¶ Intro / Opening
Home to the Rachel Maddow.
🎵 Music
Trust. Now is your source.
Learn more at MS.now.
¶ AI's Revolutionary Potential and Diffusion
There's a lot of people who really just think AI is gonna go away. But there's something here that this isn't real. The systems aren't as good as you say they are, they make mistakes. This is real. Like I talk to CEOs all the time. They are getting value out of it. Like it's funny because they'll say things like, uh, 10 million, 20 million here, nothing significant yet. But like they're getting value out of these systems.
There has not been major layoffs because of AI yet. I don't think that continues infinitely. No. Maybe people get better jobs, but like we need the systems in place to cushion it, which means we believe this is real.
🎵 Music
Hello and welcome to Why Is This Happening with me, your host Chris Hayes.
🎵 Music
Well, we're here for another special edition of our AI series with Pod the AI Endgame. And today I want to sort of focus on neither the best case scenario nor the worst case scenario.
Bye.
A trajectory in which these things really are incredibly powerful and transformational, in which the kind of more maximalist claims of its backers basically come true. short of the end all human life on the planet, which we can sort of put in another category. But what does it look like if The people saying this technology is completely revolutionary on a scale, maybe even more revolutionary than the internet, maybe the most revolutionary form of automation that's ever happened.
What exactly does that mean if that's true? My guest today is a professor at Wharton, Ethan Mollock. He's been studying AI and its implications for education, entrepreneurship, work. He's author of numerous books, including a twenty twenty four New York Times bestseller called Cointelligence, Living and Working with AI. And he has a new book coming out in the fall called Coexistence about working with AIs that are, as he says
Sometimes smarter than you. He also writes a substack called One Useful Thing. Professor, it's great to have you on the program.
Great to be here. Thank you.
🎵 Music
So how long have you been working with and working on AI?
So I mean that's always a hard question because AI is many different things. When I was in graduate school at MIT, I did work at the MIT Media Lab with the AI group at that point, some of the fathers of AI there from the original days of the nineteen fifties.
And I was the non-technical person. So I was like the business person translating stuff. And then I've used AI a lot for teaching and educational purposes. And certainly since the large language models came out, I've been talking a lot about them and how they impact work and education and everything else.
So where are you on the kind of scale of one to ten of one being it's a complete con or a bubble that's gonna completely fail? And ten being it will change the world unlike any other technology we ever have.
So I mean, one thing I'll just try and be careful about without being boring is that there is a little bit of like Changing the world is absolutely definite in a huge way. So maybe eight or nine in the longer term. Yeah. But but that this is not an instantaneous process, right? There is a process of adapting to technology and that change takes place over time, not instantly.
You know, one of the examples I think we've had is even if you just talk about the internet or you talk about other forms of digire revolution. There used to be a joke about the paperless office I remember being made in the nineteen nineties. where you were still printing out a ton of stuff and people were like talking about like joking about the paperless office as you like went to the printer and printed out ninety pages.
And it was because there was a huge gap between what the frontier prediction was for what wired networking computing would do to say a workplace. and what actually happened in real time. Now in the year 2026, there's really not a lot of printing happening. Like every time you have to use a printer is really a bummer in your life. It turned out that prediction was true. It just took like two decades longer than some people thought. Do you think that's something similar happening here?
So I think it is, right? I mean, there's sort of this weird needle of the thread because a lot of especially online discussion is either zero or one. Like either this is all fake, which it just clearly is not, or the machine god is coming next week, right? And I mean Can't 100% rule that out, but that's the most likely scenario is
some aspects of it look like other technologies. And if you work at any organization, you've probably seen this, right? No company transformed radically as a result of AI in 2025. Because to do that, let's say your programmers are a hundred times more productive. So
How are they getting instructions about what to build? Who's testing this? Who's shipping that product? How are you evaluating their work? What are they supposed to do with their time? Even if they're using the product, there's a whole bunch of other systems that have to be connected together to make AI work.
So we call this diffusion of technology, you know, in the academic world. And it's a just a constant. All technologies take some effort, some time to diffuse. So nothing happens instantly. That said, diffusions happening very quickly here, but nothing's instantaneous.
What are good analogies for that diffusion?
So
It turns out, by the way, the diffusion is the single best prediction in all of social science, basically. Every diffusion curve looks the same. It's called an S curve. It starts slow, takes off, and then slows down when it reaches saturation. So literally every technology in human history follows this path. But if you want to take
The example of the internet, 1990s, everyone's really excited about the internet. There wasn't web pages yet. So it took the invention of the web page at HTTP, and then we needed to have commerce and the speeds of the internet need to be faster and companies need to put web pages up and ad agencies need to figure out how to advertise on them. So there's lots of pieces that had to come together to make the internet a place where you could do business.
Right. And so some people adopted really quickly and some of them won out early, Amazon. Some of them didn't. Pets.com, right? Even though they turn out to be running the long term. So there's a timing issue and a adjustment issue that happens anywhere. And those adjustments are where all of the heavy stuff happens, right? Whether this helps jobs or hurts jobs, it's all about how we use these technologies.
¶ AI's Impact on Jobs and Agentic Systems
So you don't think there's like an innate valence to the technology in terms of say its job impact?
So, you know, again, we're now in the world's most complicated area of like academic debate, but there is some valence, right? Because there's this law of technology, the technologies are neither good nor bad, nor are they neutral.
Which means that like it's not inherently good or bad, but it doesn't mean doesn't have valence, as you said, right? There is a valence, but that valence also is partially based on where we are as an economy. What do we think of when we use technology? We use it for automation.
So the valence becomes an automation one, right? Use this to replace human labor. But it doesn't have to be that way. This could be augmentation. How do we make everybody's jobs better and smarter and work better and do more stuff? That those are choices that are made on the implementation edge, not in the technology itself. But we're coming into a world where the goal is automation. That would be an anxiety.
