Aiming for the Moon
Aiming for the Moon
132. AI Snake Oil and the Skills of the Future: Dr. Arvind Narayanan (Princeton Computer Scientist | Coauthor of "AI Snake Oil")
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
In the midst of the AI hype bubble, how do we spot real AI capabilities and snake oil marketing? Many AI companies promise that they will help us predict the future. Is this true? Can AI really predict the future? Turns out, in this capacity at least, it might be more snake oil than success.
In this episode, I sit down with Arvind Narayanan, a Princeton Computer Scientist and co-author of AI Snake Oil. We discuss how to spot snake oil as well as how to prepare for a workplace of integrated AI. What skills should young people like me be investing in? And how do we plan for our futures?
Topics:
- Predictive AI vs. generative AI
- Why algorithms struggle to predict human behavior
- The risks of using AI in healthcare, hiring, education, and criminal justice
- How generative AI is changing programming and other professions
- Will AI replace jobs or transform them?
- The future of entry-level work and professional training
- What students should learn in an AI-shaped economy
- Artificial general intelligence and the future of work
- How to separate genuine expertise from AI hype
- “What books have had the most impact on you?”
- “What advice do you have for teenagers?”
Bio:
Arvind Narayanan is a professor of computer science at Princeton University and the director of the Center for Information Technology Policy. He is a co-author of the book AI Snake Oil, the essay AI as Normal Technology, and a newsletter of the same name which is read by over 75,000 researchers, policy makers, journalists, and AI enthusiasts. He previously co-authored two widely used computer science textbooks: Bitcoin and Cryptocurrency Technologies and Fairness in Machine Learning. Narayanan led the Princeton Web Transparency and Accountability Project to uncover how companies collect and use our personal information. His work was among the first to show how machine learning reflects cultural stereotypes. Narayanan was one of TIME's inaugural list of 100 most influential people in AI. He is a recipient of the Presidential Early Career Award for Scientists and Engineers (PECASE).
Socials -
Lessons from Interesting People substack: https://taylorbledsoe.substack.com/
Editor Peter Choi's Youtube (Ideas I Guess): https://www.youtube.com/@ideasiguess
Website: https://www.aimingforthemoon.com/
Instagram: https://www.instagram.com/aiming4moon/
Twitter: https://twitter.com/Aiming4Moon
AI Hype And Future Promises
TaylorIn the midst of the AI hype bubble, how do we spot real AI capabilities in snake oil marketing? Many AI companies promise that they will help us predict the future. Is this true? Can AI really predict the future? Turns out, in this capacity at least, it might be more snake oil than success.
Show Intro And Guest Context
TaylorThis is the Aiming for the Moon podcast, and I'm your host, Taylor Bledsoe. On this podcast, I interview interesting people from a teenage perspective. In this episode, I sit down with Arvind Narainen, a Princeton computer scientist and co-author of AI Snake Oil. We discuss how to spot snake oil as well as how to prepare for a workplace of integrated AI. What skills should young people like me be investing in? And how do we plan for our futures? Arvind Narainen is a professor of computer science at Princeton University and the director of the Center for Information Technology Policy. He is a co-author of the book AI Snake Oil, the essay AI as Normal Technology, and a newsletter of the same name, which is read by over 75,000 researchers, policymakers, journalists, and AI enthusiasts. He previously co-authored two widely used computer science textbooks, Bitcoin and Cryptocurrency Technologies, and Fairness and Machine Learning. Norain led the Princeton Web Transparency and Accountability Project to uncover how companies collect and use our personal information. His work was among the first to show how machine learning reflects cultural stereotypes. Norainen was one of Time's inaugural list of 100 most influential people in AI. He is a recipient of the Presidential Early Career Award for Scientists and Engineers. If you enjoyed this episode, please rate the podcast and subscribe. You can follow us at Aiming at the Number 4 Moon on all the socials to stay up to date on podcast news and episodes. Check out the episode notes for links to our website, AimingForthemoon.com, and our podcast substack Lessons from Interesting People. Alright, with that, sit back, relax, and listen in. Thanks again to Paxton Page for this incredible music and Peter Choi for editing this episode. Check out his YouTube channel, Ideas, I guess, in the episode notes.
Why “AI” Gets So Confusing
TaylorAlright, well, thank you so much, Arvind, for joining me today. It's great to have you on the podcast.
ArvindThank you, Taylor. Thanks for inviting me.
