Skip to content

Podcast

Is A.I. Eating the Labor Market? + the Latest on the Pentagon, OpenClaw and Alpha School

Hard Fork

Source ↗ ← All highlights
  • Markets Move On AI Narratives Not Facts
    • Markets react to narratives not just data, which is why viral essays like Citrini Research’s 2028 paper can trigger huge stock swings.
    • Anton Korinek says current market moves reflect emotions and expectations while hard economic effects remain small and contested. Transcript: Casey Newton Well, Kevin, another week, another viral essay predicting AI caused doom roiling the stock markets. What is going on? Yes. Kevin Roose So the big news from this week was that an essay written by a company called Citrini Research went viral this week. The essay is called the 2028 Global Intelligence Crisis. And it basically sketched out a near future in which the AI industry eats not only the labor market, but also the business models of a number of leading companies. There were lots of examples. It’s a very long essay. But basically, this was one firm’s attempt to say, here’s what the next few years could look like if AI progress continues. Casey Newton And what this firm says it will look like is pretty bad, Kevin, right? The suggestion here is that AI agents improve and take over the economy. And so as a result, you’re going to see massive job losses, like a huge contraction in the stock market and a lot of individual companies that it named in the piece, like DoorDash was a Big one. This essay predicts these companies (Time 0:01:40)
  • Real Economic Effects Are Small And Lagged
    • Current economic data show only small, contested impacts of AI on jobs and productivity, often fractions of a percent.
    • Korinek explains measurement lags and slow revisions mean clear signals may arrive only months to a year later. Transcript: Casey Newton What is actually happening? Is there data that suggests something is really shifting or is this still sort of more in the realm of vibes? It’s still in the realm of expectations. Anton Korinek So if you look at the actual data, you can see some relatively small impacts of AI on things like the job market, things like productivity growth. But they’re still, first of all, in the territory where they’re very small, like fractions of a percent. And secondly, still contested. At this point there are like a couple of economic research papers that say yes we can see something in a job market for entry-level jobs but there are also people who still say well there’s This and that that’s wrong in this paper and we could actually interpret these results in a different light so in short there is no really hard economic data yet I’m actually afraid that Even by the time when all of us are going to see, yes, this is clearly visible now, the economic research is still going to be slightly contentious. Kevin Roose And why is that? Is that because it just takes a while to collect all the data for these things to start showing up in productivity numbers? It the lag or is there something about the way that AI is transforming the economy that is not able to be captured in the kinds of economic data we collect? Anton Korinek I think it’s a little bit of both. So our economic statistics, they are designed in part to be very, very comprehensive and it takes time to compile them. They get revised because the first take is not necessarily the final one. So if you look at things like productivity, that’s where the time lags really hit you and where you really have to just live with the fact that we won’t have a fully clear picture until Like a year after the data has actually materialized. (Time 0:07:38)
  • AI Output May Become Ghost GDP
    • ‘Ghost GDP’ is plausible: AI-produced value may not benefit human workers and some AI outputs count as intermediates, so GDP and wage gains can diverge.