If diffusion is irregular across industries, right? Or across different parts of the business world, where's a place that looks the most cutting edge, you know, it seems like software encoding is the place where we're really seeing it picked up at a crazy scale unlike anything in my working world or the other people I know.
Yeah, absolutely. I mean, so that's where you're seeing the weirdest experimentation. So you can kind of think of AI broadly as having a couple errors. There was sort of pre-large language model AI, and a lot of people who are critics of AI
sort of conflate that era of algorithmic decision making where we had a lot of big data and we'd use that big data to analyze things and Netflix would tell you what movie to watch or we could figure out pricing. Still very important businesses, but that that's not then there was this sort of period of time where it was this cointelligence to take the name from my book. Where you'd work back and forth with AI systems.
What's happened in the last couple of months has been a move towards agentic systems. You assign work to AI, it does it itself. And that's where this big change to coding has happened. Because if you're a computer programmer, a lot of them have stopped touching code. They will just assign work to the AI and then confirm that work exists or not.
And then the AI does all the work for them. And as the systems get better, they do more and more work with less and less errors and less and less oversight. And that's where we're seeing the most transformation. There are some companies that have been experimenting with models where Coders do not touch code at all. The AI develops the code, it tests the code, and all they do is give it a thumbs up or thumbs down to ship it.
Can you talk a little bit more about the development of these agentic systems and what the kind of friction points are, what the challenges are, and how quickly they seem to be overcoming them? Like it sounds to me insane.
And I've read the stories too of the people that have sixteen agents working overnight churning out code and then they're trying to get agents to supervise other agents. I don't actually quite understand it. So How should we understand this sort of agentic paradigm for people that aren't actually dealing with it or still just asking Claude a question?
Yeah, that's great. Okay. So the way you ask Claude a question right now, and you know, there's a billion people around the world who do. I was just speaking to some Amish people who use AI frequently just to tell you how fast the adoption's happening.
Oh really? That clears the bar they're allowed to?
There are debates in the Amish community. I feel like we're like on the edge of seventeen different academic discussions that we could have separately. But I think when we think about this, if you used AI before, you went back and forth on a chat bot. So you'd go to, you know, Chat GPT, you'd ask a question, it would give you an answer, you'd give us a question.
What an agent is is basically those same chatbots, but they're now hooked up to what are called harnesses, which are sets of tools they could use. So they can take over your computer, they can browse the web, they can write code. And now when you give them an instruction, what they can do is just figure out how to solve it themselves. So if I say, hey, create me a business idea and launch it.
It can work for eight hours independently and write the code for the website and do the research. And I mean, it may not be a good business idea, but it could do all of the pieces that you might need to make that happen. So agents are autonomous tools that can go off and do work on their own. and make judgment calls by how the work gets done and then return to you the final product.
How well do you think that's working?
So it was absolutely not working well a year ago. And it was working kind of okay in a few areas like research six months ago. And now it's working extremely well in in a lot of areas. They're not good at everything, right? I mean, first academic paper on AI, we came up this idea of the jagged frontier. The AI is good at some stuff.
Bad at some stuff. It's not good at everything. But at things like code, it is reasonable to expect multiple hours of independent work. There's this great study, OpenAI did it called GDP Val, where they brought in experts representing 5% of the US economy, average of 14 years of experience, and they had them each.
Create a test.
Then they had other experts they hired to do the task, seven hours each task. They had the AI also do the task, took'em a few minutes. They had a third set of people spend an hour judging each answer and deciding which was better, not knowing who wrote them. When this came out last year, the best AI agents were tying or beating humans forty eight percent of the time across these tasks, which is already pretty impressive.
The models that just came out a couple days ago are up to 84% of the time. Right. So if you assign them a task that a human would do, 84% of the time they tie or beat the human doing this.
Can you give me an example of a kind of task? Sure. Like what are we talking about here?
Yeah, I mean one example of this would be financial services. I have a client who is considering these three different kinds of, you know, IRA packages. This is their income. Here's what's going on in this city. I want you to create a presentation and a PDF laying out the implications of each of these three approaches.
take into account here's who the kids are and everything else. And then, you know, talk about which one you'd advise giving the legal situation for. Or another one might be, we are maker of electric cars. I need a marketing plan for our Dallas office. Here's a whole bunch of their sales information. We need PowerPoints and we need images for an ad. And then I'd like you to figure out how much it would cost to test these in the market.
I mean, that's like work product, right? I mean, that's the kind of thing that people would punch into the office and do every day.
Yes, that is the goal of this thing, right? It's called GDP valve for a reason. The goal is to find economically valuable work. And I think what's snuck up on people with agentics stuff is it's gone from AI as productivity booster alone, which is you use it and you get answers and you do a bit better in your job. And we have all these papers showing that this is what happens, right? You get productivity gains.
And you know, those productivity gains are generally captured by workers at this stage, by the way, which is kind of an interesting point because you're doing the work, you decide how much of that to pass on to agentic work where it starts to actually do economically valuable tasks. And now we've got to figure out how people relate to that.
¶ AI as Utility and Superintelligence Race
How are they going to relate to that? I mean, it seems to me like the way they relate to it is you fire the people that used to do this and you tell the machine to do it.
