Are we in charge of our AI tools or are they in charge of us?

This is the first episode of AI Vistas, a new series where I bring together people I trust and respect to tackle a major question collectively. Today’s question: are we in charge of our AI tools, or are they in charge of us? Joining me are Nita Farahany, distinguished professor of law and philosophy at Duke University and a leading thinker on cognitive liberty and mental privacy; Eric Topol, founder of the Scripps Research Translational Institute and one of the world's most cited medical researchers; and Rohit Krishnan, engineer, former hedge fund manager, and AI builder. Moderating the conversation is Nick Thompson, CEO of The Atlantic.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Azeem Azhar?

  Right now, as AI moves faster than our ability to make sense of it, I don't think any single view from where any one of us stands is enough, not even mine.

• 0:10Speaker 01

  So now we've seen a study of gastroenterologists. When they turn off the AI, they're not as good. And of course, there's the younger docs that are coming through who are going to be never skilled.

• 0:21Speaker 02

  Today, should you use AI tools to make the life-threatening consequential decision? Yes, you should. But what is it that we as a society need to invest in such that in the long term, we don't end up with humans who no longer have basic competency?

• 0:36Speaker 03

  It's not just in Twitter, right? And I think slop here is an important word because primarily what we're saying is that it has all the appearance of a fulfilling meal, but it has no calories. It's kind of empty.

• 0:47Speaker 02

  I call it hollowware.

• 0:48Speaker 04

  Writing is thinking, right? And if we de-skill our writing, we're de-skilling our thinking.

• 0:52Speaker 02

  Oh, I'm going to push back on that one for a minute, Nick. I've heard that too many times now and I think it's crap. Let me tell you why.

• 1:06Azeem Azhar?

  Hello, I'm Azim Azar, founder of Exponential View. And over the past decade at Exponential View, I've built relationships with people at the centre of artificial intelligence, the builders, the researchers, the investors, the thinkers, the doers. And right now, as AI moves faster than our ability to make sense of it, I don't think any single view from where any one of us stands is enough, not even mine. So today, I want to introduce AI Vistas. It's a new series where I bring together people I trust and respect into a conversation around a major question. Collectively, we're going to be able to see more together than any one of us alone. Now, our first question comes from my own recent experience.

• 1:51Azeem Azhar?

  I've committed thousands of lines of code this year using AI. Three of the apps I use every day I built myself, all since Christmas 2025. And a couple of days ago, I fired off a question to my AI agent before a car journey. It was quite a complex question. And by the time I'd arrived at my destination, it had gone out. Spawned subagents, found the data, assembled a report, and then I acted on it. I made a real decision on the basis of work a machine did while I was being driven. Now, I can now feel a blur. Where does my thinking end? And where does my agents' work begin? I suspect many of you know exactly what I mean by that. So the first question I wanted to bring to AI Vistas is, are we in charge of our AI tools?

• 2:36Azeem Azhar?

  Or are they in charge of us? There are three people I wanted in the room for this. The first is Nita Farahani, the distinguished professor of law and philosophy at Duke University, an author and a former advisor to President Obama on bioethics. Nita spent her career on the question of what it means to have a free mind and what happens when that freedom is no longer something you can take for granted. Eric Topol is the founder of Scripps Research Translational Institute. He's one of the most cited medical researchers in the world. Eric works where this question is not abstract. It's a clinical decision every day with a patient on the other side. And Rohit Krishnan is an engineer and former hedge fund manager. He's now building AI tools.

• 3:18Azeem Azhar?

  Rohit sees the gap between what people say about these tools and what they actually do with them. And holding all of this together is my friend Nick Thompson, the CEO of The Atlantic. Nick spent his career making sense of technology for the rest of us. And he's somebody who still engages with it with a criticality and enthusiasm every day. Nick, thank you so much for taking charge today.

• 3:42Speaker 04

  Amazing, Azim. I am so excited to help moderate this conversation. Let's get cracking. Why don't you explain what that decision was that happened while you were driving?

• 3:51Azeem Azhar?

  Well, I think it was something consequential. I had read in one of the financial papers that something odd was happening in the stock market, something they called dispersion, where the market doesn't move. And apparently some stocks go up and others go down. And it's very rare and it's very risky. And I didn't understand what a dispersion event was. So I asked my agents to go out, explain it to me, see if it was true, gather some data. And they came back and said, it's rare. Here's the data. Eight of these have happened in the last 30 years and the markets have tanked 75% of the time. And the decision I took was a decision about the amount of risk I wanted my portfolio to take. And we acted on it. So it was a consequential decision. It wasn't a trivial one.

• 4:36Speaker 04

  But it also seems like in that case, I don't see where the blur is like that. You're in control, right? That's like going to a book. So give me an example of a decision where you felt like it was starting to blur, where you felt like you're not as in control as you would have been a short period ago.

• 4:51Azeem Azhar?

  Well, I did a couple of shortcuts there. So one of the things I found in using my open claw agent, and this was a tool I used at that time, is it will often suggest, did you mean this? Or in our previous conversations over the past week, you had previously said this, and this seems to disagree with the observation you're making. And I had previously been talking to it about, you know, the AI trade and all these bottlenecks that we see. And it spotted the fact that on the one hand, I was worried about risk. And on the other hand, I was thinking of doubling down. And it drew attention to that. And we had a little bit of a conversation at that point, and it reframed the question.

• 5:26Speaker 04

  So Nita, okay, listening to this, and also knowing that Azim, at this very moment, has ran, what did you say, 300,000 lines of agent code, making all sorts of decisions for him. Apparently, I was booked by an agent, one of Azim's agents. He didn't even reach out personally. I don't know how that happened. Do you think that he's in control? Or do you think he's not? Tell me, where is Azim in control? Where is it a blur? And when does he lose it?