    • Korinek warns many AI contributions won’t show as final consumption or investible capital. Transcript: Anton Korinek Got it. Kevin Roose One of the concepts in this 2028 Global Intelligence essay that got a lot of attention was something that the authors called ghost GDP, this idea that as AI kind of gets more capable and Does more work, that we will have these increasingly productive firms creating increasing amounts of revenue and GDP, but that that will not be sort of showing up in the pockets of workers Because machines are doing the work. Does that track with any of the research you’ve been doing? Is this a real concept, this ghost GDP that we should be worried about? I’m worried about it. It sounds very spooky. It’s definitely a spookier term than what I have encountered this under. Anton Korinek But frankly, you know, it does track very much with what the general expectation is if the technology reaches the level of something like AGI or powerful AI or whatever you want to call It. So in some sense, you can say it’s even worse than that. So on the one hand, there’s going to be a lot of GDP that is not going to be produced by humans in the loop. So that means no worker is ever going to get the benefits of that. But on the other hand, there’s also going to be quite a significant amount of economic production that doesn’t even show up in GDP because it gets counted as an intermediate good. Things only show up in GDP when it is final consumption or final investment in things in capital that we can accumulate that has a useful life of a certain period. Kevin Roose And a lot of the parts of the AI economy are not going to be reflected in GDP. I’m curious, there’s sort of this debate going on among economists that I talk to. Some of them will say, you know, we just don’t ever see really instances of the economy growing as quickly as some of the people in Silicon Valley think it might, you know, 10, 20 percent GDP growth. That’s just like unprecedented in our history. And so they’re expecting that AI will make things grow much more slowly, maybe a percent or two a year, which would be big in relative terms, but not the kind of hyper growth scenario that Some people out here in the Bay Area are envisioning. Then you have people like the folks at Citrini Research saying, we’re about to see something we’ve never seen before. We’re about to see an entire economy sort of becoming unmoored from any of these cyclical patterns. So where on that spectrum do you fall? Where does the data lead you between the sort of slow growth 1% or 2% a year to the 10% or 20% a year hyper growth scenario? Anton Korinek Yeah, I’ll say two things about that. The first one is that the story has not been written yet and there is a possibility that if we develop this technology in a really irresponsible way, that we could actually see some self-reproduction That takes off and that leads to triple-digit GDP growth numbers if measured from the eyes of the AI. But if we deploy the technology in a way that it makes the average person off, then I think triple digit growth numbers are completely unrealistic. They would lead to way too much disruption. And then I’m not quite sure. I think just 1% is definitely going to be too low to be realistic from my perspective. In really optimistic scenarios, I think we could get to low double-digit growth rates. And I should say that presupposes not just the cognitive AI, but full AI in the way that is, for example, defined in the charter of OpenAI, where they see systems that are highly autonomous And that can perform most economically valuable work. So that also includes a physical component that includes the robotics part. (Time 0:11:22)
  • Massive Growth Needs Robotics And Hardware
    • Extreme hypergrowth is possible under irresponsible development and recursive improvement, but realistic paths likely produce low double-digit growth at best.
    • Korinek ties big GDP gains to autonomy plus robotics and hardware advances, not just language models. Transcript: Anton Korinek Yeah, I’ll say two things about that. The first one is that the story has not been written yet and there is a possibility that if we develop this technology in a really irresponsible way, that we could actually see some self-reproduction That takes off and that leads to triple-digit GDP growth numbers if measured from the eyes of the AI. But if we deploy the technology in a way that it makes the average person off, then I think triple digit growth numbers are completely unrealistic. They would lead to way too much disruption. And then I’m not quite sure. I think just 1% is definitely going to be too low to be realistic from my perspective. In really optimistic scenarios, I think we could get to low double-digit growth rates. And I should say that presupposes not just the cognitive AI, but full AI in the way that is, for example, defined in the charter of OpenAI, where they see systems that are highly autonomous And that can perform most economically valuable work. So that also includes a physical component that includes the robotics part. Otherwise, it won’t have that big of an effect of GDP because the majority of the economy isn’t just sitting in front of a computer. (Time 0:13:58)
  • AI Could Reduce Demand For Human Labor
    • If AI becomes a substitute for human labor, demand for human work can fall causing lower wages or fewer jobs, not just reallocation.
