An Unlikely Team Tackles AI Brain Rot

New AI paper on cognitive viruses

0:00There's a new AI paper that's making waves. It's called Large Language Models as a Cognitive Virus. Now, it's drawing attention in large part because of who the authors of the paper actually are. It's not technologists, and it's not social scientists. It's actually a group of applied mathematicians, largely from the fields of biophysics and complexity theory. And if those sound familiar, it's because that's the same group of fields that included the Ian Malcolm

0:31character from Jurassic Park. Now, what they do in this paper is they set out to model mathematically the spread of LLMs through our culture as well as the impact of LLMs on our ability to think using something known as a nonlinear dynamical system. Now, if you don't know what that means, you're in good company. Most people who tried to read this paper didn't know what that meant either. But I want to look closer at it with you today and explain what's really going on,

1:03because when you get through all the complex mathematics, what you find is actually an elegant identification of the three primary interventions that us as a culture have available right now to prevent our brains from becoming completely dependent on this technology. So we're going to go through the paper. I'm going to cut through the details. We're going to get to these interventions. We're going to discuss each. I'll tell you which one the authors of the paper like best. I'll tell you which ones I like best. And then I will leave

1:35you with some concrete advice for actually starting to take action on the intervention I like best right away. All right. So we have a lot to get into. And this is an important conversation because when it comes to the frontier labs creating these LLMs, as Ian Malcolm told us, your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should. All right. So let's get into it.

Modeling the population dynamics

2:01All right. So I want to tackle what's going on here in three parts. Let's start with the basic question of what is it actually that this paper is trying to model? I'll bring it on the screen here for a second to show you the key equation. So here's the paper. As we scroll down, we get to what are the key equations for the world that they are going to model. I know this looks a little bit complicated. So for those who are watching instead of just, who are listening instead of just watching, there's three equations. One is the DU over DT. Another is DC over DT. And another is

2:36DD over DT. Those are derivative symbols. What's actually going on here? I'm going to try to explain it. So we can take this off the screen here for a second. There's three key variables that this system of equations is using that describe the world. One is U, one is C, and one is D. And all this captures is U is the percentage of the population that doesn't use LLMs. C is the percentage of the population that use LLMs, but also are still capable of independent thought on their own. And D is the

3:13percentage of the population that is dependent on LLMs and have largely lost the ability for them to think for themselves. They've offloaded all of this to LLMs, right? So they've divided the population in this system into three parts. Now, what they do is they're going to use these equations to simulate how those proportions shift between those three categories over time. So they'll go from time step to time step, updating how many people are in the first category, how many are in the second,

3:46and how many are in the third. So it's population dynamics. How does the population evolve over time? So let's go back for a second here, Jesse, to the equation. So what we're actually looking here when we look at these equations is when you see like right here, um, only that color change. All right. So when you see right here, D U over D T, that's capturing, um, how much is the variable U changing? What is the rate of change of the variable that describes the amount of people, the proportion of

4:19the population that doesn't use LLMs? And then they have some equation you use to calculate in this time step. How much does that change? The second equation captures what's the rate of change in this current moment of the second group, the group that uses LLMs and, uh, uh, but it's not dependent on it. And then this does the same thing for the third group. I'm going to, uh, let's bring this down for a second because I'm going to shift over, God help us all, uh, to my drawing tablet here. And I'm going to try to illustrate how you use these equations to actually calculate this simulation.

4:55All right. So what I'm going to draw on the screen here is, uh, the three sets. So let's see, here is U. So we'll assume U, I draw in kind of a big circle for it. So we'll kind of assume that's big. And then maybe when we start the simulation, C and D are very small. So we start with some proportion of people in these three categories. Now, what we do is each of these categories becomes the input to those equations we just saw. So I'll just write the, you know, the word EQ over

5:28here for equations. So each of those equations, the input to it is these proportions at the current time step, how many people are in U, how many are in C, how many are in D. And what the equations tell you is how much each of these groups should change going into the next time step. So what we might get is according to those equations, it might say U gets smaller, C gets much bigger, and D stays the same. So it's the output of the equation tells you how much to change each of these set sizes for the

6:00next time step. And then we could just repeat. Okay. Now we have new sizes of U, C, and D. We feed those as input into the equations, and it will give us how much to change them for the time step down here. And then we do that again and do it again. All right. So we can bring that down for now, Jesse. So what we're going to get then is a way to go step by step. We keep updating the size of these proportions, non-users, competent users, independent users. And then we go the next time step and update it again, go the next time step, update it again. So over time, if you keep updating these proportions

6:34using these equations, you see the proportions fluctuate until if it's a stable system, they'll eventually stabilize to a sort of steady state, whereas you keep updating them, they stay about the same. That's all they're really doing here. Okay. So that's how these simulations work. Let's go back to the equations one more time, Jesse, because there's a key part of these equations, which will be critical for the discussion we're about to have. So if we look at the equations themselves, we do see when calculating these change values. As I mentioned, the current proportion in U, current

7:06proportion in C, current proportion in D, these show up in the equation. But there's these other things that show up as well. These Greek letters, like there is a lambda, there's a rho, there's a kappa. So we have these variables. Here's a mu over here. Here's a sigma. So we also have these Greek letters showed up in these equations. What are those? Those are fixed values. Okay. So those are fixed values that you have to fix as the experimenter before you run the simulation that we're talking

7:38about. These are critical because each of these fixed values capture something about the world that is being simulated. So if we go back to the equations, I'll show you briefly. This lambda, for example, captures the rate, you know, the pressure to move from a non-user of a LLM to a user of the LLM. So as you change that variable as the experimenter, you're changing how much pressure in your simulated world a non-user will feel to switch to using LLMs. We have rho over here is the disengagement rate. So at what

8:14rate do regular users go back to becoming non-users? Down here, we see mu. That's the rate at which a regular user becomes a dependent user in which they no longer really think for their own self at all. Sigma is the rate at which dependent users go back to becoming regular users. So they start thinking for themselves again. Kappa captures the degree to which being around other non-users gives you social pressure to not use LLMs yourself. Okay. So these are the, those values capture the

8:46dynamics of the world and they're just set by the experimenter. So, so what do you do? You know, what do you do if you're going to run these experiments is you set those values and then you run the experiments to see what happens. And then you change some of those values. So you change something about the real world in your, that you're simulating, and then you run the simulation again and you see what the system does. So you're really looking at when you, that, when you use these sort of dynamical systems and you simulate them, you're looking at how changing these different numerical

