How to think well with AI: signals, quietness, and the argument engine

0:00AI is not telling me what to think about. Mostly it's telling me what I don't need to think about.

0:05Azeem Azhar:Cognitive surrender is an uncritical abdication of reasoning itself. It reflects not merely the strategic delegation of deliberation, but a relinquishing of cognitive control. There is something about AI, about its allure, about its potency, that could make cognitive surrender more and more widespread. The way in which I think has changed much more than I expected. And I'm still figuring out what it means. The argument engine is built from about 100 ,000 words of my writing. House Views runs an argument against our established positions. Stylometer. And of course, we still have the golden thread check.

0:40Azeem Azhar:Great thinking traditionally happens with care and self-reflection. And in a world where there are all these tools around me, am I still doing that quality of thinking? Am I doing the quality of thinking that I could do with 10 uninterrupted days?

1:01Azeem Azhar:What I want to do today is talk about something that I've been circling around for a few weeks, perhaps even months, which is about the way in which our thinking processes change during this period where we're using AI. Is our thinking getting better? Is it getting worse? This matters more than it ever does before. I mean, AI is inside all of the processes that I use. It's not a tool that I pick up and down. It's a tool that's there the whole time. It's become completely ambient. You've seen us and heard us talk about this. We've struggled with this issue. We've tried to expose it to our readers and our listeners.

1:38Azeem Azhar:What is AI doing to that process of thinking? How can you use it well? Where does it harm rather than help? Just a couple of weeks ago, we had the first of our AI Vistas conversations. where we sat down and discussed this question, do we use our tools or do they use us? And I think that really gets to the heart of this question. And having had that discussion with Vistas, I wanted to come back to all of you and show to some extent how my thinking process has evolved and what it looks like now in a world where I'm using a lot of AI, the 100 million tokens a day that I currently use to support my activity.

2:18Azeem Azhar:Of course, we're not the only ones to be asking this question. The New York writer and podcaster Ezra Klein was on a podcast with Dave Perrell last year, and he was talking about this very issue. And I'll quote Ezra here, so I'm just going to read from my notes. Having AI summarize a book or a paper for me is a disaster. It has no idea what I really wanted to know. It would not have made the connections I would have made. I'm not interested in the thing I will see that other people would, pardon me, I've messed that up. I'll try that again. I'm not interested in the thing. I'm interested in the thing I will see that other people would not have seen.

2:55Azeem Azhar:And I think AI typically sees what everybody else would see. So that's what Ezra said. And actually what just happened there with my fluffing my lines twice is a perfect example of thinking. Ezra has his own interiority, his own world model, his own way of speaking. And so me reading out his words was difficult. I've practiced them, of course, and they saw me fluff them up twice. That is the heart of the question of thinking. And of course, we're concerned about becoming intellectually lazy as well. And there was a paper earlier this year available on the SSRN network by two academics, Shaw and Knave, and they draw a distinction between cognitive offloading and cognitive surrender.

3:40Azeem Azhar:So cognitive offloading is a strategic delegation of deliberation. For example, I used to remember all my friends' phone numbers. It was all actually their parents' numbers because we didn't have phones, because that was in the days before the mobile phone. And now we've offloaded that. Nobody really remembers any numbers. And it's not really a problem. Cognitive offloading is a tool to aid one's reasoning in one way or another. But what Shaw and Knave describe is they describe also cognitive surrender, which is an uncritical abdication of reasoning itself. It reflects not merely the use of external assistance, but a relinquishing of cognitive control.

4:22Azeem Azhar:So cognitive surrender is rather more pernicious as a practice because we uncritically abdicate our reasoning. And I think a lot of people are worried that AI will lead to widespread cognitive surrender, and for good reason. I suppose one thing I would add to that is we probably do all exhibit some cognitive surrender just in our day-to-day. We can't be critically reasoning about everything that we deal with in life. And there are systems and structures that we simply have to accept because it just makes life easier. But there is something about AI, about its allure, about its potency that could make cognitive surrender more and more widespread.

5:05Azeem Azhar:So let's come to how I think about thinking itself. My work, my job is anchored around research, around analysis, around communicating those ideas. Ultimately, it's really anchored around thinking. And I think that's true, of course, for many jobs. But you can say that a job like mine, which deals with ideas and data and connecting them together, is really all about that. And if I don't think new things, things that are challenging to you, things that perhaps you agree with, things that perhaps you don't agree with, well, I'm not doing my job at all. Thinking is the essence of it. Now, I've been doing this since well, well before ChatGPT.