• 5:48Speaker 02

  Yeah, I mean, this gets at such a fundamental question about what does it mean to actually be acting autonomously? Like, are you acting in a way that is consistent with your own desires? Or are you being steered in some way by somebody else's desires? And I think if you start from that starting place and realize there's very little that we do anymore that isn't being steered in some way. You know, if you think about how much time you spend doing anything these days, it's driven by your tools. It's driven by the amount of attention that you are able to spend. I'm teaching a class this semester at Duke on mental privacy. So advanced topics in AI law and policy. And I ran an attention audit for my students. They had to use the number of times they picked up their

• 6:30Speaker 02

  Devices, they had to record it over a three day period and what they were spending time paying attention to. Day one, record it, don't change their behavior. Day two, they weren't allowed to use any app that was algorithmically steering them, which meant basically nothing. And then day three, they had, they could do whatever they wanted, but they had to record it again. And then they had to reflect on this. And it was remarkable what like day two looked like for them. Somebody read a book. They hadn't read a book and they couldn't think of how long, these are students, right? They finished a puzzle because they thought that they didn't have time previously. And then day three, they were utterly sucked in and like worse than day one.

• 7:11Speaker 02

  And that's just a tiny microcosm. So is Azim in charge or his AI agents in charge? His behavior is in part in charge. He told his agent to go and research some particular question. But the very fact that he is reacting to information that's provided by them, that he is constantly in this loop of notification and, you know, thinking about what his AI agents are doing while he is on a drive means that it's deeply interrelated. There is no longer like clear point at where Azim ends and his tools begin. It is an extended part of who Azim is at this point.

• 7:46Speaker 04

  And immensely hackable, right? I mean, he's controlling his portfolio based on this, right? If I were a smart hedge fund, I would somehow figure out how to hack OpenClaw so that it would tell Azim to invest in a way that is most beneficial to me, no?

• 8:01Speaker 02

  Yeah, that's at like the much more direct for Azim level. And I want to just back up there to say the very fact that Azim has offloaded that question to his AI tool to begin with is the more fundamental question. Like you don't need a direct hack of how he invests to ask the question of is Azim actually self-directing or is he being directed by the technology he's interacting with? It is more subtle. It's more pervasive. It's more universal than whether somebody intentionally decides to hack his portfolio or get him to invest in something you want to invest in.

• 8:34Speaker 04

  Rohit, do you agree with Nita? Should Azim basically turn these agents off and, you know, go on a silent retreat to pure

• 8:39Speaker 02

  I didn't say that.

• 8:42Speaker 03

  I think we all should, depending on the day you ask.

• 8:47Speaker 03

  I mean, one thing that surprises me quite a lot is that, and I asked this recently, you know, obviously, these are unsecure. However, the number of negative instances of things that are happening are surprisingly rare. And this is a general pattern that we see in modern technology where the ability to cause harm is rising along with the ability to do good or ability to get benefits similar to what Azim's getting, similar to what all of us are getting. And yet, the number of events that are happening are much lower. And I feel like there is something to the order of when you wait for perhaps perfect security, you're gonna let everything kind of pass you by and you're not going to be able to stand at the edge and do the experimentation that we would want to do.

• 9:29Speaker 03

  So if you're a hedge fund, you have to make a decision, where do I put my resources? And is that going to be in hacking open claw? And how much alpha would that provide me? Getting a bunch of people's internal information versus putting that same amount of effort into something else where I can get more alpha? And this always seems to be much lower. I don't know what to make of that, but at least practically, what that means is that my open claw does talk to my email and occasionally answers them. And I am kind of okay with that threat surface, despite the fact that prompt injections, et cetera, exist.

• 10:01Speaker 04

  Eric, do you agree with Rohit's basic framework when it comes to medicine? That perhaps we talk about all the dangers of technology and giving over agency and giving over data, but the actual harms are maybe less than you might think?

• 10:14Speaker 01

  Yeah, I don't think we know really yet. There's several studies now comparing AI versus AI with doctors for performance, various types of performance. And AI did better than the doctors with AI. So you say, well, this isn't expected. It was, everything was supposed to be hybrid. And if you extend that to having agentic support, perhaps, you know, even more so. Anyway, we still don't even know. Is it because doctors have an automation bias? Is it because they aren't grounded in how to use AI? It's really fuzzy right now. So we're not as advanced in the medical world domain as a lot of the other areas that

• 10:54Speaker 01

  are adopting AI much more quickly, just because, as Azim mentioned early on, clinical decision making, you know, is much more tricky, delicate than some of the other things.

• 11:05Speaker 04

  Wait, in those studies, are the doctors, the doctors using the AI versus the regular AI, are the doctors overriding AI when it's correct and steering it wrong and then accepting the AI when it's wrong?

• 11:16Speaker 01

  Both, yes.

• 11:17Speaker 04

  They're doing both.

• 11:18Speaker 01

  Yes.

• 11:20Speaker 04

  Awesome.

• 11:21Speaker 01

  So, yeah, so the the underperformers, which is, you know, the there's fifty percent of doctors are below average, right? They're they're more likely to accept the input, whereas the experts, you know, in quotes, they're rejecting the AI good input. So, yeah, you get a you get a lot of a lot of noise, unfortunately.

• 11:41Speaker 04

  That's super interesting. Is he? And that kind of goes against something that I've heard you speak about and talk about, which is that, you know, humans using AI well should be much stronger than AI on itself or humans without AI.

• 11:52Azeem Azhar?

  Yeah, I mean, that comes from, you know, my my strange belief that somehow we can still inject something into this process. You know, what Eric described is a phenomenon we've seen in other studies, in in knowledge work, in professional work, where people below average improve, but people at the top somehow get worse because they turn down the suggestions from AI. But but I think you get to this new level, this new level of capability, because you are using a tool, where I think you can get much more done. And what I've noticed, I mean, there's a guy called Andre Kapathie, who's one of the great deep learning engineers. And Andre is a much better software developer than any of us will ever be.