9:17descriptions of your world impact what ends up actually happening. Now, this is not new. Epidemiologists and biophysicists, they've been in complexity theorists, have been using these type of dynamical systems and exploring them for, for a very long time. So you could use a similar system and people do, for example, to study viruses spreading, like what's going to happen with COVID. This is how epidemiologists were trying to predict what would happen as COVID spread. And in those models,

9:47like in a standard virus model, you have similar equations and you still have three groups, except for now the groups are not infected, infected and actively contagious, and infected but recovered. And then you run these simulations over time to see how those group change. And it matters because, for example, early on when most people are uninfected and there's a small number of infected people. The infected group grows really large, probably, if you set your equations up, right? Because there's very little immunity and there's lots of people that could potentially get

10:18sick. But for example, over time, if the group of recovered people gets really large, then you're going to see much less growth of the infected people because there's just many fewer people that are left to infect and so on. So we use these type of dynamical systems to measure lots of different real-world systems. And what really matters is the equations, but mainly also these hardwired values you put into there. So again, if you go back and look at how they use systems like this for COVID, you know, one of the hardwired numbers they might have in there captures the willingness of the

10:50population to do social distancing, for example. And so then you could change that one variable and say, if the population is not really social distancing much, how does this change the way the disease spreads versus which if they are? All right. So that's the whole setup here. These equations describe change at a given time step. And so you can apply them again and again and again to simulate populations changing status over time. And the key thing you change as the experimenter are these hardwired values that describe the world in which you're simulating. I want to take a quick break to

Sponsor break and product highlights

11:23hear from some of the sponsors that make this show possible. Look, fall is here and with it comes the chaos of the back-to-school season. Now, if you're anything like me, when faced with these new time demands, one of the first thing to fall off might be your diet. This is where Factor enters the scene. They make chef-crafted, dietician-designed, ready-to-eat meals that are prepared in only two minutes, meaning that you don't have to choose between real food and convenience. When I'm in a rush,

11:54I know I can still get a lunch I love, like the Santa Fe chicken with peppers and potatoes, that not only tastes great, but is good for my body. Now, this last part is important. With over 175 banned ingredients, every Factor meal is designed around what supports a healthier lifestyle and nothing that does not. So let's eat real. Head to factormeals.com slash deep50off and use code deep50off to get 50% off and one free breakfast item per box for one year while supplies

12:24last until October 31st, 2026. That's code deep50off, the word deep, the number 50, the word off at factormeals.com, deep50off at factormeals.com. See website for more details. I also want to talk about our friends at Caldera Lab. As I get older, I increasingly find myself looking in the mirror and thinking, hey, who's that middle-aged man? The key thing that's changing, the thing that's catching my attention is actually my skin. This is where Caldera Lab enters the scene. They create high

12:55performance skincare designed specifically for men's skin, which differs from women's skin in many key ways. Now, a great place to get started is with The Great, a lightweight anti-aging serum that's clinically designed to help reduce the appearance of fine lines and wrinkles while also keeping your skin looking firmer. This product works. In a four-week consumer perception study of 30 men aged 40 to 60, 97% agreed that the overall appearance of their skin was visibly improved. So if you're a man who no

13:28longer looks babyface, consider Caldera Lab and in particular their anti-aging serum, The Great. If you want to look and feel like the best version of yourself, go to calderalab.com slash deep and use code deep for 20% off. All right, let's get back to the show. Okay, so let's go back to this paper again. What did

Tipping points and prevention strategies

13:48they find when they ran these various experiments with different values? Well, a core value that they changed systematically was this lambda value that captured what is the pressure, if you're a non-user, to use the LLMs. And they would look at for different values of lambda, how does this system stabilize? And what they found, and I'll show you, if we bring this up, Jesse, I'll show you the main

14:18chart, which is honestly a little inscrutable. But if you want to know their main chart, this is their main chart showing their main result, where you see lambda down here is changing. And here you see how U, what U stabilizes to, which is the people who don't use LLMs. And over here you see what D and that's the dependent and C stabilizes to. I find that chart a little bit inscrutable so we can bring it down, but I can give you the high level description of one of their core findings. They found that when lambda is very small, so there's very little pressure to adopt LLMs,

14:51it stabilizes to a world where, you know, most people don't use LLMs. As you increase the lambda, there's a tipping point. So there's a particular value of lambda, so a certain amount of pressure where you get a drastic change. So before that tipping point, if you use a slightly smaller value of lambda, most people don't use LLMs. You go a little bit beyond that key tipping point, and most people are using LLMs. They're either in the C or D group. So either using and still thinking or using and completely dependent and not thinking at all. And then how, in those experiments,

15:24how the LLM users, so most people become LLM users once you get past that threshold, how they divide among people who still think for themselves and who are completely dependent on LLMs depends on the relationship between those other factors that describe the relationship between those two groups. The other thing they found is something called hysteresis, which says, okay, if you move that lambda just past the tipping point, the system will evolve to where most people are using LLMs. If you then run the simulation starting with most people using LLMs

15:57and move that lambda right back before the tipping point, it doesn't change things. People continue using them. You actually, once most people are using LLMs, you have to reduce that pressure a lot more to get them to come back. So it's easier to get everyone using LLMs in this mathematical model than it is to get everyone using LLMs back to a state where a lot of people don't. And so their conclusion at a very high level is we should look at prevention. If you care about certain outcomes,

16:28you want to prevent it from happening in the first place as opposed to trying to figure out how to get things back. All right. So what interventions do they look at? Well, first of all, we have to know what is it that they want to stop? What is the bad outcome for this simulation according to the authors? And they would say the bad outcome they care about is most of the population ending up in that final category, the D category of people who are so dependent on LLMs that they stop thinking for themselves. They said, that's basically what we want to avoid.

