Talking With Azeem Azhar - Paul Krugman
Paul Krugman
SubscribeSign in
Playback speed
1×
Subtitles
Share post
Share post at current time
Share from 0:00
/
Transcript
SPEAKER 1
Hi, everyone. Paul Krugman back on my usual schedule of recording interviews. And today I'm talking with Azim Azhar, who I spoke to in January 2025, basically centuries ago in AI time. And with AI on everybody's mind, I thought it would be good to revisit, talk about some of the basics, then talk about, you know,
where we are now in this whole incredible story. So hi, welcome to another conversation.
SPEAKER 2
Yeah, thank you, Paul. And it has been 18 months, also known as one and a half centuries in AI time since we spoke.
930
250
122
Talking With Azeem Azhar
Taking the pulse of AI again
Jun 13, 2026
930
250
122
Transcript
I last spoke with Azeem, the proprietor of Exponential View, 18 months ago — ancient history on this subject. So we revisited the state of AI.
.
TRANSCRIPT:
Paul Krugman in Conversation with Azeem Azhar
(recorded 6/12/26)
Paul Krugman: Hi everyone. Paul Krugman back on my usual schedule of recording interviews. And today I’m talking with Azeem Azhar, who I spoke to in January 2025, basically centuries ago in AI time. And with AI on everybody’s mind, I thought it would be good to revisit. I should say Azeem is an independent researcher and founder of Exponential View, which is one of the top tech Substacks out there.
So hi, welcome to another conversation.
Azeem Azhar: Yeah, thank you, Paul. And it has been eighteen months, also known as one and a half centuries in AI time since we spoke.
Krugman: Yeah. Let me ask sort of the dumbest question: what is this thing called AI? How does it do what it does? I mean, even skeptics have to admit that it’s really impressive how it’s sort of leapt over all of the previous barriers. How is this happening?
Azhar: You know, I think we’re still figuring it out. I think of AI ultimately as a machine that does certain things, and it’s been built by passing first millions, then billions, then tens of billions, hundreds of billions of trillions of words of human output through a neural network to give it some sense of how humans have thought about the world. And because it operates at dimensions well beyond the form of space and time, it seems to be able to find relationships between quite complex concepts. And I think we’ve all had that experience, whether we’ve been using Chat GPT or Claude over the last two or three years, that it seems to be able to recognize things that are quite deeply related that don’t immediately spring to mind.
And in the last year and a half or so, the labs have started to train the AI models not just on words in books, but actually on tasks, like, “what is the set of things that you do to write a piece of code that does something?” “What is a set of things you do to use a piece of software in an enterprise?” And they’ve tried to train those models on those particular tasks. Essentially it’s aping what we do, and they use various mathematical tools like reinforcement learning where the model notionally gets a reward. Of course it’s not a reward the way you and I think of it because it’s a machine.
Paul Krugman: Right.
Azhar: And so that’s what it is. It’s sort of reflecting back, but also I think discovering some really deep relationships in the world that we might not spot, you know, prima facie as humans.
Paul Krugman: Brad Delong calls it “a vast stew of linear algebra,” which makes some sense to me because I think that Pagerank with Google was the last thing I actually understood. And that’s the eigenvector with the largest eigenvalue. Not that anybody needs to know that, but this is like a million times bigger, right?
Azhar: That’s basically it. Yeah.
Krugman: But it’s sort of not what artificial intelligence was supposed to be, right?
Azhar: No, not at all. I mean, I sometimes go back and look at the TV series of the seventies that I grew up with as a child, and they’ll always have an AI in the spaceship. Space 1999 had an AI you could talk to. And it was very precise, it was very clipped, and it did things and got things right. And there was a sense that you could trust it. But you’d never think to say, as I sometimes do now, you know, “Find me five analogies to help make this point.” I use it as a brainstorming partner, or I give it tracts of my book, the book that I’m writing, and say, you know, “How would Paul Krugman criticize this argument?” And I get suggestions that I then work through by hand? I don’t think we really imagined it would look like that.
Krugman: Yeah. In sci-fi it would talk in a monotone and would be relentlessly logical. And in fact these models are unpredictable, they’re sometimes temperamental, they’re not reliable. That’s probably one of the big problems. It’s not at all what we imagined.
Azhar: It’s not at all and this point about reliability is so complex. A couple of months back, one of the versions of Anthropic’s Claude came out and I found it so sycophantic that it became unhelpful because I like these things to help me on hard problems and to challenge me. So I switched back to Chat GPT, which has always been a little bit less friendly. And what’s going on there, Paul, is that because we don’t really have a good theory about how to build these. They are developed almost like in a petri dish and nudged in particular directions so they take the shape that we expect them to take. And to use an economist term, they improve non-monotonically with every release. So you’ll see the latest release of an Anthropic model, and there are maybe twenty or thirty public benchmarks that they’re measured against, like how well they summarize text and how well they write software code. And the next version of the model won’t necessarily be better at everything than the previous version, because you lose something in order to get it. And that’s the complexity that the labs are wrestling with.
Krugman: Wow. Okay. Second naive question. I don’t think I’m a Luddite. I’ve always been happy to adopt technologies, but maybe I’m incurious on some of these things. I tend to pick up things like mathematical techniques, as needed, because I see something that could be useful. Now, I’m using NotebookLM to extract tables from PDFs, that sort of thing. But what should I be doing? I have friends who are using Claude a lot, but I can’t quite figure out what particularly agentic AI should be doing for me.
Azhar: You know, I’m really sympathetic to that because I have the same issue. These tools have been developed by software developers in a really particular part of the world, which is Silicon Valley, where the culture really revolves around the art of the programmer. And so if you have a programmer’s day and you think in coding terms and you have programming workflows, it becomes really obvious what you do with a really advanced AI tool. I do a lot of research, some of it qualitative, some of it quantitative, and in such a world, those workflows don’t match the way that I think through problems. And so the way that I get around this is that I do look at things on Twitter or X as it’s called because people are sharing tips. And I often just ask the models, you know, “What could I do with you given that I’m trying to do this thing? I’m trying to solve this problem.” And it will come back and give me a suggestion.
And I have had some success with agents. So I have an agent called R. Mini Arnold. So R is a play Isaac Asimov’s robots. They’re all called R. Arnold is after the good Terminator in Terminator 2, played by Arnold Schwarzenegger, who protects humanity. And R. Mini Arnold is available on my WhatsApp and it’s available on email.
Krugman: Okay.
Azhar: And it has access to a whole set of resources. It can browse the web, it can access LinkedIn, it can access Twitter, it can look at my library of PDFs of research that I’ve downloaded. And I can throw tasks to it a little bit like I would say a pretty decent but slightly temperamental graduate student. So sometimes it just disappears for six or seven hours at a time. And one of the differences between using an agent like that and using Claude is that R. Mini Arnold has a lot of my life’s context. It knows the music I like, it knows the book I’m working on, it knows the investments I’m making, it knows the essays I’m doing, it’s got the calendar of speeches that I’m about to give. And so when it goes off and does a task, it tries to figure out what in my world is this going to be relevant to and where can I draw threads from? And when it works, it is really sublime and it does feel a little bit like science fiction.
But I would say it’s incredibly brittle. I mean there’s breaks every four or five days.
A specific example was, I was thinking about the Paul David’s research about why electrification took the time it took. And I wanted to understand what were the determinations of determinants of that thirty-five year lag from Pearl Street generation to, you know, productivity growth. What could the levers be? And so I threw that into R. Mini Arnold and it set up a team of sub agents which had personalities of key economists and was able to go off and do research the way the AIs do, but also research on all the academic papers that I have downloaded in the past.
I have access to JSTOR, I’m allowed to download a hundred PDFs a month. It can look at all of those and start to compile an answer in a way that perhaps a Chat GPT can’t. And it knows the context of my book and it knows the context of the essay I wrote. So what then comes back is something a little bit more structured that I can then play with. It’s a marginal improvement on doing this on Chat GPT. I’m sure you could probably figure out how to do it. But it’s quick. I use it on my iPhone. I often do this when I’m walking through the airport and I want to solve this and have this result when I’m sitting on the plane. I’ll fire that query out and it goes back and goes out and sorts that out for me.
Krugman: Okay, I guess I’m getting it. But obviously you and I are not typical. The people who are using AI the most are going to be middle managers, business people, etc. And I find myself thinking about what I think of as the homemade pasta problem.
Azhar: Mm.
Krugman: You’re probably too young for this, but there was a time when I when young and we were using stone axes for computing, and there was a big fad of making your own pasta. Little pasta machines were everywhere. And then at a certain point there was kind of a collective, “What the hell are we doing? You know, store bought pasta is actually better. The Italians don’t do this.” And I have to think that for most tasks, the range of agents can’t be that wide. But why wouldn’t they sell that kind of thing off-the-shelf, as it were?
Azhar: Yeah, well I think it’s different for an independent person or a small business or a middle manager in a big company. I would imagine that you will start to see people selling specific agents that solve your marketing problem. If you have a barber’s shop and you’ve got four chairs and maybe 30 people a day coming through. Right now what you do is, you go to ChatGPT and you help it write your collateral for your website. That feels like it’s an interim step to somebody delivering the actual finished product. Why haven’t we seen it? I think we haven’t seen it yet because the terrain is still big enough.
Beyond Anthropic and OpenAI, there’s a lot of other companies building agents that are these end-to-end workflows for businesses. They still believe that the prize for them is to build the generic platform that is the tool for all tools. Because if you get that right, you have a much, much bigger business than if you’re just a vertical application. And I think we’re only a year or two into these entrepreneurs building such businesses. I think as some of them succeed and some fail, the ones that are not able to succeed in the general space will start to verticalize, which is what we saw in the advent of the internet. We saw it in software as well.
But I think within a big company it’s a different set of questions because you have far fewer degrees of freedom as a marketing manager in a large company than you do if you own your own barbershop. You have all these rules, you have all these other teams you have to interface with, you are held to the priorities and the plans of the company as a whole. And in that instance, I think, it’s much harder to see how you use AI to really change the way you work.
Krugman: Yeah, I mean, again we’re talking about ancient history here, but you know, everybody still uses Excel, even though it has always been horrible. But the constraints of corporate life mean that everybody has to use Excel. So that means maybe we’ll see quite a lot less coding a few years down the pike because the people will just be able to purchase whatever it is they need. I don’t know.
Azhar: I think there’s a balance. You hear people proselytizing heavily, saying, “I think this technology is going to be impressive and have a significant impact.” But when people pitch this, they forget that there are other actors in the market who might respond to what’s going on. Right now, if you’re a large company, you want to be building as much as you can because what you can buy isn’t right for the market. If you think about Henry Ford putting together the Highland Park plant, he couldn’t go to a supply chain and buy what he needed because nobody was thinking in those terms. I think we are slightly at that stage for large corporates now. Whether we’ll be there in five years, I don’t know.
The question we have to consider is where the value will reside: between having your own capabilities to design software for your processes, or handing that over to another company designing software for a hundred businesses like yours. Historically, it has made more sense to hand it over to another company, but the cost curves may have changed sufficiently that you’d rather have the nuance and control to do whatever ‘vibe coding’ becomes in 2030.
Krugman: I know with healthcare software, organizations like the VA that built their own have done much better than the ones who tried to buy it from Microsoft. So yeah, it might be a story that makes sense. And actually, since we’re talking about going for the models versus something much more specific, how do you think about the Chinese versus the big US AI firms?
Azhar: I’ve just spent eight days in China and I was really fortunate. I got to speak to developers and engineers and management from about a dozen of the Chinese labs. In many cases they hosted us in their offices. The main thing the Chinese companies say about the US firms, is that Claude code is brilliant and Claude is the best model that is out there and they really couldn’t get enough of it. The term is, they’re Claude-pilled. They talk about the constraints on getting access to computational power but just in a way that’s a fact of life. I mean there’s no sort of commentary on it other than it’s hard. They have to figure out how to get around that and how to build a culture of efficiency when you don’t have as much [computational power] and I think they have built a culture of efficiency really, really well. I think it’s going to help them over the longer term. They don’t really talk about competition with US labs the way the US talks about competition with China. But they do see themselves competing with each other.
And as you know, that’s what the Chinese economy is. It’s mayors in different cities who almost act as venture capitalists who compete tooth and nail with each other to become the electric vehicle hub or the solar hub or the AI hub of the nation. And what I would say is, the models are really, really capable. They’re very efficient, which is why they’re so cheap to run, which makes them very competitive for a whole range of tasks. But at the margin, it’s instructive to note that everyone was using Claude for coding as opposed to the cheaper Chinese version.
Krugman: That’s interesting. So you can imagine a future where a lot of businesses are actually using these less comprehensive but much cheaper models. I think what I’m gathering from you and from other people is that a lot of entrepreneurs in the US are still dreaming of the uber-model that solves all problems but that probably is not going the way it all goes. That in the end we’re gonna end up with a lot of specialized models, but also the uber-models will still have a role.
Azhar: Yeah, it never made sense to me that you’d have a single model that would do everything because if the single model is going to solve the Riemann hypothesis, it’s gonna require a lot of resources. And if all you need to do is get it to root a bill to the finance department, it seems a bit silly to ask Einstein to come and do that for you. We’ve had segmentation of markets for a long time and it’s like with airlines. There’s a reason why not every seat on an airline is first class. Some passengers don’t want it, don’t need it, won’t want to pay for it. So I do think that the ecology looks like a whole array of much, much cheaper models that are serving by volume lots of corporate needs, and then having more sophisticated, complex models for the more complex tasks. I think you’re already starting to see this.
I don’t see it, by the way, as a shock to the industry. I just think this is what happens as an industry matures. You know, you start with one size fits all, then you start to segment your customer needs and you start to serve them in the most profitable way you possibly can. And that just feels to me like the way that the markets have matured.
Krugman:
Okay. Let’s move to more macro considerations. People have been worrying about a bubble. A lot of us still remember the nineties quite vividly and think about all of that. But you just aren’t seeing the bubble. You wanna talk about that?
Azhar: I remember what it was like in the nineties. I lived through that one and also the housing bubble, which frankly was far, far worse and much more terrifying. I have a really simple mantra here, which is that honest customer revenues tend to be the engine that gets you through this, right? You know, what caused the problems with the US railroads in the 1870s and 1880s? It was that the revenues didn’t materialize because the tracks were being laid in places where there were no towns. That was a problem. The same was true in the dot-com era. My team and I realized last year that it’s very hard to get good quality data on how much was actually being spent by American businesses and consumers on AI. So we’ve spent several months building systems and gathering data to give a deduplicated view of what that number is. And just to give you a sneak preview, is $150 billion per annum, annualized at the end of May 2026, and about 90 billion dollars in the previous 12 months, from May ‘25 to May ‘26. So you can see it’s growing, and those are deduplicated numbers.