• 12:33Azeem Azhar?

  And he has handed so much of that over to his his AI systems. He is the ultimate experts that Eric has, you know, alluded to. And, you know, his response is he's getting more done, he's pushing the limits much, much more. And I just wonder if we have a U shaped curve here, where if you're below average, you get improvement. If you're kind of pretty good, top quartile, you might overthink. And if you're really exceptional, you're able to master this difficult machine.

• 13:02Speaker 02

  Let me let me jump in here for a minute to say, I I'm going to be the philosopher for a moment and say, like, what is it that we are measuring when we're actually trying to measure, right? So it's easier in Eric's area, right? Like, did you get the diagnosis right or wrong? And like, is the patient better off in terms of health and treatment, or are they worse off? In a lot of other areas, when we're trying to measure output and say people at the top are doing worse. Or people in the middle are doing better. Like, I'll take it from the educational perspective. If you measure, like, is the essay that the person turned in better or worse? The essay may be better with either pure AI or AI plus human. Is the learning outcome better? The answer is almost always no, right? So the human is worse off, but the essay is better.

• 13:46Speaker 02

  And then if you look at that collectively across humanity, the question is, is humanity better off? Because we have more, you know, AI-produced polished essays? Or is humanity better off? Because we have, you know, people who don't learn? And there's a big gap between what it is that they can create as output versus what it is that they understand, absorb, and do well at. And I think the problem in a lot of like, looking at all these studies of offloading and trying to figure out like, who's doing better at the top, at the bottom, in the middle? Is there a U-shaped curve? Is because we're not measuring apples to apples across this. And we're oftentimes, I think, not asking the right questions for the right context to say, what is the impact of the offloading?

• 14:26Speaker 04

  So the societal solution, then, Nita, should be that in things that are super consequential, like a medical decision, right? Like, should we do surgery? Should we not do surgery? We should have the AI make the choice. But in stuff that's like, not so consequential.

• 14:40Speaker 02

  Well, I mean, ask the question in the short term versus the long term, Nick? Which is, like, carry that out for a while. Now you have physicians who no longer have intuition about whether or not to do the surgery or not do the surgery. You have people who no longer can interrogate code because they don't actually understand it. They're doing vibe coding and they're pulling the, you know, AI slot machine for different output rather than actually being able to critically evaluate it. And over time, you know, are we better off or worse off? And I think if we're looking at it that way, instead of just looking at today, do you make the decision to do the surgery or not do the surgery? I worry in the long term that we end up in a much worse position than in the short term.

• 15:23Speaker 04

  There's no one in the world I enjoy arguing more about with ethical hypotheticals than Nita Varhani. So let me ask you this one. So are you saying that there are specific cases where there could be life-threatening surgery? Where the AI would make a better decision? And we know it. It would get it right more often than the human. We should still prefer a society where the human makes some of those choices and some people die because at least the humans will be smarter and be able to make better decisions in the long run, even if poor Aunt Marge goes on.

• 15:50Speaker 02

  Well, see, you know, one of the things I love arguing with you about Nick is that you set up these hypotheticals that the answer is not, you know, I can't answer your question directly because you created an artificial construct. That isn't what I think the right construct is at all. The right construct is today, should you use AI tools to make the life-threatening, you know, consequential decision? Yes, you should. But what is it that we as a society need to invest in such that in the long term, we don't end up with humans who are no longer have basic competency to be able to navigate or interact with AI? And we've delegated everything to AI. Like, how do we do both things at once? That's the real question to me. Not that you don't use AI, but what is it that you have to invest in to ensure that people have what I'm calling constitutive competency over time?

• 16:36Speaker 04

  Eric, you want to jump in?

• 16:37Speaker 03

  Maybe one, one thing to think about here is that, like, you know, we are, there's a set of tasks, a set of questions that were asked to the AI as well as a bunch of doctors. We saw the outcomes, you know, this work radiology, maybe diagnostic accuracy, a few of these different things, and you saw the outcomes, and you're saying like, this was right, this was wrong. Two things strike me in this. Number one is that, like, I don't know whether the physicians who worked with AI are actually any good at working with AI, in the sense that any consequential decision I have made in the last three, four years has been with this aid of AI. Sometimes I've agreed, sometimes I've disagreed. But over time, you build up this sense of when is it accurate and when is it not accurate? And you do need to kind of get some sense of its reasoning and saying, oh, yeah, it's taking

• 17:22Speaker 03

  the wrong data into consideration, so I should probably not agree with it, or it's jumping to conclusion too fast. So I need to actually press it and give it a lot more information because I'm not sure whether it understands what I'm even asking. This is not jeopardy, right? I mean, expert gives one answer, AI gives one answer, we just kind of choose one. In the real world, if you're talking to a doctor, there's a time constraint within which the doctor has to respond. You want to be as prepared as you can going into that so that the time is better spent. Doctor makes a wrong judgment call or decision, which happens all the time. Then there are multiple checks and balances along the way where somebody might look at that and go, you know, oh, yeah, maybe, I mean, should we should we actually operate? All of these things exist as checks and balances in the real world, which are not taken into account in the

• 18:06Speaker 03

  paper hypothetical. I'm not saying it is the paper. It's a great paper, but like it is important when we kind of look at it. We understand how to use the data that's coming out of it.

• 18:14Speaker 04

  Eric, can I take a thread from Rohit and ask you there? So one of the things that maybe the case is maybe the anomaly that you noticed is not something fundamental. It's just something about this moment we're in right now, where we don't actually understand how to use AI. And some doctors don't really understand it and they're confused and that once we get better at it.

• 18:33Speaker 03

  Right.

• 18:34Speaker 04

  You know, once we really learn, it'll be okay.