5:45Azeem Azhar:I mean, my last book came out in 2021. I talked about GPT-3 in it. There was no AI of any meaningful capacity or capability to help me with it. When ChatGPT launched, I realized immediately that this was going to change the way I worked and I started to play around pretty quickly. Of course, over the last three years, the tools have become, well, I mean, amazing. There have been at least three significant paradigm shifts. Of course, the first one was ChatGPT itself. The second was the arrival of reasoning models. And I think the third has been the arrival of these long context tool using systems like Claude 4.5 and Claude Cowork.

6:26They are so much more than ChatGPT running on GPT 3.5 was three years ago. And I suppose the question is, how's that changed my process of thinking? In what ways is it better? And in what ways might it be worse? In preparing for this and thinking about this over the last couple of days, I really came to the conclusion that the way in which I think has changed much more than I expected. And I'm still figuring out what it means. And one question that I ask myself, and I'll return to this during the course of this conversation, is am I doing the quality of thinking I could do with 10 uninterrupted days?

7:01I mean, it's everyone who's got an idea, their dream, their fantasy, that you could just have 10 days without the noise of the world around you to really allow your mind to get deep into an idea and play around with it. when I was writing my first book, I might sometimes get two or three days at a time. And by the time you get into the second day, it's like an intermittent fast, right? You get all of the benefits in day two. Well, the same thing was happening with deep thinking. So I'm going to go through this question in what I hope will be five brief acts. How I find signals, how I reflect slowly, how I write, how I critique, and ultimately how the words come out.

7:41Azeem Azhar:And at each stage, I'll try to illustrate how I'm using some of the AI tools, largely that we have built ourselves to do this. And when we put the essay of this podcast out, we will share some screenshots with more details. So the first stage is, you know, in a way, where do those ideas come from? What are the signals that trigger my nervous system to send a signal to my brain to say, listen, you should write about this? Of course, we use a lot of AI in this area, right? I have a signal detection layer that sweeps a very wide range of inputs automatically. Just think about all the email that I receive or that you might receive.

8:21All of these things show up as signals in various inboxes and I'll run detection across that volume of insight.

8:30Azeem Azhar:I say insight. I mean, a lot of it isn't insight, of course. A lot of it is just a news story, a funding announcement, a new chip that's been launched. Some of it is pabulum. Some of it is incredibly powerful and relevant. And so the The question is, how do you get that AI system to do the detection? So how do you get it to detect signals? I mean, typically what you would do in a signals detection system like this is that you would do something statistical. You'd go off and say, well, listen, normality looks like this signal. And if we have a signal that's two standard deviations away from what normality looks like, that's an anomaly.

9:08And draw attention to it and bring it to someone's desk so they can deal with it. And I mean, that's a perfectly reasonable thing to do, but it does depend on your being able to characterize those signals in any kind of statistical reality. But what it also has you doing is seeing things, to come back to what Ezra said, seeing things that everybody else sees, because everyone else is going to run that similar to standard deviation approach.

9:31Azeem Azhar:So, of course, I will do things like this, but I also make really heavy use of archetypes. So archetypes are synthetic personas of people that I think are interesting and distinctive. And I have these archetypes scanning my inboxes to find things that they as archetypes might find interesting together with their interpretation and give you some examples of those archetypes. One is based on me, and I'll speak a bit later about how we built that. I have another based on the Silicon Valley investor, my friend Vinod Kostler. I have another built on the hedge fund investor who spotted and shorted Subprime back in 2007, 2009, a guy called John Paulson.

10:16Another synthetic is based on Clayton Christensen, a Harvard professor who came up with the disruption theory of innovation. So each of these archetypes, and there are several others, will scan signals, dozens of items a day and hundreds a week, come back and provide a view.

10:35Azeem Azhar:What I noticed is that you do want them to look at things daily, but you also want them to look at outputs over the course of a week or perhaps even longer. So that's one way of helping me understand and build a picture of the world that is out there. I also look for cross-cutting relationships between themes. Is there a pattern of centralization or decentralization emerging across different industries? And if so, why? And if not, why not? Is there a trend towards open source? Are there differences between age cohorts in behaviour or attitude that are emerging across all of these different signals and pieces that are coming into my inbox?

11:15And the purpose of that, again, is to try to find common threads that could be interesting, that could be evidence for underlying processes that are going on. And of course, there are the prosaic summaries of industry trends and academic papers and momentum here and breakthrough there. Like that's all the hygiene stuff that is probably least interesting, but most commonly read. When we launched Exponential View Daily last summer, we hinted at what was under the hood.

