Yeah. I see you, Dylan. I see you, Dylan. Can you hear me? Yep. Okay, just stay on the line. One minute. I will introduce — I'll go live. I'll put you in the background. Just stay on the line. Thank you.
Health care in between. Physics, art, and math guide got us down this path. Historical tales come alive where science and stories dive. Blended facts with heart. Knowledge from the stars. Me blues we sing to the future memory. And science in the base. Got the fix. Complex subjects made clear that we actually digital turning into dreams through the tales of old.
Thank you so much for joining with me. It is truly an honor and a privilege. MedBicks webcast is a new kind of healthcare media channel to break the silo mentality and foster interdisciplinary thinking. Our mission is to fill that gap and to create a curiosity through different lenses, tapping the passion that guided the historical figures in medicine and bringing present-day groundbreaking research to the front lines. We interview physicians and leaders in the industry about present-day healthcare challenges on our live programming.
Hey, welcome, welcome to the MedBicks webcast. This is Ramanam Raju. I'm the host. You know, we need a tech with — while bringing the technology, we also need a bringing technology therapy for all of us. It's that much pressure nowadays. You know, technology is a pressure to think about non-technical people, how much pressure that it creates. But anyway, that's my blues for today. Let me talk about this artificial intelligence. Everybody buzzing around this topic. Actually, the first topic in healthcare I read — I interviewed on my channel Dr. Robert Pearl. He wrote a book on the ChatGPT MD. You know, he wrote a book on it. And then after that, Harvey Castro was talking about quite a bit. He wrote a lot of books on the healthcare and AI. I think, you know, it is a — it's challenging, to be honest with you. Then they said that artificial intelligence, AI, passed the USMLE test exams for the licensing exam for the medical profession for doctors. Are nervous. They're worried. They're worried about their jobs. They think that AI is going to take over. We don't know what is — what is going on. What is the potential of AI? So, you know, there is a lot of — one of the things is the — is it is a — when you talk about AI, there's two things. One is the hardware and the software. There is a special kind of hardware, graphic processing units. I, you know, I read a little bit about the — about the hardware part of it, but, you know, I have an expert now coming up. His name is Dylan Patel. You know, he — I saw him on Lex Fridman. He's — he writes analysis unbelievable. I mean, I was reading his blogs. He — he gives quite a bit of extensive analysis on the — on the process. So I'm going to bring him and let him explain a little bit. You know, he is very well known in Silicon Valley. He is — his analysis are wonderful. We — I hope he's not going to go too technical. We are all — most of the people who are watching are healthcare people. So he's going to be kind to us and explain a little bit lower version of technical knowledge, but he's going to explain a little bit about the — we need to understand what the AI and what the AI hardware infrastructure requires and what the DeepSeek is coming up. That's what they said that, you know, they changed the dynamics from China. So we need to talk with him. Let me bring Dylan Patel on the line. Hey, welcome, Dylan.
Can you hear me? Hello. How are you? Good. Good. Dylan, give me — nice meeting you, Dylan. Give me a little bit of your background. Where are you from? And where were you born and how did you — how did you end up in Silicon Valley?
Sure. Sure. So, I'm from Cairo, Georgia, which is a rural town in Georgia. And I grew up there in rural America. And then, you know, I sort of ended up in the valley as I — you know, I started my company outside of the valley, actually. But then I moved — I moved here more recently, right? So, the story of sort of me getting to the valley is actually more me being in Asia, right? In Taiwan, Japan, Singapore, focusing on semiconductors before I ever moved over to the Bay.
So, I wanted to ask you — you're from Georgia, he said. I'm from Georgia. I'm from Atlanta, Georgia. I lived in Georgia most of my life. So, what part of Georgia, he said? Well, I'm from Cairo, Georgia, which is a rural part of Georgia. Not really Atlanta, but I do know Atlanta well. Okay. Valdosta, somewhere in southern Georgia, or is it Griffin? I lived in Atlanta most of my life, actually. I moved here also like you, four or five years ago I moved here. But let me — let me — maybe you can see me that much better. Let me — yeah, let me — okay, let me — how do I do this? Yeah, let me do that. This is a little bit of a tricky thing. Okay, so Dylan — yeah. So you came — did you study in Georgia? You studied computer science or what did you — where did you go to school? Georgia Tech.
Yeah, I went to school in Georgia. And then, you know, I graduated in 2018. And I worked in quant risk for a few years before I started my company, which focuses on the semiconductor industry.
So let us talk about a little bit about the infrastructure for the AI. Doctors don't know much about it, but give us a little bit about the infrastructure, what it requires. You know, we are hearing a lot of different conflicts. Some of them say we need a multi-billion, $500 billion investment required, and then China came out with a DeepSeek with a very $5 billion investment. So what is this confusion? Is it really — what is the — a little bit, give us a rundown on the infrastructure side of the AI.
