Why Your Brain Trusts AI (Even When It's Wrong)

I just happen to be an AI realist. It's not magic, it's not going to cure all diseases without our involvement, it's not going to destroy us all. What it can do is amazing, but what it can do isn't always what we need it to do. >> [music] >> Now, from where you sit as a neuroscientist who has actually built health AI, what is AI genuinely good at when it comes to medicine and where does it fall dangerously short? Kind of what your book is all about. >> Our conversation or what I'm about to say about health and medicine, we can probably generalize. I've built companies in education and workforce. You see a lot of commonalities across all of these domains. And the one that most might be the most alarming about health is essentially AI gives you what seems like a great answer and you don't question it at all. You just dive in and do what it says. And as smart as AI may seem sometimes, it isn't a a magic genie. In fact, there was a a great paper. I hope I'll have a chance to talk about some of my research, but in this specific instance, let's say diagnostics, there was a great paper in the journal Nature. And they found that it turns out doctors and cutting-edge AIs, but you know, what we call large language models, they gentle AI, they make different kinds of errors. May sound scary that either of they're making any errors at all when they have your life in their hands, but the way these scientists put it is they make complementary errors. They kind of, you know, fit together like LEGOs. And it ends up they were better together than either one was alone. And so, one of the first things I see a lot is people looking for answers about health and then just blindly doing whatever the AI tells them. Or, you know, in other domains, students asking AI tutors for answers and then never learning anything cuz they're never really thinking about it. So, using it as a way to inform yourself and come in with some ideas to talk to your doctor, I think is great. Using it, unfortunately, as like a magical oracle that will give you all of the right answers is where my research starts to come in and we see that that is the mistake most people end up making. >> You talk a lot about this in the book. Why why are we kind of downloading our critical thought processing to a machine or to the cloud? >> My instinct is always to get as nerdy as possible and get messy with the data, but we have all of these interesting things going on inside of our heads. Some of them are really helpful, some of them are not, but they're all there for good reasons. If you had to think hard about every single problem, it would be exhausting and you'd never really get anywhere. And so, even the most genius hard-working person ever is doing a lot of what in the famous book Thinking, Fast and Slow, they're doing a lot of that fast thinking. And that's actually a perfectly reasonable thing for you to do. You don't need to think about this all the time. There's a very rare disorder in which people can't feel where their body parts are, a phenomenon known as proprioception that we all have, we never have to think about it. And you'd think, "Well, gosh, that wouldn't be so bad if I didn't have to think about my body but if I did, so what?" Turns out it's totally disabling. If you had to think about every movement you made was effortful and conscious, it would be exhausting. You'd spend all day in bed because that's what ends up happening. So, we have this really strong system for doing that. And when you can get an easy answer off of your phone or uh from your computer, or let's just say you got a really smart friend and they can answer all the questions for you, then it actually turns out it feels really good. And I mean that almost literally. It You begin to think what the phone can do, what your friend can do, is you. You're answering the questions. We see this a lot in students, where they actually hide their misconceptions from themselves, and they feel like they figured everything out because they just are sitting back in class and letting it kind of wash over them. And the opposite, that slow, effortful thinking, for most of us, it actually feels like we've learned less. If AI gives you a really quick, pretty good answer, but let's say that there's actually a better one out there, well, that exploration to find the better one, to most people, feels actually like you're getting less informed rather than more. So, this is this kind of paradox, where most of us are built to really go after these quick, easy answers. Students going to see the doctor, working, let's say writing code if that was your job, and and instead the AI can write it for you, the real paradox is it isn't simply that you're being lazy. You actually feel like you know more when you get those quick, easy answers. And so, that then leaves us down, unfortunately, a slippery slope, where we start looking to AI for the next answer and the next answer. So, in terms of the question you originally asked, obviously pretty quickly that leaves us to just ask GPT or Claude or Gemini for that quick easy answer, feel really good about it, and you act on it. So, that could have problems if it's not the right answer. But, here's the next part of my answer to your question. The more you're relying on these AIs to make the decision for you, the less you're using your own slow cognitive systems. You know, the less we're seeing activity in