Latest / In the Loop AI / OpenAI's Mind Blowing Discovery To Reverse Aging
Transcript
- Ejaaz: [0:03] Since the dawn of mankind humans have been chasing the elusive fountain of youth the elixir of life the ability to live centuries i'm talking about the twilight saga vampires reenacted in human life josh and this week we didn't get one not two not three not 10 percent closer to this reality, but 50% closer with the release of GPT-4B-micro, a model that is able to design Ejaaz: [0:35] proteins that extend your life. Josh, you're the expert in this. Tell me more about this. Josh: [0:42] Okay. Well, I'm going to pretend to be an expert for the next 20 minutes because I just, I've been learning about all of this kind of as we go. But I have some interesting information to share. This is a really fascinating topic. It actually came out, it was published a few weeks ago. This isn't super recent, but I think recently it's popped up over time because of just how interesting and novel the breakthrough was. So Ijaz, you mentioned GPT-4b, which implies the model is 4 billion parameters, which is by all terms microscopic. Josh: [1:11] I mean, most models now, the top end models are like trillions, Josh: [1:15] if not tens of trillions of parameters large. This is 4 billion. It's very small. So before we get into the actual breakthrough, I want to kind of give everyone a little biology lesson here this was the lesson that i just learned and i i asked chat gpt if it could generate me an image to one shot this this idea so then you kind of have a frame of reference as we walk through these these particles or these properties here so we start with the body the body has a small portion of it which is a cell inside of a cell a cell is made up of these things called proteins this is where these yamanaka factors are that can give you this elixir of life that can actually reverse aging and then making up those proteins is a series of amino acids. Now, I believe there's 20 different types of known amino acids. And the Yamanaka factors, which were discovered in 2012, exist of four of these proteins. So there's the O, S, K, and M protein. Each one of those protein is a series of 360 amino acids. And of those amino acids, there's 20 options. So when you do the math on this, the numbers are astronomical. This is Like 360 to like the thousand zeros, it's like more than the total known particles in the universe. It is huge numbers, which is why we've had a really difficult time understanding how these work. So before I actually describe what these do, I want to do a little test because I know you have a biology background. You went to school for this. Can you explain to us the Yamanaka factors, the proteins, how this works? Ejaaz: [2:38] Okay, so as you said, Josh, there are four of them. Kind of think of them as manufacturing machines. So one is a machine that creates bottle caps. Ejaaz: [2:50] Another is a machine that creates the actual glass bottle. And another one prints the label, right? A similar type of thing is happening with these four Yamanaka factors. They're not actual proteins that you kind of drink in your protein shake and help you build big muscles. they're kind of like the machines that build other kinds of proteins. And as you accurately said, Josh, they're made up of 10 to the 23 different combinations of amino acids, which are like the smaller particles that they're comprised of, which means that you can get not just one, not two, but millions and billions of different versions of these Yamanaka proteins, right? So no one really looks the same. There's so many infinite combinations. And if you look at this single sentence that I've highlighted from Brian Johnson, the leader of Don't Die, right? He says, astonishingly, these things are so inefficient. In fact, you get a 0.1% ability or chance to convert a cell back to its original young, unadulterated form, effectively reversing aging. So this kind of like possibility of reversing aging using these Yamanaka factors has been incredibly inefficient and elusive. So what has now been kind of worked on is, okay, well, we have these four proteins, but we never know which kind of combination is going to be successful. To add on to this, to give you a bit more context here, remember. Ejaaz: [4:16] These four proteins act differently in different cells that we have. So it's going to act differently on your. Ejaaz: [4:26] In your heart, on your liver, depending on different kinds of organs. So not only do you need to kind of like figure out what random billion combination you need to kind of form for these Yamanaka factors, you also need to apply the right combination for the right type of cell. So it's basically been near impossible to do this, Josh. Josh: [4:44] So to kind of recap what you said, we have this incredible technology called these Yamanaka factors. It was discovered in 2012. The problem is we don't actually know how to use them to apply them efficiently. I mean, like we just highlighted, they have a 0.1 conversion rate, meaning 99.9% of them, they just suck. They're not good. They don't work well. But we know that they will work in the right circumstances. We just haven't been able to figure this out. So the problem that we're trying to solve now, I guess, is how can we accelerate Josh: [5:12] the rate of progress where we can kind of understand and get that efficacy above 0.1%. And that's what OpenAI i did with this new results that they released is they basically trained a model now i kind of want to talk about the four billion four billion parameter base model because this one was super fascinating they took a base model that was trained on just the basic understanding of the human language it kind of a new text it knew probably some math but it wouldn't be able to solve very challenging physics problems and that's why it's so lightweight and what they did is they took this base model of human understanding and they trained it on a very specific data set Josh: [5:46] Around these proteins, around biology, around all of the things that you would need to know in order to generate these new protein structures and do a lot of the math around that to get that efficacy from 0.1 to maybe a little bit higher. And it turns out once they did this, they partnered up with a lab named Retro and they actually tested the