Sure, sure, sure. So I think, you know, the interesting thing on the infrastructure side for AI is really around just how pervasive it is, right? When we look at power consumption by state, right, something pretty simple, there are certain states where more than half their power is going to be consumed by AI just in the next few years, right? So while many states will, you know, won't have too many data centers, some states will just be — there'll be data centers everywhere, right? And all of this — these data centers and all this compute will be leveraged on, you know, all sorts of areas, right? Today we already see immense uses of AI in things like drug discovery, and that's only going to get stronger, right? But there's also a lot of things in genomics, biogenomics, proteomics, use of AI. And that, you know, sort of — at least in the medical field, there's a lot of use, you know, starting to come up. Obviously, it's a little bit slower because of regulatory reasons, but in areas like software development, AI's use is dramatic, right? And so when we look at the use of AI within industries, different rates, but it's pretty clear with the pace of AI innovation, right, how fast the AI models are getting better, people are going to be able to use them to solve problems faster, more efficiently, and deliver better lives for people.
So Dylan, everybody talks about GPU, graphic processing unit. Tell us what is the role of a graphic processing unit. Why is it so important? Nvidia manufactures them. Tell us a little bit about that, why that is so important. And then Google calls it a different name, tensor, TPU. Tell us a little bit about that — not too much details, but the technical — just a basic high-level description.
Yeah, I think when you zoom out from computing, there's sort of just a couple different kinds of chips, right? There's CPUs, which are, you know, generally what you find in most everyday devices and laptops, etc. And they're good at doing one thing really fast, right? And so you have CPU cores, you have just a few CPU cores, and they do some things really fast. GPUs or graphics processing units are exactly the opposite, right? They're actually quite slow at doing individual tasks, but they're really, really good at doing very parallel tasks, right? So graphics processing units were invented to draw the pixels on your screen, right? Turn them on or off, make them yellow, make them red, make them green. Which colors do you want to show? How many do you want to show to make it so you can see pretty pictures or play video games or render, you know, some radiology thing, right? These are all like graphics processing tasks. And so AI happened to, you know, just so happened to be developed on GPUs initially, right? Initially, you know, CPUs were so expensive. AI requires a lot of compute. And then everything else was very specialized and expensive. So there's these famous AI researchers, Ilya Sutskever, Geoffrey Hinton, and Alex Krizhevsky, right? These guys made AI — the first AI model that we would recognize as similar to what's today called AlexNet, right? And all it did was recognize images better than humans can. And they did this on a graphics processing unit from Nvidia, right? And this is because while CPUs were really slow because they're really single — you know, focused on a single task, GPUs are focused on doing many tasks at once. Now, initially GPUs were used for graphics and it just happened to be okay for AI. But now if you look at the best GPUs that Nvidia makes for AI, they can't even do graphics processing, right? So they're still called GPUs because they share a similar architecture, but they're not able to render pixels. You know, they're not — that's not what their task is. They're actually just focused on AI processing. And then around the same time that Nvidia made their GPUs start being focused on AI, around 2015, Google also did the same thing, but instead of using GPUs, they made their own chip called the TPU, right? And so this is where sort of the initial waves of AI hardware came from. And now today, you know, every Google search runs through a TPU.
Is it just the name itself, or is it a functionally different way it approaches the problem, TPU?
So the problem at its core is taking, you know, is multiplying and dividing and summing a bunch of numbers in parallel, right? And so in that sense, the TPU is very similar to a GPU. But in many ways, the architecture is different, right? GPUs use, you know, you use various like the microarchitecture — I don't know how much you want to get into, but it's — they're both really focused around multiplying massive matrices of numbers, right? That is the base thing that they do. And so yes, there are a lot of differences between the two because of lineage, because of how you program them, right? Nvidia's focus is: I have to serve every company in the world and help them do AI. So I need to make it really, really easy to program and I need to make it so every generation is compatible. Whereas Google is focused on: the TPU is just for me and just for my workloads and just for a number of them. So I am only focused on making it programmable for my team. Yes, I do try and sell them some, but I'm really focused on — and I'm only focused on one use case, which is AI, right? Whereas Nvidia has to focus on, yes, AI, also has to focus on image generation, has to focus on, you know, all sorts of things related to simulations. So Nvidia has a wider net that they have to target.
So the neural network is supposed to be — is not exactly the brains — a combination of a kind of a copy of the neural network of the brain, but it's not true. But they actually — these CPUs, GPUs are meant to compute these networks, right? That's a neural network. So find a solution, recognize the images or anything like that, radiology or any images — these GPUs are more useful to do the neural network kind of architecture.