17:02People using LLMs, but still thinking for themselves is not a problem. People not using LLMs is not a problem as well. So they're looking at what interventions prevent the simulation using these equations from ending up with most of the population in D. What does an intervention mean in a simulation? It's changing those fixed values, those values represented by those Greek characters. So what different ways can you adjust those values to prevent from ending up with everyone in D? The interesting thing is that for each of the useful ways you can change those values to prevent the

17:37bad outcome, it corresponds to real world stuff we could do. It corresponds to a real world intervention. So this is where the paper gets interesting. So I've written down here the three major interventions identified in this paper or the ones I thought were most interesting based on this mathematical model. So the first intervention is to try to reduce the pressure to use LLMs to a point that's so low that most people just remain non-user. So if we go to the

18:08equation, that would be making the lambda variable small enough that when you run the simulations, we sort of most people stay in you. Now the authors of the paper say that intervention is not very likely. LLMs are useful. They're already pretty ubiquitous. You know, good luck convincing most people not to use LLMs. So they sort of discard that particular intervention. I don't think we should be so quick to do so. I think there are certain contexts where we're actually seeing that exact intervention

18:41being deployed. So it's worth giving serious thought. One of those contexts is universities. Right? So a university is a place where it might make complete sense to say, in this context, we are going to have a lot of restrictions and pressure against using LLMs for a lot of these cognitive activities because the goal of the university education is to cultivate your mind, not to finish assignments as efficiently as possible. So that's an example of a context and when actually that first intervention identified by this model actually makes sense. We also see it

19:13in journalism. So there are certain really well-known publications, because I write for a lot of publications, so I sort of know this, who have very clear rules about LLMs for their reporters and their journalists. They say you can't basically don't use them. Do not write with them. Do not edit your text with them. Maybe you can do some one-off brainstorming with them. But really, when you're sitting down the right, it needs to be a human writing, which makes complete sense if you're a mainstream publication, because the main thing that you're selling is we have professional thinkers

19:46and writings expressing themselves. So that's another context in which you're effectively reducing within that context the lambda value. The pressure to use LLMs is very small because you have restrictions, rules, and a culture there to make that something that is non-desirable. So that first intervention of really reducing the pressure to use LLMs maybe is not a globally applicable intervention, but it makes sense in different contexts. All right, the second intervention

Cultural pressure and youth smoking trends

20:12that this mathematical model reveals is working is to try to increase the pressure to turn LLM users back to non-users. So there's a particular Greek symbol in those equations. It's the Greek symbol, the lowercase row, that captures the pressure of users to come back to not using it. So again, the authors of the paper don't see that as like very reasonable. LLMs are popular. People like them. But there are contexts in which we've actually seen this intervention begin to emerge and that we

20:46should continue to consider it. So if you look at Gen Z, for example, young people, they have a really mixed relationship with these AI tools in large part because they don't like the people running these AI companies. They don't trust them. They don't trust their motives. They find it kind of inauthentic. Like really, if you look at who's most enthusiastic often about LLMs, it's like middle-aged dads. You know, it's the 55-year-old CEO of some media company that's like, this is great. We all

21:18should be using LLMs to write really quick memos. And they're all experimenting on having a good time. The kids are like, I don't know about this. Like what? I don't trust you. I grew up in social media. I don't trust any of you people over in Silicon Valley telling me I have to use something. So that's an example where you could see a cultural pressure grow to retreat back from LLMs to a non-LLM world. Now, if that sounds fanciful to you, we've actually seen in other types of systems you could model as dynamical equations, differential equations. We've seen other systems where this has actually happened. This is what effectively happened

21:49with youth smoking. So the ad council got all this money from the tobacco settlement in the 90s and they invested them in high-end ads like the Truth Campaign that were made to change the culture, youth culture's relationship to cigarettes, to make it seem like cigarettes are a product that these old white guys in suits are using to try to get you hooked to make money and they're exploiting you and are laughing at you. And it kind of changed the culture where kids were like, I don't see smoking as counter-cultural and cool. I see it as sort of I'm being exploited by

22:20the Philip Morris CEO or something, and I don't want to be a part of that. So by changing the culture around youth smoking, you saw a retrenchment, a retreat from youth smoking levels got smaller as people in that culture felt more pressure, cultural pressure to not use it or to step back from using it to being a non-user. So actually that's not a far-fetched intervention. Maybe it's not going to work globally, but something to keep in mind. All right, the third intervention this model

Preserving human contemplation and cognitive fitness

22:52identifies would be to reduce the pressure to go from a user of LLMs that still thinks for themselves to a fully dependent user that doesn't think at all. There are sort of two Greek symbols in that those equations that were relevant, sigma and mu capture moving the pressure to move from regular users to dependent as well as the ability to move from dependent back. So both of those variables

23:22matter for keeping people in regular users who still think for themselves and preventing too many people from ending up in the completely dependent category. This is the intervention that the authors of this paper like best. Like they're basically saying we don't need to keep people from using LLMs. We're completely happy with a world where most people use them so long as we don't have too many people who are dependent and lose their ability to think for themselves. So they sort of identify

23:53that as the best strategy. How would you actually implement in the real world putting up that firewall between C and D? Well, the authors get into it in the paper. They said, oh, you would have to explicitly introduce human contemplation activities and preserve them. Even if there are some activities that an LLM could do for you, you have to have this sort of scaffolding, this cognitive scaffolding that says this type of thinking is important. So like humans are still going to do it. Here's another type of thinking that's hard that humans are going to do it, even if LLMs could.

24:25And by having personal habits and organizational structures around thinking in certain places where humans should keep thinking, you'll be in a world where you're effectively changing those hardwired values so that there's many fewer people actually move the full dependency. So if we explicitly preserve human cognition as an activity that we practice and make a part of our lives, you will prevent the bad case of most people ending up in D. Now, if that sounds familiar,

24:57that's basically the intervention I've been calling for, perhaps most notably in my New York Times op-ed for March about we need a cognitive fitness revolution. We need to treat cognitive fitness, our ability to use our brains to do complex thought, as a tier one skill or health or property that we should all actually be actively preserving and improving like we do with our physical health. We care about what we eat and we exercise. We need to care about what information we consume and exercise our brains. So it's pretty interesting that the mathematical model, the sort of simple

25:31three-part compartment model, as they call it, compartmental model of LLM spreading between these categories, it identified three possible interventions. And the intervention that the authors like best is the same intervention I came up with observationally, just actually working with real people and seeing what's happening and what would actually help it. So this is always interesting to me when you have different types of investigations ending up at the same conclusion. So which one do I think is best? Well, the way I see this, the way I see this is, A, it's cool that

26:07you can use mathematical models to identify interventions. B, I think this is a useful way of thinking. This type of categorical clarity is a useful way of thinking because in the discussion of AI, AI and in particular LLMs, it's muddled, it's all over the place. It's anger at the AI companies themselves or getting caught up on a particular usage of LLMs that we don't like. And in this sort of muddle or just boosters that are like, look, I can make video games with good graphics with just

