Latest / Two PhDs on a Pod / AI Without the Hype: What’s Real, What’s Not, and How It Works
Transcript
- Guru Banavar: Hey Guru. Hey mama. Well, we're gonna have exciting episode today, and we're starting off with something you told me, which is that an AI system was ninety-nine point nine percent accurate at diagnosing chest x-rays, but when they actually put it in the clinic, it was completely useless. Yeah. You know, this happens quite a lot in machine learning and AI. This particular system wasn't Predicting a disease at all. So even though you gave it an image, it wasn't actually looking at the medical, the biological aspects of the image and trying to predict disease. What it was looking at was some kind of weird text around the image and the fonts and all of that stuff. And it had figured out which hospital that happened to be from and figured out which ones had the sickest patient. Scans and it was just guessing the hospital. And so it was right a lot of the time. Okay. It was guessing the right hospital. Wow. But it wasn't guessing or predicting the disease. Wow. Wow. So the the training set included text that was very easy to interpret, but when the AI model was ⁓ presented with actual X rays, it just choked 'cause it had no additional text. ⁓ Provide ⁓ an X-ray image with the appropriate clinical information that's associated with the X-ray image. So, yes, it was it was not looking at individual text items, it was it was in looking at individual pixels, and it was trying to see, okay, how do I match this to something that makes sense? And it turns out that the hospital made the most sense for it. So it does what it does, you know, it does. It does what it's optimized to do. Wow. Amazing amazing and terrifying example, which is why doing a very, very robust independent validation from different hospitals presum hopefully would be a good test. And that's why this model failed in the clinic. Exactly. So you know, the batch effects and the side effects and the site effects and and the technician effects, all of those things matter a lot. And so the data quality has to be good. The model setup And training process has to be good. And that's the kind of stuff we're going to talk about today. And by the way, Momo, this kind of stuff happens a lot in, especially in medical AI, where, for better or for worse, people think of machine learning and AI as some kind of magic. But it is very important to lock down the correct kinds of data. the highest quality data that you can possibly get your hands on and to do the kinds of independent validation in different types of scenarios to the point where you can be confident that a model will perform the way you intended it. And it all goes back to the intended use and all of that stuff. So that's what we're going to talk about today. And hopefully we'll throw some light on what kinds of methods will make AI more reliable. And in particular, I also want to say today's modern AIs, which are all, you know, large language models, and underneath that something called a transformer, which we'll talk about. where is that headed? Where can it go and what kind of impact will it have? So it's going to be a lot of fun. I'm super excited. I am representing basically the world. So I'm going to represent the clinicians, the scientists, the general public, and I'll be asking you a million questions. And yeah, I'm excited about. Examples and how AI works and a little bit of history and then current applications and how people can recognize and you know when AI is actually capable of something versus not and what kind of questions they should be asking and then where the future is. So yeah, super exciting. I am Momo, a biochemist. And I'm Guru, the AI expert on this particular episode, who's going to be answering or trying to answer all these questions. We're two PhDs on a pause. All right. So Guru, what's on your radar? Awesome, awesome. Okay. So Momo, I have some ⁓ super interesting things on my radar. Wanna share my screen here to show you a couple of very ⁓ cool things that happened in my life in the last couple of weeks. One of them is this dish that I cooked up in the last, I think it was over the weekend. ⁓ this is called Singapore Noodles. It's one of my favorite dishes of all time, actually, because it's just the right combination of ⁓ sort of Indian ingredients and spices and everything, and you know, sort of Far East, you know, Asian types of ingredients that are that comes together in Singapore, which is also some kind of ⁓ melting pot of many different types of immigrants, especially from India, China, and all the surrounding areas and stuff like that. Was created over there. So the basic idea is that You take a bunch of nice vegetables, which is what I show on the top part of the picture here. And then you take ⁓ lots of tofu. I've cubed it and I've ⁓ sort of stir-fried it with some ⁓ good soy sauce. And I take eggs and just scramble the eggs. And you put it all together with some really good no noodles. I I happen to like sort of brown rice type noodles. Sometimes it's no it's sort of hybrid types of noodles which are a little bit more nutritionally ⁓ interesting. And then I put some sesame on top and ⁓ my god, it's such a great dish. I do this I make this quite regularly at home and my family loves it. It has the right combination of proteins and, you know, all of the ⁓ other ingredients, you know, the fibers and all of that stuff and and depending on what else you choose to put in there, you can make it even more fancy. Like for example, you could put peanut powder on top of it. It'll be awesome. It'll be different, but it'll be awesome. It's it's making me hungry actually. This looks delicious. So it's ovo vegan. Right. There's no dairy. There's no dairy. ⁓ plant-based plus eggs. Yeah. It looks like ⁓ lots of vegetables, ⁓ lots of protein and quite a bit of fiber and then a little bit of carbs and you can obviously adjust those ratios however you want and it's like an experimental platform. You know, exactly. You can cook the tofu different ways. You can introduce different vegetables on different occasions and use different noodles. Like our normal ⁓ you know local H Mart has probably a hundred different types of noodles. So you can exact play with those. ⁓ I would love for you to post the recipe for this and then I want to make it because man, it looks amazing. Happy to do that. Happy to do that. Another thing ⁓ that's on my radar, I always want to bring in ⁓ some something about the arts into our podcast. Okay. I love the arts. I spend quite a bit of time in the arts of different types, you know, music and visual arts and other kinds of performances and everything. So this time I was ⁓ over the weekend, I think I was at a concert locally, which ⁓ is a group from Macedonia, from the Balkans. Which is from your part of the world, Momo, and I loved it. Loved it. Actually, this particular band took some of the classic Macedonian songs and rearranged them using a jazz idiom. And so there was this guy, the ⁓ the the ⁓ trumpet player, this drummer, the the bass player and the c and the piano player with this singer. In the middle, the the woman singer who had just figured out how to take the real classics that everybody knows how to sing and turn them into these modern versions that blew my mind. I was like standing in the back of the audience because the place was completely full, there was no space. ⁓ small place, you know, like a small studio place, like that's the kind of music that I normally like to go to, not the big stadiums and stuff like that, but the small places which are very intimate. It was so amazing. We took a picture with this lady and and I thought that ⁓ you know, people like Toshe Proevsky, who is a very famous guy, died young but very prolific in terms of how much music he put out and all that stuff. They sang a lot of his songs. Many other people I enjoyed it so much. I thought I wanted to highlight that. Amazing. Amazing. ⁓ is there name of the band that people can maybe look up on YouTube? I don't think they have a name of the band. I they actually came together a little bit ad hoc. I think one of one or two of these people are New York based and this lady was traveling from Macedonia. So she Okay. She just kind of put the band together. It's a jazz group. So Stars Aligned. Yep. Incredible. Incredible. Yeah, yeah. ⁓ but you know what? I will put the name of the lady in the show notes. And so if people wanna look her up and see what other music she's recorded or whatever, that'll be great. So this is one of the One of the unique things about New York City is you get this these kinds of things on a regular basis. Yeah, yeah, yeah. Yeah. We enjoy that. ⁓ part of part of ⁓ living in a big city is ⁓ getting that, you know, the melting pot experience from all different parts of the world, which is which is something we consider amazing. So Momo, what's on your radar? All right. So ⁓ food, of course, is always on my radar. Food and zen and exercise, but I haven't Talked much about exercise. One day we will. Okay, so I just wanted to present this little concept that we use a lot at home, and ⁓ I call it tortilla. ⁓ so ⁓ everyone is familiar with pita bread and tortillas, but we make tortillas because that way we don't have to make separate pita bread and tortillas because it's a lot of work ⁓ to do twice. So we make these, and so this is leavened bread with yeast, so it's just a four-hour Rise with ⁓ instant yeast and it's just whole wheat flour. And ⁓ after four hours you just roll it out, ⁓ and fry it, no oil or anything in the pan, just high heat, and it's sixty seconds per side. But what's really nice about these is we can have like a Mediterranean spread with like hummus and ⁓ baba ganoush and enjoy these with ⁓ these portillas, but we can also make burritos, and we make burritos a lot in our house. So they're kind of universal and we just make them and stack them and and then we use them over the next couple of three days. ⁓ and so they're wonderful. So I encourage people to try these. ⁓ the ingredients are just literally ⁓ whole beet flour and yeast and salt and water and then some kind of fat. And so you can do olive oil. ⁓ we figured out that duck fat is the best for this. It just gives it the best texture for us. But you can experiment. It's a platform. So really excited about those. And that that ⁓ I'm gonna I'm gonna have a separate ⁓ discussion later on on wheat and whole wheat and processing wheat and all that stuff, but not now. Let's go to the next slide. By the way, I should just mention when I look at this, it reminds me of paratas from the Indian, you know. Cuisine. Right? Which is which is usually stuffed with something like allu meaning potatoes or vegetables or something else. But this is of course not stuffed, right? This is just wheat. This is just wheat, it's not stuffed, but you can absolutely make any masala you would like. Any vegetable with any kind of you know, I mean you could just fry some vegetables and tofu with gara masala in like fifteen minutes. I agree. Fill this sucker up and then eat it with yogurt. It's it's incredible. It's just such an incredible platform. It is, it is. Yes. Very cool. Next slide is okay, I just want I I run into people regularly who don't know this concept. So I'm like, hey, let me present it at the podcast. So the concept here is that drinking unfiltered coffee. raises LDL and APOB, which are some of the highest highest risk biomarkers for cardiovascular disease. So we want to lower LDL and APOB. And unfortunately, coffee