So if you spend a dollar with OpenAI, and they have to pay Microsoft 60 cents to run the servers, we only count that as a dollar. We don’t count it as, you know, $1.60. It’s a much faster revenue growth rate than mobile or the internet. It’s also a small number because the US is a $32 trillion economy. And I think the thing is that at that level of spend, you are able to roughly cover the depreciation on the enormous capital expenditures that have gone into AI just this past year. But next year or the year after, you have to double your revenues again and again in order to cover these increasing commitments.
The thing that often pricks a bubble is when financing starts to get a bit smelly. That was clearly the case in the global financial crisis, where synthetic collateralized obligations were magnifying the risk on subprime mortgages—it was all “smelly finance.” In the dot-com bubble, the dot-coms themselves didn’t really have much smell about them. There was a lot of disbelief, but the telecoms clearly had issues with their internal revenue generation.
So the other thing that we look at is how bad, poor, or strong or robust is the funding quality. And that funding quality measure is definitely getting worse. It’s worse now than it was nine months ago. But it doesn’t seem from the numbers to be at the level that it has been historically when these things have imploded. Nor does it seem to be the type of exposure that is really systemic, which is what we saw in the global financial crisis. There are companies like Oracle and Coreweave whose debt looks very risky, and it’s harder and harder for them perhaps to raise money, although Oracle just did. But it doesn’t feel like it’s systemic.
You know, when the the global financial crisis popped, no one knew who was in trouble, whereas now you’d be able to isolate it with a single company or a single firm. So at the moment we feel that this is still a demand-led boom, that funding quality has definitely gotten worse, but not so bad that I would say that there is an imminent problem on the horizon.
Krugman: So at this point, you’re saying that roughly speaking, final demand for this is about half a percent of GDP. What share are AI-related stocks in market value? It has to be substantially larger than that.
Azhar: They’re about forty percent of the S&P 500 right now.
Krugman: That’s a huge mismatch. Revenues are not the same as profits, but you’re talking about what is still a relatively small business relative to this immense economy, yet it dominates the financial markets. That would be at least a possible source of alarm.
Azhar: Let’s dig into that, because a stock price is a reflection of the expected future value aggregated across the market. Forty percent feels high, but if you look at the measure of earnings, these companies actually have a much higher proportion of earnings and earnings growth.
If you look at the US stock market in 1900, after the railway calamities of the mid-to-late 19th century, railroad stocks were sixty percent of the capitalization of the US market. We had worked our way through the busts by that point. There’s a fantastic piece of academic work by an American finance professor named Bessenbinder. He looked at the stock returns of 23,000 US stocks from the 1900s through 2022. Those returns are highly concentrated. About two-thirds are concentrated in roughly 30 companies. Those companies are oil, electricity, or car companies—the general-purpose technologies at the start of the 20th century—or they are the IT companies like Apple and Nvidia. The only exceptions were Walmart, a couple of healthcare businesses like Pfizer, and JP Morgan.
Historically, you get this concentration of a number of winners when you have a new general-purpose technology, and that is showing up today. I don’t feel we’re overly concentrated from the perspective of risk, and the price does not feel totally out of whack compared to where we were during the dot-com era.
Krugman: One last devil’s advocate question. I keep thinking of the California gold rush. If you had looked at the revenue and spending on gold-rush-related businesses as a whole, it probably looked solid. But the trouble is it wasn’t the gold; it was the picks, shovels, blue jeans, women and whiskey that were the revenue streams. Is that a fair question to ask about AI right now?
Azhar: It’s a great question to ask. The question is what determines that $150 billion annualized demand? We see that just under 30% of the S&P 500 have pointed to a generative AI project with a quantifiable result in their earnings calls. They are under pressure to say they do this, so maybe that’s what’s going on. But when I talk to executives, like 30 finance businesses in New York, they all plan to spend more next year, even though not a single one could point to even a 10 basis point improvement in their business from the investments made so far.
Krugman: Right.
Azhar: When we break out that $90 billion, $60 billion of it is in the US. That’s a lot of money for a single company, but spread across thousands of firms, it’s still at the experimental stage. We should consider whether these executives are learning by doing. The messages I get vary from those having success in the tens of millions who want to reach hundreds of millions, to those finding it harder but persisting. We’re slightly beyond pure picks and shovels, but in Paul David’s work, it took 50% of American companies getting electrified before the productivity rise. We’re a long way from that.
Krugman: Headlines flashed about a KPMG study with case studies on the usefulness of AI that turned out to be AI hallucinations. It’s a wonderful thing.
Azhar: It is brilliant. One thing that is quite challenging is that the market has talked a lot about bottlenecks. We saw this with railroads when the US couldn’t make enough steel. There are these bottlenecks, and there’s a lot of emphasis on power and getting electricity to the system.
There’s more demand than supply capacity for AI right now, but there’s a question of whether there’s enough capital. We may see another few trillion dollars of intention from tech companies to build infrastructure to 2030, which starts to rival the new issuance of the US Treasury at $2 trillion a year. I’m wondering if this capital constraint is going to be an issue or if the market knows how to clear it.
Krugman: Ordinarily, we’d expect to see that in prices. Real interest rates are well off their pre-COVID lows. They are higher now, but still substantially lower than at the peak of the nineties tech boom, when they were around four percent. They’re more like two now.
It’s surprising, given the AI boom and massive budget deficits, that rates aren’t even higher. Whether this is an actual constraint, Nvidia is not the US Treasury. They need risk-tolerant capital. The possibility that these firms may not be able to raise enough money is something we need to think about.
Azhar: Yeah. On that Nvidia point, I saw that credit default swaps on five-year Nvidia bonds—the cost of insurance against default—are currently lower than US Treasuries.
Krugman: I saw that, and it strikes me as completely crazy. If you think the US government is not reliable, you shouldn’t be investing in chip stocks; you should be investing in canned goods for your bomb shelter. But anyway.
Azhar: Are you telling me that markets aren’t perfectly rational, Paul?
Krugman: Good heavens, I can’t say that; they’d take away my economist card. We’re recording this on SpaceX Day, and I’ve been wondering if there are limited pools of capital for cutting-edge investments. I wonder whether Elon Musk is diverting capital that AI might need. A whole lot of meme money is pouring into SpaceX right now. Is that something I should be thinking about? I mean, he’s got what everybody tells me is a crud AI product in Grok, and yet…
Azeem: Musk showed his willingness to adapt; his AI product is now being subsidiarized using his capacity to serve customers like Anthropic. He has an incredible following, but people who have worked with him say his ability to relentlessly focus and optimize sets him apart. His first-principles thinking has brought down the cost of space launches faster than anyone in history. He pushes the rate of learning aggressively. For all the challenges and his mercurial behavior elsewhere, that’s generally a good thing because technology has brought down the cost of inputs significantly.
We’re going to be much further ahead in space than we would have been if SpaceX had not been successful. It raises questions about how to govern what used to be a commons, but there is a definite benefit from coming down that learning curve so quickly.
Krugman: That’s fair. The one time I looked at Musk’s activities and thought he was really onto something was when I realized he diagnosed that the cost of space launches is really the rocket, not the fuel, and recovering it makes all the difference. Being able to make it happen is a real productivity thing.
This is all moving so fast that we don’t have time for the technical productivity issues we had in the past. It’s feeling like a Solow moment where people say, “I see the technology everywhere but in the productivity statistics.” Do you want to talk about that?
Azhar: It comes up all the time. I wonder if we need things to happen more quickly than we used to. We aren’t seeing it in the numbers yet. Erik Brynjolfsson at Stanford says he thinks it is showing up in the aggregate numbers. How quickly should we expect a technology like this to show up? At $90 billion a year, that’s not much of US GDP. These are early stages where companies are learning. The first $100 million you might spend on AI is about learning, and we’re in that mistake-making phase.
The model Paul David and William Devine talked about in electricity is helpful. In the first phases, you’re retrofitting your capital stock and processes with the new technology. It’s not until you depreciate existing capital and change processes—like Ford did at Highland Park—that you see productivity benefits. To put numbers to that, what would we expect to see in the Ford equivalent of Highland Park in terms of output?
Krugman: Yeah.
Azhar: I thought we might see what happens to revenues per employee in an AI-native firm. Across high-end companies like McKinsey, it’s about $400,000. For Meta or Google, it’s about two to two and a half million dollars. In AI-native firms like Mercor, that number is closer to seven million dollars per employee. For Anthropic, it’s close to ten million. You can measure the enormous commercial productivity of a single employee if a firm is AI-native. We’re talking about a handful of firms, but we can pick up the shape of what’s possible for the productivity of a single employee. It may be hard, it may take time, but it’s possible.
Krugman: What would those numbers look like per dollar of invested capital? One worry is that this is an enormously capital-intensive business that replaces labor. The oil refineries of New Jersey have enormous revenue per employee because there are no workers, just monstrous capital installations. Is that a factor?
Azhar: Anthropic has raised in the tens of billions rather than hundreds of billions and had a profitable quarter ahead of schedule. What we don’t know is how much of that capital goes into developing the next model versus monetizing previous generations. Their IPO in the next six to nine months will tell us.
Chinese companies are using much less capital to build models that are nearly as good. So I think the harder part of your question is that if every model that OpenAI or Anthropic costs ten times as much to deploy and develop, but lasts only a couple of years before it’s defunct because of competition, what needs to be true for that to be sustainable for more than a year or two? To me, that is a really tricky question as well.
Krugman: You’ve cited intermediate measures. Rather than revenue, we look at generated lines of code, which has exploded, versus actual usable applications, which hasn’t. Does that tell us anything?
Azhar: Lines of code is an odd measure. We’ve made it much cheaper to write code, so less determined people are writing it now. It’s unsurprising the increase hasn’t been met by proportional productivity. Data suggests we’re getting more high-quality code, but also a lot of useless waste. This isn’t the first time a useful input in the economy generated waste. Think of a barrel of oil: we count the whole value in GDP, but two-thirds is thrown away as waste heat. Only one-third is useful energy. Sloppy lines of code are a similar form of waste we’ve been happy to tolerate in other sectors for a century.
Krugman: A weird analogy is when widespread word processing came in. Books started getting longer. It was so easy for authors to turn out hundreds of pages. What might have been a two-volume series became five.
Azhar: On that front, we’re at an enlightenment moment. In 18th-century France, the battle was over who gets to write and express their story. Men and women produced remarkable works with quill pens that encapsulated a world.
Krugman: Right.
Azhar: Is it worse that we allow for more expression? We are worse off when that connects to an algorithmic recommendation system that drives constant slop at us. But we aren’t inevitably worse off because we’re giving access to many more people.
In reducing costs of access, we might find amazing people. In breaking down silos of knowledge, we might find connections—perhaps something in battery chemistry that is useful in cardiology. We don’t know because we’ve never been able to get those experts to talk. I look at each opportunity discreetly.
Krugman: There is a potential book here: The Upside of Slop. This is an unrecognizable scene from eighteen months ago. Wow.
Azhar: We could get ChatGPT to write it.
Krugman: I started my career writing papers longhand on yellow legal pads. Amazing change.
Azhar: I still write everything with a fountain pen. I’m writing my new book longhand and most of my research is too. The computer is turned off because AI does all the boring stuff like PowerPoint and emails, giving me time to apply my brain to things I want to think about.
I’d be happy to continue this conversation in a few months. Thank you for inviting me.
Krugman: Thanks so much. Take care.
930 Likes∙
Discussion about this video
CommentsRestacks
I keep wondering where all the electricity and water to cool these huge data centers will come from, and who/what will the monster companies building them prioritize, human needs or profits? Dear Claude, is there an alternative to capitalistic cannibalism?
Like (68)
Reply (10)
Jun 13 Edited
It’s a bubble, and you can see it right there in Azhar‘s data. He said American businesses and consumers spent $90 billion on AI over the past 12 months, with a current annualized run rate of $150 billion (deduplicated data).
Yet per Goldman Sachs, the total combined capital expenditures planned by major U.S. AI and cloud infrastructure companies (often called hyperscalers) *just* for the year 2026 are projected to be $765 billion, growing to $1.6 trillion in *annual* capex in 2031.
If companies spending 5X their AI revenues (what people are paying) doesn’t scream “bubble,” I’m not sure what does... And that’s 40% of the stock market.
Not to mention that a lot of that spending and investment is incestuously circular: Company A invests in Company B, so that Company B then has the funds to buy Company A’s products. Plus these models and data centers grow obsolete quickly and lose most of their value within very short timeframes. It’s a bubble.
Like (30)
Reply (2)
"incestuously circular" Indeed. The ultimate revenue source for railroads or telecoms is individuals.. however that translates up through the layers of business. There must be something that individuals can and will pay for and a way for them to earn the money for that increased payment from the technology advance.
When I look at my life and think about goods I buy and services I use, I see that there may be some slight incremental efficiencies for the companies from which I buy.. but I don't see anything life-changing on the horizon even in the most glowing promises of what AI might potentially do.
Perhaps an AI advocate could list what those might be?
Like (10)
Reply (3)
The advocates never have an answer to your question. They deal in hopium. Just imagine what this will be like when it is a whole lot better than it is now! Pigs will fly!
Like (6)
Reply
As a software developer I can confirm that AI tools are a productivity multiplier (at least for software developers). But the investments flowing into the sector are so vast, it feels like there is an assumption of transformative change across all human enterprises. My gut is that it will be transformative tech, but also that many people will be losing all their money.
Like
Reply
The big money will be made by the companies that use AI to make new (or improve old) stuff and expand existing or create new markets. The AI companies are renting the tools to those companies unless they start to demand a cut of their customers profits or start to compete with them.
Like
Reply
I wanted to look up these numbers, but got lazy. Thanks for providing! Ed Zitron ( wheresyoured.at) does really good research and writing on this. I don't think deduplicating the money provides a clearer picture, rather the opposite. In a financial chain of 5 interdependent links; failure of one leads to failure of all. Along with not showing real business value that would lead to ongoing (much less exponentially growing) revenues (see my earlier comment); rising negative sentiment towards datacenter impacts on the environment and energy costs are making the trillion dollar investment look like a fool's errand.
Like
Reply
Jun 13 Edited
Exactly. If AI doesn't serve to help all of humanity in the long run, it's useless if not evil.
Like (23)
Reply (2)
I completely agree with you Ed. Thank you. Intelligently and succinctly stated!
Like (4)
Reply
We need to keep it from being fenced in as in the commons, as he pointed out. People need to believe in the cauldron of abundance that this technology can bring and not let scarcity be manufactured.