So, I mean, there's a few problems with that. One is the AI is still jagged. So it's good at some stuff, bad some stuff. And Hubids do matter in the decision making process across a lot of this kind of work. A second thing is if every company is run by Claude, there's no real strategic advantage in how you do business because everyone's business is Claude business. And like where's the advantage of the human piece there? That is
Such a good point. Right. I mean, I've heard people describe the future, even Sam Altman and AI have described it as a utility. Like we think intelligence will be utility, like, you know, water, electricity, and we charge people for it. And I thought it was such an interesting idea. I've also heard Bezos describe intelligence.
in this era like electricity. It's a layer on everything. You know, there's no like electricity industry in my life. It's just that everything's electric. Everything I use, it plugs into the grid and I mean I light a few candles here and there. But other than that, like mostly everything's electric, right? What's strange about that is like if that's the case, then it doesn't seem like you're creating a world with a bunch of different AI companies. I mean, utilities
They're regulated that produce a commodity product, right? The power that flows through the grid. No one's like sitting down and be like, hmm, what am I going to do? Which energy company am I going to choose here? And that's partly because of regulation, but also because there's a kind of natural monopoly effect.
What you just said sort of sparks that thought in me, right? Which is like, it does seem like strange that we're gonna have a bunch of these different models. It seems like there'll be one. Or it will all kind of converge in some way. And I'm not quite sure what that will mean for who that is that's making the model.
All right. So there's this sort of elephant in the room here, which is The ultimate question is how good do these models it get? How fast, right? So the AI companies, right, have sort of this two-handed thing, which is on one hand, they're making intelligence on demand. And let's call it intelligence of a sort, right? It's intelligence in quotes.
It substitutes for human intelligence under some circumstances, not in others. Sort of like electricity can be used to substitute for some kinds of human work, but not others. So yeah, we have a bad definition of what AI intelligence is, but it is intelligence of a sort on tap, right? And part of the issue also is smarter generally is better at everything, right? So a smarter model does better. So right now there's, you know,
three to five US companies that that are in the race for very smart models and another five or so Chinese companies that are about six or eight months behind and all kind of racing. The question is what are they racing for? And there's really two kinds of outcomes here or three. One of them is that Eventually, the growth in AI ability slows down and this becomes a commodity. And then the question becomes: if you are an AI company, maybe you have some advantage because
you control the models so you can make them work better with your particular software and you still sell more of it because it's not just electricity on tap. I'm not just using the AI, but I'm using clawed code. And clawed code has special value associated with this. But then there's a commodity question.
The second option is that there's some sort of specialization that happens as as ability goes off. So some are better at mouth, but it hasn't happened yet, but maybe. And then the third is what the AL lab secretly are aiming for, which is takeoff. They think that they can achieve. A general intelligence, a machine smarter than human, every intellectual task in the next couple of years. And they think after that comes super intelligence. And if you're the first person to do super intelligence,
You get all the marbles because it's a self-accelerating process. They call this recursive self-improvement. The idea that the AI helps make better AIs. So now it's like the electricity companies were all racing to create the electric god, whoever builds the most electricity fastest.
then controls all electricity in the universe forever. So that is underlying this, and that's their theory, right? And that's the thing people find most incredulous. And I think it's worth some worrying about or thinking about because it's not
impossible, right, to imagine a situation like that. But otherwise commoditization, you know, is a possibility, right? It's just a question of does this curve slow down? Because otherwise you don't want to use a year old model when you could use a more advanced model today.
🎵 Music
¶ Debating Superintelligence and Jaggedness
Artificial intelligence is moving very, very fast, and it's raising new questions just about every day about what it is, what it isn't. When all is said and done, what is the end game? I'm Chris Hayes, and as part of my podcast Why is this happening, I'm speaking with leading experts each week to help ground that conversation.
We're at right now in a situation where it's very difficult to understand what is real and what's not real.
Why is this happening? The AI Endgame, a special miniseries from MS Now. Start listening today, wherever you get your podcast.
🎵 Music
Let's stay on this superintelligence idea or takeoff idea, because to me that's the difference between what we might call a normal technology and something abnormal. If we're talking about something like electricity, which again, absolutely revolutionary in almost incomprehensibly broad way. But the internet, electricity, they're still quote normal technologies, right?
It doesn't sound to me like the story that they're telling investors about super intelligence is normal in that sense. Right? They think there's some rupture here, some break, some difference in kind, not degree. Can you just walk through that story? I have a hard time with it because I think the people that are selling it the most, they have billions of dollars at stake in selling it.
It also makes me kind of panic if I think they're right. But in a clear eyed way from someone who studies this in the business world, what do you think of that theory? Can you just push at it a little bit?
Yeah, I mean, nobody has clear answers here, right? And like this is the idea that eats people, I think was phrased at one point. Like as soon as you start wearing a super intelligence, there's nothing else worth talking about, right? Because this is the apocalypse or savior of humanity or whatever else it is.
I would say one of the problems I often see for people sort of dismissing AI stuff is they assume that everything these AI leaders say is all marketing. And there is definitely some marketing. Like you want to take things with a grain of salt, but there is sincerity there that has been there since the very beginning.
That they genuinely believe that this is what they are supposed to do. This is what they want to do. When you talk to people privately in the AI labs, this is what they're building towards like.
That's also true of cult leaders, you know what I mean? Like people could be deluded and also sincere.
Yes. And and by the way, and the fact that large language models turn out to be so good at so many things is a shock to everybody involved in this, right? And we still don't have great theories yet, you know, we have some ideas, but why they're as good as they are, right? So we also in this unusually powerful technology and so far there's been no barriers. There's all the predictions of slowdowns last year. GPD five is a disappointment. It wasn't.
they have been much more right than people saying that they're not right. Right. So I think jaggedness is something that's not going away. For example, one of the things I like to have the models do, I'm a writer, is like I actually have the AIs write long fiction. Math used to be a huge problem in AI development.