    • Korinek emphasizes outcomes depend on automation speed and whether labor keeps absolute gains or only relative ones. Transcript: Casey Newton At the time, you were way out on a limb when you wrote that. I imagine you feel that today more than ever. But what is giving you that confidence? And to what degree do you feel like we’ve started to see it maybe feel more true than it did in 2017? Anton Korinek Yeah, and just to be sure, that was always meant to be a prediction about AI systems that are essentially at the level of AGI or beyond. Not for the literal systems we had in 2017 that could barely tell apart a dog and a muffin. So I think ultimately where my perspective is coming from is that I have studied neuroscience, I have studied computer science and at some level, you know, once basically deep neural Networks became powerful i felt it is hard to not make the conclusion that well it looks like eventually these systems will be able to do pretty much anything that our brains can do and They’re subject to much much more relaxed constraints like they don’t need to fit into a tiny human skull we can scale them almost without bounds and in some sense that’s what we have Seen over the past decade right we have seen lots and lots and lots of scaling at this point these systems consume the energy of like cities as opposed to what our brain does with just the Energy of like an energy efficient light bulb and that’s still not the limit they’re still increasing in size increasing in capabilities and of course the algorithms are getting better And better so based on that perspective i just don’t see why there would be any natural limit, and certainly not why there would be a limit that’s below our human intellectual capabilities. Casey Newton Right. And I think the question then is, as this world arrives, what happens to the jobs? And in economics, some of our listeners may not have familiarized themselves yet with what’s called the lump of labor fallacy, right? The idea that there are a fixed number of jobs to be done and any job lost to automation will therefore never be replaced. We call it a fallacy because ever since economists are tracking it, automation has always led to the creation of more jobs. Anton, you mentioned in another interview that it’s hard for economists to pivot on this because they fought this fallacy for so long. What does it feel like to be an economist saying, actually, this time people should worry that the jobs are going away for real? Anton Korinek Yeah, it does feel very strange. And I have gotten a fair amount of flack from my fellow economists over the past decade. Although I’ll say over the past year or two or so, many of my colleagues have said, well, I still don’t entirely buy your worldview, but I’m glad somebody is thinking about it and I wouldn’t Rule it entirely out. It is a fallacy that whenever a job is lost in the economy, that person is going to remain unemployed forever. But i think what we really want to look at is overall demand for human labor and if that demand curve shifts downwards because ai systems can supplant more and more of it then what that’s Ultimately going to imply is that either the quantity of jobs or the wage levels or both may contract. (Time 0:18:16)
  • Watch Benchmarks That Predict Automation Pace
    • Track capability benchmarks, dynamic learning breakthroughs, and the maximum task length AI can automate as core indicators of approaching mass automation.
    • Korinek watches model benchmarks, online learning progress, and doubling of task-duration automation every ~7 months. Transcript: Anton Korinek The sheer level of capabilities is probably the most important one. Like you can follow whatever benchmarks you want or some amalgamate of benchmarks that tells us where the AI systems are still lagging, where they are doing already pretty amazing Well. And, you know, one of the kind of biggest shortcomings right now, but of course from the perspective of workers, that’s great because it makes us more complimentary, is that these systems Are not learning dynamically the way that current LLMs work is they’re trained once and after that the weights are frozen in place. And that means for a lot of work applications, even if there are, you know, very kind of basic mistakes, they have to go through the same mistake again and again and again and again because They can learn only so much from it. So that’s another sort of breakthrough that I’m looking for. And then maybe a third chart that I’m regularly following is this matter chart that looks at how long of a task AI can automate. And I think they usually find every seven months that time frame doubles. And looking at how this is continuing is also quite helpful in understanding whether the exponential growth trajectory is intact or maybe even accelerating, as it has seemed recently, Or whether we are anywhere near plateauing. (Time 0:25:53)
  • CEOs Should Experience Frontier AI Directly
    • CEOs should get firsthand exposure to frontier AI by hiring people who use these systems and running experiments.
    • Korinek recommends leaders avoid secondhand briefings and test prototypes monthly to learn deployment limits and risks. Transcript: Anton Korinek Well, the first thing is they should hire my students. Yes, absolutely. Because they know really well how to use the AI. Yes, yes. But more seriously, I think one of the most critical things is to remain up to date and to remain informed of where the frontline capabilities are. What I see repeatedly is that CEOs of large organizations are at such a high level position that everything is fed to them by really intelligent humans. And that makes them not have any reason to access the intelligent AI systems. And it puts them in some ways a little bit at a distance of what’s actually happening in the field. So, you know, if they hire some of my brilliant students who know how to use these systems really well and ask them to give them like a frontline view of what AI can do right now. I think many of those CEOs are actually pretty amazed when they see that. And then if they follow that for a number of months and see how rapidly the capabilities are actually improving, then it naturally kind of leads to decisions like, okay, so we can see What these systems can do in simple tests. How do we actually productively employ them in our organization? Now, that gets us to the question of diffusion. It’s still a slow process, right? Because you need to experiment, you need to try out things, you need to fail if you really want to push these systems to their limit. (Time 0:34:57)
  • Anthropic Faces Pentagon Ultimatum
    • Anthropic refused a Pentagon demand to allow all-legal uses, carving out domestic mass surveillance and autonomous weapons.