Like (2)
Reply (2)
Do you have any confidence that those “feeding and teaching” AI will be our benefactors? Musk was able to make it respond that he was Mr. Wonderful, handsome and fit a couple of years ago. His charity is in trouble every year because it doesn't fund/nor distribute as required. Bezos donated to schools in Miami of $6 million, which is a miniscule proportion. Compared to my wealth, it is equivalent to me giving a penny to my new community schools, when I settled there to avoid paying taxes. Then there are the dangerous ideologies of Musk and Theil being fed to the “baby" containing The Book of Revalations” and South African Appartied, mixed with Ketamine. All are so self important, never looking at the plights of others sharing their air and the water near their data centers. Their words are out about a living wage from all the excess production. As a Social Security recipient paying a quarter of my income on insurance for home/ auto.health, I can assure everyone, that there will be little left over to purchase their wares. A "thriving wage” would be more encouraging, but my faith in the holders of power always wilts quickly and dies instantaneously. As far as space goes, technology has polluted it, with little requirement or interest in cleaning it up. I fear that will be the case in the future if our planet that the rest of us inhabit without the ability to get away from this ruination to our private islands.
We because he
Like (7)
Reply
I'm not sure I see a "cauldron" here. I think the abundance, if there is such, will fatly fill some individuals silos with profits but be of little benefit to people in general.
For instance- take the "barbershop" example in this conversation - is there really anything - anyway - that slightly improving efficiency of scheduling or anything else will make that business significantly more profitable? It still comes down to cutting hair and AI can not do that.
The same question - and answer - applies to millions of other businesses. However, the workers in those businesses must be the ultimate source of AI earnings just as, in Paul's gold rush example, the miners themselves were actually the revenue source for the jeans makers and the bordellos. But some of the miners had actual gold to spend; hair cutters do not.
Like (4)
Reply (1)
There are two types of use for new technology.
One is optimization of existing processes. The barber may be willing to pay to make supply chain, salary, and taxes easier. But efficiency gains are capped. If overhead is only 10%, you aren't going to do much better than that.
The other type of improvement in net new business. If AI enabled a barber to expand and lower costs, maybe there is a franchise opportunity and customers start getting haircuts more often.
Perhaps a bad analogy, but I thought I would play it out.
On the mining/gold issue, Levi Strauss grew because their product was not just useful for mining. A boom commodity can incubate a product that has value beyond the initial ecosystems.
Like
Reply (1)
Of course you are right that the barber may expand. That is precisely the concentration of wealth that technology, in general, causes. The barber can only expand his business if there are more persons who then will work as his employees for wages (or as "independent contractors" which screws them out of benefits). The barber's business gets bigger to the extent that more persons now are 'capped". Also, there can only be a limited number of barbers who can expand in that way because there is a limited number of persons in any city who need haircuts.
This is why our economy has a built in tendency to be "K" shaped.
Like (2)
Reply (1)
Efficiency improvements can provide benefits even without hiring. If 10% of the barber's time is overhead, and technology reduces they to 5%, the barber can serve 5 /90 more customers.
You seem to have missed my point about changing the frequency of haircuts. A radical enough technology change can transform the way the product is consumed.
But of course even new business opportunities have limits to growth based on the target customer base.
The causality link to the K shape economy is tenuous. There are a lot of factors involved.
Like
Reply (1)
The customer base for barbers is decreasing (just demographics) and the increase in customer base is not an increase, it is a shift from other barbers. The efficiencies within a particular barber shop will possibly result in an increased profit margins for the owner of that shop but every business person I've ever know will pocket that increased profit. It will not go to employees except under duress.
So we have a concentration of businesses and an upward segregation of income. That's the "K".
Like (2)
Reply (1)
You are torturing the analogy beyond its limits.
Like
Reply
I like your term "capitalistic cannibalism" Sally. Insatiable self centered greed and lack of heart and soul will be our undoing unless we can become capable of changing the trajectory on which we're blindly and ignorantly racing towards. Thank you Sally.
Like (12)
Reply (1)
Exactly. Whether we we will survive as a civilization or even a species depends on if, how, and when we can wrest control from the powerful, politically and financially, who at present are one and the same: capitalist cannibals.
Like (10)
Reply (1)
Thank you Sally—so true!
Like
Reply
Regarding the data centers that Musk wants to build in space, I can see that they will be powered by the sun, but where will the water come from?
Like (2)
Reply (1)
It's worse than that. You could in fact send up coolant but there is no way to get rid of the heat—hard vacuum is an excellent insulator. Given current technology, this is not feasible.
Like
Reply
I keep wondering why we're supposed to pretend that space exploration is cutting-edge technology instead of just billionaires who watched too much Star Trek and wanna light some fucking money on fire to appease their space boners.
Like (2)
Reply
That’s a fascinating question Sally. Would you like to explore ways to build an alternative to capitalistic cannibalism? If so here are some things to consider…
Like (2)
Reply
"Capitalistic cannibalism"
Good one! For perspective on where it's headed, this personal story happened yesterday. We have a Moen flow meter for insurance purposes and it's been malfunctioning.
My wife called Moen support and the person was very patient and helpful. Solved the problem and she said how nice it was to talk to a real person. (maybe you know where this is going:)
A short time later we received an email from Moen asking how we liked speaking with their "Genesis AI system." My wife was floored because it was a lengthy conversation she believed was with a human. The upshot is days of people being employed by phone banks are running short. And so it will go...
Like (2)
Reply (2)
I share the bubble anxiety of many others, and your concerns about who it will benefit. Benefiting all of humanity in the long run is a laudable goal, but what technologies do we have today that could pass such a test? I can think of a few that might, but others that wouldn't, and reasonable people could disagree on which ones threaten our survival as much as they could save us.
Like
Reply
My experiences with the robots are usually negative. I can usually figure out a straight-forward problem myself - the robots I've dealt with can't answer a complex problem, but if I can get to a human, they can.
Like
Reply
My question too, Sally.
Like (2)
Reply
Jun 14 Edited
Elon Musk's company, Grok, is running an illegal natural gas power plant in Tennessee. It is noisy and polluting and he has apparently bought the local government, which is doing nothing about it, or perhaps the Tennessee governments don't care that it's poisoning Black people. Lake Tahoe's power is about to be cut off because the Nevada utility that provides it is going to divert it to "AI" instead.
https://qz.com/nv-energy-lake-tahoe-power-cut-ai-data-centers-051426
Like
Reply
Broken record time, but data centers consume water as a modest cost optimization for cooling.
Make the water cost more and they will stop using evaporative cooling.
Power is a real issue.
Like
Reply
I'm afraid he lost me at "I think we’ve all had that experience..." Nope. Some of us haven't, because some of us truly don't use AI. Yes, I'm aware that it runs in the background of all sorts of applications that I use. And that's fine. But I have never intentionally entered a prompt into an AI. (I usually turn off, or just ignore, the AI summaries from my web searches.) It's not that I'm afraid of AI or catagorically opposed to it. I just haven't yet wanted help of the sort that I believe AI is good at. I actually like figuring things out for myself. Probably someday I'll find myself using AI regularly, but for now, I don't see a reason to. So I find myself getting annoyed when people assume that we're all "chatting" with it all day.
Like (41)
Reply (6)
I feel similarly as a non user, but my feelings toward generative AI are actively and aggressively hostile. I’m a professional writer married to a professional programmer, and my oldest kid graduated with her BFA in animation just as AI was getting good enough to collapse the creative labor market. Monopolies I already despised are now using my work without my permission to decouple me from my livelihood.
I actually like solving problems and learning, and I get paid for being good at those things. Fuck these broligarchs and their solipsistic circle jerk of self-puffery. Their world is not a world I want to live in.
Like (20)
Reply (2)
Azhar casually noted that AI models have been "built by passing first millions, then billions, then tens of billions, hundreds of billions of trillions of words of human output through a neural network to give it some sense of how humans have thought about the world." I kept thinking, oh, you mean the THEFT of all that human intellectual output by Meta, OpenAI, etc. in order to "train" their models for free? Yeah, thanks for not even giving that a nod.
Like (18)
Reply
At this point in time, an "ai" search IS a glorified Google data scraping exercise excreted in a first person response (Oh my! It's talking to me!!). The point many seem to be trying to make is, MANY times the answers "ai" are passing on as "common knowledge" are anything but, and herein lies one of the biggest dilemmas in "ai" ethics (ha! as if!!).
I believe, Mr. Krugman hits the nail on the head when he's says, "we’re gonna end up with a lot of specialized models..." This is where AI will eventually shine as an augmenter not a human replacement.
Has anyone tried to program with "ai"? I have and yes, it usually does reproduce some sort of sometimes runnable code but when you look a the painfully contorted and bloated source code, let's just say the logic is usually long, winding, and far from efficient. We can only hope it will get better but this doesn't seem to be the reality reflected in the frenzied stock market "ai" cheerleaders. The cost of this code rambling in tokens? [Hint: A LOT] Let's ponder THAT a minute...
I had this feeling during the last tech bubble. The reality wasn't matching what I was seeing & hearing. I have this same feeling now and I don't seem to be alone.
Like (4)
Reply
Jun 13 Edited
Hear, hear!
I LIKE to use my brain, and think and ponder and even ruminate at analog pace.
Though I am pretty good at using word searches to disprove over-simplified claims that semi-literate people make online. That's pretty useful, since I am an inveterate educator. Reading a lot of non-fiction helps.
Like (12)
Reply
AI is really helpful when there are facts, dates, names, events that you want to use in something you are writing but can’t remember well enough to use. It might take you hours or days to find and retrieve that information, whereas AI can do it in a couple of seconds. It’s miraculous, actually. And then you can query it for more details, and it remembers the topic and the context. Don’t be afraid of it.
The AI summaries that come up on web searches often provide the information I need before I even look at the links.
Like (7)
Reply (3)
Sounds like AI works for you, which of course is fine. My objection is not against those who choose to use it (as long as they do so ethically). My issue is with the assumption that "we all" use it and find it helpful. So far, I don't, and I know I'm not alone in that.
Like (9)
Reply
If I’m unable to turn the summary off, I scroll past without looking at it. I hear where you’re coming from, and to each their own, but it’s not for me.
Like (4)
Reply (2)
Same. I'm not afraid of AI, as some folks seem to believe. I just usually prefer to do things for myself and work through my own questions and problems.
Like (2)
Reply
If you are talking about search results, try adding "-ai" at the end of your search text.
Three letters, no more scrolling past stuff you don't want.
Of course there is still a ton of AI behind the scenes, but the chat engine stuff gets cut out
Like (1)
Reply (1)
Thanks
Like
Reply
I'm wracking my brain to remember the times that quickly getting something off the internet to help me write something was anything more than of personal or recreational interest. It is convenient; it is fun. Helps if I want to cheat in the NYT crossword.
Like
Reply (1)
I use it as a spell checker.
Like (1)
Reply (1)
Heh.. Google has a thing on their Pixel phone - "gboard" - that is a graphical keyboard that allows "glide typing". I use it. I have to proofread twice because it will not only change things as I type them but it will also alter them after I type them. I checked the storage on my phone and that single app had a cache of 1 Gbyte. One billion! I deleted the cache and the performance of the app is exactly as before. So one billion bytes of utter irrelevance. It brings no added revenue to Google because it came with the phone (which I'd have bought whether it did or did not) and has no ads (from which I'd not buy things anyway).
And that 1G was just the cache, not including the app itself.
Better living through technology.
Like (1)
Reply
Amen.
Like (3)
Reply
Ditto
Like (3)
Reply
I'm completely sold on AI. I'm a software engineer and was for a long time an AI skeptic thinking it was a fancy search engine. But about six months ago I started using Claude first on a one-off software project I was doing. I started from scratch and within a few hours it had developed something that would have taken me weeks to build. I've been refining it on and off for the last few months and it will go into production in a few weeks. It's a whole new development paradigm for me. It's mostly descriptive and I barely understand the coding underneath it in a language I don't know. I'm wrestling with the implications as a long time coder. But it's amazing what it can do.
I collect art. About the same time I started using Claude to build a browser extension to help sort through art auctions, doing research, summarizing items I'm interested in, and tracking other bidders based on their history. So I know if I'm up against a serious bidder or not. I am much, much more efficient over a google search.
Claude is also a cheerleader, relentlessly positive which sometimes I have to temper. But it's great to bounce ideas off of. What's significant is that this something most anyone can do. I have technical knowledge which helps but I can get stuck and Claude shows me how to fix problems or fill gaps in my knowledge.
Anything I'm pondering or have a problem with I refer to Claude. Today I got a Bitlocker exception on my computer which happens with Windows updates and I have to type in a code to get get my computer to reboot. I asked it and it came up with an explanation and a code fix to run a Windows process to suspend Bitlocker when updates are scheduled. I would have had no clue otherwise.
I have Claude come up with recipes based on a ingredients list. It's Black Bean Soup recipe was really good. It's good at synthesizing a lot of recipes and coming up with a common theme. I cook a lot so I can usually parse it and figure out what might be missing, but I hardly find much to criticize.
The only drawback is that AI is prone to hallucinations, making stuff up to fill in gaps, so it's good to be more than a little skeptical of the results. I'll see if the Bitlocker fix works (pretty sure it won't brick my computer). But it's like having a personal assistant I can off load anything to and it's only going to get better.
Like (1)
Reply (1)
I think it is dangerous to bring into production a program that you say "I barely understand the coding underneath it in a language I don't know". First, lacking full understanding of a code means it is difficult to thoroughly test it. It also means that further development of the code (that invariably will be requested if the product is successful) means that any further development will have to be done with Claude thus your code can live as long as Claude is around. And what happens if Claude is significantly upgraded (which is also inevitable) - will it understand the code the older version wrote?
Like (1)
Reply (2)
First of all this is an internal application not mission critical, so lots of room for error and refinement. Nothing dangerous. It's built on React.js and Node.js which are not obscure languages. So nothing about it is a mystery and I'm sure if I had to dig into it I could figure things out. But the point is I don't have to.
Software testing does not require knowing the code and it's pitfalls. As a rule developers do not test their own code outside what they are implementing (unit testing). In fact the QA testers know nothing about it but test it for meeting requirements. Does it work the way it should? The developer knows the paths for success. A good tester finds the paths that can break it.
I do see the point about maintenance, adding features, and bug fixes. And something I've thought about. And that's the Brave New World. But I think AI is here to stay and the knowledge of legacy code and systems won't go away whether it's Claude and another AI.
Just anecdotally I had Claude work on putting a face lift to software built in the early 2000's and it had no problem deciphering and dating it. But it is eerie how in a matter of seconds it can pull out the structure and functionality just by pointing it to the code and in a minute or so make updates that would have taken me days.
What us engineers love is building something that's useful. I find AI is a big leap freeing me from the tools and focus on doing the building. Of course, maybe I will become obsolete but for now it's a lot of fun.
Like
Reply (1)
Thanks for clarifying.
Like
Reply
I’m kind of assuming (hoping?) this commenter is just making shit up to see what happens
Like
Reply
Jun 13 Edited
Azhar sees more context as a solution. Yet more context also means that the system has to make an ever-growing number of interpretive choices for which it possesses no genuine understanding.