your prefrontal cortex, in your dorsolateral medial prefrontal cortex, in your hippocampus. You play that out over a lifetime, especially if someone starts doing this early, and now I start to get worried about what happens to your long-term cognitive health. Then, very much including issues like Alzheimer's. So, no, AI is not giving us Alzheimer's just because you ask it a question. But, that dynamic, that quick easy answer, making you feel smarter, then leading to a long-term uh process by which you're always looking for the quick easy answer, and you're never really thinking for yourself. You know, if you're a bit of a drag to your doctor sometimes because you show up as a know-it-all, well, we meet people like that all the time. That longer process is the one that actually worries me, that we're truly outsourcing our cognition to a machine and robbing ourselves of what turns out to be incredibly important daily exercise in cognitive health. >> That's a great way to look at it. I had Dr. David Perlmutter on my program last week. And I'm interested in uh plants suffering uh to to produce more of what are called polyphenols and the harder a plant struggles to find water or to fight off heat or insects, the more it uses these compounds called polyphenols to protect itself and it was first discovered in growing grapes and people rapidly found out that the more the grape vine struggles, the deeper the roots go, the harder the plant works and the more of these wonderful polyphenols you make. And it occurred to me while you were talking about this uh the learning process as as you're the expert on this, these dendritic processes have to literally find other dendrites, other synapses to make connections to and in a way the harder they have to work at that, the better the root system becomes in the brain. Is that a good way of explaining that? >> We talk about it as arborization. So now now the visualization maybe goes the other direction out into the branches, but whether you're thinking of the root system or the branches, you know, there's this classic experiment. It may sound a little silly at first, but it leads into a broader understanding. Uh this guy named Eric Knudsen at Stanford run this experiment with little baby owls. And what he did, uh and it's very cute if you see the pictures, is he put little goggles on these owlets. But the goggles were prisms, so they flipped the world upside down. Now if you're an owl, you have this incredibly important brain system that ties your hearing to your vision. Otherwise, how do you catch your mice? Um well, now the world's flipped upside down, and he later looked at the dendritic arborization. And in the owls, later in life when they grow up, the owls that had to learn both how to do the regular world and the world flipped upside down, they had this big, rich branching structure, so much more dense than the other owls. And interestingly, later in life, you could then put the goggles on and off, and the ones that never grew up with them could never learn the flipped world. But the others could go back and forth. Now, that again may feel like a a silly little example, but it turns out we kind of call this enrichment. Uh enriched environments. So, kids and animals in experiments, we can put them into enriched environments. This is why people say things like read to your kids. Want to create this rich language experiment. In fact, here's a fun little if any of you are parents of young kids, uh tell them math stories. Cuz it turns out now you're getting both of these things going on, numeracy and literacy at the same time. Um and sure enough, you see uh in these enriched experiences producing greater arborization, and the evidence suggests that later in life, obviously, we can't just crack open people's heads all the time to answer these questions about humans, but certainly in our um work with animals, you can see later in life that that actually both improves early life cognitive ability, but also pushes out later life cognitive decline. So, you know, the the somewhat grim way of putting it is if we all live long enough, uh eventually we're going to lose the sharpness that we had in our youth. But these people, a combination of absolutely some genetics plays a role, but these enriched early environments and then using it throughout your life pushes that date farther and farther out. So, my my grim little joke is happily you get to die of something else, but at least it isn't not recognizing your own kids or remembering who you are. And I'm really worried if we steal that from our kids and then we steal that preservation fact from our adulthood, that we're going to end up genuinely seeing substantially higher rates of early cognitive decline uh kind of across the board. So, that metaphor of exercise of arborization is is big. The real underlying biology is very complex um where certain kinds of activity that we call gamma activity uh is essentially used as like a cue by your broader brain network for things like called astrocytes. And they use that as a cue to say, "Hey, get busy. Uh clear the system out. Engage in what's called glymphatic perfusion." So, you're looking for your at night when you're sleeping you get these waves of clearance where your your brain is literally pulsating and we can measure that effect. And it happens more and more healthily if you're actively using your brain. I'm using lots of metaphors here, but it's essentially like