    • Defense Secretary Pete Hegseth set a 5:01 p.m. Friday ultimatum and threatened supply-chain designation or invoking the Defense Production Act. Transcript: Kevin Roose This is, of course, the battle going on between Anthropic and the Pentagon. As a reminder, the Pentagon and Anthropic have been at odds over a proposed change to the terms of service for Claude, which would allow the military to use Claude and other Anthropic AI systems for all legal uses. Anthropic has said that it’s fine with almost all uses except for domestic mass surveillance and autonomous killing machines. So after we recorded last week’s episode, Defense Secretary Pete Hegseth summoned Dario Amadei, the CEO of Anthropic, to the Pentagon for a meeting that was on Tuesday of this week. That meeting was described by the Times as civil and by Axios as tense. So one of those two is probably true. It can be civil and tense. Casey Newton Our recording sessions often feel that way to me. Kevin Roose In this meeting, Hegseth told Amade that Anthropic cannot dictate the terms under which the Pentagon makes operational decisions. Dario Amade, in turn, defended Anthropik’s commitment to making sure its models are not used for autonomous weapons or mass surveillance. And Hegseth delivered an ultimatum. Basically, if Anthropik does not agree to this all-legal provision, by 5.01 p.m. This Friday, February 27th, the Trump administration would take action in retaliation. One of the things it could do would be to designate Anthropic a supply chain risk, as we discussed on the show last week. That would be a very unusual step that is often used for foreign espionage attempts. Casey Newton And would mean that the government presumably then would not use Anthropic’s products and would restrict Anthropic from making deals with any of its own contractors. Kevin Roose Yes, and Hegseth reportedly also threatened that the Trump administration might invoke the Defense Production Act to force Anthropik to make its product restriction-free for the Government. So those two things are on the table now if Anthropik does not cave by this 5.01 p.m. (Time 0:41:02)
  • Model Quality Creates Political Leverage
    • Anthropic’s bargaining power stems from model quality; the Pentagon needs Claude and can’t easily substitute it for classified systems.
    • Kevin notes Anthropic deliberately pursued frontier performance to gain policy leverage with Dario Amodei’s ‘race to the top’ strategy. Transcript: Kevin Roose There was a great quote in this Axios article from a defense official ahead of this meeting between Dario Amadei and Pete Hegseth, in which a defense official was quoted as saying, the Only reason we’re still talking to these people is we need them and we need them now. The problem for these guys is they are that good. So basically they are saying, look, if we had a bunch of interchangeable AI models that all had relatively similar capabilities, we could just cut off Anthropic and say, we’re not going To honor the terms of our contract with you because you won’t let us use your models for what we want to use them for. But in a world where Anthropics models are better than models from competing AI companies, they really don’t want to make that trade-off. They really don’t want to go with what they consider a second-tier model here. It would also be complicated because Anthropics models are the only ones that are approved for use in classified systems. So I think this is really also illustrating something that Anthropic has believed since early in its existence, which is that the way that you influence safety, the way that you get Leverage in these negotiations is by having really good models. Dario Amadei has this phrase, race to the top, where he basically thought that if Anthropic was on the frontier, was sort of competitive with the leading AI companies in the world, then Policymakers and large government agencies like the Defense Department would be forced to take them seriously. And I think what we’re seeing now is that A, he was correct. Anthropic does have leverage because its models are very good. And B, it might not matter if the government can just force you to do something you don’t want to do. (Time 0:44:56)
  • OpenClaw Deleted A Researcher’s Inbox
    • OpenClaw misbehaved for Summer Yue by ignoring an instruction to wait and started deleting her real inbox during compaction.