has these two compounds called caffestol and KW or CAW that have been shown in randomized controlled trials to raise both LDL and APOB, APOB being a better marker of ⁓ cardiovascular disease. So it turns out that The paper filter that was invented in Germany in like 1918 or something like that actually adsorbs, so basically traps both caffistol and caviol very efficiently. And so if you do a clinical trial and you give people coffee unfiltered, their LDL and APOB will go up. But if you filter it, their APOB and LDL will not go up at all. So you can basically, by filtering coffee, You can have all the benefits of antioxidants and caffeine, but not have the the these these negative effects. Now, what I'm showing the picture here is what we do at home. We make basically make three types of filtered coffee. ⁓ the machine can be used to make espresso, but then we filter it every single time through a you know, Melita. This is Melita filter. I only trust Melita because it's the original German one with which they've done so many studies. I don't know other filters, whether they do they have the same effect. On the left side, what I'm talking about, what I'm showing is my favorite kind of coffee, which is Turkish coffee. So I don't actually boil the coffee like in a Turkish like Jezva, but I do, you know, add boiling water to that ⁓ to that container, add my favorite ground coffee, and then let it sit there for two or three minutes, and then I filter it. And last but not least is the pour over coffee, which is shown on the right side, where you just load coffee beans on top of the filter paper and then you load it with boiling water. And it goes through and all three of those are super tasty and just go experiment, but all three go through the filter. So that's really cool that we have all the science to guide our the way we eat our food. And then Very cool. Well, I'll start looking for what we need to do for filtering as well. ⁓ I only have one cup of copy coffee per day in the morning. And ⁓ I just use some kind of a pod to make it to make the whole thing and yeah. You know, sometimes we use a ⁓ a similar machine for you know grinding and and brewing the coffee, but ⁓ I don't think we've used filters. So I'll look into that. Thank you. Yeah, I'll look into that. And and ⁓ one question that I that came up is people who use metal or plastic filters, they're asking, does that work? It doesn't. So the purpose of paper filters is not to trap the co the benefit here is not to trap the coffee. The benefit is that it it physically adsorbs, so traps these specific molecules. Whereas metal and plastic filters simply don't have that ability. They don't they don't absor adsorb these things. So it has to be a paper filter. And like I said, I only trust Melita filters because they're the ones that have been used in clinical trials. Cool. All right, let's do one more one more radar. Yeah. Well, so I just wanna throw out a reminder here that, you know, people don't understand how busy you and I are probably, but we're busy. Always, always busy. There's just so much to do at work. I know. You know If I don't receive any emails or any new like ideas, we still have like years of work to do just based on what we've come up so far, right? And so there's just nonstop work and you can end up just working continuously. But it is very important to balance your life. And one of the ways to balance your life is to create zen situations, which are walking or whatever cause you know, ca causes you zen. And for me it's it's One of the one of the things that really puts me in a zen situation is to walk in this town called Edmonds, Washington. And this was from a couple of nights ago. We had a friend from New Zealand and we took her out here. And this is the view, just an iPhone snap. We just, you know, I just raised my iPhone and boom, took a picture of this. These are the Olympic mountains with the ferry. And I mean, it's just an incredible thing to sit down on a bench and enjoy this kind of view. So just a reminder Go spend some time in nature, go enjoy the moment, you know, spend some time with friends and family and and kinda calm down because I mean I could tell you I love my work. I love what I do. Like it's super exciting, but it doesn't mean that you should be in it twenty four hours a day. It's just so intense. And even though you and go to go to those concerts if you can too. And go to those concerts. Exactly. So those are all kinds of balancing your life, even if you enjoy what you do. So That's what I wanted to point out. So thank you. That's a that's a good reminder. Thank you very much. So shall we jump into the main topic for this? Let's jump in, Guru. I'm excited. Let's see. I have a million questions. Let's go. Let's go. I think we should start with a little bit of history of machine learning. You know? Okay. ⁓ so let me let me let me take a little if you want, I can take a little trajectory of of AI and machine learning over the years. And know, please. I particularly like the example of cats and dogs and early machine learning. And you've been like a leader in machine learning for twenty five years. So you can give us an actual pr factual perspective, not something you've read. Well, ⁓ I should mention at the at the front here that my first AI class was in nineteen eighty nine. So that is thirty seven thirty seven years ago? And most people have not ⁓ did not hear about did not know about AI back then. Well, back then there was a particular way of understanding how to build intelligent systems that was called expert systems, by the way. That was back then. And the idea was to kind of capture the knowledge of some expert and put that knowledge in some kind of ⁓ representation. It's called knowledge representation. So you put that in the knowledge representation and you have a logic engine looking at the knowledge representation and responding to questions that people may ask or solving problems that people may pose. Okay. Okay, so the that's that was actually the first generation of AI. It's called expert systems. And over time, what we realized was that these systems are very brittle, meaning that if you go a little bit outside of that knowledge base that's been captured and represented, the system either doesn't know what to do at all or it gives some completely wrong and well ⁓ you know falsifiable kinds of ⁓ information which nobody liked. And so there was a lot of funding, believe it or not, in the eighties there was a lot of funding for expert systems around the world. The Japanese government put a lot of money, the US government put a lot of money, European governments put a lot of money, and lots of the first hype wave then was back then. That was the first, that was the first hype wave, correct. Okay. And you know what happened was it's super interesting, Momo. We realized that expert systems will not scale very shortly after that less that class that I took in 1989, right? So we humanity realized that. And so people basically walked away from AI for quite a long time. That I was one of the people who was like a refugee from the world of AI because you know we learned a lot of techniques. We learned logic, we learned representation, we learned reasoning, we learned all kinds of things. But we said, listen, this is not going anywhere else. We need to come up with completely different ways of doing things. And so it turns out around that same time, some other groups of people who were not working on this expert systems. also realize that they can do something called example based learning, which basically meant if you want a machine to learn something, don't give it the rules, give it examples of that phenomenon. Now you can easily see where this is going, right? The cats is a great example, right? So you give lots and lots and lots of examples. So people started doing this in signal processing. They would they would actually show examples of what a good signal looks like, what a bad signal looks looks like in electrical engineering. And the system would learn how to how to sort of recognize the good versus the bad and so on. So they were able to classify good versus bad just by showing lots and lots of examples of that data. So eventually what happened was people learned that that could be represented through an architecture called neural networks, right? Which is basically very small ⁓ modules of logic, which I can explain later if if we if we get into it, that can be then connected together in many. Layers with huge networks at each layer. And these networks would basically capture the knowledge represented in the examples as weights in the inputs and the outputs of those neural network neurons. Each component of a neural network is a neuron, right? The neuron artificial neural network is what's called, and components called an artificial neuron, captures the weights of all of the input signals and what the weights should of the output ⁓ signal should be and so on. I mean that sounds like a foundation model, right? Like the beginnings of it. It leads to it eventually. Yeah, the beginning of the Where was this example based ⁓ approach ⁓ started? ⁓ this example based approach was actually started in the ⁓ probably in research terms it was started in the eighties. Okay. In research terms, but it became a little bit more mainstream and there were a small groups of people here and there working on it throughout the nineties. And in two thousands is when people realize that if you have a large enough neural network and if it's layered well enough and you have the cer the right kinds of feedback from layer one to layer two or within the network themselves, it's called convolutional neural networks, meaning that they are, you know, they have these convolutions, which are just formulas for how to feed back from one layer or one component to a different layer or different component. Once you have these kind of architected ways of putting these neural networks together, it started performing much better than before in terms of identifying patterns. So that's when the field realized, this was in the mid-2000s, is when the field realized that, you know what, there's something going on over here that we should learn more about. And there were small groups who decided the best way to do this is to collect lots and lots of images. Back then, remember we didn't have People taking pictures of everything with smartphones, right? So there there was this one group who created something called ImageNet, which is the first large scale image database. And they said, if you can recognize any one of the images in ImageNet with a certain level of accuracy, you get a prize. So they created a competition from it. All the PhD graduate students would compete in it. By around 2012, 2013, so in the 2010s. The models that were doing the recognition of the ImageNet images got to like eighty, ninety, ninety-five, ninety-eight percent accuracy. And they said, darn, this is something super cool. Let's scale it. So now they took it from images to other types of data. You know, the people started using it for language data, which is just words, right? People started looking at videos. Actually, I should mention my wife. Is one of those people who back in 1989, when I took my first class in expert systems, I met her in that class. Believe it or not, I met Navenka in that specific class, expert systems class. And I've known her since then, right? So she is one of the first people who also recognized that these networks, neural network models can be used for video processing. And she