And now the AIs are among the best mathematicians on the planet will soon be better than anyone else, right? We made fun of how they could do math before. Now they do math. Writing has been a weirdly jagged space. If you ask the AI to write something, it's gonna be overwrought. Even the most recent models still have their
All of their little tricks. It's not X, it's Y. They have overwrought plot lines. They don't land story arcs well. Every character sounds the same. They're all named Alara Voss. So when you think about an AGI or an ASI, I think a jagged one, one also that has to deal with human kinds of adoption and where those jaggednesses meet human systems.
means that we don't have quite the machine god in a room. Now we might have something that's generating new drugs that save lives, or something that generates new weapons, or something similar. And the human layer and organizational around it layer looks very different than it does today. But there is reason to expect that we will both get huge growth, but that would there will also be weird jaggedness to these systems in kind of the way there are now, at least for the foreseeable future.
So what all that means is I think that this is a big deal. There's no sign of these models slowing down in what they're good at. And that means things like cybersecurity are gonna be a huge mess and Things are going to change very rapidly in software development and other areas. AIs are good, but it may not be the transformational piece to every aspect of humanity from a superintelligence that we would expect.
But let's say on that the superintelligence part of it, Altman will do the joke, you know, the idea is that we build AGI, you know, artificial generalized intelligence, then we ask the AGI how to make money. You know, that he's got a line like that. Talk a little bit about the idea of recursive self-improvement and what would be the implications if if they're right about
I mean, we're already seeing some recursive self improvement, right? So recursive self-improvement is the idea that an AI can help you build the next generation of AI. At least Google, Anthropic and OpenAI, which are probably the three leaders at this point, somewhat definitively, all are reporting some degree of the AI's helping us make the next version. But again,
Jaggedness gets in the way. They help you with some areas, but there's something that they're not good at or they don't get, or that they fall into a rut about, and humans need their intuition, or you just need a document carry from point A to point B, or whatever it is.
There are all these barriers. But require some self-improvement is about you start to see the speed up and we're seeing it, right? There's no sign of the exponential development ability of AI in the areas it's good at slowing down at this point. It just hasn't happened. It might happen. There's no reason to expect it could happen, but it hasn't. And you could see this by the way in the amount of software code these companies are shipping. So if you've played with Claude Code or Claude Cowork,
Those are shipping major product developments like every day now because they're using the tools to build the tools. So that's what recursive self-improvement looks like.
Right. I mean again, it's also the case that like they're using all this stuff to make their stuff. So they believe.
They absolutely believe in it. And like I totally understand after crypto and web three that everyone is very skeptical about technologists. But the old school view of technologists was they were very sincere. They weren't always right, but they were very sincere. I think these people are quite sincere. That doesn't mean they're right.
¶ AI in Fact-Checking and Personal Tasks
So the superintelligence question in some ways is a little too hard to even think of what the world is like. But something that is just short of that, which is like the thing you said about they're better at most human tasks. So let's say their fiction writing gets incredibly good.
Then what becomes the question on two levels for businesses and for workers. First, with businesses, one of the things that's strange to me is that you've got the diffusion thing and you're also running after this moving train. So it's both hard to put systems in place. Which takes time. But also the thing that you're putting in place is changing in terms of what its capabilities so much.
For a very long time, you couldn't even let this into a newsroom because it made mistakes, it hallucinated, and it had no like traceability of what it was finding. It's not just useless, it's worse than useless, right? It's dangerous. Now it has at least traceability. You can say, Show me links, and then you yourself can at least check the work, right? So that's an important improvement. Presumably it an agentic system will actually like learn to trust it to go check the links and like
You could even have some agency fact checker, right? I mean, you could start to build out the things that humans used to do. Again, I don't like to think this way and I would be skeptical and my guard would be very high, but conceptually I could imagine automating some of the work that we do. But also because it's changing so rapidly, it just becomes very hard to think about how you'd even start doing.
I mean, a thing I think people don't realize is nobody knows anything right now. I've talked to everyone from like Chairman Powell to the leaders of all the AI companies. Nobody really has a handle on what they're including the AI companies. Like they have a technique that builds better intelligence and it just keeps working. Right. And so like
If AI development stopped today, we'd have five to ten years of a complete change in the economy as we figure out how to normalize this and use this, right? And you could see this in companies. Companies that started two years ago have very narrow systems that are mostly like
Let's talk to every Chris Hayes podcast. Like that's what they would build. It's like it was really cool. You could never do that before. But now as you said, automated fact checking is not only possible, but the AI is often better at doing that than humans are, right? You throw a chapter of a book in and I was working a book. I threw a chapter in. Check every reference here. It would find things. I was like, oh, you're right. I probably have the nuance wrong in this. At one point it was like
Wait, you've done this. Oh yes. Wait, wait, walk me through this. Cause this to me is like as concrete as it gets in terms of my work life.
Great. Okay. GPT 5.5 Pro is the best model for using this right now, or you can use codex or code. I took a chapter, for example, from my new book that's gonna come out in October.
And I threw it at GBD5 and I did all the research myself, right? You know, I posted research on Twitter and pointed out a few issues of like, well, this reference that you're using, I disagree with some of the math. But one of my favorite bits was I told this story that I researched and read the article myself, found the article myself.
about funeral directors talking about AI eulogies. And what the AI reported back to me was it said, actually the conference that you're talking about in 2023 or whatever was in Las Vegas, not in Newark. And it turned out it was right and the article was wrong. Right.
Oh wow, so the article you'd taken it from got it wrong and it searched the internet to see where the conference announcement was.
It actually pointed it out to me. Right. And there were other examples like that.
That's what you want when you're talking about the GDP val or whatever. Like that's something for me that I really get, right? Like that's good fact checking. The fact checker did a great job there. Thank you, fact checker. And that's something inconceivable to automate literally six months ago.