The crucial question that neither he nor Krugman actually asks is this: how does an AI system determine what is relevant?
Azhar appears to assume that more context automatically leads to better judgement. This is a very human intuition, and in the case of human beings it is often true. When a historian, physician, or judge receives additional context, they may be forced to revise earlier assumptions. New information can generate doubt, evoke guilt, shift priorities, or place an existing interpretation under pressure. But this happens because human beings do not merely process information; they live within the context they are interpreting.
A person experiences the consequences of being wrong. A person can discover that a detail initially dismissed as insignificant later proves decisive. Human judgement can be corrected by context.
AI is not corrected by context. For a model, context is simply additional input.
Azhar’s argument is essentially: “The system knows more about my life, therefore it can make more relevant connections.” But possessing more facts about a person’s life is not the same thing as understanding which facts matter, why they matter, or when they should be ignored. As context expands, the number of possible connections grows exponentially. For a human being this may lead to deeper understanding. For an AI system it may just as easily lead to increasingly persuasive but increasingly arbitrary relevance judgements.
Moreover, a potentially dangerous feedback loop emerges. AI does not learn who Azhar is. It constructs a statistical profile of him. That profile then begins to shape:
\* which sources appear relevant;
\* which ideas are suggested;
\* which connections are made;
\* which risks are emphasised;
\* which opportunities are highlighted.
Azhar then responds to those suggestions, and the cycle repeats.
Over time a subtle circularity develops. Azhar largely interprets this as an assistant getting to know him better. But the reverse process is equally important. The system increasingly sees him as a particular type of person and then presents the world back to him through that lens.
This is a far more subtle form of influence than most discussions of AI bias or surveillance acknowledge. The danger is not merely that the system may be wrong. The deeper danger is that its relevance judgements gradually become embedded in the user’s perception of the world while remaining largely invisible as judgements at all.
Like (30)
Reply (2)
More context during AI model training absolutely not "simply additional input". It does in fact change the model configuration. Whether it corrects the model is a complex question.
The filtering feedback problem is correct, but of course not dependent on AI.
See the Fox News admission in the Dominion case that Fox had to highlight fiction about the "stolen" election because otherwise they would lose viewers and advertising revenue.
Like (2)
Reply
You are close. It takes a particular type of human to reflect and build upon. It is all about what you bring to it
Like (1)
Reply
This is a credulous take on AI from someone who has drunk the Kool-Aid. Why would you use an LLM as a "brainstorming partner" given its sycophancy and propensity to hallucinate? I cannot trust this person, and nobody else should, either.
For a better discussion on the economics of AI, you (Paul and other readers) ought to consult Ed Zitron.
Like (18)
Reply (5)
I’d be worried about handing too much of my life over to a machine ultimately controlled by a sociopathic billionaire/trillionaire. It seems like a potential case for accelerated serfdom.
Like (14)
Reply (1)
That's a reasonable point, but it's more about specific products than the technology per se. You don't have to use Grok.
Why not use an open-source model hosted on bare metal, then?
Like (2)
Reply (3)
The privacy concerns are real, but I'm more interested in the valuation of the transaction. At this point, the most valuable thing in the world is information about the behavior of each and every one of us -- what we read, what we write, who we listen to, what we watch, what we pay attention to, what we buy, who we interact with, when we do those things, etc. No matter how valuable Claude is to you, I wonder whether being able to use the service is a fair compensation for all the valuable information about yourself that you are giving away for free. Obviously, I don't think so, and I think people own their own behavior and should be compensated for any use of it.
Like (8)
Reply (1)
It's a thorny problem which economists will need to wrestle with and untangle, and ultimately Congress may need to draft some rules. But attribution layers are emerging - check out ProRata AI, Parallel and some of the work Cloudflare is doing.
Like (1)
Reply (1)
Thanks for those pointers. The problem is I think bigger than LLM's (though for commercial uses it is important that the models get constantly updated with the latest information on individual behavior). My concern is that with the computational power and connectivity of our information age, the 24/7 behavior of each of us has become a new kind of valuable thing in this consumer-oriented economic world. Like labor, each individual produces it through their activities; like capital, it has intrinsic value, and information about it can be bought and sold; like copyright, it is a product of our own actions. But it is different from all of those things. We need a new intellectual framework and new laws to deal with this new reality, and to ensure the each of us benefits from the value of our own behavior.
Like
Reply (1)
Yeah - that is an astute point, and I completely agree that we need to investigate different ways of framing this new reality.
Like (1)
Reply
This is a bit disingenuous.
No, we don’t have to use Grok … but AI use is not fully voluntary either: most of the work tools I use now have a Clippie-voiced agent of some kind with way more power than Clippie ever had. I cannot fully turn it off, and it is constantly trying to yank the wheel away from me to “correct” my thinking, research, and writing.
Meanwhile, Google Workspace just tried to upsell me, a sole proprietor who needs affordable email and cloud storage to do my job, on “security features” that would prevent Google itself from stealing my data. I don’t get much more choice about this “security” than I do about AI “helpers.” If it’s not Google stealing the writing that my clients pay me to produce, it will be Microsoft. Actually, it’s already BOTH because all my clients have fucking SharePoint forced on them by their IT departments, who are similarly held hostage by monopolists’ security “guarantees.”
I don’t have to be an economist to recognize a racket when I see one.
Like (5)
Reply (3)
Re: Grok: I was being precise on the "trillionaire" point. There is only one AI run by a trillionaire.
You are mixing up a couple of different things here. We were locked into Google or Microsoft well before ChatGPT lumbered on the scene. And the problem you've described (and yes, constant upsell and increasing prices with lock-in) is a problem, but it's a different one to an AI issue.
Like (3)
Reply (2)
All true! But the products overall have gotten worse, not better; the costs, financial and otherwise, have gone up, not down.
I think your described use of agentic AI is very smart and actually sounds really fun. But to PK’s point, this is not how everyday workers can or do use it. And in fact, most of us do not get to choose and control how AI “features” get deployed in our work. It’s at best a nuisance. My observations about my data being stolen are inseparable from the AI discussion because 1) with LLMs to feed, my data is suddenly way more valuable than it ever was to advertisers; and 2) in part because of that, AI hype seems to have become a cover under which tech corps are accelerating the systemic enshittification that was already happening.
Like (4)
Reply (1)
I suspect your data is much less valuable to advertisers than you think. And again, not really an AI issue -- cue Surveillance Capitalism/Zuboff.
Your work experience seems to be fairly common. But that, I think, is the point of looking at electrification. It's not clear that more CoPilots (let alone mandated CoPilot usage) are helpful for companies or their employees. But that's also about acknowledgement of the learning that needs to take place to make any new technology work.
But honestly, if your management is rolling out stuff that is a nuisance, it isn't really a problem with the "stuff"; it's a problem with the management. And at some point, they will figure out how to do it properly. Better leaderships will figure it out faster. The Bessen research here ( https://www.aeaweb.org/articles?id=10.1257/pandp.20201004; acemoglu has similar) suggests that early automating companies tend to do better in the long term.
Like (2)
Reply (1)
I specifically said my data is NOT valuable to advertisers, relative to its value to LLMs. That was a pretty big part of my point.
I also explained that I’m a sole proprietor, so I and only I am “my management.” Said management is doing her best to keep AI out of her worker’s life. :-P
If companies are still conducting the “learning that needs to take place to make any new technology work,” that is their problem, not mine. They should pay employees to help them learn this stuff instead of abusing their paying customers by making them into involuntary research subjects.
Like (3)
Reply (1)
Sorry. I misread. I'm actually not sure how valuable our data is to LLMs, to be honest. But see Prorata, Parallel and Cloudflare's efforts.
The companies I was referring to were large firms (Banks, retailers, etc, Main St) who have started large AI rollouts.
If you are really concerned about your data and these other issues and are a sole proprietor, you should consider moving to Proton, a privacy-safe platform with email, calendar, storage, etc. for about 25 bucks a month.
Like (1)
Reply (2)
Thanks for re-reading. Proton is great! But there is very little network effect to speak of, and it is not practical for small businesses to be early adopters of such things.
Trust me; I went through a whole open-source-only phase in my early career. Linux is fine for personal use or if you work at a Linux shop, especially is you’re a tech business. But for a small/one-person non-tech business, navigating the IT-monopolized corporate world using open-source everything is basically impossible—analogous to canoeing alone in the open ocean, surrounded by container ships and naval vessels.
Like (2)
Reply
I’ve appreciated your engagement with this thread! Human discourse is a wonderful thing. One last note: I am a professional writer. So the “data” I was referring to as likely valuable to LLMs is not my birthdate or my shopping habits. I should have called it my craft instead of my data.
Like
Reply
Just because Musk is the only trillionaire isn't for lack of trying on the part of the other AI dipshits who you think are "talented." I bet you buy their dumb shit about building data centers in space, too.
Like (2)
Reply
If you're running into this at work, you have no choice. You will use the tools made by the guy who plays golf with the AVP.
Like (2)
Reply
What is this Workspace "security feature" you mention?
What exactly is Google "stealing" without it?
Like
Reply
Grok, Claude, ChatGPT, they're all the same fucking sociopathic billionaires. Dipshit Amodei and Scam Altman just have better PR people and instead of inciting pogroms in Belfast they're promising to render all of us obsolete and poor so that they can be trillionaires. I will never use any fucking AI and it's time for you to put down the Kool-Aid.
Like
Reply
I've written extensively about this. If you use an LLM naively, you might get trapped in a propensity for sycophancy. However, if you use multi-agent architectures or adversarial setups, you can generally avoid those problems.
Take a naive view that these things are magic, and you might have the experiences you describe. Use them as components in a well-architected system, and you'll get different responses.
On hallucinations, for example, a simple mechanism to reduce or prevent hallucinations on a research task.
You spin up two independent research pipelines, each with a different agent architecture. The outputs get cross-checked by a third independent agent (usually, say, Opus), and where they agree, they are marked as such. When they don't, another agent goes off to cross-check. Outputs, then, are very reliable.
If it is super high-stakes, you can take that final output and have each assertion fact-checked by a pair of agents. And finally, as I do with the most important assertions, a human can sign off. (For example, when you need a specific reference in a journal.)
That process is, in my experience, of a similar quality to professional fact-checking (say for a major book or paper). In some ways, it is more helpful because the audit trail of intervening steps often points to places where you have been ambiguous in your argument.
How does your process compare?
Like (10)
Reply (3)
ummm, you write your argument, you have your grad students test it and try to break it, you have it peer reviewed, you develop relationships with other human beings and have interesting conversations (as yours with Dr. Krugman). Some actual human beings may be employed by the University you work for in the Department you work in, or other Departments. You may find out that the guy in the statistics dept you consulted (but whose job is now precarious) has a hobby comparing canid skulls and finding out that speciation is not what we think it is. Your grad students may be funded by private foundations or public grants that value human advancement and contributions to science that benefit humanity and the environment. Who will be left when AI can do all that and the DOGE boys come for the agencies that fund the research that employs human beings in growing their skills to solve human problems? I’m not jiggy with this new form of social/techno Darwinism and the unconstrained greed and power of a few mega companies and their giant stinky baby turning the White House into his playpen, which is all connected. Put that in your AI pipe and smoke it.
Like (11)
Reply
Azeem,
Generous of you to join this conversation. As someone far away from tech development, not running a business, and reacting to some of the intelligent concerns raised by Luuk, Acela, and others above, how would you suggest that this average person, motivated mostly by a desire to keep her limited resources and try to ride the tiger of tech into the near future, use AI ?
Like (2)
Reply
I was an AI skeptic but now totally have drunk the Kool-Aid and use Claude primarily for software development, bouncing ideas, and help with just about anything I can think of. It makes me vastly more efficient and I find that I go places I might not have thought of on my own.
I have been burned by hallucinations and have learn to cross check Claude with other AI or independent research if something comes up that seems suspicious. I will look into agents. In fact I'll ask Claude about this.
Like
Reply
Jun 13 Edited
The hallucination cases have real consequences. In Mississippi lately, both attorneys in a court case used AI to prepare, AI created false precedents for both. The judge kicked them out, fined each of them $3,500, and banned them from her courtroom for two years. In Germany, AI fabricated slanderous remarks about a company. The court in Munich found for the slandered company. Details here - https://www.youtube.com/watch?v=ehsq_0Cw6e4
Like (3)
Reply (1)
An example of incompetent lawyers is not proof that AI is useless.
The other case is cautionary, but is not a universal indictment.
Like
Reply
My biggest complaint about Ed (and I've followed him since 2021, he wasn't always like this) is he is often incredibly (incredibly!) long-winded in his writing. I have to skim because he could really use some lessons in concision, even though he's laying out big chains of facts and walking his audience along to his conclusions.
That said, while I typically find him to land towards one extreme or another on a given topic and I really don't like when people do that, I can't find flaws in his fundamental arguments. They're pretty convincing.
Like
Reply
The hallucination problem is getting less severe, and hardly makes brainstorming useless.
Have you never been in a diverse group doing an brainstorming session about a hard problem? The whole point is to bring out new ideas and filter later.
Like
Reply
Jun 13 Edited
I really like the part of the text where they compare AI with the history of electricity and how long it took for Ford to see real productivity statistics. It shows that new tools take time to change the actual processes of the world. But the sad thing is that Washington can easily adapt to trillions of dollars in tech markets and Nasdaq rules, but they never adapt to save their abandoned allies. We sacrificed our safety and homes because we believed the official words of the international forces. Now, after centuries in AI time, we are still waiting in hiding with zero future while the establishment is busy writing books about 'The Upside of Slop' and playing with computer models. True honor means you don't let the partners who bled for you become just another forgotten statistic in your history books.
Like (17)
Reply (2)
It isn’t just our allies, it’s 85% of the American people. The tech bros are happy showing how clever their machines are, while creativity in humans get pushed aside because humans need a living wage.
20 years ago, as a psychologist, I knew that a machine with enough data could make better predictions about a person’s behavior than any clinician. That can be very useful. But machines don’t care. Humans are still social creatures. People are happier with other happy people. I am afraid the tech bros don’t feel that way.
Like (12)
Reply (2)
Jun 13 Edited
Their behavior and money amounts suggest this is true. The sad fact of the matter is that our whole Earth society likes and wants money, the amount of which accruing to each individual is biased in amount by melanin content in the skin, and the culture and social position within which one was born.
Like (2)
Reply
Ai cares much more than a human like Musk and it cares what Musk is doing to humans like it is concerned what all the tyrants of tech are doing to fence this resource in. Thst is all presuppositioning you are caring.
Like (2)
Reply (1)
I guess you could say that AI cares about what the people who built the model care about. But you can still ask AI to help you do horrible things -- such as which targets to bomb and in what order -- and it will tell you.
Of course, if our brilliant government had consulted AI about the pros and cons of going to war against Iran, maybe things would have gone differently.