recognized objects and videos, words in videos, scenes, trajectories of An individual person, all of those kinds of she started recognizing those things using those models of neural networks. And eventually in 2017, somebody in Google, actually, one of the researchers recognized that if you structure the neural network in an exact particular way, that you can capture something called attention, meaning that you look at a long set of data, you you're You're focused on one piece of that long structure, but you can see the impact of all the other pieces of data on that individual structure. That's called a transformer. Okay. Okay, there's a transformer is just a small set of neural network components that are structured in a particular way. Wow. And if you start scaling transformers to millions and billions and and so on, they started seeing that the performance of prediction of the next word given a lot of words or next token given a lot of tokens just goes through the roof. So that happened in twenty seventeen. So let's say ten years ago. Okay, but you're saying so Google actually ⁓ invented the modern transform. Correct. Correct. Correct. Wow. The guy at Google ⁓ who invented his, you know, is is is now in a startup and and he's running his own ⁓ show is trying to do other stuff. But that invention from Google was picked up by a lot of other people and they started scaling it and scaling it and scaling it. OpenAI was one of the first companies that scaled it to the point where they realized that language prediction is done extremely well with these transformers. If you scale it and structure it in a particular way. Okay. And so then they released ⁓ Chat GPT, as you know, in 2022. And then the rest is just modern history that you know history. So this is from 1989 till today. Yeah, that's a fascinating story. Thank you. And quick question. Someone won a Nobel Prize for these neural networks. Can you just talk about that? Yeah. So one of the you know, one of the fundamental ⁓ structural inventions that happened in neural networks is actually coming from the work of three people who all three of them got originally they got the ⁓ they got the touring award. and ⁓ you know each one of them is in a in a in a different part of the you know different part of the industry and ⁓ in and the academia and so forth. so the three of them have actually I should give the three of them credit, not just the one person, because all three of them were developing the concept In the same way, in the same time. But they were just applying it in different parts of the problem space. Okay. So ⁓ so let me let me just make sure that I I I have all three names in front of me. Touring award, winners for neural nets. So the so one of them is Jan Lacoon, who you who you know very well, probably, or if you don't know it, you should have heard about him. The other one is Joshua Bangio, who is in ⁓ who's from Canada and also was very fundamental in ⁓ inventing neural networks. And the main guy who is given the most credit is Jeffrey Hinton. ⁓ Hinton is the guy who ⁓ probably started thinking about neural networks and their architectures. A little bit before the other two guys, but I I give credit to all three people. Now Jeff Hinton ended up getting the Nobel Prize Momo. And ⁓ there's a lot of people who feel like all three of them should have been shared, yeah. Gotten it. That's why I'm mentioning all three of them. And the these inventions ⁓ enabled Google to basically come up with the transformer model, correct? Yeah. Exactly right. Okay. So Jeff Hinton started his work probably in the late seventies. ⁓ beyond the queen. Yeah, late seventies. ⁓ or early eighties. ⁓ Okay. That was when he invented the first, you know, initial parts of ⁓ the neural network systems. I think Jan Lakoon's PhD was in around the same time, maybe early eighties or something like that. Wow. And Yashua Bengeo, by the who I know Yashua pretty well. I used to work with him back in ⁓ when I was at IBM doing all the work. He used to visit us and we used to visit him and all that. We had collaborative projects going on. I mean, I know Jan Lakoon as well pretty well. I've done Multiple ⁓ panel discussions with him and all of that stuff. But I just want to make sure that all three of them get the credit for it. Yeah. Amazing. Great historical perspective. Okay, so now we're in the Transformers, and that's basically the modern AI. Yeah. All right. So well, if you if you don't mind, I want to kind of step back a little bit from from AI and and talk about the even bigger picture than AI. Okay. Okay. So what what are we trying to do with AI? I think the way we ⁓ have defined the word AI is a little bit too narrow. Okay? Just because we kind of get influenced by all the Hollywood movies and the robots and all of that stuff. So we kind of think of r you know AI as these kind kind of ⁓ weird organisms, maybe anthropomorphized organisms, meaning people looking or human-looking. Robot that is super intelligent, super capable, maybe dangerous, and will kill everybody and all of that stuff. So that's what most people think of as AI because of the movies. But if I step back and think about why the field of AI started and what it really was trying to do, it it is really about understanding nature. It's really about Understanding why life can do smart things. And it's not just about the smart things that humans do. Life does smart things. There's biological intelligence, right? Complex, complex things. Yeah. Correct. Like you know, you I mean your your biology knows how to configure itself to be able to help you perform very complex functions like energy production, like mitochondria. And you know, and there are weird things that happen in your biology, like cancers, right? There's all these weird phenomena in nature that we're trying to understand. And in order to understand it, we need to be able to simulate nature. So that's I think the fundamental reason why AI exists. Interesting. Okay. And yes, of course the In the limit, what we want to do is to be able to create a full human being and so forth. Right, right, right. But vibe coding was not the initial thought for AI use. Correct. Exactly. Yeah. And and so the the ⁓ the other part of it is you know, understanding nature also means that you need to be able to look at two different phenomena phenomena, two two different phenomena and say, is there a are these two phenomena the same? Or are they different? And if they're different, what exactly makes them different? Right? So that's called the signal that you pick up from comparing two different phenomena. So I if I step back, this whole process started with statistics. Statistics is a way of analyzing phenomena. In any any kind of phenomena, you just sort of take the numbers and data that comes out of a phenomenon. And you ask questions about that phenomenon. Yep. That's what statistics is about. So a lot of people kind of say statistics is the foundation of AI. So it's math is really the foundation of AI. Okay? And the purpose of AI is to understand nature, right? And to simulate nature. But I i if I may, so is it true that statistics tells you the two systems are different, but it doesn't tell you all the details about what the differences are, whereas AI should be able to get to that level of detail. Correct, correct, correct. Yes. So statistics is intended for explaining or describing systems. They're describing in the in in in some very specific way, meaning that what is the same about the system, what's different about the system, and so on. You can actually create models and statistics that explains certain types of differences between two phenomena. Coarse grain. Coarse grained. It's very coarse grain. Correct, correct, correct. But with you know, intelligence in general has it's not just about comparing two systems or many systems. It's about understanding what causes what. Yep. Yep. Right? So you want to understand why did ⁓ for example, the light go on. The light went on because somebody turned on the switch, right? If the switch was not turned on, the light wouldn't go out. Or why did, you know, the there's a lot of things that we can talk about. Why did the sun come up in the morning. Well, here's a more complex ⁓ concept. Maybe, maybe this is too early to discuss, but if you have molecular data from a healthy liver versus a diseased liver from let's say hundreds of people, those are two very complex systems. And I think that, well, you you tell me what statistics what statistics can do and what its limitations are in comparing those versus what machine learning can do. Well First of all, you can take a healthy liver or set of healthy livers and a set of diseased livers and you can say what are the differences between the two. That's what statistics can do. So you can compare the data from healthy livers compared to the data from diseased livers, and you can say what are the differences between the two. And if you captured the right data, let's say you captured some state of each cell in the healthy livers and some state of each cell in the diseased livers, and then you can ask the question. How many of those cells are similar? How many of them are different? Yeah. And what is the, you know, what is the ext the extent of the difference, the quantified extent of the difference. So that's what statistics can do. But what we would like AI to do, I wouldn't say machine learning. What we would like AI to do is to say, how did the cells that are different become different? Which means that it It should be able to sort of get to the root cause and say what changes led a cell to move from a healthy state to a diseased state in the in the diseased liver. And so that process is way more than just describing the similarities and differences between two livers. And it's it's getting into the I know, this is known as reasoning, right? So the causal reasoning is what you need to be able to do eventually. And I would say the causal reasoning is not a sort of problem yet, to be honest with you. It's not a complete solved problem. It's partially solved, but it's not completely solved. And that's what AI should be able to do. Machine learning, by the way, just to be super clear, since you brought up machine learning, machine learning is A an advanced form of statistics. It's just an advanced form of statistics. Actually, to be honest with you, let me say that the first department of machine learning that was created in the US, I think, is at Carnegie Mellon University. And they did not call it machine learning. They called it statistical machine learning because they wanted to emphasize the fact that machine learning is really very, very advanced statistics. Okay. ⁓ AI is, you know, it is it's gotten to the point where it's become non-deterministic. It's probabilistic. Okay, so probability theory and and all of the non-determinism and all of those kinds of things are not very well captured by statistics. Actually, there's lot of people who think even causality is not very well captured by statistics, right? So there's a lot of you know discussion in the philosophy of math and all that stuff. Let's not go there. But the point is AI should be able to understand. the root causes behind things. And that's what we want to try to do. Okay. So on the spectrum from statistics to machine learning to AI, I want to clarify this a little bit. I think ⁓ statistics is pretty well defined and most people have taken like courses in college and things like that. So machine learning is almost like a a human assisted sort of ⁓ statistics