That's exactly right. And by the way, my latest version, just for fun, was I used dispatch, which is a version for my phone and asked my home computer. I was like, could you read over my book for me? And give me notes on each chapter, right? And like fact check every piece of it. And a couple of the issues were genuine issues, a couple where I'm like,
It's being too pedantic, which is always the issue with a fact checker. You're like, you're actually right. I shouldn't have multiplied it this way. But like like there's one piece every time I feed fact checking and like opening gets really mad at me because it's like you're missing the nuance here. I'm like, they don't care about the flop calculations on An IBM series 360, and I understand you're saying a 1955, but it's a 1957 model, but
No, that's like a real fact checker almost doing too much where you're like, it's fine. Right.
Like nobody cares. I probably and like I'm not saying it doesn't miss things, right? Like I actually use three different models to fact check everything and look at them myself. And that's my human judgment about what to use and what not to.
But that, as you said, was inconceivable. I would have expected completely made up stuff. Now I still paid a human to go through everything and do the checking. I did the things I'm supposed to do, but like the human spent a lot less hours this time than the last book. And, you know, it's a big difference.
Yeah, so that's a great example, right? That's a real productivity gain. And that's just one little world, right? The other dimension of this, I want to get to workers too. You know, there's businesses, then there's just like
There's certain obvious use cases for this agentic stuff. You know, my favorite is If you have young children, whether it's daycare or it's classes, a lot of parents I think have gone through the thing where like the registration for like the jazz class opens at twelve oh one.
There's all kinds of things, make doctors appointments. There's all sorts of things you're managing in a household that would be amazing to delegate, right? And this is not a labor force. I mean, often this is because of the dynamics of American patriarchy and gender, but You know, there's a million things like that just in household life that would be super useful to be able to outsource where you're not like taking anyone's job away, but would probably be a life improvement.
Yes. And by the way, you can pull that off just barely right now. Which means by the way, in six months everyone will pull it off. You could pull it off by using the right combination of OpenClaw, which is like an open source AI agent, and some of the scheduling features. It's all very tenuous, right? Like I've done this, like I've had the systems do this. Although funnily enough,
I tried to get them to fill out a regular file I have to fill out for the University of Pennsylvania and it wasn't that it couldn't do the system. It said, I feel unethical doing this because this is certification from you that you're agreeing to do this, you know, whatever the minor thing was. And so Claude refused to help me. Yes. Which I thought was pretty funny.
It is funny and also very on brand for like their whole thing is like we have the most developed superego of all the models.
It's like, come on, please, just really. In any case, I think I would say by the time this airs in a month or two, it's gonna be much easier because it already works. Like I know people doing this with all of their family work. I know people who have Two versions of Claude, one for the husband, one for the wife, and they actually negotiate with each other about stuff to do.
It's messy. Well, or cute, maybe. I don't know. But like the point is it is messy right now, but that is a hundred percent in the realm of zooable, right? But then again, there's harder problems like travel planning turns out to be harder than filling out forms. But 2026 is not gonna be a problem because I've already done that and I've done that by shoestringing stuff together. If you want to make a weekend project of a tool that looks your Google calendar.
and signs your kid up for stuff and pre-writes or email. Like that's all doable right now. It's not a hard problem.
There's one aspect of that is that there's a huge cybersecurity and trust question. OpenClaw, you mentioned, was this program that sort of allows people at the kind of bleeding edge of AI use to kind of say to a computer
have at it. You got everything in here. Use my computer like you or me. And they can do all kinds of stuff with that. And that opens up all sorts of security issues. This question of getting to the place where it's super useful requires trust. This is really what I keep coming back to.
I gotta trust the fact checker, or if you have someone booking travel for you, you have to trust that person, right? And I mean, obviously there's ways you verify, you see the confirmation email, but you know, here's my social security number, here's my passport photo, here's my credit card. So it just seems like there's a huge amount of personal data and interaction between systems and trust that has to exist to get to the point where it truly is.
¶ AI and Organizational Evolution
doing the kind of thing a human could do.
So there are a few things there. I mean, one, you're right. I mean, everything is lagging, right? The models are smarter, in quotes again, then we have uses for them to be smarter about, right? So like We don't have the whole infrastructure on them. Like, how do agents talk to each other? What are they allowed to do? Amazon still blocks your shopping agent from going to the website.
Right. I asked Claude to go and buy something to delight me based on it has access to my email. Right. I trusted enough to do that. And I give it an extra credit card. I'm like, check all my stuff. It bought me a very nice art book actually on Purinacy, which I had mentioned before in a couple of posts. So it was very nice to get that. You know, my money. It'll be nice if it was Claude money.
Okay, but then here comes a really interesting part to me, right? We're talking about these self-contained Work is coordination. Work is people doing stuff together. And people doing stuff together isn't just coordination. It's conflict. And it's Solving conflict, dissolving conflict, or routing around conflict, or so much of what work is is. People interacting with each other.
And so then if you take a step forward on this, if we're sort of on the agent line, right, which is like the agent's gonna go do these things. The agent's gonna be out there and it's gonna be hitting either other humans it has to do stuff with, other automated systems or other agents.
Now we gotta figure out some ways that all of the stuff that is it's why we have meetings, it's why we have committees, right? Like all of these things have been built up over time, not just in the workplace, but in social structures and institutional structures. which are ways of coordinating activity between groups of people. Well, how does the AI do that?
That is the what trillion dollar plus kind of question. There's like a fifteen ways to get at your question, right? Like one of them is the way we organize work today is the org chart, which was invented in eighteen fifty five to organize how railroads operate. You live in one, right? Everybody does. Like if you work in an organization, because we had to organize work for the first time in real time across complex organizations with the telegraph. So we still use that. We still use
the assembly lines that Henry Ford invented, time clocks for Henry Ford's invention, we still use agile software development from 2002 because they were built around human limits and human abilities. So we need new methods. And their companies experimenting with some really crazy approaches to how you reorganize work with AI.