• 18:37Speaker 01

  Yeah, that's the hope, Nick. I think the medical community likes that explanation the best. But I don't know for sure. You know, I I know some of colleagues who think the AI is going to prevail. The other issue in medicine that we've just seen the first confirmation of is de-skilling. So now we've seen a study of gastroenterologists who let the AI find the polyps. And now over time, when they turn off the AI, they're not as good as finding the polyp. And of course, there's the younger docs that are coming through, who are going to be never skilled. So these are other competing issues, but you're right. I mean, the hope is that over time,

• 19:19Speaker 01

  this will settle out that the hybrid that everybody expected to be the the best will will be the the winner, but we don't know.

• 19:26Azeem Azhar?

  You know, I think, sorry, Nick, I think de-skilling is is is real. And what Nita and Eric draw attention to is a gap or a shift or a change in in education. And today, in these early stages, where there are so few practitioners, it's up to us to figure out pedagogy around how we maintain our skills. So I'll I'll share one of mine, which is this. This is my my fountain pen and I spend a lot of time writing in fountain pen with the computer turned off and thinking about the experiment Nita ran with her her students, you know, yesterday I had a session. With one of my colleagues where we jointly worked on a problem in silence, on paper, no computers, no phones for a couple of hours and then reconvened after that. And that's something

• 20:10Azeem Azhar?

  we do in my team quite a lot because maybe it's the wrong thing to do. Maybe it isn't effective, but it's our gesture at understanding exactly this, this de-skilling, this sloppy vibing that has you moving further and further away from, you know, your own mental capacities and what you can try to achieve.

• 20:27Speaker 04

  But aren't you also doing the opposite, Azim? I mean, you've been writing about how you're trying to get your AI to write for you and how you're trying to get it to be a better stylist. And like, writing is thinking, right? And if we de-skill our writing, we're de-skilling our thinking.

• 20:39Speaker 02

  So Oh, I'm gonna push back on that one for a minute, Nick. I've heard that too many times now and I think it's crap.

• 20:44Speaker 04

  Crap?

• 20:45Speaker 02

  Yeah, here, let me tell you why.

• 20:47Azeem Azhar?

  A quick note. If you want to support us in bringing more of these conversations to the world, please consider subscribing to the show.

• 20:55Speaker 04

  The writing is thinking or the?

• 20:57Speaker 02

  No, that, that, that, like, that's not everybody, right? I mean, so, so, for example, for me, when I write, I actually give a talk before I write it out. I think in public speaking more than I think in written form. And then I, like, I think in a keynote form of like, how do I tell this story? And then once I have told it and interacted with people a number of different times, I can reduce it into writing. My thinking doesn't happen primarily through writing. And I do work out some nuances and details in writing, but that's not where my primary way of thinking through a problem is. And I think that's one of the problems is that if we don't start to actually recognize that each person's generative capacity might happen differently. For Azim, it

• 21:42Speaker 02

  might be writing in fountain pen, or it might be talking it out with a friend, or maybe it's painting on, you know, a canvas. But how people do their thinking comes in different forms. And I think once people figure out where is my generative constituent of competence live, and that's the thing you're trying not to offload and preserve so that you're still a thinking human. I, I think we move away from this idea that, no, no, no, you can't, you can't have writing be replaced. Maybe for Azim, that's not where he does his best thinking.

• 22:09Speaker 04

  So then your prescription or your advice to individuals would be figure out where you do your best thinking, right? Maybe for me, it's for writing. Maybe for Azim, it's for painting. Maybe it's Rohit while he's, you know, sailing.

• 22:20Speaker 02

  Sewing or, I mean, who knows, right?

• 22:22Speaker 04

  Then take that thing and don't de-skill yourself there. Don't offload that.

• 22:26Speaker 02

  Yes. Preserve your constituent of like the core thing that enables you to form your capacity to be flexible across different areas, novel situations that you encounter. Don't offload all generative thought.

• 22:40Speaker 04

  So you're not going to offload speech creation, but you might offload writing?

• 22:44Speaker 02

  I mean, I, yes, I will offload some writing, like my emails. If you get an email from me, Nick, it's probably an AI agent or it's probably, you know, AI. I don't, I don't like to write emails at all. If I write something that is a recap of a lecture, that might be AI. In fact, it often is AI that I will use to recap my lecture. But if I'm writing something that is a keynote or a novel idea, I'm not going to use AI to do it because that's how I do my thinking is to actually figure it out in spoken form, then reduce it into written form. And I won't do the reduction into written form, but maybe I'll do editing. Like instead of having a copy editor that I hire, I might have AI be my copy editor.

• 23:25Speaker 03

  One thing that I might say is that like, you know, the way I look at it is I have more capabilities at my disposal. And by default, when you have more capabilities at your disposal, you end up spending time on what you want to do and you don't end up spending time on what you don't want to do. You know, it's endogenous, right? I mean, I end up spending less time on things I care less about and more time on things I care more about. And that might be different for all of us. Part of it might be writing, a specific type of writing. Part of it might be specific types of research. Part of it might be specific types of public speaking. Each one of these will be different for us. And I am reasonably sure it's different for me today than it was like a year ago or 10 years ago. Cause like we change over time. So considering that, one way that I think about it is less

• 24:06Speaker 03

  as like, hey, what do I want to hold off on before handing it over to an autonomous agent? And more that like, considering this autonomous agent exists, or I can make it exist, what can we do together? And if I get more confidence than it worked, that it works well, then I hand off more. Just like, you know, if you have a really good PA and like you understand what he or she is really good at, then you might say like, okay, you take care of that. I'm great. I will do this other stuff. It's the same kind of dynamic that I find most useful.

• 24:37Azeem Azhar?