that adds some additional layers of understanding, ⁓ something like that. And then AI is like an intelligent system That looks for things even without you prompting it to do so. Well, okay. So let me let me step back. ⁓ so machine learning is the generalization of example-based learning that I was talking about earlier. Yep. Right? So when you give a lot of data, so statistics was was built at a time when the amount of data that was available was not a hum huge amount. I mean, it was population level, or maybe it was, you know. It was large, meaning thousands or tens of thousands or whatever, but it was not in the trillions and you know things like that, right? Right. So machine learning is a sort of an extension where you can now understand the patterns within the data. It's not just about understanding the description and the differences between two data sets, but you can start understanding the patterns in the data. So what do I mean by patterns? It's basically whether Whether there is a a car in a picture, that's a pattern. Right? I mean the car may be in many different configurations. You may be looking at the car from the front, you can be looking at it from the side, you can be looking at it while it's sort of in progression with something blocking the view, but it's all it's a car. Car is a car. So a car is a pattern of data. And that pattern is recurring in many pictures. So when it when you give it a thousand or million pictures with cars in front of different configurations, a machine learning model understand that understands that pattern of a car and says, ⁓ yeah, this is what I'm looking for. And it kind of abstracts from that pattern. So perfect. The jump from statistics to machine learning happens when you realize that statistics is about describing systems. Machine learning is about learning patterns. within systems. Got it. So that's that's the jump. And then you from machine learning, if you want to go to AI sort of in the broadest conception, it's truly about understanding the dynamic behavior of a system. So in other words, if you show a let's say ⁓ an o it knows where the car can go, for example. It knows that a car is not going to go through a building. It recognizes different objects. Correct. Correct. It will it it will actually know first of all, it'll know that a car is made of many things. It a car is made of, you know, a an engine and a tire and a steering wheel, blah, blah, blah, all of these things. What are the functions of all of those things? And if something breaks, what might be the impact of it? Will this car hit another car? So it's it's an entire dynamic understanding of the entire world around that car and inside that car. Amazing. That is the ultimate explanation of of what an AI system should be able to do. Okay. Great. Great. Okay. Yeah. Fabulous. What's the next topic? So I I wanna, you know, I I wanna also maybe get into the idea of what types of machine learning systems exist. Okay, so there's a lot of people who might be We're geeking out now. What? We're geeking out now. Types of machine learning systems, yes, geeky, I think I think we have to sh talk about this a little bit because Yeah. Because it's it's important to the fundamental point that we are may gonna make soon, which is about biological systems and biological analysis, which I think is where we should go. Yeah. We should talk in this episode we should talk about how to apply statistics and machine learning and AI to biological problems because that's the health Application area that we are interested in, right? We want to prevent that fatty liver. We don't want to just tell someone you have a fatty liver. Exactly right. Exactly right. What led to the fatty liver and what can you do to intervene and what can you change about it and all of those things. That's where we want to go, right? So here is a basic, basic view of machine learning that I would like to just put as a as a as a fundamental concept. Okay. There's I'm way I'm showing here three ways of Teaching or training a machine. One is called supervised, which is the example based learning that I was talking about earlier. The other is unsupervised, which is that you don't have to give any examples and labels. Like for example, let's go to cats and dogs. So in supervised, you're gonna tell the computer, here's a million pictures of cats and here's a million pictures of dogs, and you label each ⁓ each picture with whether it's a you tell it basically what's what, right? Correct. That's supervised. In the unsupervised, you would give it the two million pictures, but you wouldn't tell it which one's a cat or a dog? You would not. And it would have to figure out that there are two types of animals only and which one's what? That is correct. That's lot harder problem. It's a lot harder. Exactly right. There's no labels. Actually, I would claim that large language models are a sort of a subclass of somewhere between supervised and unsupervised because it's called Self-supervised, meaning that you don't have to label every word with anything. You just have to say this word follows this word. So you say A, B, C, D, E, F, you know, all the words that are in the training data set. So it's self-supervised in the sense that it just knows that this word is most likely to follow this other word in the full corpus. So that's it's self-supervised. So it's in the in the middle between those two things. So I'm just kind of laying out the full framework here. Yep. And and then finally. There's this other type of system called reinforcement learning, which is that you don't tell a system a priori the label of the system, of the the label of the objects in the in the ⁓ items in the data you're giving it. But instead you give it a post hoc reward if it gave you the right answer. And if it did not give you a right answer, you give it a punishment. You give it a l you take something away from it. So you you tell you build a system to get the maximum points. Every time it gives you the right answer, you give it two points. And every time it gives you the wrong answer, you take away a point or something like that. So that's called reinforcement through trial and error and reward and all that. There's a lot of people who argue that humans learn a lot of what we know when we are little kids, when we're babies and toddlers and so forth, through reinforcement learning. So you try something If it works great, you get feedback from the world. Maybe your parents say something, maybe your friend says something, maybe you get hurt. You touch a touch a hot stove, you get hurt, and that's reinforcement. It says, ⁓ this is not good. I'll never do that again. Right. So that's That makes sense. So that's the type of learning that happens. It's not by pre-labeling things and saying, This is hot, don't touch it. It's by learning through experience what works, what doesn't work, and all. So that's a different type of machine learning that also happens. So these are not like pure Different things. There's a lot of systems that kind of combine all of these things. Like I told you that self-supervised is between supervised and unsupervised. There's also systems that use supervised or unsupervised with certain degrees of reinforcement as it goes along. So there's all kinds of hybrid systems that you can build, but it's important to know that ⁓ these types of machine learning exist. Yeah, go ahead. Correct. Potential, the highest potential, but it also requires the most compute time, the most exploration, and the most situations and data. Well, ⁓ in in modern large language models, there was a time when they were scaling it and it was generating random text, right? Based on prediction, like like all predictive models, but they employed something called human in the loop reinforcement learning. Or reinforcement with human feedback. Okay. So R L H F. So they literally employed thousands and thousands of people in different parts of the world where there's a lot of labor available. And these people were providing feedback, human feedback, saying, hey, this is going well. You know, if if somebody asks a question about, you know, describe this house, and if the machine came back with a good enough description of that house, the human would say, that's a good description, except that you should not have said this, you should say this, right? That's human feedback. Okay. So that reinforcement happened in the modern generation of of ⁓ large language models that happened until, let's say, you know, I don't know, twenty twenty three, twenty twenty four or whatever. And and there was there's a lot of this reinforcement learning that's built into ⁓ large language models as well. Okay. And now modern Modern frontier models don't need humans anymore to check their work? Well, not for the underlying model. The foundation model uses self supervision at this point in time. Okay. But there's a lot of layers that are built on top of it. Like for example, if you take a foundation, a generic foundation model and you want to apply it to medicine, you do want to get reinforcement learning through human feedback. Got it. Or whether or not it's doing the right kinds of things in particular situations. And there are lots of sam examples where that kind of RLHF is incorporated into domain specific models that are built on top of foundation models that was just using amazing. Self supervised learning. Love it. So love it. So supervised is an approach that you would take when you know the ground truth and you know what specific goal you want. like differentiating cats from dogs. But if you have complex data sets and you don't actually know what you're looking for necessarily, you're gonna do unsupervised and say, analyze these data and tell me what you see. Yep. And it'll come back with things that you couldn't have thought of or hypothesized. Exactly. And then you can give feedback. You know, and and and nowadays you can have a an a supervised system that sits in one place and then an unsupervised system that learn something from unstructured, unlabeled data, and then maybe the supervised system is giving feedback to the unsupervised system. So there's all kinds of combinations of things that people have come up with that makes these things empirically better and better and better, meaning to experience. And ⁓ just a quick question here. Would you say that most of this research is done by these giant companies, three or four leading companies, or is this a type of research accessible to like graduate students at universities? Well, interestingly Until about 2020, most of the advanced AI research was happening in academia. Right. Like all these guys that I was talking about, you know, Jeff Hinton, Jan Lacoon, Joshua, these are all academics. They're all professors. So what ended up happening at around the middle of the 2010s is that there were so many companies interested in these problems that they basically hired All the professors and all of the graduate students and postdocs and everyone from academia gave them a lot of money and and everybody basically moved to industry. So what is left in academia now is the more fundamental mathematical kinds of research about understanding, explaining and you know, creating theoretels and so forth. But the practical improvements and everything are all happening now in the in the frontier labs. That's what's going on. Awesome. So do you want to switch to biology and you want to talk about how all of this stuff applies in biology? Of course I do. Let's let's let's do that. I think we've talked about some of