11:43Azeem Azhar:What we were attempting to do was sort of show our ability to sort of make sense of this and how that feeds into the kinds of decisions that we make. The end result of all of this signal processing is something that shows up in my email or it shows up in my WhatsApp. I read maybe half of it. What it does is it gives me situational awareness. That's all it is. It is situational awareness. I'd say it's a haystack. And occasionally there's a needle in it. Out of these signals, very rarely do they actually trigger me to write something. What they might do is reinforce a belief, just provide me with a bit more of an evidence or a critical view.

12:21So what's going on here is that the AI is not telling me what to think about. Mostly it's telling me what I don't need to think about. So I'll give you two concrete examples. So we've just seen these incredible results from Anthropic where their revenue is approaching$20 billion per annum. On an ARR basis, they added about 5 billion in February alone. It's kind of staggering. Now, that could be a hot take, right? I mean, if I had time, maybe I would have produced something that people would have read on that subject.

12:50Azeem Azhar:But in a way, if you've been reading Exponential View, you know that we've already predicted that, suggested that that could be an outcome relative to open AI, because we've spoken extensively about how we felt that open AI was spreading itself thin. And it didn't have a coherent way of tackling a market. It was choosing a ubiquity strategy rather than a focused beachhead strategy. And on the other hand, we had been talking about how Dario's strategy at Anthropic was so specialized that it likely lended itself to the kind of classic land and expand that has made so many startup companies successful.

13:26So in a sense, I feel like I don't really have all that much new to say right now on that particular angle. So I'll pass. Another one is prediction markets. There's a lot being written about prediction markets. Should insider trading on prediction markets be allowed or not allowed in a sense? They're working better if it is allowed. Should prediction markets make their way across the economy more widely? There are some really interesting debates there, but actually I don't have a framework to exploit, so there's no essay. The way to understand this is I get the situational awareness from the AI, but the writing triggers are still very, very experiential.

14:01They emerge from conversations I have. They emerge from listening to people. They emerge from a lecture or just some kind of moment of friction with an existing idea. Look, this isn't quite Kekale doing some lucid dreaming and imagining the snake eating its tail and voila, there is the benzene ring. But it isn't a process that I think gets systematized. It gets supported by these tools more than anything else. And signals are noisy. So let's get to the second act. Quietness. Like what's the value of quiet when you're thinking? I wrote my last book without AI and there was Google, there was Wikipedia, there was hours of writing and reading.

14:41There would be stints of 10, 15 hours where I wouldn't leave my desk, leave my chair and be reading and taking notes, note after note after note quite often. I mean, it's very hard to do that, right, because you're not living any other part of life. And I did wonder whether AI would create more interruptions the way the smartphone and Twitter and so on has. But as many of you know, I have felt that AI has given me more time back and more time to think and time to do it without interruption. That isn't to say that I'm not using AI tools to summarize long documents. I do that all the time. And let's be honest, much of what you read, much of what I read deserves to be summarized and should only be summarized.

15:25But the best stuff doesn't. That gets ordered from Amazon or downloaded and read. But where the summaries do help is that they actually de-risk. They help me understand whether I should dedicate four hours or eight hours or 10 hours to a section of a book or a book in its entirety, whether there is a likelihood that it will challenge me sufficiently to develop my thinking. And so there's lots of reading that is going on as a consequence of this, really, of the space created by AI. Quietness is not a practice. I wish I was that disciplined. It's a capacity. The discipline isn't in scheduling that quietness.

16:02My team knows I've got these blocks in my diary that says quiet writing time, which no one pays any attention to, least of all me. But what it is, it's about resisting the urge to colonize those gaps at every opportunity. Some weeks I don't have any time. Some weeks I will have literally entire mornings and afternoons where I can be with a pen and paper. And you all know about the fountain pen. Well, here's the other fountain pen I use. And actually it leaks a little bit. So I've got inky fingers today. When I'm thinking about writing and doing that quiet work, I will use a fountain pen and I will use A4 paper in landscape mode.

16:37By and large, not small books. Small notebooks means small ideas. They're to-do lists. They're not for your thinking. And what happens with the pen, writing with the pen, it's about flushing the internal cash in my head. It gets things out of my brain differently than typing. There's connection making that somehow reaches further back in time. It's more associative. It's more surprising. and when something does need to trigger me to check a fact honestly at that stage of your thinking you're not thinking about facts they're irrelevant i just put a little asterisk with the thing that i need to check and put it on the side of the piece of paper and in theory i'll go back and check that fact in practice i normally forget but the pen sets a much higher standard for what

17:20Azeem Azhar:counts as worthy of an interruption a quick note if you want to support us in bringing more of these conversations to the world, please consider subscribing to the show. So we're getting into this idea of writing now. Been all over X over the last few months. This notion that writing is thinking. And I'm going to quote Ezra again, and I'll probably fluff the lines again because I'm reading someone else's words, not mine. Here's Ezra Klein. You can have an epiphany through writing, but weirdly, I think you have to be careful with that because sometimes writing is a process of persuading you of what the piece needs you to think.