26:38a prompt and that seems really cool. And it's also muddled. I like the clarity that the mathematics spring into it. There's different types of interventions that work in different places. And this type of intervention might make sense in a university. This type of intervention might make sense in youth culture. This type of intervention might make sense in people's individual lives. That clarity of what are the different types of things we could do? Why might we do them? Where might it apply? Why might it be a problem? Why might it be good? That's the right way to think about it. All right. Then C, obviously I like this final intervention because I've been

27:13writing about it a lot in non-mathematical terms. And the reason why I like that final intervention intervention is probably for different reasons than the authors of the paper. So the authors of the paper discard the first two because they say it feels unreasonable on a societal wide level to sort of keep people from using LLMs. As I just argued, I don't really believe that. I think there's specific groups in which that still makes sense. So I'm not discarding those, but I think they're organizational interventions and they apply for some organizations and not others. I like the final

27:44one because it's something you can do right now. Cognitive fitness, like physical fitness, is something that you as an individual can take action on right now. You don't have to wait for your school district to make a change. You don't have to wait for your employer to make a change, like with those other interventions. You can start implementing those interventions right away. And when it comes to techno selectionism, I'm a big believer and you have to have that personal substrate, stuff that people can do right away to gain autonomy over technology in their lives. The other levels matter, organization, societal matter, but the fastest moving one is what

28:18individuals can do. So I wanted to end this discussion by giving you three quick ideas about if you wanted to implement this final intervention in your own life, that is to build a scaffolding of cognitive fitness in your life that prevents you from falling into the dependent category of LLM users. What are three simple rules you can apply to your LLM use in your own life that will help you get there? I have three. I think about this list a lot, but I want to give you the three I'm thinking about right now. First of all, when using LLMs, your goal should be

28:50to eliminate tedium, not cognitive friction. So if you're coming up to a use of an LLM where, hey, this is going to take a tedious task that I'd rather not do. I just need it done so I can move on to do something more interesting. That's a good use of LLM. You're automating tedium, which has in itself no real benefit. But if you're using an LLM to avoid the friction of having to think, I don't want to come up with an idea. Can I just go back and forth with an LLM and get one?

29:21Writing, that's kind of hard. Can I have it write a draft and I could just edit it a little bit? If you're avoiding cognitive friction, that's a problem. Cognitive friction is not bad. That's you using your brain in difficult ways makes your brain smarter, gets the best results. Do not outsource hard thinking or creativity to an LLM. Friction is good. You want to feel the strain in your brain the same way you feel a strain on your muscle when you're trying to get stronger.

Three rules for maintaining cognitive fitness

29:44If you're automating tedium, that's okay. All right, second piece of advice. Do not let LLMs write on your behalf except in instances where it's clearly marked as LLM output. I do not agree with the trend of let me let an LLM write this email for me because it'll do a better job and it's kind of hard to write emails. Or this memo that I'm sending out to my employees. If I have the LLM write it, it'll probably be nice and clean and it'll save me some time. I think that is psychological fraud. I think when we read text that we believe a human wrote, we simulate the mind of the reader in our

30:20mind. We're actually like interacting with this sort of conception of another entity. If you didn't actually write those words, there's a sort of psychological fraud going on there. I think the rule should be any LLM generated text needs to be identified as such. So if you want to have, for example, an LLM summarized, like here's like the three action steps we should take based on the meeting we just have, you should write an email that says, good meeting, below is an LLM summary of the meeting and the three action points we should have.

30:50And then you can paste that text. Or if your memo is written by LLM, be like, here is an LLM generated memo based on the following sources. So now the human recipient knows this here is not from a person. This was written by an LLM. This means you're going to write more. And that is a key way to keep up cognitive fitness is when I'm writing on behalf of me, I have to do that writing. I have to clarify my thoughts and try to put into clear words. That's a very important scaffolding. If you want to avoid falling into the fully dependent, like the easiest route in the full dependency on LLMs and

31:23losing your human birthright of original thought is to outsource writing to LLMs. Like, ah, that's hard. And then once you do that, you're, you're seeding your brain. All right. Rule number three, read human writing, not AI summaries of human writing. All right. The thing is, if you're reading an article, what is an article? It's a summary of a very complicated reality that an expert came to understood that reality and said, I'm now going to summarize this down to the level of compression

31:55that I think gets the main points across as accurately as possible. Do not add another layer of summary on top of that. That's all. It's already summarized. I had this conversation with someone recently. We're like, why would I ever read a book? I'm going to, I run my books through an LLM to get a summary of the main points of the book. The book is already a summary of the points. This is an author who's an expert on that topic saying, I am getting this down to the level of detail that's enough to have mastery of the subject. In four or five hours, you can be a master of the subject.

32:27I've compressed it for you. I've spent years doing this for you. You don't need to summarize it further. You're over-summarizing. You're over-summarizing because maybe there's cognitive friction in the actual reading of the summary level that the author chose and you want to make your life even easier. So I would say avoid that if you want to, again, if you want to avoid going from C to D, from use to dependency, read things at the level of summarization that their authors have already determined to be appropriate. I would only use LLM summaries where, you know,

33:01it's something, A, it's very long and there's only some points in there that matter to you. Like, here's a massive report and I just need to know these conclusions. You could have the LLM pull that out for you. That would make sense to me. You don't actually need to have all that information. I think that would make sense to me. I guess if you have a terrible communicator in your organization who's sending you these super long emails that are wordy and complicated, but there's only like one point in there, I guess you could summarize that. But otherwise, for the most part, the text you're reading has already been summarized down to the level that the writer thinks

33:33maintains the optimal ratio of maintaining useful information to not having too much. So don't run away from reading human writing. That strain is not a problem. All right. So those are three rules I've been implementing if you want to make sure that you don't fall into that final category. So there we go. It's an interesting paper. You know, there's a lot of controversy around it. I don't know. I think this is just sort of standard. The controversy around it is people saying, how do you know those are the right equations to describe LLMs? And how do you know the right

34:05values for those hardwired values? You didn't go out there and empirically measure those and stick them in. But this is really common with, you know, biophysicists do this type of thing all the time. These mathematical models can capture general population dynamics. And then you mess around with the parameters to understand how does changing this part of the world affect these general dynamics? It just gives you insights about how a world might unfold. They're coarse, but it's a way of mathematically getting at dynamics that, as we're seeing here, we're seeing show up in the real

34:36world. The interventions suggested by this math are actually seen in the real world. So I thought it was a useful paper, but probably my summary is enough unless you really want to try to understand, you know, some of these pages. Like, here, Jesse, let's load this up one more time. Unless, like, I spend a lot of time with pages like this, where it's, this is what complexity theorists do. They set up these differential equations and they try to solve for different equilibrium points, which is actually, like, not conceptually that hard. You're basically trying to find

35:09where does the change become zero because that represents an equilibrium point and you do algebra. So, but you don't need to know all that. I think I got to the key, the key conclusions.