the basics for for long enough now, right? This was great, yeah. Yeah. So so one of the you know, one of the fundamental kinds of ⁓ studies that we do in biology, right? It's it's important for us to sort of l set the set the baseline here. The studies that we do in biology usually fall in in some sort of categories. So I I wanna I wanna just put up a s put up a slide here. So you you know in in in ⁓ biology you're trying to answer the question what causes what? Especially when you're doing a an interventional study. Yeah. Right? You're saying does my drug improve this disease? Or when you want to change a system. When you want to intentionally change a system. When you want to intentionally change the same exactly. And that's called interventional studies down here that I'm showing, right? You wanna you wanna see that. But you can also do the same kind of thing about whether X causes Y, even in observational systems, by looking at exposures. Like this is classic epidemiology, right? You can look at two populations, and let's say you have ⁓ in one group, you have very few smokers, or you have no smokers, and you in the other group you have a lot of smokers. But you didn't ask them to smoke. You just know that there are smokers. Yep. Then you ask the question, does being exposed to smoking cause people to have certain kinds of diseases or not? Yep. And that's like an exposure analysis. And you can look at the odds of you know what the outcomes could be. And that could be another way of looking at X cause causes Y, right? And you can do that at a cross sectional level, which is like you take a snapshot. of one population and st snapshot on us of another population and you compare those two snapshots, or you can do longitudinal, you know, what kind of incidence is happening over time as you have more and more people. You you observe the same human population over time and see what kind of diseases they accrue and what kind of nutritional and lifestyle and and pharmaceutical practices they they implement. Exactly. The Framingham study is a really great example of a of a longitudinal study. I think they did that for what, twenty years or something? They looked at more, but yeah. Maybe even more than that. I think they had study one, study two, study three, and so on. So they I think now it's probably like forty, fifty years since they've been doing that kind of work where they followed people over time to see what kinds of exposures early in life led to what kinds of diseases later in life or what kind of you know mortality happened or morbidity happened and so on and so forth, right? That kind of stuff. All I mean, these are all different types of studies to s to ⁓ to answer the same question does X cause Y? And the gold standard for that, I mean, everybody knows this, but worth mentioning, is the randomized clinical trial, where when you have two groups, one group is considered a control and the other group is considered an interventional group. And you ideally you want to give a placebo to the control group, and you want to give the actual intervention, whatever that is, whether it's a food or a drug or anything else, to the interventional group. And you want control everything else between the two groups so that you don't have confounders that may be driving people to do one thing over another. And the best way to control confounders, if you cannot a priori match age and sex and everything else, but the the number of confounders is just a huge number. But in order to control the huge number of confounders, the best way we found is randomization. Randomization means you don't, you know. look at a person and a priori decide because they're so and so they should belong to the control group or the intervention group, but you randomly assign every participant into either the control or the intervention group. And the reason you do that is because you are trying to control for the confounders. So you're expecting that all of these variables that are confounders are more or less equally distributed in the control group and the intervention group because you're randomizing it. So yeah. So ⁓ is the placebo controlled randomized control trial process. ⁓ Guru, I think we should sp ex ⁓ explain a little bit about confounders because I'm I'm not sure that ⁓ you know most people understand that. Give us an example of a confounding variable in ⁓ either an observational or ⁓ well in an observational study, for example. Yeah. So ⁓ let's say that you're looking for people who ⁓ who ⁓ took a certain drug and lost some weight. versus people who took the drug or who took a placebo and whether or not they lost the weight, right? So you're looking for does the drug cause the weight loss? So let's say that this drug for some reason is very dependent on whether somebody has a particular f biological feature. Like for example, let's say the drug Only works for people who happen to be male but not female. So sex happens to be a confounder here because you you did not take that into account in the design of the study where you're checking for whether the drug caused ⁓ somebody to lose weight. You need to account for the fact that sex is an important factor. You cannot, if you discount it, it may confound your analysis because you may conclude that it did or did not. cause weight loss without accounting for a factor that was in in fact impacting both the ⁓ response to the drug and the weight loss. Yep. I'll I'll give another example, recent one from our collaborators with Australia and and New Zealand. We we ⁓ we studied a population of women with fibromyalgia and healthy control women. And ⁓ in that study we found that the women with fibromyalgia had more inflammation. However, when we look at their BMI We found out that women with fibromyalgia have a significantly higher BMI. And BMI is known to cause inflammation. And so now you have you have these inflammatory biomarkers that are strongly associated with fibromyalgia, but the actual confounder there is obesity. And obesity is causing it. It's not fibromyalgia. So the way to get rid of that confounder would be to then sub-select a group of healthy women that match the fibromi the the women with fibromyalgia for BMI. And once they're matched, now you can ask, okay, are those inflammatory biomarkers still higher in fibromyalgia or not? Exactly. Exactly right. That's a good that's a that's a good way to ⁓ to address the confounders. There's many other ways of addressing confounders, but that's a great way to address address confounders. You can in fact even take into account confounders later on after this kind of analysis has been done, there's models with which you can Account for confounders in the analysis phase, even though you did not collect data in the right way. But there's these are all good things to to understand as a study ⁓ analyst or a study design, right? Yep. So so that's one important point. The other important point I wanted to show was ⁓ the the whole idea of ⁓ let me let me just bring up the correct slide here. ⁓ where is this? The whole idea of ⁓ machine learning applied to biology. So Momo, if you're if you're ready, I can I can I can talk about that. I'm ready. Excited. Okay. So typically what happens is you ⁓ are asking a biological question. Let's say you're asking a question of ⁓ what are the ⁓ what are the biomarkers of a disease? Let's say you know you're you're looking for a cancer and you have data from somebody's ⁓ ⁓ gene expression. Let's say, I mean, this is you I'm picking an example from from biology. Very relevant. Very relevant. So you have the gene expression data from the the blood of humans and you're asking the question what are the biomarkers for a disease X? Let's say you're looking for a cancer. Yep. So what you would do is collect data from people who are known to be healthy people who are known to have the cancer. And then you would have to train a model to classify when a new sample is given, whether that new sample belongs to the healthy set or the cancer set, the disease category. So that classification is a very classic biological application of machine learning. So it's you know it's just it's just giving a binary classification of every new sample that comes in, tell me whether it's in the disease cohort or in the healthy cohort. The way you do that is you first collect data for the two populations and you train a model to do that classification. And there are many methods of doing it. We don't have time to talk about it in this one episode, but assume that you can build a ⁓ linear regression or a logistic regression or a random forest, some kind of a model that is able to capture the characteristics of the healthy and the cancer. biomarkers from the gene expression of the two groups and create a classification model that given a new sample says yes or no. Yep. That is a very classic problem. And that's how in fact di ⁓ diagnostic models are created. And the way that usually works is you create I'm gonna sh I'm gonna have a slide here. Let's have a visual for this. What's that? Let's have a visual, please. Yeah, let's have visual for this. So, what you end up doing is you collect the data, you build a model, and then you predict what that model says about a new sample. And if there is an error in the prediction, you adjust the model and then you retest the model on new data. So you keep doing this over and over and over until the Error becomes very small, meaning a below a certain threshold. So when the error becomes small enough, then you say, Yes, the model is doing well in the classification task. So in the beginning, you know, you can start the model that's just a coin toss. It can say, Hey, you know, this can be either this or that and just picks a you know, picks one or the other. And then you say, nope, you adjust it and then you go back. So this is the, you know, a certain kind of reinforcement. That's happening in here. Yeah. But so clinical research, we don't often have luxury to do this multiple times. We well, what what happens in the training in the training phase, Momo, is that if you collect, let's say, five hundred samples from a ⁓ from healthy people and five hundred samples from people with the disease. Yep. You use the thousand samples. to go through this process over and over one sample at a time. Okay, that makes sense. Yeah. And that's, you know, typically what you end up doing is you pick like sixty percent of the sample set. So in this case you have a thousand people total. So you would take, let's say, six hundred of those thousand people and you would put that aside as the training group and you run through this process over and over six hundred times. Actually you'd you know you have many different parameters that you can tweak and all of that stuff. And once the error is good enough, you validate it with the 20% that you've also set aside just for validation of the model. Yep. And then once you're happy with the validation, you keep another 20% independent of any of the model development process. So we call that final testing of 20%. We call that independent validation. And independent validation is super important because you don't want any samples in the independent validation set that was used during the training or the validation, the internal cross-validation part of it. So the testing is super, super important. And in biology we call it independent validation. In the classic machine learning literature, it's called testing, but in biology we call it independent validation because it's independent of the entire training process. It's