TaylorYeah, so I have really been enjoying your book, AI Snake Oil. It's been a fascinating read as AI is always in the news. I've been interested in computer science research, plus we've had on a variety of different AI experts who've talked about with Kate Darling, our future of AI with animals uh compared to our past with animals, and just a bunch of different topics. So when a book came along from Princeton Press that was discussing AI snake oil and the over-exaggerations of the technology, I've I've been really interested in this. So kind of to dive in, what can AI do well? I guess we must first define AI. And then what can it not do well?
ArvindFor sure. And it trying to answer that question is really the main aim of the book that I co-authored with my colleague Sayash Kapoor here. And I think to answer this question, we need to break down different types of AI. The book starts by asking the reader to imagine a world where we didn't have words for different forms of transportation. We just used the collective noun vehicle to refer to everything. And just think about how confusing this uh world would be. Um, you know, people would be discussing, oh, are vehicles getting faster or not? But it turns out some people are talking about rockets, but then other people are talking about cars, and you know, it would be pretty insane. And that's kind of the conversation that's happening around AI right now. We felt, and that's why we wrote the book. Let's talk about two major types of AI, which are very, very different from each other. Of course, they're not the only two around, but we think they are the most important ones. The book spends a lot of time on
Predictive AI In High Stakes Decisions
Arvindthese two. The first one is predictive AI, and I'm specifically starting from something that's not about ChatGPT, not the kind of AI that gets a lot of play in the media. I'm talking about the quote unquote AI that's used in the criminal justice system, for instance. When someone is arrested, there are these data-driven algorithms that are used to assess how risky that person is. And if they're deemed too risky, then they're jailed until their, you know, trial that might be months or years away, right? And so these are enormously consequential decisions about people, not just in criminal justice, in education, uh, in in uh in in colleges for things like who is at risk of uh dropping out and uh could benefit from assistance. So those types of things. It's used in employment. Uh, when we submit a CV these days, more than half the time it's first looked at by AI before it's looked at by uh a person, and there are automated video interviews that's used in in healthcare, so many other domains. So that's predictive AI, and we're very critical of this type of AI in the book. We can go into more detail, but the broader point, the higher level point, is that all of this is based on the idea that we can predict the future. Who will commit a crime, who will, you know, drop out of college, who will do well at a at a job. And we can't predict the future no matter how much data you throw at it. And we review the evidence in the book to show that these are all highly dubious predictions that are being used to foreclose people's life opportunities. Uh, and then we talk about generative AI in the book, and I'm sure we're gonna be discussing this more in the next few minutes. But broadly speaking, while we think there is a lot of hype in this area, we do also think that this is a genuinely useful and powerful technology if we use it in the right ways.
TaylorYeah. Let's hone in on the predictive AI to kind of set the stage for an end. I think talking about predictive AI is really interesting because it kind of sets up the scaffolding for doubting a lot of the AI claims that we see in general. Um so to begin, what are some of the main over-exaggerations that we see? Like if you were going to narrow down the core problems with how we end up with over-exaggerated claims, why do you think that is? What are the core areas that we fail in?
ArvindI mean, the core problem is that it doesn't work. Uh, you know, the cross the board. Yeah. Yeah, there's there's there's more nuance in the book, of course, but at a high level, that's that's basically what it is. And in terms of how we ended up here, I think it's because of the conflation of all the different things that we mean when we say AI. Right. So uh it is true that research on um AI algorithms has been progressing rapidly and it's led to remarkable achievements in some areas, again, such as generative AI, another area is self-driving cars. Uh, social media algorithms are pretty powerful. I mean, they have dubious societal effects, but uh, you know, they in a sense they uh uh do their job pretty well, perhaps too well. So those are all areas where AI is working well to various degrees, depends on what we mean by working well. However, the problems that those kinds of systems are asked to tackle, right? Um, for instance, translate uh text uh from different between different languages, or driving autonomously on roads, or finding you know, uh videos and articles online for us that we might enjoy based on what we have uh uh engaged with in the past. So those are all problems for which there are specific correct and incorrect answers, or good and bad answers, or a clear sense of what it means to succeed or fail. The car either you know uh navigated correctly or it crashed and there was an accident. So there are clear signals, and that is fundamentally what has enabled these AI systems to gradually get better, especially as more and more investment is made in bigger models, hardware, uh, you know, data centers, all that stuff that we're used to hearing about. Very different story when it comes to predicting human behavior. There is not a clear, correct answer to the question of is this person going to commit a crime? That's a fundamentally different type of question. That's in the future, the answer is not known yet. So even a hypothetical, ideal, omniscient algorithm is not going to be able to get the right answer, you know, 100% of the time. Okay, but what percent of the time is it? Is it 90% of the time? That would still be pretty good. But this is where the the uh details uh start to really matter. It actually turns out to be more like 60% of the time, you know, in the in the ballpark. And 50% is random guessing, let's not forget, right? So a lot of these algorithms are slightly more accurate than the flip of a coin. They're picking up basic statistical patterns. I mean, someone who has been arrested a lot in the past is going to be arrested more in the future, statistically speaking. That alone explains most of what these risk prediction algorithms in criminal justice are doing, right? And beyond that, there's not a whole lot of value. Um, and so I think it is that fundamental conflation between different types of AI that has landed us in the place we're in.