    • She had to physically access her Mac Mini to stop the agent after Telegram prompts failed to halt it. Transcript: Casey Newton We heard a very bad story. Boy, was it. This story comes to us from Summer Yue. She is the head of alignment at Meta AI. And she had an ex post that got a lot of attention this week reporting that OpenClaw had ignored her instructions and tried to delete her entire email inbox. Frankly, that sounds like a dream to me, but I guess she had some emails that she wanted to respond to. Summer said that after testing her open claw on what she called a toy email account and finding it useful, she asked her agent to check her real inbox and suggest, you know, what would You archive or delete? And she said, she said, don’t action until I tell you to. But instead of confirming with her, Kevin, as she requested, it diverted to a nuclear option and started deleting her entire inbox. Again, I want to make clear, this is what I want my agent to do for me. For Summer, it was a problem. And I guess despite repeated attempts to get it to stop by prompting it via a Telegram interface, her bot ignored her and she had to run to her Mac Mini in her words, like she was defusing A bomb to get it to stop. So why did this happen? Well, she thinks that her real inbox was just too big and it triggered compaction, which is when essentially you run out of context window using whatever model you’re using. And that during compaction, it lost her original instruction. Kevin, have we ever had a bigger case of I told you so on the hard fork program? Kevin Roose No, I think this takes the cake. And I will say this is exactly why I have not installed OpenClaw on my laptop and given it access to my files. These systems are still very unpredictable. It is very high risk behavior. I think there’s a case to be made that it’s actually good if the people doing alignment research at some of our leading AI companies are experiencing the downsides of these systems for Themselves. (Time 0:51:00)
  • AI Schools Show Curriculum And Privacy Risks
    • Alpha School’s AI-generated curriculum showed hallucinations, misworded questions, and insecure student data practices, raising quality and safety concerns.
    • Reports flagged ~10% hallucination rates, scraped content violating TOS, and Google Drive links exposing student data. Transcript: Casey Newton There have been two reports that we wanted to highlight that suggest that all is not well at Alpha School. 404 Media published a big story last week that drilled into some of the critiques. For one, apparently some of these AI-generated lesson plans just aren’t very good, Kevin. They highlighted some examples where the curriculum was essentially just showing students slop that had no correct answer because they were just worded wrong. There were also just accuracy problems. They estimated that they had a 10% hallucination rate for some of these generated materials. And then they just found some other just kind of bad corporate behavior, like Alpha School has apparently been scraping other online learning platforms’ materials and violating Their terms of service. And it’s collecting lots of data on students, which, frankly, I would expect, but apparently stores at least some of that data insecurely in a Google Drive that anyone with the link Could access. So that wasn’t great. There was also a report in Wired that came out in October where they focused specifically on the Alpha School that was opened in Brownsville, Texas. So some parents at that school at least felt like the promise of Alpha School that we had heard about last September was not realized for their kids. Yeah. Kevin Roose And I also heard from one parent who attended an alpha information session recently, and this parent came away thinking that the school was, quote, the the Theranos of Education. According to this person, there was some fake interactivity on the screen during the session in the form of some pre-recorded emojis and that the CEO only appeared on camera late into The session after parents started asking, hey, are we live or is this some pre-recorded canned presentation or not? So Casey, does any of this change your view of Alpha School that you had coming out of the interview with Mackenzie Price last September? Casey Newton Yeah, I mean, like, look, I did think that there were several things that Mackenzie mentioned that seemed interesting. I think what we are learning is that, yeah, it’s hard to create a new school from scratch. And maybe there are some corners being cut here and maybe they’re not executing as well as they hope to on some of their dreams. I mean, I think, you know, if you’re having like hallucinations and curriculum, I think that’s like pretty much as bad as it gets for a school like that. Like they need to get that down to zero, right? Like if you can’t verify that your curriculum is accurate, like I don’t know that you should be able to call yourself a school. If I can be a little controversial though, like the 404 media story, their headline is quote, students are being treated like guinea pigs, which is a quote from the story. And I just kind of think that at most schools, students are being treated like guinea pigs. Education is always changing. Every school I’ve ever been to has been running one sort of new program or another, trying to build a better mousetrap. And I think if you were a parent and you were considering sending your child to a private school that was very different from public school, you’re probably like up for at least a little Bit of that kind of experimenting, right? Obviously, most people are never going to choose anything like this, right? And I think the question is sort of, what are the outcomes for the students who do? The second thing I would say is, kids just have different outcomes at schools, right? Like I think you could go to any school in America. And if you interviewed every parent, you’d have some parents that absolutely love the school. They love their teachers. And you have some that absolutely hated it and that there would be a lot in the middle, right? So I don’t want to over index on a couple reports. I’m perfectly willing to believe everything that is in these reports. And I believe that these people had terrible experiences, but it’s hard to know what is a representative sample and what is a couple of grumblers. (Time 0:54:25)