  And can I just add personal experience to that? So I've had so much more time in my space of generative, constitutive, constitutive, Nita says it so much better than me, competence. I spend much more time writing, I spend much more time reading, I write lots of my, my essays that go out on exponential view, longhand first, and I read them out, then I get them transcribed. And it reminds me of this idea that these tools perhaps could give us more, more time, right? Which is a concept that Eric first introduced to me a few years ago, that doctors would get time back in order to be, you know, better doctors. And I certainly feel in 2026, I'm doing more good thinking than I was four years ago, when I was essentially battling with an inbox and Google searches.

• 25:22Speaker 04

  This is amazing. Azim, you seem to spend part of your days in like 2100 and part of your days in like 1450. This is a beautiful, beautiful to hear. Eric, do you agree with this framework? And what do you, what do you consider your essential skills that you would never cognitively offload?

• 25:36Speaker 01

  Yeah, I do agree that it's very personal. The one cognitive skill, right? If I'm writing something, whether it's a paper or a blog, whatever, I feel like if I used AI to write it, it would be cheating. It isn't right to share with other humans if it's not my product. It probably would be better if I had AI as the copy editor and making tweaks, whatever, but I just feel bad about that. I just couldn't go there. But you know, maybe I'll move into the 2100 with, with Azim, I don't know. But otherwise, though, I do think it's changing, it's dynamic, it's a very fluid thing. And offloading things more and more, as Nita

• 26:20Speaker 01

  mentioned, you know, things that are just to make your life not just more automated, but much more practical, getting time back. Getting time back in medicine is probably the greatest goal, just because, as Azim mentioned, it's squeezed so badly to try to care for patients that if you can get rid of data clerk function, a lot of the mundane tasks, you can be a much better doctor and the patient doctor relationship can get restored.

• 26:45Azeem Azhar?

  I am curious about one thing, though, that one thing I struggle with, I really battle with, is that I'm still moving at speed. And a lot of great thinking comes from attention. It comes from observation. It comes from chewing it over in your head, in whatever mechanism you use. And many of the examples we've given are in the kind of practical day to day where there is a time constraint. You know, you have to see this patient now, or you have to, you know, ship this for the client by then. But really, really great thinking has traditionally happened with attention and care and self-reflection. And I would also just, you know, in the world that we live in, I don't think I get enough of that. Yes, I'm doing a lot of pretty good

• 27:28Azeem Azhar?

  quality thinking. But am I really doing the quality of thinking I could do if I had 10 days without an interruption? And what do I give up for that?

• 27:37Speaker 02

  So I want to, I want to say that there is a way to use AI for that, too, though, right? And it's not just metacognition of reflecting on a particular issue. I came across this blog post somebody wrote where they described what they called their potato prompt. And I adopted it and put it into Claude for me. Their potato prompt is like, whenever I type the word potato, followed by an argument, I want you to tell me the three ways in which the argument I just made could fail, two counter arguments to the argument, and like some other blind spot that I've missed. And I want you to be a hostile critic, like, like shed all of your niceties and be hostile. And the first couple of times I ran this, I was like, Oh, my God, I'm going to shrink into a little hole and

• 28:18Speaker 02

  cry because it, you know, it's so unkind to you and devastating. But it's made my thinking so much sharper that like, every time I like, throw up an argument, and I'm like, Oh, on self-reflection, that was brilliant insight, Nita, right? And then, like, it's interrogating this prompt over and over and over again. And, you know, yes, I have the benefit of time as an academic to be able to sit around and do that with my potato prompts. You would bounce these ideas off of friends. I still bounce the ideas off of friends. I still turn to other humans and say, Hey, can you peer review this? And can you give me your reaction to it? But that instant, real time feedback to push your thinking is A way in which you can use AI to challenge you and to push you further rather than just to offload.

• 29:00Speaker 04

  Azim, can I go back to the blur that you started with? So you started with this very enticing point that we're in kind of a blurry stage right now. Like, we clearly controlled these tools, or we most, like, we were manipulated a bit by Facebook a few years ago, and there has been manipulation. But like, we were mostly in control, and now it's a blur. Your brand is all exponential because these tools are improving exponentially. Aren't we basically cooked? Like, don't we have, like, a very short window of time where there's a question of whether we're in control? But a couple of years from now, it will be all agentic manipulation of us all the time. And like, we'll have moments where we can disconnect and go into our cave and write with our pens. But things are going to change very quickly. What do you think of that?

• 29:44Azeem Azhar?

  Well, you know, part of this comes from what Nita said much earlier on, which is that we are always making decisions within the constraints and frameworks of the things around us, many of which predate Facebook, believe it or not, you know, the advertising industry, the greasy pole in the corporate job and wanting to move out of the cubicle. So I think we're seeing something that we have previously seen before. There is a deep challenge, Enik, which is that the things that we've discussed today suggest that you can counter some of the negative force of AI. You know, the doctors who misdiagnose, or we can find ways of educating and training people, even though the junior ladder has closed.

• 30:23Azeem Azhar?

  But it does require a lot of self-knowledge, and the four of us, the five of us, really are at the point where we do spend a lot of time introspecting, and we do work with these tools professionally. Do people really have that self-knowledge as a tool within themselves, but more widely? I'm, you know, I'm not sure, and I'm not sure that schools, for example, have necessarily been set up to teach that. You know, if the thing that you need is self-knowledge and a grip of meaning and a certain degree of self-confidence about where you do and don't have agency, I don't think any high school anywhere in the world has thought that that's so critical it needs to be, it needs to be taught.

• 31:03Speaker 02

  Actually, there's this wonderful German school. That has trialed this mindfulness program, this breathing program. And then there's this huge program that is nationwide called Learning to Breathe. And what these programs are designed to do is to develop interoceptive skills in children. And what they've shown is that in doing that, it has a direct impact on things like math scores and other kinds of scores. Because the embodiment and learning to develop the grounding in oneself and the grounding in one's body and to develop those intuitions become very powerful ways of doing so. So there are some examples, I agree with you, it's not the norm, Azim, but there are some wonderful examples that we can draw from and build from

• 31:47Speaker 02

  to start to recognize that in order to change our educational systems, to start to orient toward the felt capacities that are necessary to develop those kinds of intuitions, there are some that are working and that can be scaled.