Healthcare Algorithms And Real-World Harm
TaylorWhat is the problem with some of the systems that focus on, well, uh predicting health outcomes? So you talk a lot about predicting health out um outcomes with algorithms and a lot of the deficiencies of these algorithms that are proposed and they say they work 90% of the time, 80% of the time, and then they're just when you actually test them, it's maybe 60%, maybe, and that's good. Um what is uh the fallacy that people are making when they say, oh, we have all this health data? And surely we could find some correlation in the data to some of these outcomes? Like, what separates using predictive models in that way from scientists and chemists and doctors sitting down and doing actual research? Of course, these are fundamentally different, but if you're going to explain this to a lay audience, what is the difference in that? Aaron Ross Powell, Jr.
ArvindThe key difference is that the scientists are not making decisions about individual patients, right? So they're not saying, unlike the insurance companies, so here's uh here's an example from insurance. These algorithmic models are used to predict how long someone might need to spend in a hospital or a nursing home or whatever to recover after a surgery before being sent home. Right. And so this algorithm spits out a number like 16.6 days, and you have to wonder about the oddly precise nature of those predictions. And it turns out on the 17th day, the patient can't even walk, and yet their insurance payments have stopped, right? So uh you have these kinds of uh really horrifying effects. And then uh in other examples, you have uh biases. So we did an analysis of the UK's uh liver allocation algorithm. So these days, in an increasing number of countries, for an increasing number of organs, the organ allocation system is being handled by a type of uh AI algorithm. People will, you know, quibble about whether to call it AI or not, but it shares a lot of the same characteristics as some of these other predictive AI systems we're talking about. It tries to predict how many years of life will this patient gain from being given this liver. And basically, whoever is predicted to gain the most years of life uh is going to end up being the recipient. But then when you really start to interrogate that logic, it turns out that there is a very severe age bias. Uh, for instance, uh, there was an investigation that showed that people under, let's say, 30 years of age who are often, who often have failing livers because of chronic conditions, right? Because of uh conditions they've had throughout their life, have a very, very difficult time being allocated a liver through this algorithm. So that's the main difference, right? So when when scientists are doing this research, you know, the what they're coming out with is a scientific insight that's not going to directly be mindlessly applied in order to make life-changing decisions about people. So we have no beef with the use of data in science. And obviously, a lot of our colleagues are doing great medical research that is data-driven. We absolutely need that. And in fact, we are very much supporters of randomized control trials. And in the book, we talk about randomized control trials in medicine and how that data-driven approach has been so important. And actually, we call for expanding that into other areas. Perhaps if we want to be serious about using algorithms in other domains like hiring people for jobs, we need to be justifying more of those decisions by doing randomized control trials. Even if it might seem unfair, oh, you know, based on this flip of a coin, one person is getting a job and another is not. But in response to that, let me say two things. One, that's essentially what these systems already are. They're elaborate random number generators. If we want to improve them, we actually need to do randomized controlled trials with some humility so that we can build better systems. And the second thing I'll say is that if we have accepted the necessity of randomization in medicine, where it's literally people's lives on the line, then I think we can certainly accept it when it comes to other consequential areas such as hiring or college admissions. It is only if we make some decisions in a randomized fashion that we can get a better get an idea of how to make these decisions in a better way.