35:20All right, before we move on to opening my inbox, I want to take another quick break to hear from the sponsors that make this show possible, but stay tuned because I have a pretty fiery response to an article coming up next. So let's talk about our friends at Notion. Let's be real. In knowledge work, AI shouldn't be just about going as fast as possible and maxing out token budgets. It should also help you do the important things better. This is why I like what Notion is building. Notion is the AI workspace

35:51where your team's knowledge, projects, and agents all come together in one place. Teams using Notion move faster, cut friction and cost by consolidating tools and staying aligned. Because everything lives together, Notion has the context it needs to answer questions instantly, automate busy work, and keep work moving. It's AI designed to strengthen teamwork, not replace it. With over 100 million users, Notion is trusted by 98% of the Forbes Cloud 100 and 95% of the top AI companies, including

36:23OpenAI, Cursor, Lovable, and more. Learn more about how Notion can support your business at Notion.com slash Cal. Now you've got to type that all in lowercase letters, Notion.com slash Cal to try Notion today. And when you use our link, you're supporting our show. I also want to talk about our friends at My Body Tutor. Do you want to get in better shape but struggle to follow through? Here's the solution. My Body Tutor, T-U-T-O-R, an online coaching program that actually delivers results.

36:54Now, how does it work? When you sign up, you're assigned a coach, and they help you devise a nutrition and exercise plan custom fit to your specific goals and your specific circumstances. And then, and this is the key part, you quickly check in each day on the My Body Tutor app to let your coach know how things went. This accountability really helps you follow through. It also puts an advocate in your corner who can help you adjust your plan to whatever life throws your way. Or maybe they'll notice you're struggling and can offer to help. You need someone in your corner,

37:24and your coach can be that person. So if you want to get healthier, go to My Body Tutor today. And if you do, mention I sent you, and they will give you $50 off your first month. That's MyBodyTutor.com. All right, let's get back to the show. All right, well, because it's a Monday,

Addressing reader questions and critiques

37:42I like after we do our deep dive to open up our inbox to hear what you're interested in. Again, if you have questions, comments, or reactions, you can send them to podcast at CalNewport.com. All right, what are we doing first this week, Jesse? We have a question about a recent article that characterizes your stance on AI. All right, so someone sent a note, actually a lot of people sent similar notes, but someone sent a note specifically that says, what does Cal think about how Jesse Singal describes his AI views in the dispatch? So a different Jesse, not this Jesse.

38:21Jesse Singal. Here's the article. I read this, actually. I'll load it up here in the dispatch. It's called, Will AI Kill You? And then it gets into it. Okay. It talks about me in there. So I figured it's probably worth getting into. I pulled out the key quote. I'll read it off here. So that, I mean, the key quote related to me was the following. So I'm reading here. People who should know better are letting stereotyping get the best of them. The computer scientist and podcaster Cal Newport, a leading anti-doomer, I'm a leading anti-doomer now,

38:56Jesse. I like that. Recently had a column in the New York Times in which he decried the doomers and their strange ideology that certain leading AI figures talk in apocalyptic tones about the power of their technology, he argued, doesn't mean that they've seen evidence we haven't, rather that it's representative of how rationalists always think about AI in these circles. It's taken for granted that AI capabilities will rapidly accelerate and completely transform the world and to talk about it any other way would be considered uninformed. All right. That was quoting me. That's a little confusing for people who are listening. He contracted, now we're back to Jesse

39:26talking. He contrasted this supposedly out there movement with the cooler, chiller stylings of the Chinese government, Jensen Wang, and, quote, American AI leaders who have minimal connections to rationalism, end quote. My only response to the claim that we should defer to the judgment of the CCP on trillion-dollar matters of technology and industry and human flourishing and or extinction is bafflement. As for Wang, it's extremely silly to treat him as unbiased just because he says we shouldn't freak out. But he does have the cool jacket. He's got that cool leather jacket. I told you, someone told me, was it Ed Citron? He told me

40:02what type of jacket that is. It's like a motorcycle racing jacket or something. It's very expensive. I got to wear that. I overheat easily, though. I don't think I could go around in like a heavy leather jacket. All right. So this is probably worth talking about. For what it's worth, I like Jesse, actually. He's a good reporter. He actually does reporting. A lot of people don't do that anymore, like the reading reports, talking the sources. He's a good reporter. It's also surprisingly funny if you listen to his podcast. He's a funny guy. A lot of people have had similar reactions,

40:32though, that Jesse had there. So I think it's worth getting into it to try to explain what's what or what I'm actually trying to argue here. Okay. So again, you can go back and watch my episode from last two Thursdays ago. I got into this in more detail. Last Thursday, I also talked about this. There's plenty of recent podcasts of me trying to get in more detail about this. But as a really quick reminder, the complicated thing about the current AI conversation is that you have these two related movements, which are essentially utopian futurist movements. So you have rationalism

41:08and large parts of the effective altruism movement that subscribe to a utopian futurist belief that superintelligent AI, if shepherded properly, i.e. by leaders of their movements, will bring us to this utopia in which like a chosen few will be able to fuse with this superintelligent AI God, and we can transfer and expand these digital consciousness throughout the stars and the galaxy.