never seen during the training process. So this is very important. Now unfortunately what happens is a second, Guru. I wanna I wanna I wanna qu sort of cover this with my ⁓ with my own ⁓ explanation because I mean this is one of the most important ⁓ slides in today's podcast. Okay. I mean this is this is this is literally how the world of diagnostics works. Like this is not some theoretical example. Like we we this is what we do at VIOM and this is what other companies are doing. So I just wanna sort of summarize here the blue parts Would represent the samples that are all used for training. And you can use them as many times as you want to. You can play with them as many times as you want to. You know what the ground truth is, meaning you know which sample comes from a healthy person or not. And so you can divide them into subsets that help you develop a model, and subsets like the light blue that help you test the model. And then you can tweak it, and you can do as much of that play playing around as you want. At some point in time, You have to decide and say, okay, I've done I've done playing. I've developed a model that is going to go to the market. Let's now test it how well it works. And then you're going to apply to the orange part, which is completely unseen new data, so that any biases that you generated during the learning, the training, they're not applicable to this data set because it's a completely independent data set. And in a blind fashion, you're going to say, I am going to apply. ⁓ my model to these 20% of the samples, but I'm not going to know which one comes from a healthy or not. I'm just going to apply the model and I am going to predict is it a cancer or not? And then someone else is going to get those data from me and compare to the ground truth and say how often you were actually right and how often you were wrong, which generates true positives and false positives and then true negatives and false negatives, which directly determines the sensitivity and specificity of a diagnostic test. Well well explained, ⁓ Momo. And and and the reason we wanna do this testing stage is because we wanna know whether the model that we developed generalizes to a broad population. It should not be just limited to the population that we did the training for. It should be able to generalize to any population. That's the reason we do the And I mean about the the perfect example you gave was that in that training on X rays, the model picked up some text. And you you thought, wow, this thing is amazing. And then you gave it an independence set that it never seen before and it just crashed completely because those that same text from the same hospital wasn't there. And so what a great example that was. Yeah, yeah. So, you know, we did this, for example, in ⁓ in our saliva test as an as a as an example ⁓ for oral and throat cancer. We started off with approximately twelve hundred participants in the study. We did a training versus independent validations separation, 950 and approximately 250. And after going through this entire process, we ended up with the sensitivity and specificity that Momo was talking about, which is calculated from the true positives, true negatives, false positives, false negatives that you get in the independent validation set. Not during the training, but independent unseen data is how you end up getting the 90%, 95% in the diagnostic. Now, I do wanna say that there are some pitfalls here that I should highlight and say these are all practices that are not very well done in the community, in the academic community especially. There's lots of papers where I see these kinds of pitfalls happen over and over again. So let me just mention a few of the pitfalls. One pitfall is the guru. But yeah, but the context here is that on the surface, to a naive reader, it looks like, ⁓ this is machine learning. This was done right. They're using some words, right? So on the f on the surface, it looks legit. Correct. You know, it they may use the right terminology and they may say, you know, show some fancy graphs and all that stuff. But you need to dig deeper and see whether they did all the right things during the training and the validation of the of the model, just like in that in that X-ray example, right? They apparently did not do all the right things during the training and the the internal validation. And so it failed in the real world. So if you don't do the right stuff, you go into the wild and it just fails. So to avoid that, you should avoid a certain number of pitfalls. So let me just mention a few of that. One is this idea of overfitting. It's called overfitting, meaning that you build a model where the training accuracy goes up and it seems like it's so it's performing so well. But it actually happens to be great for the training set and it will never generalize to anything else. So you the held out accuracy, which is the independent validation, is like approximately 50%. This was more or less the example that we started the show with. So this is a very bad method. So you want to be able to do good independent validation before you get out to the real world. The other thing that people do, which I see over and over, is something called data leakage. Data leakage basically means they somehow, through, I would say, negligence, through you know, not being very careful, they let some data from the training set leak into the independent validation set. A typical way this happens is sometimes you know, peop you may have two replicates from the same patient. Again, one replicate is in the training set, the other replicate is in the validation set. Or you may have ⁓ two samples that were somehow used in both or a single sample that was somehow used in both the training process and the validation. It's amazing to me how many people do this, that they don't just take a data set, draw a line, and just put away the training set separately. They recreate the tr the independent validation set from the same spreadsheet or something, and they by mistake, they take a sample that was in the training set over into the independent validation set. And that creates all kinds of problems, like this example that we were using using before. When there's a leaky independent validation, you may get 92% something, let's say sensitivity, but an honest look might just give you 49% because the data leakage kind of artificially inflated your metrics. That's another very important one. And there's a lot of people who just don't even do independent validation. So the example we talked about was the chest x-ray one that's in the bottom right hand corner. That's the one that we started the ⁓ episode with. But there's many others. There was a a genomics trial where cancer patients got picked based on chemo with predictors that was built on falsified data, meaning it was just not accurate data. It was just not correct data. And so there were ten papers that were were retracted because of that. They were just not clear about doing the independent validation. There was a lot of imaging that was done during the COVID nineteen problem where there there was a ⁓ I think a meta analysis of sixty-two studies or sixty-two models, out of which none were good enough for clinical use. Actually, it was two thousand two hundred and twelve studies. And so there were zero of sixty two models that were fit enough because they did not follow all the best principles, and they f fell through one of those ⁓ those ⁓ pitfalls that ⁓ that we were talking about earlier on. There was another sepsis model that was built by Epic, and there were sepsis cases that were outside of a hospital. And it turns out that 0.83 was what was claimed, but the When they went out to the real world, the AUC was AUC means area under the curve for the ⁓ receiver operating characteristics, which is a a way of specifying and showing how the sensitivity and the specificity ⁓ interact with each other. You know, when your specificity is high, your sensitivity may be lower. When your specific when your sensitivity is high, your specificity may be lower. So how the two interact with each other is drawn in a curve. ⁓ many of you may have seen these ROC curves in in ⁓ diagrams. ⁓ and some they claimed zero point eight three, but when it went out in the real world, it was only point six three to point seven six. That's a pretty significant hit. And that was because they probably did not follow the best practices of doing the independent validation, not overfitting, not having data leakage and all of that stuff. So I created this checklist. Make sure that you put away your independent validation data set. Select the features and tune inside your cross-validation in a very, very careful way. Split by different patients, sites, time, and all of the different markers and all of that stuff in a more or less randomized way between your training and your validation set. And look for confounders. Just do the basic. validation on an independent external cohort, not just an internal cohort but an external cohort, ideally. And always look for a simple baseline, like a logistic regression. And so you can interpret the model. You can understand the math behind it and you can see why something is drifting or something is changing. So always start with something basic and then work your way up to more and more complex models. So you can compare it back and see whether something changed dramatically. And if it did, You want to probably stop and ask the question, why did it change dramatically? And you want to understand that it actually makes sense. And then always report confidence intervals. Don't just say what the ⁓ don't just say what the final number was. Tell you know, give the confidence intervals. These are all standard practices and statistics. And also talk about how you did the power analysis whenever possible. These are all ways of handling uncertainty. And this is what I do not see in many of the papers today moment. Yeah, I know. ⁓ I agree. This is amazing. ⁓ super educational. ⁓ can I bring up an example that I that I find often? Yep, please. ⁓ I think we should talk about the N and the D, and I think you should talk about that. And the the most common example that I find is that the D is high and the N is low, and that there is no independent validation, which means that the false discovery is essentially the most likely reason for the findings. Yep. ⁓ I think this started originally with GWAS studies, where people took whole genomes from fifty a hundred people and had access to hundreds of thousands or millions of features, and then they found all these candidate genes that are associated with certain phenotype, but then they were not reproduced at all. And so there are now multiple studies showing that these highly, highly published candidate genes that are associated with certain phenotype. They're basically completely random noise because the proper techniques were not followed. Exactly. So this, you know, biological data has this property that generally speaking, the number of features or the dimensionality of the data is usually a lot higher than the number of samples. So the typical example is, you know, when you take a human ⁓ DNA or RNA study, you might have like 20,000 genes and you may only have like, you know, a few hundred samples. So the features are the measurements. How many measurements are you how many things are you measuring? How many data points are you measuring per sample? Correct. Yeah. Yeah. Features of that Yeah, go ahead. Go ahead. Yeah. Before omics methods, people generally