• 32:01Speaker 03

  So in general, I disagree with the notion that like education can get us out of this. Educational systems were set up to do something fairly specific. And we tried to push and prod them and to try and tell them like, hey, you need to teach the kids to do meditation better, or you need to teach the kids to how to use AI or how to use computers back in the day. The kids who learn how to use these tools better are primarily the ones who've just spent a bunch of time and tried to do stuff with it. That was true of the internet, that was true of mobile, that's true of computers and programming in general, as opposed to saying, here's a structured course whereby you actually use it. And specifically with AI, I think it's exacerbated because ultimately it is the easiest thing in the world to use. Because you open up something and you type something in and you get something back.

• 32:45Speaker 03

  Like my eight-year-old accidentally discovered Canva as an AI, which I didn't know. So he discovered it and he's been making little website stuff for himself based on dinosaurs and different things that he cares about. Nobody needed to tell him because it's a text box, right? He knows how to type stuff in and he can actually respond back and forth with it. I feel like that's gonna get Blown up. I mean, that is the thing that is required. We need a lot more whimsy and people, we need people to actually play around a lot more, which is probably not something that education system can solve. Maybe we can solve it individually, but I don't see that there is a systemic solution to any of this.

• 33:24Speaker 02

  But Rohit, there's a difference.

• 33:26Speaker 03

  Sorry, go ahead.

• 33:26Speaker 02

  Can I just, can I just ask you and push back on that for a minute?

• 33:28Speaker 03

  Please.

• 33:28Speaker 02

  Which is, the example I was giving was not education in the way you were just describing it, right? Was, so it was not education about like, how do you use the tools better or anything like that. It's recognizing that our educational systems primarily have focused on output, rather than how do you build like, kind of core competencies in individuals. And the reason that I think the breathing program is, is not about education traditionally, it is recognizing that there is something eroding in human capacity for thinking. And if we, instead of focusing on output, focus on some of those constitutive elements like embodiment and how to actually develop intuition, for example. How you use tools comes later,

• 34:11Speaker 02

  but whether or not you develop those kind of core capacities for thinking is something that the educational system could do better. And I'm just, I'm, I'm wondering if you think that that's possible.

• 34:21Azeem Azhar?

  I'll be the relentless optimist on, on that point, because we've not even tried to do that in a consistent coherent way over the last thirty or forty years, which is why we see, you know, in the professional services, you know, lawyers who have terrible patterns of how they interact with their clients. And of course, the bedside manner of doctors varies enormously from the ones who you like to the ones who just freak you out. And we've got here without paying any attention to those types of values or skills or how to cultivate them. And perhaps now we'll be forced to because of the, the nature of this technology, but it, it seems like it's more of an opportunity than a, a deep, deep problem.

• 35:02Speaker 04

  Eric, when you look at people going into the medical profession and you look at the skills they need for a future in which AI will be ever present and everywhere, do you think we're training them the right way? What do we need to do differently?

• 35:15Speaker 01

  Oh, we're definitely training. Them the wrong way. They're first selected by their test performance scores. AI can do better on that anyway. And their grade point average. And they're not for their interpersonal communication skills. That's a problem. And then as far as the training per se, there isn't one medical school out of one hundred and sixty now in the US. It's actually has AI in the core curriculum. So it's almost like it doesn't exist. It's like in a denial state. So whether it's the selection or whether it's the education, we haven't even acknowledged that it's the transformation of medicine in the future.

• 35:54Azeem Azhar?

  Eric, to what extent with a doctor today are we looking for that high reliability to the differential diagnosis? Which is actually a kind of mechanistic algorithmic thing, right? Then mechanistic algorithmic things are what machines do better than than people. So we've needed to do that because the knowledge was really scarce. So it's hard to get that knowledge into an intelligence, into a human skull. And maybe now what we're able to do, because that knowledge can be offloaded, is develop the other part that delivers patient outcomes, which is about, you know, this kind of constitutive awareness that's embodied understanding of what it is to be a person and a person in the world.

• 36:37Speaker 01

  Yeah, absolutely. I think the diagnosis accuracy will be one of the really big jumps that we'll see because of what you're saying, Azim. We already have some good indicators of that. But it's relying on the right way to deploy AI, which we've already mentioned, you know, has some of its issues at the moment. But eventually, I do think that's absolutely going to be how this plays out. Accuracy is going to be one of the greatest gifts of AI in medicine for that reason.

• 37:09Speaker 04

  Can I, I want to ask a big question of the group. So we all seem to agree. We're all pretty optimistic about AI benefiting us, AI benefiting our fields, different degrees. And we're all intensely aware and reasonably concerned about this descaling offloading process. We're about to enter a stage. Where AI is going to become much, much more powerful. What is the thing you're looking for? Like, what is the bad thing that could happen or the good thing that could happen? That will put us on a better track with our relationship with AI. What is like, what is something that? Either will go this way or that way in the next year, but we're not sure about.

• 37:43Speaker 02

  Eric said that he feels like he's cheating if he puts something into the world that, you know, AI has written. And that's the tipping point that I'm actually looking at, which is there, we've been talking a lot about. Like selves, the building of selves, the capacity of selves. What that ignores is that we, we exist within an ecosystem of information, of content, of. Things that are produced, like think about that as like the corpus of human knowledge that is out there that we draw from and. The more that becomes synthetic. And AI output or slopification or and shitification, or however people are describing it, right? The less there is like felt human experience in the things that are out there.