41:38Right. So this is this sort of futurist ideological view that comes out of futurism movements from the 80s and 90s, from which leaders of EA and the rationalism emerged. The flip side of this futurist utopian view is the dystopian fear that if we, you know, superintelligent AI, which they again believe is going to be the Godhead, if it is not shepherded properly, so if these leaders of these movements aren't allowed to do it right, it will kill us all. So it's eschatological, right? We are going to have

42:09this sort of apocalyptic deliverance from superintelligent AI that will either kill us all or give us this utopian future where the select few are going to spread our consciousness stored in digital forms throughout the light cone of the universe, right? So they believe very strongly. Here's what's confusing about this. Most of the debate right now is between people in the same group. So anthropic and open AI are heavily influenced by those ideologies. This is what

42:40I was trying to capture in that column that Jesse was talking about, but has now been extensively reported on by many technology journalists at the New York Times, at the Wall Street Journal, at The Economist, very extensively captured. Those two of all the AI companies working on hyperscaled LLMs. Those two frontier labs, as we call them in particular, are heavily influenced by that ideology. They come out of that ideology. It is essentially their raison d'etre is to ensure that we get the good superintelligent God and not end up with the bad God. That's why they were formed.

43:13The people who are yelling at them now are also from the EA rationalist movements, but all they're disagreeing on, it's a family squabble. The EA rationalist, so you have the Yukowski and all the different EA institutes, they're arguing, you guys, being the labs, we're not ready yet. You're going too fast. We're going to end up with the bad God. We're not ready yet to go for the good God. And then Dario Amadei or, you know, Greg Brockman over at OpenAI are like, no, no, no, we can do it. We'll get the good God. So it's a family squabble. So all of this debate going on is all from people who are believing the same ideology. And I don't

43:49think people recognize the degree to which especially the effective altruist version of this ideology is incredibly well-funded. They're funding fellows that are on congressional staffs. They're funding straight up just funding journalists, giving fellowships to journalists to make sure that their point of view will show up in the journalism. These are very powerful groups. And they're they're in the same group. They're just having a family squabble about whether we're ready or not to do it. And also, you know, there's now profit motives. The people in the frontier labs are

44:21discovering you can buy things with money. And so there's, you know, these interesting tensions going on. So this really matters if the entire debate over a technology is being shaped on both sides by people from a particular ideology. Well, we really need to understand that ideology so we can understand how to parse what is being said, what we should believe, what we've shown it, what it really means. So it's very important. This is not an unrelated lifestyle quirk of otherwise just sort of

44:54normal engineers trying to build this technology. It is central to understanding what's going on in the AI debate right now. It's similar to, you know, if you're if you were to ignore it or to say it's just a quirk would be like in the 1950s, you're trying to cover, you know, a school district in the South and the leader that the school board is really against desegregating schools. And they also happen to be a member of the Klan. And if you ignored that, like, well, that's what does that have to do with, you know, that's just some, yeah, they're weird in their beliefs. What does that have to do

45:28with their work on the schools? Like, no, that the ideology of the Klan would really impact the way they're thinking and talking about school segregation. We need to understand that. So it's important that we understand these ideologies and their impact. So, for example, there's two things I really care about. One, if you subscribe to these ideologies, it's really going to change the way you talk about the technology. So for decades, these groups, you know, since really the early 2000s, these groups have been talking about super intelligent AI is inevitable. It's going to be fueled by recursive

46:00self-improvement, which is an idea from their movements. It's not a computer science idea. And they're going to rapidly grow in power until they're dangerous. They've been talking that way for years, long before LLMs came along. So now we have to ask the question, if you're listening to Anthropica OpenAI or to their friends, EA Rational's friends who are on the other side, who are yelling at them, to what degree are they seeing things that we should really be scared about? And to what degree are they just continuing to say the same things they've been saying since

46:31well before LLMs came along? Like, that's an interesting question. Now, how might you resolve that question? Well, one of the things I suggested was the following. There's other AI companies that have hyperscaled LLMs. There's other researchers that know this technology inside and out, right? There's other AI industries in other countries that are also working at the cutting edge of LLMs, but that are not connected to this particular ideology. So a good litmus test would be if you hear the people in this group saying a fear, let's see if the people who have access

47:08to similar technologies but aren't in that group are saying the same things. That seems like a good litmus test. And guess what we're finding? They're really not. That's interesting. So when I mentioned like Mark Zuckerberg, look, I'm not a fan of Mark Zuckerberg, but Meta has hyperscaled LLMs. They have top researchers there. They've been working on this for a while. And what has his reaction been to the two particular companies that are within the futurist utopian ideology? He's like, guys, you got to chill out. There's not an inevitable thing here.

47:39We just don't build dangerous things. We're building tools over here. We're just not building the dangerous ones. Like, what are you talking about? He pushed back, why are you doing, why are you talking so much about recursive self-improvement? What does that have to do with building useful products? So it's basically a kid from the Silicon Valley lunch table. You have the EA rationalist over here, and then you have the other kids over here. They're like, dudes, chill out. Like, don't tell the teacher that we've been, whatever, stealing the French fries. That's a weird analogy, Jesse, but I'm trying here. Why do I care about Jensen Wong? Because they are deeply involved

48:13in LLM industry since they provide the technology that all of these models run off of, and they also run their own hyperscaled models internally. But Wong comes from an old school hardware world that has very little connection to the utopian ideologies that are big at OpenAI and Anthropic. And all year, he's been like, chill out with the sci-fi guys, right? Again, they have their own biases. Wong doesn't want export controls. They want to sell as many chips as possible to China. I know this, not be naive about it. But what matters is we need to double check what the ideologically oriented

48:47labs say versus the non. That's a useful check. They all have their biases, but that's a useful check. Why do you point to China? Is it a celebration of China? No, they're doing terrible things with AI over there. They're using it to surveil and persecute minority groups with a level of Orwellian rigor that would turn your hair gray, right? Like it's terrible stuff they're doing over there. I'm no fan of the CCP. But what's interesting is the CCP, if they believed for a moment that this technology was going to create a deity that was more powerful than the government, people would be

49:22going out the windows, right? They would just take over those companies. They would not tolerate that for a moment. But you don't hear that rhetoric out of the CCP AI industry in part because they have no connection to those futurist ideologies, right? So I think these are important data points. This is not a moral judgment on these groups are somehow morally better. I just think this is important that if you have this group believes to this ideology, which has long been saying AI, whatever form it takes, is going to improve the recursive self-improvement and become

49:54super powerful. And they're saying that, and they've always been saying that. We want to check that against other groups who work with the same technology. Are they saying the same things? And right now we're seeing, not really. I think that's an important data point. The other thing I worry about here that makes these ideologies important to understand is that I think it is actually impacting the specific experiments that those two labs are running. So it's relevant. Notice OpenAI, they ran these very dangerous hacking experiments. We know they're dangerous because