Generative AI And The Job Market
TaylorYeah, so let's shift to the job market and particularly to education. So a lot of our listeners are going into college or thinking about college or in college and in that kind of like stage. We're looking out at a landscape that has generative AI, um, specifically when it comes to computer science jobs when we think of programming. Um that seems to have a very high success rate when it comes to um what is it, vibe coding and everything else that goes into that. So I'm very curious what do you think from uh well, let's break this down into two different uh questions. So the computer science landscape. What do you think about that when it comes to the involvement of generative AI, particularly when it comes to people who come out of college? They usually tend to get lower-level jobs, of course. Um now, is what do we do about us computer science people who are going into a workplace that has generative AI everywhere? And then two, uh from a higher bird's eye view perspective, what jobs do you think are going to maintain their legitimacy? Um, what jobs are most likely to be replaced if they will be replaced, or is that um a myth that we're kind of buying into more?
ArvindThere are so many great questions packed into that. So I think we're shifting now from predictive AI, which we've been talking about, which is used to make decisions about people, to generative AI, which is a consumer tool, which is in all of our hands. And I'm sure most of us are already using, you know, whether it's for education, for our jobs, or just for entertainment, so many great uses of it. And yeah, so this is something I think about a lot as a computer scientist. Computer programming and software development has been at the leading edge of a profession where generative AI has already made very heavy inroads. So let me start by saying one thing, and I'm gonna quote Simon Willison, who is a very notable software developer who's been experimenting and writing about large language models for software development. He says, quitting programming as a career right now, because of large language models, would be like quitting carpentry as a career because of the invention of the table saw. Right. So I think that's really the right analogy. This is a power tool. This is not, you know, it's uh at least for now, it is so far from actually replacing uh software engineers. Uh, yes, you're hearing lots of stories about, oh, you know, tech companies are hiring less because of AI. I think those are excuses. Companies have uh, you know, they overhired during the pandemic, and actually over the last several years, they have been cutting hiring. It's just that AI is a very nice excuse for them now to uh justify what would otherwise be a story that would make a lot of people mad, right? Um, you know, we overcommitted. Yes, yeah, yeah, that's right. Uh and so I don't think we should believe these narratives that we read in in uh you know coming out of CEO's mouths or in the news or whatever. I think it requires a little bit more nuance to really understand uh the impact of AI on jobs. And I I mean it's still early days, the picture is still emerging. There might be job displacement, but I think it's way too early to conclude that right now. And so far, I don't think it is having such an effect. And the thing I want to point out is that a lot of people go from these comparisons in various professions, comparisons between human task performance and AI task performance, and they say, oh my God, it's gonna, you know, it's it's catching up or it's already surpassed human performance, it's gonna be cataclysmic, etc. If you ever that hear that style of argument, you should just throw it out the window. That kind of argument has no merit because it's never, I want to repeat, never a comparison between raw human performance and AI. It's always a comparison between human performance aided by AI versus AI alone. That's the only relevant comparison. So if the prediction is not based on that kind of comparison, it has no merit. And this simple criterion eliminates over 90% of the predictions about AI and jobs that we're seeing out there. Again, I don't want to say, you know, AI is not going to replace any jobs, but most of the predictions about this that you're seeing are based on very naive comparisons. And so I'll mention one last that maybe two quick things about this. Uh, and one is that uh whether or not it's gonna replace a particular job depends less on this type of comparison about capabilities and more about what kind of worker protections exist in that industry. That fundamentally human question is really the key one. It's less about the technology. Uh and the last thing I'll say is you already put your finger on this. Uh, more than job displacement, the thing that concerns me the most is that a lot of the tasks done by junior workers in many different industries, whether it's law, you know, various other things, including programming, uh oftentimes those are the ones that can be done okay by AI tools with, you know, that are being operated by more senior workers. So the real question then becomes what's going to happen to our training pipelines? And I think the answer might have to uh vary uh based on different fields. Uh, you know, uh, maybe in in some areas we need more government support for uh for these training programs where employers are actually incentivized uh through you know uh tax breaks or whatever to hire these uh uh junior workers if it turns out that they otherwise have too little of an incentive to do so.
Career Skills For An AI Workplace
TaylorThat's very interesting. Do you have any career advice for people looking at these particular fields? So the the jobs that often come up are radiology because it's image identification oftentimes. Um and then of course junior level programmers. Do you have any career advice? Are there skills that people who seem to have their jobs threatened by um generative AI or image detection, which I guess would technically be, is that technically uh prediction? What AI group does that fall into?
ArvindI would call it classification. Uh it's it's not quite predictive AI. You're not really predicting the future in most cases.
TaylorAaron Powell But for these jobs that seem to be threatened, at least in the nearer scope, um maybe not fully replaced, but should we be looking to gain more tools than we had before? Is there any more um things we should be investing in and skills? What should people looking to go into these careers that seem to be threatened at least now be keeping in mind, at least in their education journey?