results of the model. So the most fascinating thing to me is that, again, this is a text-based model. So they just fed it some text. It popped out some text and they gave it to a lab to actually test out and see what happened and this is where things got really interesting is because for the first time ever the model was actually able to generate net new protein structures that actually worked and actually increased the efficacy of this age reversing protein into a place that is is much better right you just do you have the results here do you want to share them Ejaaz: [6:34] Yeah, yeah. So I'm highlighting the key statistic that's getting spread across headlines here, which is they delivered a 30 to 50% functional hit rate. Ejaaz: [6:44] Now, remember, originally it was 0.1% success rate. I don't know, again, what the multiplying factor of that is, but that is absolutely insane. That's like a 300 to 500x better turnout rate. And they managed to do this by creating AI model, like you said, trained on all of these different amino acid sequences, and they just asked it to run through as many sequences as they can and apply the knowledge that it has to all the different types of human cells and livers and organs and come up with the best potential, I don't know, top 10 combinations. And they took that top 10, gave it to the labs, and they ran it through and ended up having such a high functional hit rate. And for those of you who are still trying to grasp or understand what this actually means, I'm going to read you through what this results in and why it's so important for the rest of humanity, which is if you nail the combination of these four different proteins, you could result in either full reprogramming of cells, which means that it completely erases the cell memory and it creates a, think of like a natural born human or sorry, a newly born human or baby that has all the memory and knowledge and capabilities that you have right now sitting here. So what if I told you, Josh, that you could reverse your age by 25 years and be, I don't know, in your athletic prime. Maybe I'm aging you. Ejaaz: [8:08] I think you might be five years old after that. But let's go back 15 years, right? And you're in your prime athletic squad at high school, but you know everything that you do right now, right? And you're able to kind of like jump higher, sprint faster. And that's full reprogramming. And then partial reprogramming, which is kind of like keep your age, your looks, but you're metabolically healthier. You might have the functioning liver of a 10-year-old, but still look like you are a 30-year-old, right? Ejaaz: [8:37] So the effects of this should not be under-exaggerated, basically. This means you can effectively reverse aging. You can have the metabolism of a young teenager. You can burn calories well into your late 50s or 60s. And as Brian is highlighting here, we might be the first generation who won't ever die, which is just an insane thing to contemplate, given that literally a decade ago, there was like, you know, cancer could never be cured. We were struggling with a lot of medical stuff. We were actually increasing at a rate that was menial compared to this. Just insane. Josh: [9:14] Yeah, a lot of it stems from these things called stem cells, which I was also learning about. They're kind of, you can think about them like shapeshifters, and they can become many different cell types. And what they do is they'll actually regenerate a lots of parts of the cell. And like you said, you reverse aging because aging in reality, it's a disease. And you could actually just turn the dial backwards. And as we improve on these models, we'll be able to literally apply these and turn the dial backwards. A cool analogy that I was thinking of when I was going through this is, it's kind of like if you're playing a video game and like an RPG and you have a skill tree where you get a certain amount of points and you could kind of allocate points to specific subsets within your character. And you can, you try to optimize for the best builds based on your play style, but you never really know because there's so many options. It's impossible to test all of them. What this does is it actually tests all of them and it can try hundreds of these changes at once. Instead of just one at a time and what that results in is these significantly improved i mean in my case the video game example a significantly improved character that's genuinely optimized because it's gone through so much trial and error versus my like dumb human version and that's kind of what this is for these proteins it's just this highly iterative version on research that would previously have taken decades that we're now able to do in weeks of time to test i Ejaaz: [10:32] I think we should actually emphasize that point, Josh, which is the timeline of capability that we had here, right? So let's start from like a level one, which is humans, scientists discover that we're made up of cells and they find out all our genes and that we have a lifespan, right? They understand that how genes work and how they degrade. Step number two is they realize these tiny little things called amino acids actually dictate how long we're going to live. So if we can kind of create brand new versions of these amino acids and proteins, we could live longer. Step three, oh my god, there is a gabillion different versions of these proteins. Step four, let's manually sort through these protein combinations ourselves as humans using our measly little brains and come up with the successful ones. Obviously, that did not work. Step five, wow, these computer things are actually pretty cool. Maybe we could run these different combinations in a computer and maybe then we'll have a higher success rate. And that's what got us to 0.1% success rate. And finally, step six is this genius supercomputer called AI and these AI models that can not only sort and pass through all these different combinations, but are able to astutely figure out which ones are most likely to be successful on a better magnitude and order than our brains can. Josh: [11:52] So, Ejaz, we've kind of highlighted that aging is a disease that can be reversed but if it's a disease then how does that work why do we even age in the first place Ejaaz: [12:05] It's a good question. And