50:25it was negligent to the degree to which they let these agents run for very long amounts of time with minimal supervision and plenty of evidence that they were going to do unpredictable, dangerous, or hazardous things. And they said, we don't care. We need to keep going. Other AI labs aren't doing this. Meta had like a little bit of an incident with one of these tests, and they were like, shut it down. We're going to, okay, obviously don't do that. But OpenAI just kept going. Why were they running this particular band of tests that was clearly hazardous and was probably going to cause unpredictable results? Well, if you come out of those futurist ideologies,

50:59you say collateral damage along the way is the price you pay to get the good God. We got to win the race because we have to summon the right God to prevent the destruction of humanity. So if you believe those ideologies, it affects the experiments you run. Anthropic is obsessed with recursive self-improvement. That is not some computer science idea. That's not an engineering idea. It's an idea that was latched onto by futurists in the late 80s and early 90s as a way of being able to make predictions about superintelligent AI without having to do engineering. I mean, if you're Eliezer Joukowsky, right, if you're Nick Bostrom, you want to talk about

51:33superintelligent AI and utopia, but you're not engineers. So what do you do? They latched onto the idea of what they called recursive self-improvement. We don't have to tell you how those systems would work. They'll just make themselves better until they get there. So it's an idea that came out of futurism, not out of computer science. Anthropic is obsessed with it. So is open AI. We have to do this. We have to do this. Now it's a really bad idea. Why? Is it because we're going to have a takeoff and get a digital God? Most computer scientists I know say, no, that's not how that works. But what will you get if you try to increasingly have AI agents modify model training

52:08runs or the actual code of the agent harnesses? What are you going to get as obfuscated systems? Systems that it's harder to figure out how they work and they're less predictable. Why would we want that? The only reason why you would do that is if you've been immersed your entire adult life in an ideology that told you recursive self-improvement is how you summon the God. And again, Mark Zuckerberg, who I am no fan of, but he's useful because he has hyperscaled LLMs and is not a part of that ideology. He called them out on this recently. He called out Anthropic and open AI.

52:42He said, why are you working on recursive self-improvement? What does that have to do with building useful AI tools? Maybe, I'm paraphrasing, but he basically said, maybe if you spent less time chasing recursive self-improvement and more time trying to build useful, safe tools, we wouldn't have this problem. So we're seeing this inner scene war happening between the future ideologist and the non-futurist ideologist in Silicon Valley and beyond where the non-ideologists are increasingly vocally calling out the ideologists and saying, this is your Dungeons

53:15and Dragons weirdo stuff, guys. Build useful products. Why would you try to do recursive self-improvement? It's not a way to build a useful product. That's part of just like the sequences told you. That's part of how you're going to end up with a digital God. So I think these ideologies really, really matter, not because of weirdness. I'm a computer scientist. My people are weird. Spoiler alert. We're all weird. We all do weird things. This is not exactly a cool crew. And I don't mean to throw shade on me and my fellow computer scientists. We're not cool people.

53:47We're all weird. Mark Zuckerberg is weird. Jensen Wong is weird. He wears a motorcycle racing jacket. Why is he doing that? Because the engineers are weird. It's not about weird or not weird. It's about if your industry and then the primary critics of your industry are all from a very specific ideology that affects how you talk and how you act, then the rest of us have to understand it. In fact, I don't understand how you can be reporting on AI right now without knowing that deeply through and through. Tim Knight-Gerbo has a good analogy here where she said, look,

54:20if we're in a world where most of these two AI labs were run by Scientologists and the loudest voices that were yelling back to AI labs were also Scientologists and they had a disagreement about like when the lizard gods were going to be revealed or this or that, those of us covering AI would really, really want to understand like how much of what they're doing and what they're saying is coming out of Scientology versus like what's really happening with the technology. This is the same thing. We have to understand the futurist ideology that is completely shaping the discussion right now. Like you can't understand Jacob Croxon without understanding. He comes completely out of

54:52that ideology and talks that way. He makes no sense to people. If you don't understand, him and Anthropic are from the same group and they're having a family squabble. And Jacob is leaving the members of his family who think they can summon the good God and going over to the other groups, the other cousins who are like, Anthropic's not ready to do it yet, but they're all from the same group. They all have the same utopian vision. You have to understand that to be able to cover what is happening in the AI doom debate. So that's why I think I'm seen as an anti-doomer.

55:25I'm a big believer in various AI systems. I think it's a very powerful, interesting technology. And I'm very worried that the discussion around AI and two of the leading labs are caught up in this ideology, which is completely affecting lots of things negatively. And so that's why I really care about that ideology. We're all weird, but that goes beyond weirdness. It's what explains your behavior, the way you talk, the actions you take as a company. We have to be able to at least understand that so that we can have as much clarity as possible in understanding the landscape. So I

56:00don't know. I hear the same question a lot, Jesse. So I figured I'd give some clarity. Um, you know, there's like major journalists right now. We're taking fellowships from effective altruist organizations. I don't think people really understand the, how all these things connect together. Uh, what else do we have? Seven listeners sent in their vision documents

Reviewing listener vision documents for fall

56:23for the fall. Yes. I asked about this. All right. Completely changing gears here. Um, earlier this fall I did an episode about what's a lightweight system for organizing your efforts so that given all the distractions and digital whirlwind and all the stuff that we have to face in our typical lives, what's a simple way to make sure that you're still making progress on things that are important. I talked about creating a vision document that captured the key things you were working on in the fall and then the key tools, practices, or projects you're going to use to

56:53make progress on them. I asked people to send in, um, their own sample visions and some people did. So I thought we would, um, look at a couple and I've anonymized. These are sufficiently anonymized. All right. So here's a sample vision, um, looking at fall semester, 2026. So this, I guess, comes from a student and we see a, a, a summary of the goals up top. My goal this semester is to complete my coursework and go into December, ready to study for the bar and begin my legal career under projects

57:27and goals. We have fall coursework, empire, post-grad job search, move out apartment, Marine Corps marathon. So like these are specific things to be accomplished in the fall to move closer to that vision. And then under methods and tools. So what will you be using day-to-day, week-to-week to organize all these efforts? We see, for example, they'll use things for task managements, Google calendar for events, Garmin connect, uh, for marathon training, weekly planning, daily planning in the morning, daily planning, the PM with a shutdown routine, and they will track,

