Talking With Azeem Azhar

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… Get full access to Paul Krugman at paulkrugman.substack.com/subscribe