Yeah. Yeah. So neural networks are, you know, the way you can think of them is it's a major — you know, huge numbers of numbers that you're multiplying. You're doing a simple operation: you're multiplying, you're adding, right, and then you're normalizing, right? You're either taking an average or a sum, and then you're maxing the number, whatever it is. These are the simple operations that you string together, and you do billions if not trillions of operations every time you create — you go through one of these large language models, right? You go through billions or trillions of calculations depending on the size, right, of the model, and then it generates a piece of text, right? And that piece of text is, you know, if you've watched any of these AI models, they generate them in sequence, right? They don't just give you the whole thing. They give you, you know, just small pieces of a word and it slowly streams out, right? So if you go to ChatGPT, that's what it is. Every single time it's outputting a little piece of text that's going through the full model, right? And so there are trillions and trillions of calculations needed to just make simple statements like 'the sky is blue,' right? Because, you know, if you ask what color is the sky and it says the sky is blue, it takes trillions of calculations to calculate this because you're going through the model many times. And so that's why these parallel processing units are very useful, is because in the model every parameter, or you could think of it simply like a neuron, needs to be multiplied, right? I'm trying to simplify, and, you know, so there's some complexity loss, but each of these neurons basically has to have multiple operations associated with it, and they can be done in parallel. And so that's why GPUs are really useful, but also why TPUs are useful for it, and why CPUs are not good at it.
Yeah, the training data — a lot of AI depend on a lot of training data. So the training data — the GPUs are fundamental to the training data and the architecture of the transformer network, they call neural network, for example — the GPUs are fundamental to the — is that right, if I understand you correctly, that these GPUs are fundamental to the training data?
Yeah, so it's interesting. These large language models are trained on two types of data. One is literally everything on the internet: every book that's ever been posted on the internet, every scientific article, every post on Twitter, every post on Reddit, right? So all the way from social media to, you know, informational content, transcripts of YouTube videos, everything on the internet. It's trained on that first, right? And then it's trained on a second set of data, which is much smaller. The smaller set of data is specifically labeled, right, by people, right? And it's designed to give you instruction following and safety and all of these things, right? Like if you ask the models to say something bad, they won't — they'll refuse because that's been — that's in that second set of data. If you ask them a question, oftentimes it won't say — it won't just ask more questions, right? It won't answer with more questions because that's, you know, if you look up questions on the internet, oftentimes they're in lists of questions. Instead, it'll give you the answer, right? And so this is like some features that are trained into a specific set of data. And when you look at specific applications, right, this is where AI can fall on its face, right? Because you have the entire internet. And everyone knows the internet has every opinion you could ever want, right? You could have, you know, pro-X, you know, anti-X, anti-Y, anything, you know, politics or, you know, food, right? A discussion about what the best bread is, right? There's going to be so many discussions about stuff on the internet. But then the specifically labeled data is the sort of like what you want the AI to act like. But the challenge is getting the specific data is very costly, right? It's cheap to just take everything from the internet, right? It's very expensive to have very specific data labeled by experts, right? And so when we look at applications like the medical field, well, yes, the model has read every single medical journal ever, right? That's awesome, right? And so it generally knows enough to be dangerous, but sometimes it's inaccurate because it doesn't actually have, you know, notes — the sort of the notes from the doctor's visit, right? It doesn't have all these other details that you would want to have. And it's never had — it's never been presented with these sorts of things because this data is all hidden away under HIPAA and all these other things. And there's no doctors writing what, you know, like, hey, here's this, here's the symptoms, here's the meeting, here's the notes, here's the issue and how it was solved. This isn't — this isn't in the training data set, right? This isn't in the specific small set of data, right? And likewise, you know, you take a radiology exam and it's like, well, actually models are getting pretty good at radiology because there are some data sets where people have paid for radiologists to look at, you know, the readings and then say what's wrong, right? And so models are getting somewhat good at that. So it's all about where the data is too, right? So that's another aspect of these.
You know, the doctor — you know, during — before the pandemic, there was a lot of talk about AI. You know, that pitch AI — Google said that we can predict the flu season three months in advance. So in fact, they could not even predict the pandemic, which is a storm and killed eight million people. So there is a lot of skepticism among the medical community about the AI's prediction. What changed that prediction now? You're all confident AI guys. What changes that prediction now? Tell me that a little bit.