57:59they have a list of metrics they're going to track. So that's like an example of a good vision document. One page, you can keep that in your sock drawer. That's where I suggest you have it. And every day, remind yourself, what am I trying to do? What are the big projects or goals that'll move me closer? What are the tools and methods I'm running daily so that I can actually make progress towards these things? All right. So then we have one more we can look at, uh, this one I've, I've redacted some of the details just to keep things sufficiently anonymized, um, because this is a vision plan from a CIA assassin. So interestingly it has, no, it's not a CIA assassin, but I wanted to

58:33get rid of identifying details. All right. So the high level summary of this sample vision says this fall, I will implement the new CMS and something redacted because it was too identifying. I will create protocols and procedures to ensure staff understand how to operate these new systems under major projects and practices. The ones I haven't redacted includes organizing requests and feedback, identified required protocols, create video and documented procedures. So, um, these, the thing I blacked out is the bigger picture project, and these are sort of the sub things to get done. And then under daily, daily

59:07practice, they have organized folders and files. I'm going to organize one folder daily. Then under materials and methods, we have time block planner using a task list, loom to capture processes that, uh, so if you haven't used loom before, it's a way to create like videos describing how to use a digital tool that you can then use for other people to learn it using whisper flow to capture discussions and brainstorming. It's an AI tool that can create transcriptions from voice and a remarkable tablet for meeting notes and drafting sketches. So we have goals up front and then we have materials and methods below. Um, this individual had a separate vision for the

59:42personal life. So that was professional for the personal life. They said this fall, I will move my family to our new home. So our existing home can be rented for additional revenue. The major projects and practices include moving the new home, prepare current home for renters and have a daily practice of reading one chapter daily in terms of materials and methods relevant to this vision. They have OmniFocus for tracking projects and tasks. Um, and then they have a summary of some other stuff below. All right. So those are just examples of like what a one page

1:00:12vision document looks like. One sentence summary, major projects or goals you're working on, then the methods and tools you'll be using to help organize your time and effort so that you can make progress on those particular goals. So thank you, the two of you who sent in their sample visions. All right. I think we have time for one more question, Jesse. What else do we have? Our final question is about applying your advice in a parenting situation.

Applying time blocking in parenting

1:00:38All right. This is going to be volatile. All right. So here's what the message says. My wife stays home with our eight month year old baby. I've been seeing how much time blocking helps me. And I'd like my wife, uh, I'd like to help my wife organize her days, but I'm struggling. Her time is mostly shaped by the baby's naps and feedings and anything can change in a moment. Should a stay at home parent with an infant use time blocking at all? If so, what would it look like when so much of the day is out of her control? So I don't know, Jesse, why? Here's my question.

1:01:10Why isn't the baby time blocking her day? Right? I mean, there's a lot of efficiency that baby is leaving on the table. If you optimize your feedings, it leaves more time available for moving blocks around. Um, and naps and naps. So you got to optimize your naps. They all should be time blocking. Um, let's step back for a second. What's the goal of time blocking? Well, it was invented largely for the knowledge work context where we were, it was trying to help solve a particular problem, which is due to persistent overload. We often have endless things coming

1:01:45at us as knowledge workers, um, things we could always be reacting to things that could always hijack and take over lots of our time. And if we're not intentional about how we use our time, we get caught in a reactive loop and we make very little progress on the things that actually move the needle. So time blocking is a way of taking back control of your time and saying, I'm just going to face the reality of all the things I have to do. And what's the best way for me to balance reaction with proactive work? That's what it was invented for. So I don't think it's relevant for someone who's at home trying to raise an eight month year old baby. There, the goal is pretty

1:02:19clear is like, keep the baby alive and try to maintain your sanity as the person. It's not a problem you're going to solve with time blocking. You don't need to time block, right? So time blocking really is something that is deployed for complicated professional circumstances. It's why I recommend don't time block your evenings. Don't time block your weekends. It's not a way to go through all of your life. It's a way to tackle the very difficult challenge of the modern knowledge work workload and communication style. So my main recommendation is if you keep telling your wife to time block her baby care, you're going to end up with a baby bottle smashed and

1:02:54stabbing you in the throat. So I would, I would cool it on suggesting that I think whatever she's doing is just fine. Your job is to, you know, do whatever you can to make her life easier because that ain't easy dealing with infants. I like to conclude on Mondays by talking about what I'm up to.

Book review and closing thoughts

1:03:11Read a new book recently, Jesse, Simplify by Jim Vander Hei, Mike Allen, and Roy Schwartz. So if those names are familiar, they are the founders of Axios. They left Politico to find Axios and they wrote a book about simplification in work or in your own life. They know about it. Axios is famous for having the smart brevity approach to articles where it's bullet points and bold and like just get the key information as quickly as possible. People are more likely to read it. So they're sort of

1:03:42expanding that theory to how you can apply it to the rest of your life. You know, obviously I'm a fan of simplification. I think that book's out. It might not be. They sent me a copy. So either it's out now or it'll be out soon, but I'm always happy to hear how like different business leaders are thinking about things like simplification. They embrace a lot of ideas I would agree with, you know, simplifying priorities, deep work, minimizing stuff that's taking up time that's shown in the very anti-shallow work, very anti-suit or productivity. So those are definitely kindred spirits.

1:04:16What would an LLM summarize an Axios column? Do you like? I mean, an Axios column is basically already like an LLM summary. Yeah, I think they're down to it. They also heavily recommend in the book that their employees use LLMs for like a lot of things, including writing, which I wouldn't typically recommend. Um, yeah, they were like, they were like the LLMs before their LLMs. An Axios column is a political column put through an LLM. Maybe that's the way to think about it, except for LLMs get things wrong

1:04:48and hallucinate in Axios. You have local professional reporters. So it's like an LLM that doesn't hallucinate, I suppose. All right. Um, that's all the time we have for today. Uh, we'll back soon with another episode always here on Monday. Sometimes we have other episodes as well. Uh, and until then, as always stay deep. Hey, if you've made it this far, you must be ready to join my fight for depth in a distracted world. Now, the best way to do this is to join over 125,000 people who received my email newsletter each Monday. You can sign up at calnewport.com slash ideas. And when

1:05:26you do, I will send you a free guide to my seven best ideas about cultivating a deep life. Sign up today at calnewport.com slash ideas.