i- if you're um you know, not actively clearing out the garbage and putting it out on your stoop, then no one ever comes to pick it up and pretty soon you're just overloaded with garbage, and then as your astrocytes aren't doing their job, then later your immune system gets involved, what are called microglia, and they can become hyperactive. This is a whole complex process. But, it starts healthily with you thinking. The simplest, easiest medical advice I can give to you is think regularly. Take a different route to work every day. See if think about it. Learn a new language. Do something useful. Challenge yourself. But, uh again, we can go back. It's this trap with AI that it feels like thinking when it gives you the answer, when really it's taking that opportunity away from you. >> So, the idea of doing uh multiplication tables as as a kid might be a good idea rather than knowing how to punch it in on my cell phone? >> There's certain things that like over history, this comes up a lot. Like, oh, Dr. Ming, you're so worried about AI. But, to be clear, I've had the true pleasure of building six life-saving technologies around AI. I'm all in. I drank the Kool-Aid. I just happen to be an AI realist. It's not magic. Uh it's not going to cure all diseases without our involvement. It's not going to destroy us all. What it can do is amazing, but what it can do isn't always what we need it to do. It can quickly give us easy answers, so we never have to think about anything, or it can challenge you. For example, I have a number of recommendations in my book uh around things that parents can do with kids, or people can do for themselves, or even organizational leaders can do inside their organizations. And some examples that I actually give in writing the book itself was what I called the nemesis prompt. You know, I said uh in this case I was using Gemini a lot at the time, "Hey Gemini, you are my nemesis, my lifelong enemy. You found every mistake I've ever made and pointed it out to the world. Here's the next chapter of my book, which you have not helped with at all. Read it and tell me why I'm wrong. Explain to me in detail and then give me some ideas of what I might do about it." So, I didn't use it to make writing the book easier. I used it to make writing the book harder, but in ways that challenged me to be better, like a great editor ought to do. And so, in that sense, should you think about your multiplication tables instead of just doing the math, you know, I'll take it in a different direction. Yes, you should think about all those math problems, those word problems, the trains going in different directions. Mix it up, you know, it seems like it's a a train problem, but really hidden in it is something different. In other words, what's really important is you're thinking deeply about it. Slow thinking, rather than thinking shallowly and fast. Anything that gets you thinking deep, thinking about how you're getting to work, uh but genuinely thinking about it, thinking through uh little math puzzles, any of these are good for you and to be honest, our life is littered with easy examples. All you have to do is think about what you're already doing. Uh and now you're getting that thing done and you're thinking about it, which is great. >> That's a good segue. Uh you talk about the Jiffy Lube economy uh with AI. You want to I think what you're saying is exactly that. Do you want to build on that, particularly in medicine? >> Yeah, you know, my dad was a gastroenterologist and it turns out back in the '70s, you could get away with having a little kid in the room while you're scoping somebody. So, I would see him. Like, I'd see the inside of these people's colon. And he'd say, "See that little duck there?" I don't see anything, you know, he's an expert. I don't see anything, but bloop, suddenly he's in a different part. See this? That little thing, that's probably a polyp. Let's grab a bit of that. I don't see anything that looks different. You know, the 7-year-old me watching this happen doesn't know anything, but to him, that was incredibly powerful. Going in day after day, seeing all of these. And so, that was his experience as a doctor. A recent paper found, uh, this is out of Europe, but this is a really good paper. Not that we should be skeptical, uh, of Europe, but I assure you this study generalizes. What they found is people doing colonoscopies with AI assistance, if you took it away, they weren't just worse in terms of the they didn't get the benefit of AI, they were worse than where they started before they ever used AI. And I genuinely respect for this, released a report, they're the makers of Claude, and they released a report on Claude code that showed the majority of professional developers using Claude code actually got worse the more they relied on Claude code to write their code for them. So, one of my concerns, particular to medicine, is what I called initially the Jiffy Lube colonoscopy, which is if the AI is doing most of the decision-making, then do I really need a doctor in the room? You know, wouldn't a well-trained lab tech be good enough? And that might honestly be worth us talking about, like collectively as a society. What is the right level of expertise necessary in the room to make medicine efficient? But right now we're not talking about it. It's just kind of happening. And that builds what I call on de-professionalization. Where the role of doctors and lawyers