Yeah, so I think it's, you know, I do want to note, right, it's hard to get the data for the medical field, and so models are generally going to lag in capabilities for medical versus other fields, right? I think that's just, you know, this is something that I think if we can get the data out there, you know, if we could get every single medical note ever, right, from every doctor that's been digitized and every EMR and train on it, I bet these models would get a lot better, but we're not able to, right? So just to be clear, because this is like privacy concerns, right? So the reason — or sort of the reason why we're so excited about AI is not because, you know, the data — you know, the data is getting there slowly. You know, people offer up data, you know, hospital systems will work with a model company to train a model just for them. Like, these sorts of things will happen over time. But more importantly, the internet data has been there, right? And people have had that for years. But in 2020, no model could write code, right? No model could even talk to humans well, right? By 2022, 2023, they were starting to get okay. Now we're in 2025, and software engineering — the software engineering market is horrible, right? If you know any software engineers who are just graduating, it's very hard to get a job. If you know any, you know, people that are, you know, new — you know, who have been laid off and the market is tough, like it's very — the wages are not going — their wages are in fact not even going up, they're kind of flatlining because AI is so useful in software engineering today because so much — the models have improved so much in capabilities and scale, their ability to learn and the data sets provided to them, they're able to learn from them and actually do parts of the job. Are they replacing programmers outright? No, but they are making programmers much more efficient, right? And so this sort of — is the hope that AI can, you know — it's — there's some bad ramifications, right? People not being able to find jobs, but, you know, the flip side is software — developing software is much cheaper now. And the medical field has real big problems in America with how expensive the medical system is, and there's got to be ways to help improve the cost. Now, most of the cost is in administration, right? I think, you know, that. And so, you know, there's a lot of administrative tasks that people are trying to push models into, and that's really where I'm most excited about it, is because administrative tasks have less regulatory burden, right? There's less risk if something goes wrong. And so this seems like the right spot. And so when we look at like the — the insurance claims industry, right? There's a lot of AI integration already happening there. And maybe there's like, you know, for the doctors out there, you know, it's like, oh, well, my claims are automatically getting rejected by AI instead of a human. You know, this is possible. But I think that things will get better and more efficient with AI, and there's a little bit of — there's a little bit of, you know, going to be growing pains. People are going to hate it, but at the end of the day, society itself is going to change, right? AI is, you know, the intelligence level of AI is huge. Yes, it gets stuff wrong. It'll get stuff wrong less over time. It's not going to happen in the blink of an eye, you know, like many people in Silicon Valley believe. But it is going to happen, and we are going to get more efficient as humanity.
Dylan, you used a word called 'aha moment' — that, you know, AI kind of — it doesn't need to be trained on a particular data. It started thinking on its own without training with that particular data. That the trained data is actually never trained on those data, but still is answering the questions. That is really very strange, and it is a kind of intelligence — is exhibiting a kind of intelligence, right?
Yeah, so I think there's two different things, right? Like, humans don't need to read the exact text or see the exact thing to be able to predict what happens, right? Now, obviously, if you've studied something a lot, you're going to be a lot better than someone who studied it a little bit, right? But, you know, I could read a bunch of papers about diabetes and then all of a sudden start opining on what the best type of insulin is for a patient, right? That doesn't mean it's the right answer, but I can definitely give opinions, and hey, a lot of the times I might even be right, right? But I'm definitely not a doctor, right? And so this is sort of the case, right? Like, you know, there's doctors and they're trusted with a certain level of stuff. PAs are trusted with a certain level of stuff. Nurses are trusted with a level of stuff, you know, sort of, and then, you know, CRNAs are trusted with a level of stuff. They've all had different levels of education, but they've learned some stuff. They've generalized and they have opinions. And this is the way models work too. I think, you know, yes, it hasn't seen the exact same thing, right? That's because it's generalized and learned generally what's up with that field. It's read every paper about insulin, right? So it has opinions, right? Even if it hasn't read every paper about insulin, it still has opinions because it's a model, right? You know, humans are just like this too, right? But what's interesting now is that models are able to — are starting to develop the ability to think in a chain, right? Before they give an answer, right? So before, models were: put in data, get answer.
You think that's it? Is it thinking that what you're saying? It is.
Yeah. So this is something — this is something that's been developed in the last six months that's been really exciting for software and math. It hasn't made its way to the medical field yet, but it's something that's been really exciting in software and math. Basically, college-level math tests were initially 4% before this innovation happened, to now they're scoring 70, 80%, the best leading-edge models, right? And this has just happened in six months. And the same applies to various certain coding tasks. The same applies to — yeah. And so this feature or this ability for models to not just do input data, give answer, but instead input data, think on its own for a period of time, and then give an answer, has been tremendously powerful in increasing the capabilities of models over the last six months. And sort of this is something that it's called reasoning, right? You know, what the industry calls reasoning. This is a huge capability that's been very helpful for math and coding, and the hope is that this can be extended to many domains, not just math and coding.