So one option is the AI just pretends to be you, right? And you have a lot of permissions in the world, right? The soccer company assumes that you're signing up your kid for soccer. If the AI has your email and your permissions, it like it could do The world on the internet, right? Because people don't have to know, but conflict's important, right? If the AI just produces output, but there's no conflict of views, that's bad as well.
Right. No, that's what I'm saying. Like conflict is actually like part of the generative friction of collective human endeavor.
And so that is one of the things I emphasize people a lot is like when the output of your work is a PowerPoint and the PowerPoint is the only goal, there is a decent chance that AI may be able to produce that PowerPoint. If the goal was the PowerPoint to go into a meeting so marketing people would yell at legal people. And then the person who was doing the interviews would say, actually I spoke to the people and you kind of are wrong about that.
And and the client has this view and there's a legal like that is harder to replace, right? And that's another reason why I don't think that the absolute destruction transformation of the economy happens the way people think it does, because there are bottlenecks everywhere.
Right. And we place high value on people who solve bottlenecks. Jobs of lawyers is mostly bottlenecking and solving bottlenecks. So the bottlenecks we have will change, right? Software development, bottleneck used to be writing as much code as you could.
Right. Now it's being a good manager of software. So now the value is going to flow from coding being the high value thing to managing software being a high value thing. So I think we're going to see a ping pong effect throughout the economy as AI gets good at something, the bottleneck's going to change. to different human tasks and that's gonna change what kind of jobs we value.
¶ AI's Transformative Role in Education
Well let's talk a little bit about teachers and students. This is a place where I think a lot of people have a lot of concern. I do. I mean, I really don't think it's great if the next generation of people don't learn how to write. I think writing is thinking and I think argumentation is a way that a person develops their critical faculties. I think if it becomes the case that everyone just goes to the AI to do their schoolwork.
in four-year undergraduate institutions, which I think is almost inexorable force that we're driving towards absent in intervention. It's just obviously going to be the case that a bunch of nineteen year olds are going to go to the II. What are the interventions you think are necessary there to get something useful out of the technology without essentially like destroying the minds of all future generations?
Yeah, great topic. I mean, and I should say, educator myself, I've been building educational tools long before AI came out. So I think about this a lot. I mean, honestly.
I predicted something right after ChatGBT called the homework apocalypse, which turned out to be true. Like all homework becomes invaluable. But like it's actually not that hard a problem to solve, right? Classrooms are actually the easiest place to solve this problem because we've already solved it before. In the 1970s, we had got calculators.
And we made choices collectively, and teachers did and curriculum designers did about what we want to keep and what we don't want to keep about doing math by hand. And we do different like you still have to learn math by hand. And how do we conforce that? You are in a classroom and you take a test and we can watch you take a test and we know whether you could do this by hand. And now we can teach things we couldn't teach before.
And also, you know, I have a I have a second grader, right? It's both that and it's also that we've all bought into this norm, right? So like you could give your second grader a calculator to do their homework, but parents don't do that because we
But when they do the tests, you can't give them a calculator because I control where they are in a classroom. I control their grades, right? As a second grade teacher, assessment is not that hard in a sealed classroom setting. I can test you. Tests suck, but they turn out to be really useful educationally. It turns out frequent testing is actually one of the few interventions in education that makes a huge difference. Like nobody likes it, but that works.
So there's this concept in education called the flipped classroom that we've known for a long time has big advantages. The flipped classroom is the idea that you do your learning outside class and your activities and active learning inside of class. So
As opposed to going to school, getting a lecture or coming home and do homework, you go home and maybe watch some module and then you come in and you're actually doing the work there with the teacher and other people. And
Examples are on the board, or it doesn't have to all be math work, right? We're doing a role-play simulation where you're George Washington. You know, we're gonna write short stories. Like it can be many kinds of things.
And that has been educationally advantageous for a long time. And by the way, lots of your classes are doing this. If you're kids in middle school, there's almost certainly gonna be a class soon where they're watching con academy videos and then doing math problems in class, right? Like the worst version of this.
So we actually know AI tutors, very bad if people just use AI to answer questions because the AI is designed to be a helpful assistant. It will tell you the answers, even if your kid is trying to learn.
They'll think they learn because learning requires friction. It's annoying. You can't learn without effort and effort sucks. We try and short circuit it. So if you just ask like, how do I solve this math problem? You will think you learned and you didn't learn. And we know this is controlled studies. However, if the AI acts like a tutor, if it doesn't give you the answer, if it quizzes you on the questions, if it personalizes to you,
There are a bunch of studies showing exactly what you expect to find, which is large educational gains. Because when I'm in a classroom lecturing to everybody, I'm not lecturing to the most advanced student. I'm just lecturing to everybody. Even in a class of 20 people, I can't. personally communicate with them. AI can be a personalized tutor. That's not an instantaneous thing. It's not magic. It doesn't happen without effort to do.
But to me, the education problem is one of the more solvable problems because I control people in a classroom setting, I can evaluate their ability, and I can build tools that they have to use outside of class that are actually good for them as opposed to all the junk ed tech out there.
So two thoughts on that. One is one place that I found AI as a technology really useful and sometimes even thrilling is if there's something I'm trying to figure out. Some new area of knowledge. Like quantum computing the other day was in the news. And I read a few articles and then I went to Claude. I was like, all right. Let's assume I'm like a pretty bright undergraduate. You know, not a physics major, not a computer science major, but I really want to understand quantum computing.
And a series of iterative questions, right? Like, okay, well, what about this? Well, why does this do that? What's the implication for cryptography, et cetera? I kind of get there's this third positionality. It sort of dissolves the binary, but like, how does that actually relate to the physics? By the way, I had read the Wikipedia entry on quantum computing. I had read a bunch of articles. And do I understand quantum computing now? Not really. Um, but I got a little closer with it.