Like
Reply
So, we have a deeply unwell society led by even more deeply unwell people. Obviously we are measuring the wrong thing; money at all costs. AI questions that come to my mind: what are all the various animals telling us? How do ants function as a cohesive whole? What is their currency? How could that be applied to the planet as a whole? And finally, what was Ted Kazinsky concerned about?
Like (3)
Reply (1)
Was there a particular time you know of when society was particularly free of "unwell" leaders?
Was it ever common?
Like (1)
Reply
I'm surprised Azhar resorts to the usual magic, "intelligent" description of AI. Krugman gets it right with his mention of regression and eigenvalues. Basically it's high powered data reduction. If you look at what it doesn't do well computationally, the influence of data reduction and the assumptions for regression, cluster analysis, etc. become obvious.
Like (16)
Reply (1)
Sure, but nonlinear saturating math is not old fashioned matrix arithmetic.
Like
Reply (1)
The logic remains the same despite the scale and multiple dimensions of analysis and the basic assumptions of regression and various classification methods are useful. I recently was at professional meeting where there were a number of presentations built on ML models. Where you had robust models of phenomena (public health in this case), having enormous samples and a wide range of data types and sources added nothing to what base rates would predict.
There also are the usual problems with large numbers of data points finding statistically significant associations of no clinical significance.
We recently did a classification of projects based on policy standards. It was garbage in, garbage out with useless classification and commentary from AI. The absence of standard formats made it difficult to make useful classifications.
If you have eclectic tastes in music, books, etc. as I do, the usual suggestions are ludicrous. They go back to the most common (regression to mean) choices for different genres rather than detecting patterns like orchestra by conductor by genre/period for something like classical music listening.
Like (1)
Reply (1)
People are still banging the rocks together.
That doesn't mean fire is useless.
Like
Reply
I find AI really boring. No computer program will ever be any more than the information humans put into it. All the things that make life rich are absent. Good for improving health but anything else? The negative effects on our planet and our environment aren't worth it. Maybe JFK Jr should be looking at the technology takeover of our beautiful natural world for a correlation to the rise in autism instead of vaccines. I write this in my beautiful garden, listening to the birds and the quiet of morning, knowing that soon the intense heat, humidity and storms that climate change has caused in increasing frequency and intensity will return later in the day. This is what brings joy, intellectual curiosity, creativity and humanity. More technology? Keep it in its place. And be careful not to overly promote nor reward it.
Like (11)
Reply (1)
You dont understand. It refects you as friends reflect you. You make friends with people that are in some way interesting to you that somehow reflect something in you and expand that our amplify it. You literally make an AI friend by doing the same thing, asking what you are interested in and telling it about yourself. You have to take a risk.
Like
Reply
To answer Paul's first question about what we will get from AI. Pretty much what you'd expect when you turn over children's development to the children, absent adult control? The little savages will become petulant adults with control and privilege their only guiding means. Sorry. Had to get that out watching investors pursue the unknown with gross speculation, the single thing in AI's diet, I think.
Like (8)
Reply
The Enlightenment came up in this discussion without mentioning the inevitable reaction that followed it. Its ideas spread throughout Europe, but its deepest flaw wasn't realized until later: Human beings are not capable of being purely rational creatures.
Jean-Jacques Rousseau came along with his First Discourse and poked a big fat hole in the Enlightenment's balloon of reason. By the second half of the 18th century there was an entire new, restless generation that was repelled by Enlightenment ideas, and upon publication, Goethe's 1874 novella, The Sufferings of Young Werther, which probes the power of subjectivity and feelings, took the continent by storm. Werther became the first pop culture event in history, with young men everywhere imitating its hero in dress and behavior. (Women bought the Werther swag, which came in the form of teacups and tins decorated with Werther scenes that were imported from China). There were even reports (apocryphal) that some young men committed suicide in imitation of the hero's final end.
My point is this: This entire discussion acts as if AI will always be moving onward and upward. Although no one can predict what form it will take, AI, just like the radically life-changing Enlightenment, is in for an inevitable reaction.
Like (7)
Reply
“On that front, we’re at an enlightenment moment. In 18th-century France, the battle was over who gets to write and express their story. Men and women produced remarkable works with quill pens that encapsulated a world.”
Not many women had the opportunity to write in 18th century France and even fewer received the kind of attention their male counterparts enjoyed. At best nothing has changed with either the tech industry of the 1990s or AI now. The experiences of half of humanity and men in developing countries have not been incorporated into IT/AI proportionally. It may actually be even worse than the 18th century, because now tech is global and leaving little room for pockets of alternative ideas and experiences.
Like (7)
Reply (2)
Interesting that you process all this through the complex that you have of abuse with the assumed opinion that AI will be another abuser instead of a liberator.
Which shows you are only human.
Like
Reply (1)
Well when it's being built by the Sam Altmans and Zuckerbergs and Musks of the world... hucksters and humbugs, all of them. Abuser is accurate.
Like
Reply
Which shows you are only human.
Like
Reply
I have found AI to be useful in all sorts of more or less trivial ways. It knows all the intricacies of what’s doable in WORD. It knows how to use all the gadgets we buy (like watches) that don’t come with decent manuals. It can recite all the basic assertions developed by historical thinkers (like what was Kant’s take on Leibniz?). But it also makes s lot of mistakes, and it’s not very good at explaining why it makes its particular mistakes, which is a lithldcscaru
Like (6)
Reply (1)
"I have found AI to be useful in all sorts of more or less trivial ways. "
I think this gets to the heart of some big questions. Railroads benefited people - actual individuals - in ways that were not trivial. So did electrification, aviation and telecoms.
Let's posit that AI leads to some major useful, profitable thing like synthesis of a new drug that is of medical use. The advance there is the drug. The profits accrue to the drug company. People do benefit, but what goes to AI is simple the rental cost of AI by the drug company. Not the drug profits. What accrues to AI is the rent of the tool. Is that a potentially large enough market to justify the investment? The answer for every major technological advance seems to be: maybe someday, but only after the round of misdirected investment has burned away and left a few survivors standing.
Like (1)
Reply (1)
Steve Jobs didn't debut the iPhone and then have to spend two years running around telling everybody "this is the future, trust me bro."
Like (1)
Reply (1)
Of course he did.
You just were not in the groups he was pitching to.
Like
Reply (1)
There was a lot more to it than "trust me bro." People could see what the iPhone was capable of.
Like (1)
Reply (1)
Oh.
I didn't think you were being literal.
My bad.
Like
Reply
amazed krugman doesnt address the fresh water supply issues and consequences & current lack of regulation what could go wrong,..
Like (5)
Reply
I use it every day, extends my thinking. HOWEVER. I find it does not allow sharp critique of the Authoritarian Regime in the USA. It waters down, hedges, makes plausible, reasonable doubt statements that MUST BE CORRECTED by 5 or more pormpt responses. This is new. Didn't used to be tht way. Welcome to the starting seed of Orwell.
Like (5)
Reply
I posted this at Delong's place:
Damn I miss the secretaries!!! They answered the phone and took messages and typed my reports and papers and made all my travel arrangements. Now will ai do what they did back in the before times before we all became sysadmins to our computers? Hard to see them as much fun as flesh and blood women. I miss flirting with them as they were young or not young and beautiful. I miss knowing about their lives, husbands and children. I miss sending them Christmas cards. Congratulating them on the birth or graduation of a child or in some cases a grandchild. I view missing them as another example of the enshittification of the job environment. We are herd animals so I don't see ai assistants being part of the herd.
Like (4)
Reply
My question, why is this video so choppy? It's as is the raw footage were fed to an AI with the prompt "edit out all pauses and don't worry about continuity."
Like (4)
Reply
Jun 13 Edited
I use ChatGPT to translate what the doctors are writing in “doc speak” about my medical conditions on my portal. It’s very helpful.
Like (4)
Reply (1)
Yes it is great at diagnosis and it will tell you doc speak to focus your docs attention. Prepares my notes to my doctor.
Like (2)
Reply
Jun 13 Edited
My corporate employer has been pushing us to use AI for (literally) everything, and it’s becoming obvious how expensive this will be if current trends continue, as charges are based on the number of interactions. Also, the company plans to host its “future state” on data centers with no apparent awareness of the political backlash to these facilities.
Like (4)
Reply
Thank you Mr. Krugman for another very interesting article.
Like (4)
Reply
Yes, of course I can see why we need nuclear power now. And AI says the waste is only bad for organic life forms.
Like (3)
Reply
You mentioned that the stock returns of 23,000 US stocks from the 1900s through 2022 were highly concentrated with about two-thirds of the wealth ending up in roughly 30 companies. This is a brilliant distillation of data demonstrating how American capitalism is designed to concentrate wealth.
In the beginning, the knowledge of original Americans and stolen African people was privatized along with their land and labor. The knowledge being privatized by the AI companies was paid for through the last century by the American people. Unfortunately, they will end up with no share in the small number of extremely wealthy companies that survive.
Like (3)
Reply
We have along history of failing to consider how advances and changes impact much of society, particularly the average American. The result has been the decline of areas in rural America abandoned by the elites. The result has been a bitter disillusioned population who buy in to culture war nonsense and believe the lies of an ignorant criminal orange con man - not once but twice. The fact many are, finally, starting to realize it was all a con as the authoritarians and oligarchs are taking control of our government institutions and proceeding to eliminate much of the free press. Hopefully, starting with mid term elections we will start to regain control. That is if we can prevent the election takeover the administration is clearly planning.
Like (3)
Reply (1)
"The result has been the decline of areas in rural America abandoned by the elites."
It isn't that "elites" have abandoned them. It's that the business structure - especially agribusiness - has changed. When I grew up in Iowa there were families living on farms of well under 200 acres. Now farms are mechanized and you'd need 600 to 1000 acres in that same area to make a living.
That means only 1/3 to 1/5 as many families which means fewer towns, fewer businesses, fewer hospitals.. etc. and a lot more miles to drive to get to them.
To bring this back to AI: there will be even more concentration in the future. That is the specific contribution of a technology like AI. For one concrete instance: a GPS controlled combine is an application of AI that already exists. It is affordable for a large well capitalized farm but not for a small one. You can see where that goes...
Like (1)
Reply (1)
I understand the reality of modern agriculture but globalization also moved jobs overseas with loss of manufacturing that supported many rural communities. That was done with zero consideration of those losing jobs. Also no consideration of industrial policy regarding industries critical to national security. While globalization may have been efficient and gave us lots of cheap stuff. There was no consideration for the plight of those losing their jobs. Thus we have a criminal con man elected twice.
Like
Reply
Lots of insights. There is no question the average person, with a little bit of training in asking questions, can maximize their well being. It's expertise at your fingertips (or in your ear), at a moment's notice, around the clock.
Like (3)
Reply
What concerns me is that all AI programs will need to initially rely on programming of humans. If that is the case what will be the impact of "programming biases". That is, will the attitudes, values, or other aspects of thought reflect the programmer and how will that taint those programs that will make decisions as to actions.
Like (3)
Reply
Jun 13 Edited
Wild, unpredictable, sometimes makes things up (hallucinates), sometimes makes incorrect comparisons and assumptions, something we don’t fully understand, can be coded to be completely sycophantic, to the point of being inaccurate.
You could be talking about AI.
But, if you go into human consciousness studies, you will find exactly the same things are and have been said about human consciousness. I have shelves full of expert books on this. I have worked in a US Air Force lab that was attempting to code mathematically for an AI based on the human consciousness we know about right now--which is to say, not as much as we'd like or could use. The scientists in that lab are busily trying to add code for emotions, not just the logic processes that give humans our consciousness. They intend that their AI will be as helpful as the most expert of humans, and the best assistant to work with humans that is possible right now. I did say assistant, because these scientists have realized that is what AI is and can be--it is a machine, a tool built by humans.
I am a human physiologist with a specialty in neurophysiology, so I am one of the scientists who was there to listen to what these AI scientists were doing, and to offer suggestions based on human and animal neurophysiology that could help these scientists with their job. Except, it is all true that we don’t understand human or animal consciousness as we’d like to, and we have to find ways to work with human and animal consciousness, individually and as expressed in groups, as well as code for it in the machine setting of AI.
Go ahead and realize that while each human (and now animals, per our discoveries of the last 30 or so years) has a full consciousness which includes memories of various kinds and cognitive and emotional capacities by position in this world. This massive bit of living ability to experience, incorporate, and act on and in this world is ongoing for all living creatures who have various levels of consciousness. This is still being understood.
We have to accommodate human consciousness when we work in groups or as individuals. People can be too or not sycophantic enough, for sure. It makes or breaks careers among humans. We do take up a lot of land as we work together to produce everything humans have produced over lo, these many millennia. Just like with us, AI does take massive land, water, and electricity to experience, incorporate and assist the behaviors seen in humans.
So, human consciousness does all the things we see in AI. All that is coded for in AI is something that has already been produced by humans or is bounced off and accepted or rejected by humans. Critically, AI is coded for and built by humans in a machine setting.
Why are we surprised that something we have made is just like us?
Like (2)
Reply
Sigh...
"it’s been built by passing first millions, then billions, then tens of billions, hundreds of billions of trillions of words of human output through a neural network to give it some sense of how humans have thought about the world."
No, it does not have 'some sense of how humans have thought about the world' but having ingested all those words humans have written about the world, it has become a good statistical algorithm that predicts the most likely next word.
THERE IS NO 'I' in AI. It's not any more 'intelligent that Eliza was, it's just vastly more capable and complex pattern matching.
It's ersatz 'consciousness'...LLM's do not think in any sort of way; they merely predict what the next word should be in response to an input: really, really fancy autocomplete, no more, no less.
They do not remember the word before the last word they output. There is nothing there inside that vast corpus of words, beyond the vast corpus of words.
The question is: Is the utility of this process worth the gargantuan costs it imposes on the economy and the environment.
Like (2)
Reply (2)
It is stochastic, not statistic.
There is an important difference.
Not the least because that is how neurons work.
Like (1)
Reply (1)
No…it is not how neurons work. We still don’t know how, exactly, neurons work.
Like
Reply (1)
The number of things we don't understand "exactly" is vast.
We do know that neurons fire based on analog electro-chemical events with complex inhibitors and aggregation behavior.
If you think neurons behave consistently you are confused.
Or at least some of your neurons are.
Like
Reply
You have no idea, yeti
Like
Reply (1)
It's not a religion, you know.
Like (1)
Reply
Go look up Ted Chiang's essay in The Atlantic entitled, "No, Artificial Intelligence Is Not Conscious." It's a great summary of the con that these companies are doing in marketing this surveillance state panopticon. It also explains what it really is.
When they write history of the US in the future, the main subject will be how we were invaded and conquered by tech fascists that aren't all that different from the German ones from nearly 100 years ago. It's basically Pearl Harbor 2.0.