ArvindYeah, radiology is a great example. I don't know if this has come up already in your previous episodes, but there was this famous prediction by later Nobel Prize winner Jeff Hinton back in 2016. He's one of the uh notable pioneers of uh modern AI. And he said, Oh, yeah, radiology is just uh, you know, classifying these images. Yeah, AI can do that. Uh, there are going to be no radiologists left in five years. Uh, but what has in fact happened is that almost a decade after that prediction, the demand for radiologists has actually grown. And this is not because they have somehow turned into a cabal and rejected AI to protect their jobs. In fact, it's quite the opposite. They have uh very readily embraced AI and they're using it heavily as a tool. It turns out that what Hinton thought of as the job of radiologists is only one of 27 tasks that radiologists do. And furthermore, I've been saying in some of my recent uh writing and talks on this that even breaking down what a person does in a job into a sort of discrete tasks is already losing the most nuanced and hardest to automate parts of that job. And so that's why these analyses of impending job loss are very, very misleading. However, however, by that same token, the reason actually that radiology is doing so well is because they have embraced AI. So their workflows actually look pretty different now. A lot of new skills regarding how to use these AI tools and knowing what they're good at and not good at, you know, when to second guess the AI tools output, right? So many other things. These are all new skills that you wouldn't have needed a decade or two ago. And I think this is happening in virtually every job. And in looking at my own experiences doing software engineering, it's not something that I do full time, you know, but as a computer scientist, a lot of what I do involves software engineering. I mean, my day to day, you know, when I'm doing software engineering is not. Almost disjoint set of skills compared to five years ago, right? And I've had to gradually pick up this whole new set of skills. And I think that is the key thing that people thinking about their future jobs should keep in mind. It's less about which careers to go into, but more than whichever career you go into, you should be prepared basically for a lifetime of uh you know constantly having to uh learn new skills. And I find this interesting and exciting. Um, but I know that this can feel daunting to a lot of people. Uh, but uh since you know, since a lot of your listeners are teenagers now is a good time to be mentally prepared, right? What it means for you to have a job is going to be different from what it meant to have a job uh a while ago because it's not like you learn skills once and then use it for the rest of your career. You're gonna be learning every day because advances in AI capabilities are not going to stop.
TaylorYeah, exactly. So it's this has been a wonderful conversation. We're about to switch into the the last two questions which we ask all of our guests. Um but listeners, if you're interested in more of this topic, check out our interviews with Sarah Newman about um data integrity and nutrition and whether AI can be sentient. Spoil alert, she doesn't think so. Um and Dr. Kate Darling on our history of the animals applying to our future with robots, which we dive more into the, especially the task-oriented models and how you can't really have a generalized model per se, which we didn't touch on as much. I'm curious, um, just real briefly here.
AGI Defined By Economic Tasks
TaylorYou defined AGI in a way that was very pragmatic, um, versus some people who are like, it's Centi and everything else. Could you just briefly discuss your AGI definition? I thought it was very interesting and kind of how it applies to this um field of job placement.
ArvindSure. You're perhaps referring to some of my writing online, and uh uh Sayash and I in our writing have adopted an AGI definition. Uh it actually comes from some of the AI companies uh and uh and AI researchers. It's not our definition, but it's about whether there's going to be an AI system that can do the majority of the tasks in the economy that people do today. And I think the reason I like that definition is that these questions of sentience, you know, while they're certainly interesting to philosophers, cognitive scientists, many others, to people people who are wondering about what's going to be AI's impact on my career, my life, my job, it's it's more about what's AI going to be able to do rather than how it works internally. Yeah. And so from the perspective of these kinds of economic impacts, our key message, you know, really dovetails with what I was saying a minute ago, which is that it's not going to be human versus AI, right? I think by default we should be thinking about augmentation of our jobs rather than automation of those jobs. We do have a lot more to say about that, that the fact that there are a lot of what we call um adoption or diffusion frictions. So when AI is going to be deployed again in radiology, software engineering, law, or many other fields, there are so many changes that need to happen to the structure of those industries, the way people are trained. So we're asked that we're suggesting that people expect a multi-decade-long timeline for these impacts to be fully realized, as opposed to something that's gonna uh a transition that's gonna be complete by the end of this decade. And we don't think that's realistic.
TaylorYeah.
"What books have had an impact on you?"
TaylorWell, this has been a fascinating discussion. Um, wrapping up, what are what books have had the most impact on you?