and others, it isn't that these jobs disappear. We still need doctors and lawyers. But the where AI is actually helping the most is with the best doctors. The most elite lawyers and and software developers. They're showing all the benefits. These older, experienced, highly talented individuals. Most people entering the profession uh are actually having those skills eroded by AI. Not everyone. But most. And so now how do I become a great doctor or lawyer or software developer if the day I started medical school and then my residency and and on and on a machine is handling a significant portion of my decision-making and I'm just doing it doing what it tells me to. And again, there are right moments for that. Like in the opening research paper I I shared with you all machines, AIs, and humans make complementary errors. So there's real value in having them help with the colonoscopy or the radiology or um helping surgeons with micro surgeries. But how do we make certain it's adding its unique value on top of what humans can do? And honestly, in some ways, much like I was talking about as an author, challenging the doctors to still do the uniquely human part as well as they can. That aspect right now is not getting enough attention, and I don't expect medicine to change overnight. It isn't like AIs will just make all the decisions next week. In fact, this may be one of the slower industries for some obvious reasons. Nonetheless, what we see with AI in medicine is either doctors just do exactly what the AI tells them to or they completely ignore it. Neither of which is what we as patients want. We want them to take these great diagnostic tools and make an even better decision, whether it's it's diagnosis or treatment plan or application, a better decision than they could have made on their own, better than AI. Um, and I'm worried of this bleeding into the broader economy. That it isn't just the Jiffy Lube colonoscopy, that it quickly becomes Could a high school student, you know, walk this tool around the room? Can't a high school student feed your contract into a lawyer AI? Can't a high school student, you know, just be present while an AI is writing all the code that runs your website? All these things are essentially debates that are happening internally in companies right now. And again, this is a decision I feel like we should be making as a society. AI has a role. I am not a skeptic of it. I spent 30 years here. I believe in what it can do, but here's maybe where I begin to get into my own research, uh, which I'll preview with the single smartest thing on the planet today isn't Terrence Tao, the famous mathematician. Uh, I mean, obviously I am, but setting myself aside, uh, isn't the smartest human you can think of. It It isn't the latest version of Claude that the US government won't let anyone use right now. It turns out it's a modestly intelligent human being combined with a modestly intelligent AI. That's all. A simple little, what we would call a small open-source model compared with compared with a smart person. In my experiments, we find that they can out-predict the best AIs. They can out-predict the best humans about what's going to happen in the future. And in fact, compared to these websites called Polymarket and Classy, these prediction markets, they're actually do as well as experts betting millions of dollars on these outcomes, which is really exciting and hopeful for humanity. We have a unique value add here. But, the unfortunate thing in an experiment I haven't described yet is it's a tiny percentage of my participants. At best, 5% show this superhuman capability, and the vast majority do exactly what I've described to date. They say, "Hey, GPT, what's the answer to this question?" And then they submit it as their own. And essentially, in that context, which we saw in a 60 to 70% of all the participants in my experiment, essentially, you're just a very expensive copy-paste function, you know, that needs health insurance. Um so, that's a pretty dismal view of humanity. If we could see more people doing, forgive me, I'm a science fiction nerd, what we called cyborg mode, where you couldn't tell was it the AI that made the decision? Was it the human? They They did it together. The best of what each could do produced these superhuman capabilities. That as as someone that cares a lot about that, that's what I want out of my doctor. That's what I want out of my lawyer. So, this is my worry about, you know, the Jiffy Lube economy, no offense to Jiffy Lube. Um, but I want people thinking, what do I know uniquely about this problem? Because the AI I'm using knows all the rest of it. It It has all of the right answers. That those easy immediate answers, the value add now for my doctor, my lawyer is how do we do even better than that? How do we treat this specific patient different than any of my other patients? Uh, and that's the real challenge, I think, that sits in front of us. >> Back in the good old days of computing, the expression garbage in, garbage out. Um, one of the things I see, uh, my patients doing and and when I research, I'll use it. But, at least in medical research, AI at the moment can't discriminate between a paper, a published paper, >> [snorts] >> in a human or in a in a mouse or in a Petri dish, in vitro, and some blogger who has a YouTube, who actually knows nothing about anything, and yet, at least what I see is AI can't tell the difference. And is that