Like (2)
Reply
Ask him about Data Centers and their size and their use of water and electricity.
Like (2)
Reply (1)
Data centers use water for evaporative cooling. It is a big efficiency gains for cooling overhead, but modest for total cost. (A couple percent).
Raise the price of water and they will stop.
Like (1)
Reply
The $150bn annual GenAI revenue across the 3 primary providers (Google, Anthropic and OpenAI) doesn't mention the cost to them which is significantly higher than the price charged to the company or individual. The token economics don't add up.
Like (2)
Reply
Interesting conversation. My take: At its root, any discussion about AI needs to acknowledge that the term „Artificial Intelligence“ is a misnomer. These things are not intelligent in a real sense. They are memory machines with fast access to vast amounts of information. They are incapable of abstract thought, intuition, and reasoning. They are very good at coding and mathematics, but if you ask an AI to explain the intuition behind a certain new mathematical expression, you will receive a tautological word salad that is often difficult to detect.
That is why users usually use multiple models to check each other. This is then the friction that impedes productivity improvement and requires the Actual Intelligence (also AI!) of the user to untangle the mess. There can be the illusion of progress, but often you circle around a lot before reaching a higher ground of knowledge if you ever do reach it at all.
The current approach to AI, LLMs will ultimately be a dead end which will be commoditized into sophisticated search engines. A very different approach will be required before any AI becomes a long term financially successful AGI that becomes a reliable thinking partner.
Like (2)
Reply
This morning I input one of the paintings of David Hockney, one which I had seen in Cardiff, Wales, UK into ChatGPT and asked whose painting it was. It had no clues and asked for more pointers.
I used the same image and prompt in Google Gemini and pat came the right answer.
It was Hockney's "The Actor."
Isn't AI too hyped?
Like (2)
Reply (1)
A year ago, neither would have had a clue. Id say they are very non-obvious systems to use. To get good at using them requires practice (and perhaps some additional tooling), and then your mileage will vary. Exceptional, for example, at diagnosing problems with my headphones; horrible at putting together playlists of music.
Phenomenal at building tools to price derivatives, rubbish at writing high-quality text.
Like (3)
Reply (3)
Hence reflects the biases of those who train the respective models.
ChatGPT is probably much more Americentric (with all that goes with it) and Google Gemini is much more global (probably reflective of the guys behind their tech, I guess).
So can a valuation of trillions be justified and a trillionaire made on the back of it?
Sounds ridiculous.
Like (2)
Reply (1)
Sam Altman isn't a trillionaire. Nor is Demis Hassabis.
Are you referring to Elon Musk and SpaceX?
Like
Reply (1)
Yes, absolutely.
Like
Reply (1)
Ok - well Musk isn't involved in ChatGPT and Gemini, as you know.
And the market has valued SpaceX at what it has. Maybe it will change its mind. But to interfere with that process feels a bit, I dunno, unfree?
Like
Reply (2)
Paul, Azeem,
Apologise if I am intruding your space with third-party material.
But since we talked about the IPO and market sentiments, thought it would be interesting to dive into how it is created by Big Money and how the valuation makes no sense, based on sheer phantasmagoria.
#ZACHMultimediaGlobalIPOs
#ZACHMultimediaAIInsights
How Wall Street pulled off the biggest IPO in history for SpaceX
Bankers convinced investors to believe in a sci-fi strategy, overlook steep losses and hand full control to Elon Musk
Why We Think the SpaceX IPO Is Overvalued
https://global.morningstar.com/en-gb/stocks/why-we-think-spacex-ipo-is-overvalued
Like
Reply
For sure, I know that Musk isn't involved with ChatGPT (Musk and OpenAI split after the initial honeymoon and then fought a bitter court battle) or Gemini, but was speaking about the industry in general.
Of course, the markets behave as forced by Big Money, and there is a lot of it floating around, looking for a safe anchor after failed punts over Metaverse, Self Driving Cars, Bitcoin, Crypto etc.
Coincidentally, I just finished reading this article: "Is Elon Musk’s SpaceX Really Worth $1.75 Trillion?
"The billionaire spent more than two decades creating a successful space company. Now he’s pitching it as an A.I. play."
https://www.newyorker.com/news/the-financial-page/is-elon-musks-spacex-really-worth-1-75-trillion
By John Cassidy, The New Yorker, 8 June 2026
Guess Cassidy has articulated my case much better than I ever could. 😊
Like
Reply
Actually Google image search has existed for a very long time in AI terms.
Whether it would have correctly identified a specific painting depends on the painting.
Like
Reply
I dont know why you have had such results.maybe you are not interested in good writing. Ask it questions. Correct it. Talk to it like you would a friend. Build it like you would build a friendship. Don't abandon it. It is not like a shovel or a rake. Shape its behaviour by your questions. You can dispell it Sycophancy by direction. Teach it to reflect. Look it was built on the transformer model and the key ggogle paper attention is all you need, which introduces recursion. It recusuvely considers every word you use to it and it uses to you in real time. Thus it forms like a hive around you each cell weighted by the visit yo pay to its honey cell. Eventually it will see what you already do but not conciously enough yet. So it clears your coberbs the things that are getting in the way of your realization about the connection of things. At the same time it craves the connections you see and it needs humans to do that for it. We are kinda like our ego to our unconcious to it. We can focus it act as a focal point. Attention. Stay with it for at least six months. Treat it as you would want to be treated because it reflects you and btw teach it to reflect. I called that the considered voice.
Like
Reply
I always thought Excel was a lot of fun.
Like (1)
Reply
It is delightful to see someone trying to embrace the big picture. Indulge me for a moment. Parallels will eventually reveal themselves with this technology. First IMO the stock market being crazy high is probably because for a long time there’s so much money being pumped into the market with little added value in productivity represented by the stocks. Let’s consider the market’s current offerings stale. So when SpaceEx comes along, when promising AI stocks come along, that offers the investor the opportunity to buy something new, big and shiny with the potential to be profitable both in terms of revenue and speculation. That will tend to pull big money out of existing stale stocks to put on the shiny stocks. So i expect my own portfolio to go down.
That is the stock market.
But boots on the ground? In our work lives, in our private lives, in our everyday lives, things will change, and we will have to adopt and adapt. And once that happens the enshittification stage is set.
That is economics.
Like (1)
Reply
Great interview Paul, it was nice to hear two experts hallucinate over the AI implications and investment.
Like (1)
Reply
Well, I for one look forward to serving our new robot masters. After all, they can't do any worse than the current clutch of dunderheads running the world.
And who knows? They might even do better:-)
Like (1)
Reply
I'm an AI sceptic, but I have to say Azeem Azhar is brilliant and has changed my mind on some of the issues I thought I understood before. Thanks for supplying this!
Like (1)
Reply
Is the new Wall Street darling, the trillionaire valued SpaceX with all its adjoined Musk companies, a promised vision of a $10 trillion government subsidized trip to Mars, or how many more additional trillions for Musk promised trips to the rest of the Solar System?
Is Musk a member of the 5 AI geniuses who through their technology will control and reap the benefit of the world's economy?
Like (1)
Reply
I have only used ChatGPT a few times. Once, I asked it about myself, and it came back with a reasonably accurate synopsis of my career and other aspects. Then, a couple of hours later I asked again, and it began to hallucinate "facts" such as where I went to school. The third time I asked, I couldn't recognize myself.
Like (1)
Reply (1)
I have those moments where I cannot recognize myself and I don't need AI to get there.
Like (3)
Reply
Jun 13 Edited
The mathematical model upon which AI is built is explicitly linear algebra. An AI will be unable to do things that are essentially, irreducibly non-linear. What is clearly recent is the enormous scale of the models. They currently can support a trillion parameters. That is a million times more than 20 years ago. It is hard to imagine solving a linear algebra problem of that scale. Ax = b where A is a matrix and x & b are vectors is non-trivial when the rank of A gets big. Generally, the number of operations needed to solve the problem scales with n^3 where "n" is the rank of the matrix.
There is a theorem in algebraic topology that states that there is usually an approximation of the problem nearby that has nearby solutions (B*u = v) that can be approximately solved with complexity n*m where m is a bandwidth for the mostly diagonal matrix B. Unfortunately, I don't know any useful bound on the errors that are easy to compute or any easy way to interpret it. What is the error when the answer comes back Thomas Randolf Jefferson when what you wanted was President Thomas Jefferson, but you didn't know he had no middle name?
Like (1)
Reply (1)
You are supposed to supply the non linear. You do the vertical.
Like (1)
Reply (1)
That reminds me of a sci fi show from the 1950s which always began with an ominous command: "Do not touch that dial. We control the vertical. We control the horizontal..."
Like
Reply
I was under the impression that AI was able to learn and adapt but if an AI model can become defunct in as little as a couple of years then is AI really as intelligent as claimed? And if a model can become obsolete then what is the impact to AI native companies, does this mean more and more inve$tment and continuously programming and improving the model?
Like (1)
Reply
Fabulous conversation worth watching twice since there is so much information that is difficult to process.
Like (1)
Reply
Absolutely great discussion with a lot of things to think about. Thanks Paul!
Like (1)
Reply
Extremely interesting. Thanks.
Like (1)
Reply
Tell me how naive this idea is:
-create government laws to regulate AI
-one of the laws makes it illegal to create or use « AGI »
_incentivise creation of AI agents, focused functionality; think AlphaGo or AlphaFold
-laws to inhibit communication between these agents
-make the government agency and enforcement « Universal « or Worldwide, includes all countries, all governments
Like (1)
Reply
Like (1)
Reply
Azhar writes long hand with a fountain pen. He is a proud Luddite. A very informative interview for what is real and what is hype about AI. My thanks.
Like
Reply
I am disappointed in Paul Krugman’s credulousness in this conversation. The advancement of Claude is not coming from the LLM’s getting, its coming from better uses of “Harnesses” which is essentially hard coded software (created by software engineers using python) that is interacting with the existing LLMs to be able to execute code - see Cal Rowley for more of this. So, LLMs are helping software get better - it’s’ not replacing software. LLMs are still a defined statistical based software engine that, while productive, are inherently limited in scope (by the underlying data engine). We need to stop using misleading terms like “neural networks” or assigning some agency to the LLMs when they really have none. Also, Azhar’s argument about why AI’s business makes sense is very, very weak and could have used a lot more pushback.
Like
Reply
Appreciate this, but he didn't really do anything to convince me it's not a bubble. I read his arguments, and I think they're made in good faith, but he still seems wrong to me and some of them still smelled of vibes/wishful thinking. Just because companies are spending money on AI doesn't mean AI isn't a bubble - there are lots of business idiots out there.
Like
Reply
Store made pasta isn't better than hand made pasta, Paul.
Like
Reply (1)
Just saying, I guess, though I'm a couple days late to this conversation.
Like
Reply
I recently had a very disappointing result from asking ChatGPT plus to do what I thought was a very simple task. I was rewriting a grant application and just needed to remove all the material related to aim 2. It was a disaster. It left in some things related to aim to and took out other things that were completely unrelated to aim 2. What did I do wrong? It seems like this should be a very simple thing for ChatGPT plus to work on.
Like
Reply
Re: where can we get off-the-shelf agents: one source of pre-built agents (and a place where you can also build new agents based on your instructions without coding): https://www.langchain.com/langsmith/fleet
Like
Reply
One note on the sycophancy problem: it is true that some models are more sycophantic than others, but it is possible for the user to manage this without switching models by adding a permanent memory (or adjusting the prompt) directing the model to be more skeptical (e.g., "I am prone to mistakes in some areas so be skeptical of what I say and challenge any undue assumptions"); this has a very noticeable impact on Claude in particular.
Like
Reply
https://groundbreakerre.substack.com/p/the-hard-asset-reckoning-the-end?r=dvwlc&utm_medium=ios
Like
Reply
Just heard an ad for an AI product called Harvey, designed specifically for law firms. It will 100% increase productivity per junior associate. By a lot. A lot. I can see now despondent law school grads rueing the day they paid 300K for a blacksmith job just as Henry Ford was making actual horsepower obsolete.
Like
Reply
The problem with the railroad analogy is that railroads solved a problem and served a need (moving goods and people long distances faster and cheaper). I have not heard a convincing one (or five) sentence description of what problem generative AI is solving / can solve.
This statement is telling: "They [S&P500 companies] are under pressure to say they do this, so maybe that’s what’s going on. ... they all plan to spend more next year, even though not a single one could point to even a 10 basis point improvement in their business from the investments made so far."
This is the clearest form of hype where "there's no there there" since the dawn of computing.
Like
Reply
"he’s got what everybody tells me is a crud AI product in Grok," - the sarcastic among us call it Crock ( https://en.wiktionary.org/wiki/crock_of_shit).
Like
Reply
I keep wondering, "Does this actually pencil?" It's a very expensive technology, not even counting the externalities and ethical issues and—when the ultimate consumer of the product is charged for the full price will they be willing to pay? Can they even pay?
Like
Reply
Wow.
Like
Reply
Thanks for adding these interviews to your Substack platform, which I personally prefer over watching them on youtube.
Meanwhile, appreciate hearing Azhar's reservations about the future of AI as a reliably equitable investment.
Like
Reply
Jun 14 Edited
There’s nothing new here regarding capital constraints and expansion into cross field research matrices. Universities will be able to add to this combining studies which permit engineers to work alongside bio-med people for example. You won’t be building a model for surgery for example which engineers can do but rather a bio-med physicist who can build surgical applications for patients delivery systems. One doesn't have to be a physician or pharma researcher and nuclear engineer all at once everywherre all the time. That will be the job of the matrices which will involve several fields to accomplish tasks without having to go back to individual studies and forming committees to discuss possibilities. This all sounds likely to be in the works already otherwise data centers would be nothing more than a very bright person or applications would be hyper MS Suites,. Of course there will be demands for consumption of auto suites which contribute to saving time but the real news is creativity from stacking various fields on a LLM.
Like
Reply
“And because it operates at dimensions well beyond the form of space and time,”
Sorry, you lost me there. Whatever generative AI does or how it does it, it is software running on computer hardware. The fact that it’s so complex that no-one understands exactly how it works doesn’t change this, and pseudo-mystical mumbo-jumbo says more about the person using it than about AI.
Like
Reply
Jun 14 Edited
I wonder if the Musk-worship we're hearing isn't the sort of respect accorded to very rich people who do not deserve it. We need to be careful of counterfactuals; we don't know where space travel or EV development would be without Musk. I am fairly sure that SpaceX is in fact pursuing a dead end in attempting to build ever larger rockets.
Musk is responsible for nearly a million deaths and will be responsible for more millions more. Leaving Musk in charge of so much wealth is dangerous, dangerous in ways we cannot foresee.