And it did feel like that really thrilling feeling I had in nineteen ninety-five when I would start to search on the internet, you know, where it's like, oh, here's a thing I just wanna know about. And so that does feel great. The write my paper for me feels the opposite of great.
And the one thing that I would say about the asterisk about the controlled environment, again, I'm obsessed with this because this is my craft, right? This is my life's vocation, which is long form writing. The research paper is a tough one to simulate. So that seems to me like a real unsolved problem because You can write in class essays, but research papers are actually a hugely important part of my education, I think generally education. And that one seems like a tough one to crack.
I mean, there are losses. No teacher asked for all of our homework assignments and a hundred years of teaching pedagogy to be thrown out the window. And uh by the way, OpenAI didn't know they were doing it either. Not that this is a defense of you know releasing whatever, but
I think that that's right. On the other hand, by the way, pedagogically, we don't actually know much what happens in essay writing. Like a lot of us professors assign them. Hopefully it's like you and me. Like, but you and I are writers. So writing is think that's not
true for everybody, right? So there is a little bit of like kind of privileging our angle of like, you know, I break through a hundred page paper and other people are not literate enough to write those, you know, but I think you're right. And by the way, there's more profound questions. What do we teach? What do we not teach?
But I think that the idea of learning basic skills and we know how to teach them and I think AI will be helpful to them while being destructive in other cases. Because we have enough evidence for that. But I also agree with you that there are losses of stuff.
That there are things that we used to do. I mean, it's funny because one of the things you're talking about is like intrinsic learning, right? Like you're intrinsically curious, you want to learn something, you go out and learn on the internet. I think a lot of us who were in the early days of internet education thought
Oh my God, Wikipedia, education is gonna be transformed. Now everyone has access to all the information in the world. Ignorance will be defeated. And of course, like that looks so stupid now, but in the late 90s, early 2000s, this was a real thing, we believed, right? And we learned how complicated it is after.
And it did transform people's knowledge of the world, for sure. I mean, I it didn't make people more informed on average. It was transformative. But yes, it didn't it didn't solve at scale for a whole bunch of reasons.
For a lot of reasons there's polarization around technology. But look, the early data from pretty good econometricians, right? Like this is not like this sort of like messy case study, studies in Taiwan and in Kenya and at Harvard and Stanford.
AI tutors seem to make a very large impact. I mean, large in an educational research setting, which is a third of a standard deviation, which is, you know, a grade or so, which is huge still. But they do seem to matter. And like they could make learning more interesting, they could tailor it to more people. We're seeing less.
demographic dispersion, like there's less biases. We're seeing racial adoption across the board. Gender seems to now have evened out. So there's a lot to like about AIs as tutors, but like we still don't know the answers. But we're sort of in this giant experiment anyway.
We'll be right back after we take this quick break.
🎵 Music
Artificial intelligence is moving very, very fast, and it's raising new questions just about every day about what it is, what it isn't. When all is said and done, what is the end game? I'm Chris Hayes, and as part of my podcast, Why is this happening, I'm speaking with leading experts each week to help ground that conversation.
We're at right now in a situation where it's very difficult to understand what is real and what's not real.
Why is this happening? The AI Endgame, a special miniseries from MS Now. Start listening today, wherever you get your podcasts.
🎵 Music
¶ Future of Work: Policy and Human Value
So now let's talk about workers.
Because, you know, what is gonna happen to workers is a central question that I've talked about in other parts of the series. So we don't have to specifically talk about that. Because I think in some ways that's kind of a political policy institutional question as much as it's a sort of market forces question.
or a technological question. But, you know, one of the things you see all the time is like, there's a little bit of this relic from the late 90s, early 21st century of like, you need to understand how to use this stuff. And train up on it, or you're going to be left behind. Right. And this was a big part of the late nineties and early aughts. You would put that you're proficient in Microsoft Word on your resume, you know, or Microsoft Excel. You know, like, oh, great.
You know, to me, applying that framework here seems totally wrong because the whole point is that there's no interface. You just talk to the thing. Like if the technology is as good as everyone's saying, which is the point at which it presents all these revolutionary possibilities.
It's as good as people saying because you just I don't know, you just interact over wherever. You don't have to like figure out where your file is buried in something. The whole point is to sidestep that. So like why do people have to like get on board or be left behind? It's good, you know, you're gonna use it or you're not gonna use it.
So I mean I think the broad picture is exactly right. I worry a lot about bright line sort of teaching. Like be good at AI is not something we know how to teach people. Like now there do seem to be some people who are better at AI than others. Those tend to be people who are good at their job.
Right. Because they understand the AI makes mistakes kind of like a human does. Like if you use as a fact checker, you would probably pretty quickly be like, No, no, you're too pedantic about this and you're not paying enough attention to this. And you could give it the feedback that you'd be good at AI because you're a good manager of people. And I agree, that's a concern. I mean, it's a part of this general like reskilling for what is the kind of question.
I think you said the real right thing, which is in some ways this is a policy argument. This is about incentivizing organizations to do augmentation rather than replacement. It's about building the right incentives to make sure that there's support. Right. We have a whole bunch of semi broken we're going into a crisis.
There will be a prolonged long-term crisis, right? Let's say, by the way, that the most optimistic economist view that this all works out in the end, right? We've got two to three industrial revolutions that we have data on, depending on how you count them. And they all worked out in the end, but they kind of suck to live through.
Right. Like Charles Dickens is all about how much it sucks to live through an industrial revolution. Yes. Right. Like we want to cushion that. We're better at this stuff than we used to be. And I think we need to be thinking about the policy implications of this, which is yeah, again, you know, you're sort of saying, where am I in the hype scale zero to out 10.