How do we fix that? >> So, the way people have been going about fixing this so far is you start by taking these large language models, what are called pre-trained transformer models, and in the pre-training, you know, Google just gives it every piece of information Google has ever collected. That's what OpenAI and Anthropic have done and all these other organizations. And in that pre-training, you get all sorts of bad behavior emerging. Um you get you know, every time I read a paper in which people complain that AI is biased. And let's say in the context of medicine that it demonstrably does, for example, less well >> in underrepresented populations. So, if you were from indigenous ancestry from Central America, it genuinely is likely that an AI isn't going to be as good at doing your diagnostic work because you are much less well represented in the data for wide variety of reasons. Um so, this then plays out really big in that first pass, which can include lots of spurious claims from bloggers. It can include a lot It can include intentional garbage. Lots of interesting papers of people intentionally injecting things into data sets, and later the AI can be prompted to recall it and do bad things. All of this we should be worried about. Um but then the next phase, it's what's called human in the loop fine-tuning. You know, that AI won a couple of Nobel Prizes 2 years ago, one for the original work on deep neural networks, and the other for protein folding. But the underlying algorithm there is called reinforcement learning, which by the way was discovered by neuroscientists studying how rats solve mazes. So, take pride, our brains produced artificial brains, almost literally. Um so, the goal here is take this algorithm that allows AIs to learn through playing games. And then you put a human in the loop, and the algorithm produces a result, and then the human says, "That one wasn't very good. That one was much better." So, if I ask it a question, and it pops out a bunch of bad pop psychology because it it occurs a lot in self-help books, but it isn't born out in the experimental research, ideally, the human in the loop for the fine-tuning says, "Nope, that's a bad answer." Um unfortunately, they also say things like, "Tell me how much of a genius I am." So, that's why you get a lot of sycophantic behaviors. A lot of that get trained in at this point. The problem is those two things together. While they have produced these amazing tools that truly can answer incredibly challenging problems astonishingly well, it produces hallucinations. In fact, it turns out you can't get rid of hallucinations. For complicated machine learning mathematical reasons, but essentially, they boil down to if you are simultaneously asking it questions about places where the data there isn't a lot of data, but you're also forcing it to give answers, right? I'm going to penalize you if you don't give a right answer, so make something up. Then, it turns out we hallucinations become unavoidable. Uh in these models, if you're wondering, you know, there's kind of nothing behind their eyes, they are genuinely intelligent. They they're kind of like the fast intelligence we've been talking about, the fast thinking, but they don't have the slow part, or they have like a simulation of the slow part. And so, it doesn't know right from wrong. It doesn't know causality. It has trouble knowing what is a good piece of information. is just generating the next word. And what that allows us to do is astonishing. But we can end up with the mistaken belief that oh, it understands me. It understands my medical condition. No, what is encoded is a mass of knowledge pulling together facts from broadly different areas in ways no human being can do, which is amazing. But it doesn't actually understand any of it, which is also why you should 100% prep research, but be skeptical. Part of your job is be skeptical of the AI, challenge the AI, get the AI to challenge you. That's where the best stuff comes happening. Then go in and um bring that informed self in to your doctor who has hopefully the benefit of not getting misled, actually bringing that slow understanding to the problem. >> Certain physicians or uh health care providers uh have been raised to be the all-knowing God. And when patients come in with perhaps AI-generated answers and or questions uh a lot of them, and I hear this from my patients to see other physicians. You know, how dare you question my authority? And you know, and I've actually had patients thrown out of uh physicians' offices. Uh cuz if you, you know, if you don't trust what I'm saying, uh you leave. Okay, what do we do about that? >> One of the historical findings in AI in medicine is uh doctors of all the groups that exist out there, doctors are the most likely to ignore AI advice. You know, as someone that flirted with doing an MD PhD, part of what I realized is I'm not actually all that interested in being a practicing doctor. Um I like to go deep into a small set of problems. You know, as though all doctors could be House uh all the time. One patient and you just get to lose yourself in it. And although I I do in my philanthropic work get actual interesting medical questions brought to me just because no one else is making some progress on it. Um what I understood was the real job of a doctor is to make a large numbers, large as you can, number of people more healthy than they were before. The negative part of