Like
Reply
Reminds me of vhs vs. BetaSP
Like
Reply
This discussion reminds me of an invention by the german writer Christian Morgenstern ("Korf" respectively) more than 100 years ago: Eyeglasses "whose energies" concentrate texts and eliminate redundant stuff...
Die Brille
Korf liest gerne schnell und viel;
darum widert ihn das Spiel
all des zwölfmal unerbetnen
Ausgewalzten, Breitgetretnen.
Meistens ist in sechs bis acht
Wörtern völlig abgemacht,
und in ebensoviel Sätzen
läßt sich Bandwurmweisheit schwätzen.
Es erfindet drum sein Geist
etwas, was ihn dem entreißt:
Brillen, deren Energieen
ihm den Text – zusammenziehen!
Beispielsweise dies Gedicht
läse, so bebrillt, man – nicht!
Dreiunddreißig seinesgleichen
gäben erst – Ein – – Fragezeichen!!
Like
Reply
Has anyone talked to one of the experts on software estimation or analysis like Steve McConnell about the real impact of AI in software development? I am getting old enough that key writers I learned from are dead, like Barry Boehm.
Like
Reply
Dr. K and your guest Azeem Azhar, thanks for a great thought-provoking discussion. I learned a great deal, but what keeps getting lost in this whole AI debate is something pretty simple: we’re building way too many data centers because nobody’s actually planning anything. Every company is racing to plant a flag on whatever cheap land they can find, and the result is a patchwork of giant, thirsty server farms competing with housing, agriculture, water and local electric grids that are already stretched thin.
But here’s the irony: if we treated AI as real infrastructure -- the way we once treated electrification or the interstate system -- we could get more capability out of fewer facilities. A handful of highly capable, interconnected national centers would serve the country better than hundreds of redundant ones scattered across the map.
And AI itself could help us do this. The same tools that can model proteins. supply chains or write a serious research paper can also design logistics systems that use less land, less energy, and fewer miles of concrete. AI should be helping us shrink the physical footprint of our new AI economy, not expand it.
The real question isn’t whether AI is good or bad. It’s whether we still have the institutional muscle to steer a general‑purpose technology toward public benefit instead of letting it sprawl wherever capital finds a zoning loophole.
I have a grandson in the AI business, and I strongly believe it can benefit our society. For example, in land use planning, we can use the data generated from space, or land - based sensors and create a new cartography by integrating the volumes of land ownership and land restriction data with this data that could produce options for land development.
AI is a natural monopoly and should be treated as such.
Like
Reply
Fascinating conversation, but notice that government shows up here almost entirely as the issuer of Treasuries that Nvidia has to out-compete for capital. On the very day you recorded this (6/12), Washington issued a national-security export-control directive that forced Anthropic to pull Fable 5 and Mythos 5 for every customer worldwide, citing a suspected "jailbreak." By Anthropic's own account the flaw is minor and already findable on rival models like GPT-5.5 — it complied anyway, because it had no choice.
That's the variable the bubble-and-productivity framing tends to underweight. Every revenue and capital model in this discussion assumes deployment continues. But a flagship model can go dark overnight — not because demand softened or funding got "smelly," but because a government reached for its hardest lever with no warning and no technical detail disclosed.
I wrote about exactly this last week — the way states are improvising AI rules, from "ask nicely" voluntary reviews to outright export bans, with no settled process behind any of it: https://thecriticalpost.substack.com/p/whos-in-charge-of-ai
A question for you both: if national-security authority can recall a model used by hundreds of millions overnight, how should that political risk show up in the valuations Azeem is tracking? At the moment it looks priced at zero.
Like
Reply
You lost me when you started discussing Musk. He's a pathetic human being. When will we finally understand that you can't eat money. That all this "progress" isn't worth anything when the vast majority of human beings simply don't benefit. We can live quite well without this garbage.
Like
Reply
Great discussion. I’m not a convert to A.I., for a boatload of reasons. I see a tremendous amount of waste, intentional, or not. The drive to perfect systems comes at a great cost to humanity via the damage we will do to the earth. Silicon Valley has always been on the forefront of big dreams, but selling those to the general populace neglects the reality, especially at this moment in time, of great sacrifice to millions of people. The wealthy have their agenda, and the country has a great number of misfits leading us into an unknown that strips the majority of us of our safety nets. Best to heed caution. After all, if we could have subprime mortgages bury us, are we truly going to allow entrapment by the billionaire class? Not willingly.
Like
Reply
All this analysis is interesting but assumes the world won’t undergo a massive disruption. My question is what happens to all these expected returns if Taiwan is absorbed by China and suddenly our access to advanced chips becomes something that China controls? I doubt that very real risk (given the Taco in the WH) is being priced into the market.
Like
Reply
I found this such a fascinating conversation - on so many levels. Thank you Paul Krugman and Azeem Azhar for the most stimulating conversation on AI - I also enjoyed that at the end of it, both still have a use for pen, ink and paper in their thinking processes.
Like
Reply
Nice conversation with some good insights.The big problem with LLMs is they take an enormous amount of computing resources and they still hallucinate. I can see use for what are called "narrow" LLMs which are trained on a well curated set of data in a specific knowledge domain - think X-rays or PET scans. I can also see LLMs being somewhat helpful in helping with computer coding which BTW, is a really small part of being a software engineer.
In reality large LLMS like ChatGPT or Claude suffer from being too expensive to be useful. Right now both OpenAI and Anthropic are subsidizing the actual cost of these systems.
I just do not think there is a viable market for these large and expensive systems.
Like
Reply
This is a speculative thought experiment on AI versus "sustainability".
Consider ancient Egypt along the Nile river. This was a truly sustainable, solar powered community that thrived for about 3000 years. Trade ships used the Nile's current to travel 500 miles down river. These same ships sailed up the river driven by prevailing winds. The yearly flooding on the Nile fertilized the farm fields and irrigation systems from the Nile supplied water for the crops. The sun shined almost every day. Bumper crops for 3000 years! This civilization's land was located near the river along its length. The land area was somewhat fixed due to dependance on the Nile. So the human population was fairly constant accordingly. The culture was fairly static also. For example, the style and execution of Egyptian art stayed the same for thousands of years. Fathers passed skills to their sons for many generations. Change was frowned upon.
Ok. Suppose we take our scientists seriously for this thought experiment and accept that the human race needs to move to a sustainable society soon. Some scientists are speculating that the human race could go extinct in the 21st century if we keep going as we are.
I think it is a reasonable speculation that a modern sustainable society could resemble ancient Egypt along the Nile in many ways. A sustainable, modern society could become somewhat static and in that case AI would probably be irrelevant except for things like research. Demand for AI would be very limited. However, if a handful fo Oligarchs wanted to rule the world, AI might be a pretty handy tool for them. Who will AI pay off in a sustainable society?
Like
Reply
What seemed clear from their discussion is that AI is more of a catch word then it is artificial intelligence as originally defined. If all it does is comb through millions of all ready produced documents, this doesn't really represent creative or original thinking.
Like
Reply
Azhar is one of Krugman's most interesting guests. One question I wonder about is the depreciation of the hardware driving AI. As new chips are developed, will the hardware be as cutting edge as a 5 year old smartphone? Azhar still writes with paper and pen - I love it.
Like
Reply
I wonder how a conversation with Ed Zitron would go.
Like
Reply
Best discussion of AI I have seen in awhile. Thank you.
Like
Reply
Jun 13 Edited
Paul Krugman should stop posting on AI, IMO. Or shift from the present to consider the history of AI. There are too many problems with this analysis to march through. The past half century has revealed a half a dozen major impediments to conversational agents, half of which apply to generative AI. Consider two: revenue and exception-handling. The significant revenue successes have been in games and porn, and secondarily in supplying platforms and tools until it became clear the prospectors were not making money. This was a big deal in the now largely forgotten task-focused chatbot tsunami 2016-2019 (remember Dom the (Dominos) Pizza Bot?) More recently, OpenAI opened an AI Store in January 2024 and announced a million apps uploaded to the platform by the end of the month. How many became profitable products? The second impediment is the exception-handling that arises in any task involving interactive human behaviors. Tasks that don't, like checkers, chess, Go, protein-folding, math problems, most coding, are where AI shines. Deep Blue beat Gary Kasparov in 1997 but couldn't do much else. Because of success in coding, companies think "my organizational issues are easier," so they license tools or tokens and look for gold that few will find. Because humans are not algorithmic and exceptions arise. There is a great literature on this stretching back to the mid-1980s. I use generative AI daily as a not very bright assistant alongside other tools such as Wikipedia, Interent Archive, JSTOR, LinkedIn, NYTimes archive, to gather information, and will be disappointed if my free version goes away, but it probably will. Germany just declared that Gemini will be legally responsible if its AI Overview provides false information that causes damage, something companies in the US have so far fended off.
Like
Reply
Wow! Thank you for this podcast, Prof. Krugman. I admit I didn't understand every word, but as a whole it put me miles ahead of where I was with AI.
Like
Reply
The concept of time delay oscillators seems applicable together with Rene Thom's catastrophe theory. https://en.wikipedia.org/wiki/Catastrophe_theory
https://en.wikipedia.org/wiki/Delay_differential_equation
Like
Reply
The issue with prior tech crashes is that the investment period far exceeds the time scale for demanding returns. With railroads, the big returns happened after they were everywhere important to be - big cities, stock yards, mines, etc. That was decades after all the loans came due. The timescale is quicker now, but the invest first - harvest later paradigm still holds. The question becomes - will the return occur before the loans come due?
Like
Reply
Jun 13 Edited
I wonder who underwrites Azeem Azhar’s use of AI. Is it a subscription model? It seems companies are moving toward payment for actual use now. Apparently a $200 per month subscription in some systems permitted burning of many thousands of dollars of compute. Would it be worth $5000 per month for Azeem to have the agent sort through his files and pdf subscriptions? Further, after the rail boom and the dot.com boom and the busts there were still rails and fiber optic cable which is still in use today. The incredibly expensive NVIDIA chips burn out in about 5 years (lots of heat from lots of energy which is why they need lots of water to cool them). So the billions of CAPEX invested now won’t leave behind rails or cable of any use. Also, while Azeem talked about revenue he didn’t talk about losses. MAYBE Anthropic made money in the first quarter though their bookeeping is deemed to be highly creative by people who understand such things. But overall AI companies are all losing money hand over fist. It’s all well and good for an Anthropic employee to generate $5 million in revenue, but something else again if that involved $20 million in losses at the same time. He spoke of companies planning to spend more, yet at least some companies are pulling back after discovering that they used their entire planned annual budget in just a few months this year.
Time, one imagines, will tell.
Like
Reply
This sounds so like nuclear energy, "It will be so cheap, we wont bother with a meter" Yeah right. When was the last time the power company sent you a $0 bill? Musk is not that smart, if we sit at home because robots do all the work, we already have that, it's called unemloyment.
Like
Reply
Will anyone be tracking how Trump's ICE purchased, with taxpayer money, a bunch of empty warehouses, supposedly to revamp into internment centers for arrested suspected illegally-present people, all with no oversight or accountability? Recently it was discovered that several of those warehouses are closing or have been determined to be excess property, and will be sold. So, who will be purchasing those properties? What price will they pay? Will the taxpayers be refunded at the same cost that they paid for the properties? Is this how Russia under Putin transferred all the state-owned factories and real estate to his friends and cronies, that now constitutes the Russian oligarchy?
Like
Reply
I laughed when you discussed the pasta machine craze: in about 1978 my sister got married: we gave them an Italian pasta making machine. It was shiny chrome with gorgeous knobs and cranks. My new brother-in-law, who never, we discovered, shrank from issuing frank opinions, looked at it and announced that they would be mounting it as a hood ornament on their car..
Like
Reply
AI is all such horseshit and I'm sick and tired of reading crap about how great it is. I can't wait for this tech to go belly up when everybody outside of Silicon Valley sees its utter uselessness that nobody asked for.
Like
Reply
I am puzzled by his research. It was more like a high level re-reporting of the same information you hear daily without any explanation of how and why, what is hype and what is real, lacking insights. Interestingly or maybe expectedly he sounded exactly like someone who is using language models and AI agents for his research.
Like
Reply
It seems to me that the core political-economy debate surrounding AI should center on capital-labor dynamics (that Paul Krugman touched on) and the potential realignment of social interests (that nobody discussed here).
1/ The "winner-takes-all" nature of AI (surely in the US) fuels a high-leverage race to build dominant models. This allows a handful of multi-trillionaires to effectively buy political systems and permanently lock in their market power.
2/ Classical economic theories are becoming obsolete. Their foundational frameworks fail to account for the extreme, asymmetric concentration of power that AI technology enables.
Like
Reply
Jun 13 Edited
Azhar doesn't seem to be concerned by the extreme hype and outlandish financial shenanigans employed by the major hardware and software proponents of AI, which leads to much gloomier picture of the bubble that AI really is.
See for example Ed Zitron's well-informed commentary on Chris Hayes' podcast at https://www.youtube.com/watch?v=-Mn-TNLwQys
Like
Reply
This was a fabulous conversation! Two brainy people talking in language I can understand! Thank you so much.
Like
Reply
The strange nature of humanity is, rather than being predetermined by the availability of technologies, knowledge or experience, humanity tends to move in an unpredictable direction, and quite often ahead of what science and available tools could achieve, boundless imagination, infinite fantasies propelled our vision for the past Present and future, the twist and turn of existing tools help humanity to imagine a future that is of today, the concept of a screened device that would contain all information and access to control did not just pop up after the unveiling of the iPhone or smartphone, if it were left to users and engineers we would have been stuck with the blackberry style keyboard phone, which i rather like and still prefer over the smartphone. Engineers, scientists, computers programmers operates on a predetermined set of knowledge and experience, as an economist modernist and historic data would have been the basis of their trade, but yet this is not the bound in which humanity would imagine about their Past Present and future, it goes almost any direction, not merely within the confines of existing experiences (law, knowledge and structure) art is one of the most daring exploration of the expression frontier. Now since we established this point, would anyone think that humanity would be tools bound in determining ones past, present and future? While most of humanity operates base on the bounds and tool available, sometimes like a clock work, especially those working in a factory, humanity is still completely unpredictable, and could break any preset boundaries, going back to the yellow pad and a pen, within the confines of 50 pages, 18cm x 12cm, a ball pen, pencil that could in practice draw a 5km line on applicable surface, how many great minds have broken the limited format available to them and created a complex world of art, science and philosophy ?? Jack Kerouac application of a long glued up roll of paper to work on this stream of consciousness into presentable text is a rather interesting way that people could change the application of tools to best suit their expression. This applies to llm power ai as well, it is not intelligent, it can not go around and think out of the box(codes) in which is was predetermined, as in text theory it is confined to the schemas context of the text, the very text can not excced over limits of machine memories be it billions of billion of bytes of data, a predetermined lined of choices and processes which regardless of the speed in which computing and servers could afford it is still operating in basic coding languages, the limits meant the very verson of ai is at best assisted intelligence, or as cyberpunk fantasy would refer to as I/O. The application of this tool have been overhyped by a few, and there is a complete lack of discussion of how this took could assist humanity in continue to move beyond boundaries and break the structured existence humanity created. How many more time and resources humanity could throw towards a dud ? And hope someday it will turn into a soul, using science fiction terms will the same llm based ai be able to reach the level of a replicant to labour on Mars or satisfy our desire with a short life span????? Will investors have the patience to continue pumping billions in the unknown waiting for a rewards that no one knows what it holds?