17:57You've got to be careful not to become accidentally persuaded by the formalism of whatever your own assignment is, which is really, really a very, very nuanced reflection by a great writer on that process of writing. And on the AI Vistas discussion that we had a few weeks ago, Nita Farahani dove straight into the writing is thinking meme. She said, I've heard writing is thinking too many times now, and I think it's crap. When I write, I actually give a talk first, I think in public speaking, more than in written form. And I agree with both of them in that sense there, that it's not as simple as saying writing is thinking.

18:32Not all thinking is writing, mathematical thinking isn't, certain types of pattern recognition isn't. If you are wandering through an art gallery and you're looking at paintings over a period of time or from a particular artist, you're thinking, you're pattern making, your connection making, but that's not writing. The writing process itself is a fractal. It operates at lots of different dimensions. It's not just the words on the page. There are a lot of people who are writers who say, that book is entirely my own work, and LLM didn't write a single word. But the words themselves are just the final surface layer of the writing.

19:12There are many other layers, and they all have different degrees of importance at different parts of the process. There's, you know, why am I writing this in the first place, right? What's its purpose? What's its provocation? How will I know if it does whatever I want it to do? That is not about the words that get put down. Because the next decision I can make is, how do I actually want to argue this? Do I want to make it in a very directed way? Do I want to make it quite Socratic? Do I want to embrace ambiguity? Do I want to write it in an essayistic style that shows it doesn't tell that gradually reveals.

19:47Do I want there to be narrative tension? Where do I want that tension to be? This is all the art of writing and it's nothing really to do with the words. And then there's the nature of the structure that you'll put down with your writing. If you've been trained at one of the strategy consultancies, you're probably familiar with Barbara Minto's pyramid principle, situation, complication, resolution. It's useful for certain classes of business writing. It's a structure that we'll often use when we're doing a very analytical piece within exponential view. But that is also a part of the process of writing.

20:19And the words themselves really are often the final surface layer. And sometimes there are words that are great that don't make it into an essay. And I do find myself pulling words out from the off cuts of a previous essay and wanting to drop them into a new one. And I think that it's important to understand that the way in which an AI can help at this level operates also at those many levels. So let's go back to the purpose. What are we writing? Why are we trying to make a claim? Well, one of the tools that we've built is a golden thread analysis, the idea that there should be a single golden thread that exists in an essay and that all of the sections of the essay are in service to that golden thread, all of the paragraphs in the essay are in service of that golden thread.

21:07All of the sentences in the essay are in service of that golden thread, except when they're not in service of that golden thread, because writing and thinking is all about the exceptions rather than the rules. And so the golden thread analysis that we will do, which is a pretty sophisticated, I think now that's on

21:24Azeem Azhar:iteration 50 or whatever else, is really designed to help me or whoever's writing to think about the deliberate nature, the intent they had when they set out to write that piece through the golden thread. And of course, we can do similar types of structural analysis on the pacing of an essay as well. Again, not to be told what to do, but to be given things that we can critically reflect on. And when you finally get to the words itself, you know, when I was writing my first book, I had a subscription to the Oxford English Dictionary, which has got the, you know, the best dictionary that's out there on the web if you enjoy words I would go and get that subscription it was only about 25 bucks and you know it's got the best thesaurus so much better than thesaurus.com but of course yellow lens themselves are also quite useful thesauruses so we we're getting through this process now we've had the signal detection we have had the idea of where do you reflect to start to develop the idea we're getting into the writing but there is this middle stage, which is a circular loop, because my writing process is explicitly not linear.

22:29And that means that I go round and round this iterative circle. But the starting point is often that I just have this idea, maybe it's come from the signals detection. No, it never has come from there. But maybe it's just come from something that I have been thinking about. And it will happen on a walk or in a long shower or walking to the tube station. And I will start to build that argument up in my head. And then almost all of the time, I will outline it on paper, handwritten like this. In fact, entire sections of my book are written by hand. And I'll just show you, this is my notebook for when I travel.

23:05I have these pages, you're going to find them hard to read. This will show up in my new book at some point. It was the start of a chapter I'm working on. So effectively, I will generally always start anything that I'm writing, whether it is for exponential view, whether it's for my book, whether it is a proposal, a talk I have to give by sketching something out by hand. And if you remember the Magnitudes of Intelligence essay that we published a couple of weeks ago, that again started like that. I had this idea of like powers of 10. I thought about the film I had seen. I started to write it out by hand.