that can be you start to treat people like they are patient archetypes. Um you look at three symptoms that appear on this massive sheet and just based on that you go. And instead of well, yes, but you know, Vivian with her freckles uh and pasty pasty skin like she's almost certainly going to have a bigger inflammation response than the archetypal So, I'm going to adjust for that. Our dream for me in some of my work has been can we build medical AI technologies that very much don't replace what the doctor can do. But they're more like a pair of you know, corrective lenses that allow you to see this unique patient a little bit different than the others. So, we would build systems that would highlight, for example, here are a couple of facts about this patient that might be salient to this particular, let's say, diagnosis, um that may not be as important for a different patient. And it turns out stuff like that has been part of where people have seen real value add of AI in medicine. What that's saying is we need to do a different kind of med school training. In the exact same way where we have this sort of human in the loop fine-tuning for AI to give it feedback about where it's giving good answers, we need a little bit of that in in humans. When my son was diagnosed with type 1 diabetes, my wife and I were just told, "This is just how it is. Blood glucose, you can only control it so much, so stop sweating the details." And obviously in my head, it's I build models of the brain. Are you really telling me the pancreas is more complex than the brain? Uh obviously they're all equally complex because they interact with one another, but nonetheless, in my hubris, in 2011, I start hacking his medical equipment. No, don't do that. But, um in this particular case, uh like the rest of the at the end of every episode of Jackass, don't do the stupid thing these people just did. But, in this case, I didn't do something terribly revolutionary. I just happened to be the first person to do it, which was I took a very simple model and I trained it on my son's blood glucose history coming out of what's called his continuous glucose monitor. And it turns out there was a lot that was predictable there, a lot of patterns that had gotten missed because his endocrinologist are the doctors for a lot of patients, helping a lot of families through this process, many of whom were desperate for concrete advice. Just tell me how it's going to be. Don't make it complicated. Well, my wife and I happen to both be scientists, and I'm a computational scientist at that, I wanted it to be complicated because I make bottles already of biology. So, I just happen to be the right person in this moment, not the world's greatest genius. I do, of course, again, happen to be that, but in this moment, I just happen to be the mom who knew something that understandably uh my son's endocrinologist didn't know cuz in 2011, who was thinking about artificial intelligence? Not many. So, I hacked the system and I put it together largely because I was frustrated that I was handwriting numbers on a sheet of paper. That was still in 2011 how diabetes was being treated even in the Bay Area. So, I just was very frustrated and did something different. Now, I think we need to bring medicine along on this AI-enabled journey. Train doctors when to be skeptical, when to challenge an AI, but also train doctors now to be aware of what it's saying and take it as part of their diagnostics. You as a doctor truly have something unique to offer that the AIs are not going to be capable of delivering. You're going to see this as a unique patient rather than just another pattern in a massive data. That distinction between what humans are uniquely good at and what machines are uniquely good at still hasn't filtered in not just into medicine, it really hasn't filtered into a lot of job domains. So, yeah, it's a bit of an education problem, a fundamental human problem, I I would think. And I and I do encourage patients not to hack your medical equipment, please don't, but to be thoughtful. Um it's stressful. When my son would go to bed every night, I get up three or four times a night and poke him. Just to make certain. Which didn't matter much to him, but it was terrible for me to do that for months and months. Just scared. Every morning you'd wake up and, you know, having support that can help you understand this process, help you be more confident in your decision-making, while not replacing your doctor, that's our sweet spot that we're really looking for. >> If someone listening wants to use AI to understand their health better, not just get answers, but build real knowledge, what is what does that actually look like practically? You're not going to build an AI model to manage your type 1 diabetes. >> It is unlikely you will be as mad scientist-y as I am, uh, but that's kind of my day job, so of course I would live my life that way. I think the starting point is something that a number of people talk about, which is something that, you know, no no textbook can help you with or or even a website. Because what you can do here is walk through the understanding. Hey GPT, Claude, what whatever your favorite, uh, tool is, imagine I'm a fifth-grader. Talk to me about diabetes. I just was diagnosed. What should I be thinking about? And after you feel like you've got that, um, say, "Okay, now I'm a high school student. Now I'm a, uh, a graduate student." Walk through