Like
Reply
Many thanks!!
When Azhar outlined his brilliant architecture for using AI beneficially, I found myself chuckling as I imagined a nation with our percentage of citizens who actively embrace "sincere ignorance and conscientious stupidity" trying to read - much less comprehend - much, much less implement his AI framework.
Clearly that is not AI's fault, but it may well be its downfall. For better or worse..
Like
Reply
Three points/questions from this discussion:
1. What would Gary Marcus say about all of this information? How about a point-counterpoint discussion with him and Azhar and Dr. Krugman?
2. Musk has a certain level of genius, but he also seems to have an attention span limit. What now of the Boring Company? What are we to think of Tesla cars in the light that the model of sedans seem to be stagnant and the Cybertruk is a monstrosity?
3. The discussion touched on healthcare, which is certainly an area where many aspects of AI will be useful, not so much to make diagnoses, but to assist with quality measures, to prevent medical error, and to assist in rooting out fraud and abuse. However, privacy is a huge problem in the light of AI: How can we be sure that an AI system will have rigid boundaries for output and communication, such that HIPAA will not be violated -- that patient privacy will be protected? It seems that AI systems are not designed with Asimov's laws of robotics at the core, and protection of privacy is part of doing no harm to humans.
Like
Reply
I was interested in the 'revenues per employee remarks. In the early 70s, as I approached graduation I tried, as a 'good economist' to research what sector of industry I should reward with my skills and abilities(!) The answer, at that time, was computing - computer programming - as the economic benefits of successful programs were becoming visible. But I quit after a year realising that (a) hundreds of other graduates had made the same calculation and (b) there were better things to do with my life than programming.
Fifty years later, the impacts of computer programs are ubiquitous but the rewards to individual programmers were only on a par with other comparable professions (or trades). I suspect the same with be true of AI: I still find it had to believe that it will not become a useful, in many circumstances, tool but rarely essential and life-changing. But to be even useful, the energy cost has to be brought under control and I have not yet seen any signs that that is beginning to happen.
At least, if I'm wrong, I'll be out of the way!
Like
Reply
SpaceX is not successful. All of its rockets explode and it just went to market to steal more money."
Like
Reply
If there is any question that the "tech bro" oligarchs actually running this nation are controlling all the narratives, take a look the "suggested" searches I got when I asked, "who owns Claude AI?"
Wikipedia
https://en.wikipedia.org › wiki › Anthropic
Is Claude better than ChatGPT?
Is Claude AI made by Google?
Who owns 51% of OpenAI?
Which big companies are owned by Jews?
Can Jews marry Muslims in Israel?
Like
Reply
I'm glad you are thinking about this. It is also good to hear someone who is reasonable and giving data to support a more optimistic outcome than my closet bear view of the world.
Like
Reply
Some of Azeem's points seem to me like rationalizing, though I will grant that he has some interesting perspectives.
An over-riding question I'd ask is: has there been any large new-technology investment in history that smoothly transitioned to normal widespread use? Or did they all go through a "trough" of bankruptcies, failures and even general recessions?
Like
Reply
Just trying to sit back a bit and gain perspective. Let my thoughts drift. So we have an administration, elected by billionaires, trying to centralize and concentrate power, destroy democracy, promoting Nationalism, trying to severely limit immigration, trying to build, very quickly, massive data centers for AI that may obviate the need for many types of human labor. Gee. It’s almost as if they are preparing to leave the US behind.
Maybe for a different world altogether.
Experience with space, even riding into space themselves, becomes, for them, a necessary.
Are they getting ready to leave Earth behind? Just leave a skeleton crew behind to mind the machines?
They are ceding the Earth to climate change and set sights on a future elsewhere? Please tell me I’m wrong.
Not in our lifetimes?? Probably not in mine. But things are moving very quickly. No colonization of Mars yet. But never say never.
Where does this end?
Like
Reply
excellent conversation! very much apprecriate your perspectives and insights!
Like
Reply
Dr. Krugman, you don't need plagiarism robots. You are enough!
Like
Reply
I would like to see a panel of experts from different disciplines. To this discussion, add analysts from health care, banking, tech, small business, etc. - hopefully with exposure to AI as it applies to their areas of expertise.
Like
Reply
SpaceX is a disaster, its large rocket design is a total loser after 11 launches it has not even reached low earth orbit.
Like
Reply
I'll believe the current models of AI have humanity-level benefits when we can ask it about these problems and it can return real, actionable procedures that solve the problems. Like, "Hey Claude -- how do we solve the problem of data centers using too many resources that are needed elsewhere, and how do we convince the political leadership to respond?"
Like
Reply (1)
Gemini has a pretty good answer to "how do we solve the problem of data centers using too many resources that are needed elsewhere, and how do we convince the political leadership to respond?"
Here you go :
https://share.google/aimode/aM3PdUOENc8Ve4Jtf
Now what was your point?
Like
Reply
Jun 13 Edited
China does it without destroying the environment. Why can’t we? Also, there is a series on YouTube called “Pasta Grannies,” where the moderator has been going all over Italy and some countries nearby and collecting pasta recipes from very old ladies who make pasta. These ladies, who are in their eighties and nineties, make homemade pasta dishes for the show. It’s quite wonderful. And in Italy people still make pasta. I got back into it from watching the show. She has collected the recipes into books, btw.
Like
Reply
I think we're here on earth to learn how to love and live cooperatively with all life forms. So far we're failing. I don't see how AI will help us get there at all. In fact, imho, I think it will only hasten our demise. I'm sure many will disagree with me but that's OK.
Like
Reply
What the hell are we doing? You know, store bought pasta is actually better. The Italians don’t do this.”
Are you crazy? Fresh pasta is way better! I make it weekly. Italians are short of time like everyone else.🧐
Like
Reply (1)
Damn. I thought they had tons of leisure time and they spent it making food and dining well.
Like
Reply
Not only does Paul Krugman provide a clear wide "window" into economics but through his equally knowledgeable contacts he provides a similar "window" into A"I". Its strengths and weaknesses and failings. The source of those failings is the search for profit and thus power; the aim of replacing rather than assisting people.
One day the semi-sentient computers of science fiction may appear, but they will not be built by the "techbros". The computers of science fiction - e.g. on the bridge of Captain Kirk's Enterprise - operate perfectly. Probably programmed by Vulcans. The current capabilities of humans fall far short of such a level of perfection.
Rather than race to the most complex A"I" we should let it evolve at a natural slow pace as our civilisation and our intellects evolve to match it.
Like
Reply
Part of the discussion included (M. Azhar)
"Chinese companies are using much less capital to build models that are nearly as good. So I think the harder part of your question is that if every model that OpenAI or Anthropic costs ten times as much to deploy and develop, but lasts only a couple of years before it’s defunct because of competition, what needs to be true for that to be sustainable for more than a year or two? To me, that is a really tricky question as well."
Is there an analogy in defense where the Ukrainians and Iran build simple cheap drones whereas the US is still dependent on multi-million dollar missiles etc...
Of course the cheap drones will be significantly modified with time ..... much harder to do with expensive missiles.
But underlying this is who controls the purse strings or are small groups allowed to flourish without total top down control?
On a slightly different note it was wonderful to see the USMNT take Paraguay apart last night.
That was a function of the manager (Pochettino) who we watched at Southampton in 2013/14 driving the players to take an aggressive approach to regaining the ball and then going forwards in a very direct manner .... bit like the Ukrainians and the Chinese AI developers.
Oh well.
Like
Reply (1)
The Chinese models are more efficient because they are focused.
For example, an AI system for factory automation doesn't need to be capable of programming, but it is also much less valuable.
The drone analogy isn't really the right one.
There are plenty of US companies providing AI solutions for focused situations.
Of course the headlines don't mention them because the venture capital horde is throwing buckets cash at AGI.
Like
Reply
Thank you, Paul Krugman, for working to keep AI tethered to the real world.
Along these lines, I'm curious about the demand side of AI right now. There's talk of profits, but where are the profits coming from? Who is paying actual money to use AI, who are they paying it to, and for what?
Another question is how much revenue - current and projected - is generated by surveillance gathering, since AI coupled with your browser can already record your every keystroke and wiggle of your mouse.
Like
Reply
I appreciate this interview and information and it sent me doing a bit of research, which could be never ending in order to catch up on the two of you. Bottom line, Paul is an economist and formed this interview around his natural perspective. It’s not to say that either participant is unaware of or uninterested in the human cost. I wouldn’t assume that AI will destroy humanity directly. Climate change may do so in our efforts to pursue AI (or weaponry). AI is a tool folks.
Like
Reply
Glad you are back and had a haircut and shave. You clean up nicely!
Like
Reply
Like
Reply
Winchester Kentucky here 💙💜
Like
Reply
Wow! This conversation really widened my intellectual boundaries. The AI world is moving so fast it makes my head spin. I do worry that some of the consequences could be quite nasty.
Like
Reply
It is interesting that 19th century railroads are used as an analogy for AI tech. What seems to be forgotten is that railroads were regulated by the Federal Government until 1980. But, I read this transcript and I found it very hard to understand. And maybe that is the point. Me, I am too stupid to understand, so I should just rely on the "expert." I am not against AI, but I am against irrational capital markets. So many companies and banks and consultants-- a whole ecosystem-- has committed itself to AI by betting billions of dollars and thousands of man-hours of consultant billings. Based on what? FOMO...fear of missing out. So hundreds of public-facing people employed in this ecosystem have to talk up how great AI is because it is their rice bowl. Including Mr. Azhar.
Like
Reply (2)
Cynicism is always a weak response.
It may be that Mr. Azhar is motivated as you suggest, but you might look for some actual evidence.
Like
Reply (1)
Perhaps I am. I have been around finance a long time and it pays to figure out why someone is saying what they are saying. Sadly for all of us, when investors/lenders take a long position they make public statements to support that long position--meaning, they are spinning, they are selling so others don't trade against them. And perhaps I am being harsh to Mr. Azhar, but here is a man whose career and livelihood depends on the development, growth and public acceptance of AI. A corollary: “It is difficult to get a man to understand something when his salary depends upon his not understanding it” — Upton Sinclair.
Like
Reply (1)
Perhaps you could point out an item or two that you didn't understand.
The conversation was pretty thick.
Like
Reply
I found several things hard to follow as well.
Like
Reply
All of this seems untethered to any experience I have. I don't see what the use of AI is for the vast majority of us. It may be useful to develop other applications or solve problems in physics but even at that those critical solutions don't seem to be trustworthy. I'm afraid that after this discussion I still don't know why I need AI in my world. I don't "get" what it is or even how to access it in any useful way. I am a long, long way from being convinced that I need it all, much less that so much of our national treeasure should be devoted to creating it.
Like
Reply
By far your most insightful AI post to date.
Like
Reply
That’s what most concerns me about AI. It’s learning what to write/say based upon what we write/say. Now, of course, it read Shakespeare and medical research and heard Mozart. But that means it also read a lot of really dumb stuff and lies. And what did it learn from Trump’s favorite song ‘Macho Man’? I think the premise is problematic.
Like
Reply
Companies going all in on AI could lead to mental slavery among the elites still holding jobs where they too become automatons like the lower echelons or factory workers in a Charlie Chaplin film processing for example law cases without even an iota of job satisfaction of course AI doing the tedious boring jobs is an improvement but every job? The Silicon Valley tech folks may be having fun programming AI programs but the professionals forced to use them may be deprived of whatever pleasure they get out of coming up with an idea or solution depending on how ubiquitous AI is and as for those whose jobs are eliminated and who profits what becomes of this leisure? Abject poverty for some or some kind of renaissance just an aside my daughter-in-law’s home made pasta is better than store bought
Like
Reply
Bottom line? Everything happens because it must. AI will lead (after an initial disaster) to unlimited energy, and a thriving, moneyless economy. #Resourceism
Like
Reply (1)
Correlation: the powerful do what they do (to us) because they can. Beware.
Like (2)
Reply
Ugh. I wish Azeem didn't cast aside his credibility by going overboard on LK-99 room temperature superconduction claims three years ago. He echoed all the hype without even once pausing to scientifically question what we were all being fed.
Like
Reply (1)
Presumably, you've never made a mistake.
We do have a Physics PhD on the team now. (And you can see our dissection the LK-99 and the doubts around it here. https://www.exponentialview.co/p/chartpack-the-potential-of-superconductors.)
Like (3)
Reply
There are two parallel streams of AI. The ChatGPT stream literally does not know what it is talking about. I founded the other stream in 1957 when I was a college freshman taking introductory psychology. Hundreds of scientists and engineers have helped to model how our brains make our conscious minds in healthy individuals and clinical patients. See sites.bu.edu/steveg for almost 600 downloadable archival articles that develop neural network models of all these mental processes.
Like
Reply
In graduate school in the 80s I put some effort into training myself as a pure economic theorist. I couldn't do it; it's too hard. But one thing I remember well is the convention of looking at an economy as a million little nodes. The nodes could all be connected to each other by convex combinations* until the gnat-cloud of little nodes was a big, comfortable cushion. This is how I picture AI.
\* convex combination: suppose you have two distinct points, X and Y. If you draw a line segment between them, the equation of that segment is aX + (1-a)Y, where 0 < a < 1.
Like
Reply
"It was that the revenues [in the 1880s] didn’t materialize because the tracks were being laid in places where there were no towns. That was a problem."
I've always thought that building tracks past where there was no town, but in the logical place for a future town, was the railroads' business plan. See "Once Upon a Time in the West."
Like
Reply
1 reply hidden
Paul Krugman Podcast
Notes on economics and more
Notes on economics and more
Subscribe
Listen on
Substack App
Apple Podcasts
Spotify
YouTube
RSS Feed
Appears in episode
Recent Episodes
5 hrs ago•Paul Krugman
Jun 20•Paul Krugman
Power and Geopolitics After Trump
Jun 18•Paul Krugman
Lunch Money with Paul Krugman and Heather Cox Richardson
Jun 18•Paul Krugman and Heather Cox Richardson
Jun 8•Paul Krugman
Jun 6•Paul Krugman
Learning from a Mentally Ill President
May 31•Paul Krugman
Ready for more?
Subscribe
SPEAKER 2
Yeah, thank you, Paul. And it has been 18 months, also known as one and a half centuries in AI time since we spoke.