• 38:27Speaker 02

  I'm worried about that the things we're ingesting. Start to erode us in other. Different ways. And so the kind of tipping point that I'm looking for is. You know, there are people who are out there trying to track like how much of the content on the internet is now produced by AI? How much of it is produced by humans? How much of like the corpus of what we're drawing from is human and how much of the corpus of what we're drawing from is no longer human? And what's the impact of that on our ability to to think? Eric was thinking about it from the kind of perspective of contributing to that corpus and that he feels like he's, he's wronging us all if he does something that contributes to it. That's not. Fully human.

• 39:06Speaker 02

  Most people don't feel that way. And so what we have out there is a lot of non-human content and we have no idea what the impact of that is. On like brain rot of kids or anything else like that. So I'm looking for that tipping point of like both what happens? And there's some really interesting early studies that are looking at like, how do humans react to AI created music versus non-AI created music? How, what is their emotional response to art that is created? What's the impact of labeling of that information? Can they tell? The difference or not, I think it's going to be a lot harder to track and more subtle than that to to actually understand what that does to us.

• 39:43Azeem Azhar?

  There was a moment in the in the atomic weapons race when you can actually measure when it started in geological or atmospheric measurements because there's not a certain amount of radioisotope before and then there is afterwards. And I think that there is this this moment that we'll start to see as more of us get exposed to text that is written by AI. That's what we will see as text and start to to follow up. And if you're still on X now, what you will see is a very large number of these essays that appear on X. They're no longer 288 characters long and they're being produced at a rate of knots by very busy people who aren't writers. And I'm, of course, I'm reading them because there are insights and

• 40:25Azeem Azhar?

  there's data and there's analysis and argument in there. So you then get exposed to that. And I am quite curious whether that that is sort of what Nita is identifying, which is what looks good in our minds starts to change because what we're being exposed to constantly. And, you know, perhaps that's how our learning process will work and we won't notice it's happened.

• 40:45Speaker 04

  But what if the AI gets good and the AI slot becomes better than what the humans put out, right? Like, what if the average writing quality of Twitter is higher in 2028 than it was in 2023?

• 40:58Speaker 02

  Yeah, it's not. It's not about whether it's better or worse, though, right? Like, the reason Eric feels like he's cheating to put it out into the world isn't because, like, he recognizes it might be better in terms of copy editing, like, it might be more precise in grammar, it might even sound better, and people might even react to it better. And that's, in fact, what the science shows is that people react to AI-generated information more positively. But there is something subtle to Azim's point about like, when it when it shifts, we don't know what that does to what it means to be human. We don't know what it does to have information that is not created by felt and bodied humans that we're feeding on. How that subtly changes us.

• 41:38Speaker 01

  Yeah, you described it better than I could. Thank you, Nita.

• 41:42Speaker 03

  I may just add that like the average quality of X post is higher with AI slop today already than it was in two zero twenty three. But what we are realizing collectively is that despite the average quality being higher, like Nita said, people dislike it a lot more. There are people who kind of still push it. I use Pangram quite a lot to detect sort of AI slop under. It's everywhere. It's not just in Twitter, right? I mean, like a journalist, like a lot of articles that are getting published are basically AI slop to sort of some extent. And I think slop here is an important word because primarily what we're saying is that it has all the appearance of a fulfilling meal, but it has no calories. It's kind of empty.

• 42:23Speaker 02

  I call it hollowware.

• 42:25Speaker 03

  A couple of points in it there. It is fairly hollow. And, you know, to me, that is an example of like AI doing something really badly. Like if it it can't write a really good article just yet, but what the things that it can do really well, like create the apps that a Zim uses every day, for example, or that I use, that is a functionality that is a capability addition to me. Whereas if I got AI to write one of my articles and then I had to rewrite it, that's a whole lot more work.

• 42:52Azeem Azhar?

  Yeah, and I would say that it also tells us something about how mechanized, how much through Adam Smith's division of labor, a lot of the production of words has become. Because, you know, many things like certain types of research reports, even certain types of academic papers, let alone the text on many marketing websites, are essentially produced by algorithms that were being carried out by humans, a bit like the female computers' calculators during the early days of NASA, right? Executing an algorithm that we now hand over to a computer. But but the real, you know, for those of us who are writers, the real writing happens somewhere else, and it's extremely difficult. I mean, all

• 43:32Azeem Azhar?

  my efforts to build an AI-enabled style guide, I'm sure The Atlantic has a style guide, many newspapers have a style guide. I wanted a a style guide that used AI tools to help. People adhere to the way I I write across my surfaces. It simply cannot replicate what I do, except in certain trivial ways, like the fact that I prefer Germanic roots to English words and Latin roots, the fact that I write ninety one percent active sentences and I vary my sentence length in paragraphs. You know, those are very, very surface because what's missing is what pops out of me and my embodied and lived experience.

• 44:11Azeem Azhar?

  And I guess that the challenge that we we might face is that as these tools become pervasive and many of the experiences we get, whether it's on Netflix or TikTok, or frankly, the products are in the shops, are mediated through those tools. My experience is going to be that much more blunted and therefore much more aligned with what an external set of AI systems are doing than what I might have done had I experienced them a little bit more raw.

• 44:40Speaker 04

  Nita, can you explain a little more what you think might happen when we cross this point where the majority of information we take in was created by AI or when ninety percent of the information we take in was created by nothing from the Atlantic, but of all the information outside of the Atlantic?

• 44:54Speaker 02

  Is that true still, Nick? Nothing from the Atlantic?

• 44:57Speaker 04

  Certainly, we have a rule against it. The same rule I had with my book, which is not one word written by AI. I used it, you know, constantly as a tool to help me think through things and to analyze transcripts and to, you know, measure the chronology, but not one word. And that's the rule of the Atlantic. Now, whether writers ever written word, I don't know, but certainly we care a lot about it, in part just because we want to hold this line for brand identity, right? Like, we were founded by Ralph Waldo Emerson. Like, you can't be founded by humanists and then have AI write your essays.