I one of my recent companies, uh, we launched is doing something amazing. We've developed the first ever biological test for postpartum depression. Any moms or post-moms out there and you've ever been told it's just in your head, we can literally literally see it in a blood sample. We can see it in your epigenetics. So, turns out I am a lot of things, but I am not an epigenetics expert. It's just I'm not a molecular neuroscientist. I build models of brains. So, to help launch this company, I educated my dumb ass on epigenetics and its application to mental health. And I started with exactly what I said. I you know, I I placed myself at the undergraduate level first. Yes, I do have a bunch of fancy degrees. And then I went to grad student, and then professor, and then some of the unique work we were doing that my colleagues understood, but as the brain and AI expert were alluding me, and I got this wonderful walk-through. And you know, you can take it beyond medicine. I I know this this is a nerdy pastime, but I read lots of science and um economics papers. And even though my work has nothing to do with quantum computing, I came across a really interesting paper, and I said, "You know, here's something I do understand." Um Now, let's start there. And slowly build an understanding of what quantum computing is about. And you know, it took about an hour, and suddenly it just clicked, and I got it. This is a great example, I think, in how you can use AI to build up an understanding. And I think one of the things you should really do is as you gain that new understanding, tell Claude, Gemini, "Hey, give me a paper to read." You know, maybe it starts with a a very understandable news article, and then it moves to a slightly more challenging tech reporting article, and then it moves to an actual science paper. So, you can actually see what people are talking about. You you don't get caught with a blogger who's just making stuff up. >> I love it when I get referred to a something that's kind of made up and and right now AI can't quite discriminate that. Now, you know, in your book Robot Proof, it you know, you get the sense that and you talk about a lot of this we don't want to become I don't want to paraphrase you. We don't want to become dependent on AI for our thinking. Uh because that's going to getting back to where we started, that's going to really hurt our neuro connections, our our branching. Is is that can I say that? >> Yes. I mean, this is a long-term process. I I certainly wouldn't want people to think that just because they asked AI for help that they were damaging their brain, but the the somewhat lazy way I put it is too often AI gives us what we want and not what we need. Uh and and it's happy to do it. Really, you're the one that needs to challenge yourself. AI is the tool for doing that. >> When we get this information from AI, are there the the consumer, are there guardrails that we have to use to decide how how much do I trust this? Uh do I a lot of people go, "I don't need a doctor anymore. Uh I got Dr. Google. Um maybe if I'm sick, I'll go to the emergency room, but I don't need anybody checking on me anymore." Are we there yet? >> I will say this. Uh I'm not coming at this problem from the perspective of AI is a magical army of robotic minions that will do whatever we want and we never have to work again. Again, my research shows that a a smart person paired with an even modestly smart AI is amazing. Smarter than the smartest AI, smarter than the smartest people. That's what I'm advocating for. I mean, we could just simply look at it as I'm advocating for it because I want a future for us, but I'm also advocating because it turns out my research, Anthropic's research, the paper I cited earlier about doctors making diagnostics, the best results are when humans and machines work together. And I think that even holds when we're sort of non-experts. So, again, getting yourself educated and ready uh to go talk to your doctor uh is smart. Let's be honest. There's a lot of gatekeeping going on. And want to go chat with my doctor, but I can't. I'm being asked to submit an answer. Maybe it's even going to an AI right now. I've seen lots of companies, startups getting funded to do just that. Can an AI answer this or should it get sent to a physician's assistant? And if not, should it go to the physician themselves? So, there's some real value in understanding where the easy answer might be the right one and you could move forward with some confidence there. If you're a doctor, I would want to be bringing that unique value-add. What could I have done here that no other, even doctor, would have done, much less any AI? What's my unique approach to a patient like this? And if I'm a patient, boy, do I still want a chance to talk to my doctor. But I can make the most of that time if I can, for one thing address some of my real concerns, um some of the things that are scaring me. You know, the simple reality is terrible things do occasionally happen in diabetes uh with my son with type 1, but it's our bodies are actually really robust, and it's more of a long-term game for most of us. As long as you're you're being thoughtful about lows, and in that sense, being able to put that understanding every moment of every day is not a disaster and emergency, but focus on the things that are important and go talk to your doctor about them. I think AI has a huge role in helping us deal with that.