My worry about the closer to zero is like this is that there's a lot of people who really just think AI is gonna go away. That there's something here that this isn't real, the systems aren't as good as you say they are, they make mistakes. This is real. Like I talk to CEOs all the time. They are getting value out of it. Like it's funny because they'll say things like, uh, 10 million, 20 million here, nothing significant yet. But like they're getting value out of these systems.
There has not been major layoffs because of AI yet. I don't think that continues infinitely. Maybe people get better jobs, but like we need the systems in place to cushion it, which means we believe this is real.
I also think, you know, we talk about reskilling, like to go back to the learn to code thing, right? Like there's pilot programs up and down this country of like the laid-off factory worker. There's a program called Trade Adjustment Assistance, which is sort of a notoriously failed program in many ways.
As is every reskilling program, honestly.
Exactly. As is every reskilling program, right? So what are you even reskilling for? Like if it hits the points that we think, like, I don't know, I mean, tomorrow you could create an AI avatar who could deliver the news and you know, do what my job. I mean, I think I'm gonna be better at it. I think there's like advantages I have as a human because I'm kind of sort of producing a kind of form of human trust with the viewer. I think but
There's no like conceptual boundary at all to like fully automating my job. There's not. Whatever it is. If there's some innate human charisma or there's something like that that matters, I don't know if there is or not, but there's nothing conceptually limiting it. And if it does get there, it's like, what am I supposed to do? Like
So a few things. Again, we're in sort of a zero-one realm. There's all these interesting political currents, right? Every previous realm of adjustment was aimed at blue-collar workers. When lawyers and news hosts lose their job, they have a lot more pull over the economy and over policy responses. And they're not gonna sit down, right? It's not gonna be smashing the loons with hammers, it's going to be regulations saying there needs to be a human approving each other.
decision by the AI and can't have AI on news because it's untrustworthy. So it has to be a human delivering the news. I mean, we'll see that reaction. But let's go back to the Jagged Frontier idea. I mean One thing among economists who are really savvy and I have been thinking about a lot is what's called the O-ring model for work, which is suppressing nominated after the space shuttle where everything worked great except an O-ring failed, right?
This is in the Challenger. Famously, the Challenger, the sort of point of failure was the O-ring, which had a temperature invariability that was not.
And a million other systems worked beautifully that day, right? And one thing failed. And like there is a lot of work like that where a single error is multiplicative. And it's not because AI is high error in that spot, but it might just be that's the jagged frontier. It's you need this last bit of writing, otherwise everything sounds claudy. You need a bit of randomness in your decision making. And the value of that job then flows to the choke point in the O-ring.
Right. Jobs are not one thing. They're many tasks. You are not just reading the news, right? You're making editorial decisions. You're talking to me. You're doing research. Five those areas. Right.
Now you sound like Claude, you know?
Right. It was also three thing lists. This is why I promise I'm a human watcher.
You're not just a newsreader.
Let me think with this for a little bit, right? Um Oh my god. Claudisms are now spreading throughout written and oral talks just because people are picking that up. Like I loved M-dashes. But anyway, moving all that aside, part of this issue though is that there's still then all the value of the news job still now rests on a different part of your job bundle. So I think that people overestimate how much replacement, especially for complex white collar tasks,
changes or going back to coding, being an engineer manager suddenly becomes an incredibly valuable part of your job. Now, does that get automated at some point? Maybe, but then there's another part of the O-ring. So I think that people do have a little bit too sanguine of a view of what replacement looks like, especially when parts of that O-ring of tasks
Some of those tasks in the task chain are things you want a human to do. Like I think people do want a human newsreader. They'll want somebody who is making editorial decisions. They'll want to relate to you. You know, and by the way, the imaginative version of destroying your job is not replace you with a video version of you. It's everybody who logs in gets a personalized version of Chris.
that is subtly tuned to their interests. It's the impossible version, not the easy substitute. So I do think that there's more to hope for for this kind of work change. Doesn't mean there won't be disruption, but I think Replacement is a harder task to imagine. Unless we have superintelligence, then we're all out of jobs and we're all gentlemen and gentlewoman scholars of the Victorian era where we go to parties and play ball or something.
Agents do the work, produce all of the sort of value that then produces GDP, that then flows out somehow in a way and is distributed.
Taxes. That's the way it flows out.
We'll see.
We tax the hell out of all use of AI systems.
Ethan Mollock's a professor at the Wharton Business School at the University of Pennsylvania, where he studies AI, its implications for education, entrepreneurship and work. He's got a new book.
Coming out in the fall called Coexistence about working with AIs that are, as he says, sometimes smarter than you. His last book, New York Times bestseller, Cointelligence, Living and Working with AI. And he writes the Substack, one useful thing. That was a really great conversation, Ethan. Thank you so much.
That was terrific. Thank you for having me.
🎵 Music
You can get in touch with us by emailing withpod at gmail.com. Why is this happening? It's produced by Donnie Holloway and Brendan O'Melia. It's engineered by Hazik bin Ahmed Farad, Bob Mallory, and Jake Loomis. Katie Lau is our senior manager for audio production. Our coordinating producer is Franny Kelly. Aisha Turner is the executive producer of MSNow Audio. Watch us on YouTube by going to ms.now slash with button.
🎵 Music
Hi, I'm Jane Coaston, former writer and podcaster for places like the New York Times, The Atlantic, and National Review. And now I'm here to hang out with you five days a week on What A Day, Crooked's Daily News Podcast. That's right. Now who's respected on both sides? Okay, yeah, it's not me, it's probably still just Dolly Parton. Moving on.
These days, it's hard to separate what matters from all the noise, especially when the noise somehow also feels like it's made out of knives. That's why every weekday, the team and I scour the headlines to bring you the stories that matter most to the way you live, in just 20 minutes. Listen to the show wherever you get your podcast.