• 45:25Speaker 02

  So I've been, I mean, so I've been looking at this and and kind of all the studies that are out there so far, and in particular, trying to look at meaning making for humans. And, you know, it was interesting to hear Eric's reflection because I think the impact will be more on the contributor than on the receiver for what the impact will be for humans, which is. Your contribution to humanity, to that corpus of knowledge, it turns out is part of like the act of becoming, like the brain needs that act, like through architecture in the brain that we can get into if you want to. But like the, the act of contributing itself is the act of meaning making. And when it gets replaced by AI that does it,

• 46:10Speaker 02

  you don't experience that same effect. And so kind of meaning and purpose for the contributor looks very different. The receiver right now doesn't seem to experience it that differently because if it's not labeled as such, sometimes you can feel the hollowness, but sometimes you can't. Sometimes it looks identical and increasingly as humans start to write more like AI, you might not feel the difference, but sometimes you can. I think it'll be more pronounced for each of us individually contributing than it will be necessarily for the receipt.

• 46:40Azeem Azhar?

  I disagree with that perspective. And I, because that idea that you're not contributing if you do that work.

• 46:45Speaker 02

  Oh, I'm not saying you're not contributing. I'm not, I'm not at all saying that. And in fact, like, I think you can inject humanity using AI tools. If what you're doing is saying, like, I am working iteratively with AI to bring my voice and my human and lived experience, but I'm using this as a vehicle to channel that. I think that's different than the wall of text that somebody posts. Like, they get to a LinkedIn post and then they generate a response and don't even think or put themselves in it. There's a qualitative difference between those things.

• 47:13Azeem Azhar?

  There's a qualitative difference. But you know, when I build something for my team, I'll give you an example of something I built today. So I've written millions of words. And what I wanted to do was, was take out our house view, like, what do I believe about certain things? And essentially, I had the AI run across all these words to fill at a concept level, what are the 10 or 12 beliefs I've had and how they shifted over time. And I was then given something that I personally reviewed and then shared with the team. The bulk of that work, the human work that I spent an hour and a half would have been a thousand hours. So I did less than 0.1%, 0.2% of the work. But I still had the sense that I was gifting something that was useful and usable and also distinctly, distinctly me.

• 47:58Speaker 02

  One of the most influential books for me in life is the book Flowers for Algernon. And I read it way too early. I was in like second grade when I read it. And, you know, crying as I'm reading it, not, you know, I have no clue. I'm a seven or eight years old at that point. Now I look into the history of it and it is fascinating. It took him 14 years to write it. And it started as a medical student. Later, he has an encounter with somebody who in fact has intellectual challenges and wants to be in kind of special education. There's all of these felt experiences that go into his writing of this book. Is that different that he brought 14 years of experience and iteration into it than the hour and a half you spent instead of

• 48:40Speaker 02

  the hundreds of hours that you spent in both what you experience and but what I experience as the reader from what you bring to the table? It may be, we don't know, right? And that's what's changing. We have to acknowledge that that there is something qualitatively different that has happened in the 14 years of writing that book, maybe. And what impact it has on me and what impact it has on him and the act of becoming than the hour and a half you took to boil down your thoughts?

• 49:09Speaker 03

  Rohit, it seems like you want to respond. We just have two minutes left. So both respond, deflect, destroy, augment, and help us sum up. Oh, I'll be very quick. All I would say is like, I think anytime AI gets good enough, we start seeing it as a tool because it becomes dependable. When it's not good enough or when it's at the edge, we start looking at it as a destruction of some intrinsic humanity that we hold ourselves very dearly to. We have done it in every field so far, whether it's in like radiology, finance, supply chain, whatever. So art is a rather unique human endeavor. I don't want to start upon it as the kind of first thing that will go. I don't think that's

• 49:51Speaker 03

  true. But whenever it gets good enough to actually do a piece of work that we ourselves wanted to do, but now we can depend on us to do it. We have generally been pretty happy to hand it off and then say, we now know what, you know, this little box will do. Let it go do it while I go do other things and maybe write way more books like Flowers for Altonon.

• 50:11Speaker 01

  You know, it's a little bit like Quebec and the rest of Canada. That you have to have deliberate intent to preserve your language. Otherwise, it's crowded out by AI. If we don't have deliberate intent to keep AI as a tool, which is where I think it belongs, we'll lose a part of humanity, which is essential.

• 50:31Azeem Azhar?

  There is so much in here, and I want to just pick up on what Eric described, which is that deliberate intent, that deliberate intent about our own capabilities, that deliberate intent that these things are tools. And however we interrogate them, how we anthropomorphize them, they ultimately still are tools and should stay as such. You know, today's question is about the ongoing process of self-reflection. The deliberate intent is so important that understanding how to maintain the self is really, really important. I opened the session by saying that together we can see what none of us can see alone. And I really think that with Nick's help, we've just proved that. So that's been the first of AI vistas, or perhaps AI vistas.

• 51:16Azeem Azhar?

  I have to figure out how I'm going to pronounce that. I don't know where the vistas came from, it just snuck in. It won't be the last one. We'll be back with a new question with new people in the room, same premise, real experience, different viewpoints, and the view that you can only get when we all meet together. Now, if something landed for you today, you'll be able to send the output of all of this to someone who needs to hear it. We're also going to turn today's exchange into a written synthesis and a video on YouTube. Now, if you're not already subscribed to Exponential View, this is a great moment. Exponentialview.co And I also recommend that you follow Eric, Nita and Rohit on Substack.

• 51:59Azeem Azhar?

  And Nick, of course, buy The Atlantic subscription as soon as you can. Thank you, everyone, and have a wonderful weekend. Thanks for listening all the way to the end. If you want to know when the next conversation is released, just hit subscribe wherever you're listening. That's all for now, and I'll catch you next time.