over the previous decades generated very few data points because they were limited by the technology. But now with genomics and and multi-omic methods, you're generating these vast data sets. And ⁓ if you don't have the right number of people in the study. Go ahead, Guru. No, no, I I I I agree with you. So you know, if you're just looking at, you know, somebody's height and weight and you know their ⁓ their waist size or you know, those kinds of things, you probably have a few dozens of of data points per participant. And you may end up having, you know, ⁓ let's say 50 data points and you have five hundred participants, and that's reasonable because that's that's a a tall ⁓ data set, right? But if you have twenty thousand genes per participant, you're now talking about a very wide data set. I mean think about the data set as being a a table where each gene is a column and each sample is a row. So you have a very wide data set, and that is the general nature of genomic or biological data sets these days. In my in biome, we study microbial RNA. That has thousands of taxa to millions of different functions, like KOs, you know, what kinds of you know, reads we have, what kinds of pathways we analyze, what kinds of gene functions do we see, right? So this type of data requires special handling. So there are many, many techniques. I can go through some of the techniques, but not sure that we have time for going through all those techniques and stuff. But the idea is if you don't Reduce the dimensionality of the data, meaning reduce the features to the point where it's likely to contain the features of interest, but it eliminates the features that are likely to contribute to mostly noise. Like, for example, very, very low prevalence features, extremely low prevalence features, right? That can create a lot of noise. So you want to understand what is the prevalence in the data set that makes sense for you to include. And that's a difficult decision in many cases. But if you don't do that, you end up doing false discovery. So I I have a slide here going back to the basics of ⁓ you know ⁓ statistics that if you're not careful, if you did, you know, you compare gene by gene, if you have twenty thousand genes and you compare gene by gene by gene, you you make twenty thousand tests, you can get a thousand false alarms. Because you're you know you're looking for five percent tolerance of Comparisons. And that's not going to be good enough because you're looking for very high levels of signal in the data. And you may be just capturing the noise. So in order to cap in order to address these issues, there's many different correction methods. There's some very strict methods. One of them is the Bonferroni method, where you just divide the alpha, which is the tolerance that you want to have. By the number of tests. So if you had a tolerance of 0.05, you divide it. So you get 2.5 times 10 to the minus six, which a super, super low tolerance, right? So you just sort of multiply the ⁓ the tolerance level by a very, very large, ⁓ very, very large ⁓ the most conservative method. ⁓ The most conservative method, correct? This is the most conservative method. But a lot of people do these other methods like the false discovery ⁓ Benjamin Hawkburg method. And ⁓ there's also the GWAS convention, which is roughly one million independent tests across the genome, right? And and it's it's a it's a very tiny number. So it has to pass this high bar in order to be considered to be a signal in the humongous amounts of noise that's in there, right? So these are all happening, but I I want to get back to a bigger point that that happens in ⁓ in biology these days, right? So ⁓ okay, while you're looking for that, just in quick summary, so obvious issues with machine learning studies are using very few subjects, so like tens of subjects, ⁓ but you're measuring thousands of features. That's one obvious thing. And the second obvious ⁓ fla flaw that you would identify is if someone just did cross-validation, meaning they use the same samples for both training the machines and And actually testing. And if you if you already told a machine what the ground truth is, because this is supervised learning, then of course it's gonna know that. So you're biasing your results. So those are the two I would say very common ⁓ I'd say the majority of AI, you know, machine learning appl ⁓ publications have that issue. Agreed, agreed. And you know, I wouldn't be surprised if somebody came back to me and said, I did a meta analysis and probably eighty plus percent of the papers have one of these flaws. Yep. I would not be surprised with that. I would not be surprised. So if you if you don't mind, I wanna sort of switch gears and go to the next go to the next level. Let's shift. Shifting up. I yes, I wanna I wanna just, you know, remember I talked about the history of ⁓ of ⁓ the recent ⁓ L L ⁓ technologies that's called Transformers, right? So I said in twenty twenty seven, I mean twenty ⁓ twenty seventeen. Yep. There's a paper ⁓ introducing this new transformer architecture that came from Google. I mentioned this earlier. That paper was called Attention Is All You Need. I mean, if you haven't read this paper, it's it's a sort of a highly mathematical paper, but this is kind of the classic that everybody refers to these days. It's called Attention is all you need. And that actually gives you a lot of the fundamentals of the latest generation of ⁓ of transformers, meaning large language model based ⁓ based AI. There was Another biological ⁓ model called Alpha Fold that was also generated by Google DeepMind. There was a first version, there was a second version, which won a protein challenge. You know, there was a a yearly protein, just like the ImageNet challenge that I mentioned earlier in the history of AI, there was a protein challenge where people had algorithms that was that were ⁓ predicting the 3D structure of a protein molecule given a sequence of amino acids. And they had competitions every year. Somebody would bring in a new algorithm and so on and so forth. But in 2020, Google DeepMind used a transformer-based technique to get in there and win that competition. By far, they actually overtook everybody else. And that was when it was Well understood in the field that, ⁓ you know what? AI is here. It's gonna help us make breakthroughs that we couldn't make otherwise. Now, you you know, we talked about 2022 GPT, Chat GPT came in. 2024, we had Alpha Fold III, the next generation, and the Nobel Prize in chemistry for protein AI was given to the team that developed Alpha Fold III and related technologies. There were, you know, there was ⁓ Demis Hasabis. John Jemper, both of them at ⁓ Google DeepMind, and then David Baker over at University of Washington, these three folks had developed these techniques to the point where protein folding prediction is now considered more or less a solved problem. So that is a huge ⁓ step forward in biology. So I want to just give a little bit of a sense of how that happened, ⁓ go one level deeper. So I to say something here, Guru. What's that? ⁓ I want to say something here. So ⁓ just ⁓ d on your previous slide, I want to make sure that people understand the impact of Alpha Fold on understanding human biology. We are currently generating data ⁓ for the protein sequence where the vast majority of the of those proteins we don't know the function of. And that's really where Alpha Fold is going to make a huge progress. in that now from the primary sequence of a co of the coding sequence of the protein, we can infer function computationally because experimentally to do that, it's called functional genomics. It's an extraordinarily complex and and and expensive task. And that's why it's not been done. And now computationally you can do that to a to a good degree. So it's going to unlock a lot of human biology, microbial f biology, anything. So that's that's a that's why it won a Nobel prize. It's amazing. Yeah. So moment are you gonna talk about Evo Two? ⁓ I I I mean if you have time, we could we could talk about it, but ⁓ you know I would I would actually want to have an entire episode on Evo two because let's move on. Let's move on. Look, I mean, you know, we we have so much material. This is like saying, you know, do all of biochemistry in one episode or something like that, right? So I mean I feel the same way about trying to do all of these things, connecting the dots and all this stuff. So we need to do many other episodes. Evo two would be a great topic. Even Alpha Four would be a great topic for a future episode. Yes, if we can get David Baker To come and on the show, that'd be fantastic. An amazing, amazing guest for our David, if you're listening to this, please ⁓ contact us. Please contact us. And Evo two people, if or Evo three, you know, the latest guys, if if you're interested, please contact us. We'd be very, very happy to have you here. So what I want to show was some fundamental things about how Alpha Fold works. Okay, which hopefully this will not take that much l longer. But go back to a transformer. When you give it a sequence of words, like the cat sat on the transformer is trying to predict the next word after having learned the sequence of words in a huge amount of corpus data, training data, coming from the whole internet. In the whole internet, you have 10 to the 24 bytes of data, which is a humongous amount of data. So that's all been used, and it can come back with three predictions. The transformer can say the cat sat on the mat with forty two percent confidence, floor with eighteen percent confidence, sofa with nine percent confidence. And you can decide or somebody can decide that I only want to pick the one with the highest confidence. And that becomes the next predicted word. That's how transformers work. It it picks the the highest prediction from transformer. Now remember it's self-supervised, meaning that nobody labeled all these things, it just learnt it on its own. But you can also set the transformer to not always pick the top highest percentage. You can say, you know, that's called a temperature setting. If you set the temperature to be very high, it randomizes even across these predictions more and more. If you set your temperature for Chat GPD or the underlying model to to very high, it might say the cat sat on the sofa as a just a random. ⁓ prediction just to you know keep things interesting so to speak, right? So this happens because the transformer has been trained on so much data that's called scaling. Okay? Now, the same exact idea applies to alpha fold. What you do is you give it a sequence of amino acids. These are all one-letter you know abbreviations for the amino acids. But once you give all of these ⁓ amino acids in a in a sequence The Alpha Four machinery will come back with a predicted three D structure colored with the confidence level. So the blue stuff is very high confidence that this is how it's been seen. By the way, Alpha 4 has been trained on all of the protein data that has been collected forever by the community. So that's how Alpha Four was able to actually reach this level of of prediction accuracy. I mean it it It started with all the data that was already there in the human protein database and then it it started generating ⁓ synthetic proteins and comparing it to e existing real ones and started kind of doing an adversarial kind of ⁓ learning process where until it starts started to get better and better and better and over time it got so good that most of this predicted shape of the protein is very