reasonably familiar with, but I'm not an expert on. You are an expert on this. Tell me what I'm getting wrong. And one of the reasons for that is just a lot of my readers are software engineers and are eager to tell you when you're wrong. Yes. Extremely eager. Like I sort of measure people's ability as software engineers in particular based on how quickly I get an email from them that there's a typo because the really good ones like one of the skills that they have is just looking at a lot of text and immediately seeing what's wrong. Now that skill is kind of obsolete right now. LLMs are better at it, but still the mindset of I'm going to look at this and kind of absorb something coherent and I'm going to look for any little issue with it. That's still quite valuable. In fact, since there's more code being produced than ever before by a huge margin, it's really really valuable to be able to read it fluently and understand what it's supposed to be doing. So, I use AI a lot for research. I have only in the last few months have I actually gotten ideas for things to write specifically from ChatGPT or only in the last few months has it made some original points I would not have thought of that were just like clever insights and not just it's reciting a fact that I did not happen to know.
David, talk to us a little bit more. So, this question of progress, right? We all know it's getting better. That's obvious, but the more interesting question is, is it getting better at a pace faster than what people had previously anticipated? And talk to us about like, okay, where we are now with the technology and where the people that you were talking to, and maybe I'll throw this to both of you, where would they have said we would be in is it March? March 2025.
Yeah. I think that the big surprise of the last year has been, you know, the rise of vibe coding, the rise of tools like Claude Code. If you just went back a year ago, I think that nobody really expected the extent to which these things would be able to do large scale complex autonomous coding problems. You know, I think in a lot of ways these models have become useful faster than they've become smarter. And I don't think that that's something that people expected. Well, you know what's interesting is that every year there are a bunch of AI experts who then go and make predictions. And one of the biggest surprises has been the revenue growth of these companies, which I think is in many ways the most important benchmark. I think Anthropic's revenue last year was something like 2x what experts who already I think were quite AI-pilled predicted. And so I think that's probably the biggest surprise. And you know what I would say just on the broader transition is just that tools like LLMs have been now adopted faster. You know there are a bunch of graphs that are like how long did it take to implement radio or electricity or the internet. And this is much faster than any of those things. And so, you know, I really worry like when we talk about past transitions of like telephone switch operators or factory workers, you know, all of these things happened over an extended period of time and only impacted certain sectors of the economy. And this technology really threatens to upend every single job at the same time. You know, at a point when, you know, people's overall views of the economy are not good. And so I really worry just politically the extent to which our political system can handle this.
It is true that just when you look at these measures of how broadly technology is adopted. I think the thing that's really hard to measure is at what point does it become just part of your baseline expectation and what point are the impacts so obvious that they're unspoken. So if you look at electrification like I think anyone who talks about AI is just required to say that it took like half a century to go from the first electrified factory to most US factories being electrified. But the reason for that and there's a lot of fun economic history on this is that you have to run your factory in a different way. You actually have to build a different kind of building and you also have to finance it a different kind of way or you can finance it a different kind of way. So if you have a factory that has some kind of mechanical power source, you tend to expand your business in these discrete factory-sized increments. And what that means is that your investors, you pay out most of your earnings as dividends because there's nothing to do with retained earnings. It's like if you're going to grow, you want to issue a bunch of stock, you want to issue a bunch of bonds and then grow in like one shot. But once you have electrified factories where they can expand, they can just add another assembly line or they can upgrade this machine to a newer machine, etc., they can actually expand more organically and incrementally. And so that's when you start to see dividend payout ratios come down and that's when you start to see the growth company as a concept emerge. Like if you read investor accounts from the 1920s and they're talking about the stock market, one of the weird things is they're very fixated on whether the stock is above or below $100 a share because that was the par value that was just like the standard value for the stock and it was supposed to be the book value etc. then people kind of viewed equity as just the most junior claimant like the way you view the equity slice of a co or something like that and now we view equities completely differently and so we deploy money completely differently like it would be incomprehensible to an investor circa 1926 to say I'm going to invest in this company that has no operations it's just people and I'm going to put the actual money into this business but most of it will be owned by these people and then the business my stock could be worth 50 times as much in a few years if things go perfectly like that wouldn't make any sense to them. The idea of a stock going from par value to 50 times par value is just insane. So but that change actually was part of I think it was like causal and in both directions with America just having a really flexible financial system. We also have a really flexible labor market relative to every other country which means we will be patient zero for like all the job loss stuff like it'll hit us before it hits anyone else and it'll hit us harder than anyone else. On the other hand, one of the unique things about this particular general purpose technology roll out is that it is happening much much faster than before. But the thing that slightly offsets that is that the specific technology actually gives you access to information and cognition such that you can ask ChatGPT like here is my job like I've been selling insurance for 20 years. How should I reskill? Like what should I do? And it will ask follow-up questions. It'll ask, okay, well, what other problems do the companies and people you work with have? Or, let's break down the skills that you have. What makes you a particularly good or bad insurance person? And then what are the other jobs that might be less AI exposed that you could do? Or, as the models get smarter, you could just tell it, hey, I want you to invent a new job for me, like a bespoke job I was born for.