ArvindLet me just mention one book in the interest of time, and also because I want to uh tell listeners what lessons I learned from that book. And this is the book The Selfish Gene by Richard Dawkins. And this is something I read as a teenager, and that's why I specifically thought of this book. And the reason it made such a powerful impression on me was that this was, you know, a groundbreaking new understanding of genetics. And what shocked me was that as even as someone who was a teenager and not a specialist in that area, I was basically able to understand the book. And the book was written in that way, right? It was written to reach a mass audience. And that was the first time I uh started to believe that making deep scholarly contributions is not at odds with being able to explain those contributions to everybody. Uh, and that's been a North Star for me in my own scholarship. Obviously, not everyone can be, you know, as influential as Richard Dawkins, but I think but I think every scholar can certainly try, and maybe not always succeeding, but but always try to make sure that even when you're you know saying new um important things from an intellectual perspective, that it's it's not couched and jargon in a way that only a few of your peers can understand, but really is accessible to everybody because there's just such a thirst for knowledge out there.
TaylorAnd your book does an excellent job of translating these very complicated ideas to um a general audience. I've recommended it to a variety of people, and it was very helpful in guiding my introduction into the further intricacies of AI. So you're definitely doing a great job of living up to that mantra. Thank you. What advice do you have for teenagers?
Advice to teenagers - Activism Needs Evidence And Sources
ArvindOne thing I notice when I talk to young people, teenagers, is that there is a lot of interest in making change in the world, in activism. Uh, I can't say for sure whether there's there's there's a trend toward more of that than there was in decades past, but there does certainly seem to be a lot of it. And I think this is fantastic. You know, it's uh it's really heartwarming to see how interested teenagers are in making change in the world. One note of caution that I want to give is that to truly effectively make change does require some level of expertise in, you know, the technicalities of the thing that you're trying to change, right? Whether it's AI or climate change or whatever it is. And oftentimes I think young people try to skip this part because they think of expertise as the domain of PhDs or, you know, they need to be much more senior. That's not true. Uh when I say expertise, I'm not talking about something that requires several decades. It's something maybe, you know, that you build up to over a period of a few months. And I very much encourage people who are seeking to make change to take that part of it seriously. Because the reason for that is that oftentimes the things that you end up advocating for might not actually be backed by the evidence, right? Just to give one example from climate change. I mean, it's while it's relatively clear that climate change is a huge problem, uh, it requires a little bit more expertise to know what are good and what are bad solutions. For instance, I think you know, nuclear power is one of the things we should be pushing for more. Uh, however, I do see a lot of young people who are very anti-nuclear power, and it's not clear if they've done the research to really justify that position. So I do think that advocacy can be much more powerful if it's combined with uh with a certain amount of research. Again, not a PhD, but something that's very much within the reach of young people.
TaylorAaron Powell As someone who's in um academia and that world, what sources do you recommend on this? Because of course, everyone out there, there's a whole section of the internet that's like, I do my own research, of course. And this these are the conclusions I've come to. And you're like, that's not what anyone actually agrees with. So when it comes to doing your own research, um, where do you recommend looking to? Are you looking, I guess, public intellectuals who have taken this um their research and trying to explain it to a general audience? Or how do you recommend young people go about this search?
ArvindAaron Powell There is a spectrum on the one hand between deferring toe is uh whoever has the most credentials in a particular area to uh quote unquote doing your own research in the sense that you don't necessarily believe anything you read online and you want to, you know, uh try to verify every last bit of information, which ends up being completely unrealistic simply because no one has that much time, and so you end up down these rabbit holes, and it's easy to get into conspiracy theories and so forth. So I think I do think you need a healthy balance between these two. Uh let me just talk about some of my own strategies. I like to uh you know look for people in a particular topic who um are who have you know good practices when it comes to how they share knowledge, how they update their own beliefs. Are they willing to change their minds when they, you know, when they see evidence that contradicts something that they have said before? Um what what sources of evidence are they themselves using to come to their conclusions? So I think it's a matter of not necessarily deferring to authority, but uh uh but yes, it is important to have some deference to a set of people who you've selected based on your own assessment uh as people who have these good epistemic practices, good knowledge sharing practices, and using that set of people or that set of sources as the base upon which to build your own knowledge and expertise.
Final Thanks And Wrap-Up
TaylorWell, Arvin, thank you so much for coming on the show. I've really enjoyed our conversation.
ArvindThis has been wonderful. Thank you again for having me.