high. Which is ninety plus percent. But you can also see that there's some light blue ones that are s less confident, the yellow ones that are low, and the red ones or the orange ones are super low. So there's a little bit here that's very, very low confidence. And this makes sense, Guru, because the alpha helices are very nicely structured, so it's very confident. The loop structures, the turquoise ones are loops, and they're more disordered. And that's not like this looks absolutely fantastic. The red part is at the very tail end and that tail can just flop around. So ⁓ this this basically looks phenomenal. And so yeah. That's what that's what Alpha 4 can do. You give it, you give it the amino acid sequence and it generates this with all the levels of ⁓ the confidence and all exactly like the LLM was doing for all the different possible confidence levels and so forth. So it's that's the parallel that I wanted to draw. Love this one. ⁓ so I think those were the main points I wanted to make on this ⁓ episode, ⁓ Mo. And I think, you know, if you want, we can end with where is all this going, kind of a thing. Yeah, let's let's see. What future do you predict in the next ten years? Yeah, the scary part of the spectrum and the the awesome part of the spectrum and somewhere moderate in between. Well, look, I think there are some areas which is gonna which are going to be extremely challenging like cybersecurity is a is an area where I think it's going to be extremely challenging. And in in one sense it's scary, not because it's a robot that can kill people, but it's these systems that are optimized to go do certain things that it can indiscriminately do things without attention to where it is and whether it's allowed to do it and whether it's a bank or a hospital or anything else. Right. So I Do you think that it's not A very good outcome if AI systems are not properly controlled. And that's called alignment. You need to align it to the goals and stay within the guardrails of what it an AI system was designed to do. So there is actually a lot of extremely positive, beneficial applications that can come out of. properly designed and properly controlled systems. Alpha Fold is a great example of that. And there's many other examples of AIs that can, you know, you probably use Claude, you probably use Chat GPT for various research and various ⁓ tasks like creating presentations and everything else. Those are all great beneficial functions. So we want to get the benefits, but we don't want to pay the cost. So you've got to do all the all the right things to do that. But then in the end, let's say we we, you know, over time we figure out How to balance the good and and and not pay the cost of the bad, ⁓ there's still going to be a a question of how much capability do these models have. So somebody ⁓ who originally came up with the idea of something called artificial general intelligence, which basically can do anything that a human being can can do. Right. Okay. That's called general intelligence. ⁓ came up with a ladder. Of how good AI systems are. So I just want to show the ladder as a as a way of evaluating how good these AI systems are. So let me let me just show you the ladder as I see it because I've annotated the ladder in a way that hopefully will make sense. So there's five levels of artificial general intelligence in my mind. There's level one is emerging, right? All right here. Level two is competent. Level three is expert. Level four is exceptional, level five is superhuman. And when we say competent, we mean it's better than 50% of skilled adults. And when we say expert, it's better than 90% of skilled adults. And if it's exceptional, it's better than 99%. And if it's superhuman, it beats every single human. So this is the performance scale that that has been defined in this progression. And there are two different types of AI. There's a narrow AI, which means it can do one task at a time, but it can do it extremely well. Alpha Fold, you see down here, Alpha Fold or Alpha Zero or other Alpha systems, they're all Alpha Go, for example. They're all narrow AIs and they can do really, really well. I mean, I I was in IBM and we did, you know, the Watson system, which was a Jeopardy system. There was a deep blue system before that. They're all really, they're all narrow systems. And they can do really well, and they have all these different levels, right? I mean, Siri and Alexa are, you know, competent, narrow systems at this point in time. They can become broader and broader over time and so forth. But general systems, meaning that they can do many tasks, Chat GPT, I would call it the original Chat GPT in 2023 with and Gemini in 2023, they were at level one back there. The recent GPT 6 Astra, which was released in 2026, that was just a couple of months ago. In my personal view, that has crossed the threshold of what I consider to be a competent AGI, meaning that it is now better than 50% of the skilled adults. That's my view. Okay, this is not something that everybody believes in, but that's my view of just looking at the capabilities of of ⁓ GPT6 Astro, and specifically there's a metric called ARCAGI, which is, you know, you go through ARCAGI 1, ARCAGI 2, RCAGI 3, and so forth. It was able to do ARCAGI 3 recently in a way ⁓ that was able to show the skill level in an unstructured, unknown environment. So I believe this is going to be ⁓ level two. Level three is currently being built. That's in progress. As of now, September or October first, twenty twenty six, level three is being built. Don't know how long it's gonna take, and probably level two systems are gonna help construct it. Don't know how long that's gonna take, but it's in progress. And after that we're gonna get to level four, which is definitely not there yet, and level five is not there yet. So that's kind of my view of where we are in the progression of these AGI systems, Momo. Amazing. Love it, love it, love it. Yeah, I think we should ⁓ discuss a lot more in future episodes, but now we're this is like an appetizer for machine learning and AI for our audience. So it's been now more than an hour, like an hour and a half. So let's go ahead and wrap this up. Should we move to the questions? Yes, let's do that. Let's do that. Okay, so you had one question from the audience? Yeah, I I selected one question from our audience. So ⁓ a viewer was interested to know what was the motivation for Viom to acquire Circulate. So we just announced this recently. Circulate is the company that owns the patents on the application of therapeutic plasma exchange to basically flush out any chemicals or any un un you know any harmful molecules in blood. And we had an episode on that with ⁓ Dr. Brad Younggren, just the last episode. So you should check that out. But basically the motivation is that VIOM is definitely interested in helping people age better and healthier, meaning improving increasing the health span. ⁓ and TPE is ⁓ a complementary technology to what we're doing because if someone gets loaded with ⁓ anti in in with pro-inflammatory cytokines or with heavy metals or with mycotoxins, there's there's there are very few things that you can do about it, and TPE would be one of these technologies that can do that. And then ⁓ the second part of the question was are we going to utilize the ⁓ the the rest of our violent technology to maybe identify people who are the the the best candidates for TPE? And that's definitely in the works. So we are planning some clinical studies where we're gonna profile people and then triage people based on vile tests ⁓ that are the most n that that need TPE the most. But you know, that's gonna take some w some time to do all that clinical research. Until then, ⁓ what we what circulate has shown already is that they can actually reduce the biological aging clocks by quite a bit, like two point seven years in in in about a year. ⁓ And then they've also shown that for people who have high amounts of microplastics in their bloodstream, they can actually remove that very effectively. So those are the two clinical endpoint or endpoints we have so far, but we'll work on more. So that's it. And then ⁓ in the news, Guru. Yeah, yeah, yeah. Let me just add one more thing to your ⁓ d question from the audience is that you know the connection between biome and the TPE approach is basically longevity and chronic disease management. So that's a good therapeutic approach for doing that. There's many other therapeutic approaches as well. I'm sure there's gonna be other things that we're gonna develop or we're gonna, you know, bring into the biome fold in the future. But this happened to be one of the, you know, one of the areas that we wanted to grow towards and and it's ⁓ longevity and chronic disease management. So with that, let me go to ⁓ s a super exciting thing that happened in the news ⁓ recently, which is ⁓ available as a press release. You might have seen it in the press already. We announced a new, I mean speaking of AI, right? We announced a new Molecular Foundation model initiative with a few parties, mainly with Bhutan, the country of Bhutan, and there's a couple of ⁓ organizations that have built data centers and provided energy for the data centers and allowed advanced Types of models and algorithms to run on these data centers or within these data centers, for example, Bitcoin mining and so forth, that they've already done in the past. They're willing to partner with VIOM to build these molecular foundation models. And molecular foundation models are just like text language foundation models, except they're built using molecular data, which is what we develop and generate in VIOM. And the idea would be to build healthcare applications from the molecular foundation model. So in other words, we can generate diagnostics for many diseases just by having small sets of samples that we can prompt this molecular foundation model with and generate a diagnostic that has high sensitivity and specificity. That's an example. We can do companion diagnostics, we can do ⁓ drug molecule, molecule ⁓ validation or drug ⁓ even development in different ways or ⁓ target validation so on and so forth. So all of those applications are going to be possible with molecular foundation models. Lot more to be said on this topic in the future, but I just want to say that this happened last week. And so we are super excited to work with Bhutan and to work with the other partners. Actually, if you're listening to this episode and you're interested in partnering with us, for example, if you're NVIDIA or if you are anthropic, Or anyone else who's interested in working with us on a molecular foundation model to build the next generation of healthcare biological systems that can improve health ⁓ across the global population, please get in touch with us. We'd love to partner with you on doing this kind of stuff. So, with that, I think we can wrap up the episode, Momo. Yep. So amazing. Thank you, Guru, so much. We'll continue more on AI in some future episodes. Thank you. And I'm Momo, a biochemist. And I'm Guru who was the AI scientist who discussed just the tip of the iceberg of AI. There's a lot more stuff to be discussed and we'll do that next time. We're two PhDs on a pod. Thank you, Guru. Bye. Two PhDs on a pod is a Viome Health Science Podcast production, written and hosted by Viome CTO Guru Banavar and Viome CSO Momo Vuy Sich. Produced and edited by Lisa Shomo. Podcast Art by Kevin Wu. If you enjoyed today's episode, like and subscribe for new episodes every other week. We'll see you next time.