actually, when I speak to a lot of people about aging specifically, and I did this during my university degree, where I actually studied senescence, which is the act of aging, most people just think it's like wrinkles, you know, they just assume like, yeah, I'm going to die when I'm like, I don't know, 80 to 100 years old. And that's just, it is what it is. But very few people actually understand how it actually happens. So to give a kind of loose description, your body is composed of many different parts, right? Organs, and each organ is made of various different types of cells. And within these cells, the core component of a cell is something called its nucleus. It houses all the DNA, the genetic material. Now, what fewer people probably realize is these cells, they die and they kind of regenerate. They reproduce, they create offspring, similar to like us creating kids as humans, right? But when they create these kid cells and, you know, all these kid cells then grow up and they create cells of their own, the genetic material, the DNA, gets a little older. Ejaaz: [13:12] Specifically, there's a part of the DNA, so if you imagine a little double helix DNA piece, there's a piece of it right at the end called the telomere. And the telomere dictates how long the DNA and the cells actually live before they, like, die out completely and their lineage basically doesn't live on. And with every successful cellular regeneration, the telomere gets slightly shorter, Josh, right up until you get to the end, and then the cell dies. That represents basically your heart getting older, your skin getting older, so you start to see more wrinkles, your hair falling out, so your cells are kind of dying. and the question has always been, can I reverse this? And that's when stem cells kind of became the thing. They were like, wait, stem cells are the original versions of all these different types of cells. If I can turn my old aging, dying cells back into its original form, then I've effectively reversed aging, right? And then you're probably thinking, well, what about cancer? What about all these other different types of diseases? The main reason why you're susceptible to all these different kinds of diseases is because it's a by-product of your cells getting old, your immune system failing. Think about it, right? So you become more prone to, you know, or susceptible to some of these different types of diseases. So why this is such a big deal is. Ejaaz: [14:30] If you can apply these Yamanaka proteins, these four different proteins in an astute personal way to you, Josh, for your specific organs, for your particular case, and then I get a different type of combination injected into me for my particular case, for my ethnicity, my background, my genealogy, it means that now you can get a personalized solution or medical treatment that could extend your life 10, 20, maybe even 100 years. And what that has as an impact to society, to the economy, to the workforce, to great-great-great-great-great-grandfathers meeting their great-great-great-great-great-grandkids is just a crazy concept to think about and one that I think is being very understated with this. Josh: [15:13] Yeah, to me, the exciting part is it now feels like we have a solution where we previously did not. So we were just like, we kind of had an understanding that aging was a disease, but we weren't quite sure what to do about it or how to solve it. It would appear as if now we we have discovered a solution being these yamanaka factors at least to some extent and the problem is making them not bad so we kind of we know what it works we just don't know how to make it work well and currently it works really poorly but with the advent of llms and then the integration of these llms into forming a model that can solve these we've made a ton of progress in going from 0.1% to significantly higher. I mean, 50 times multiple in improvement in some of these instances, which is huge. And I have to wonder what the natural acceleration or what the natural curve of this looks like, because I mean, this is the first try. This was with a 4 billion parameter model. And I assume part of the reason why it's so small is because we probably don't have a ton of data on biology of humans relative to the general data of text on the internet of just english but i assume that's something that will change and one thing we've been seeing a lot with humanoid robots and robotics in general is training them on synthetic data and artificial data Josh: [16:29] And if these AI models are able to start to understand the foundations of like the genetic makeup of humans and are able to start generating more of that data themselves, then you can very quickly see a world in which this progresses from a 4 billion parameter model to 40 billion to 400 billion. And it can really start doing some serious damage because the numbers are huge, but the numbers are also going up very quickly in terms of efficacy and how well this is working to reverse aging. So this seems to me like this super exciting really optimistic outlook from open ai and also aligns with what we hear a lot from sam altman i think sam frequently when he speaks about the future of ai he always references sciences and biology as being the place that he's most excited about and as a company this is really open ai's first time that i've seen of them pushing some material noteworthy research in this category that actually makes a big difference and they tested it with retro biosciences and it works and it works really well so i'm just it left me really excited for the the future of this subcategory within ai being sciences and biology because this is really i mean this is the first big breakthrough we've seen from a company like this and i imagine now that the floodgates are open it's just going to continue to to improve from here yeah Ejaaz: [17:47] I mean i mean i remember when sam originally announced gpt5 on their live stream He spent a good 20 minutes pushing the kind of health theme. And I guess we're seeing the fruits of that labor kind of like in reality right now, right? With this new kind of launch. And it also got me thinking, Josh, that whilst this is so impressive, it is yet another marker of a growing trend, right? Because this is an isolated event. We have spoken about Google before this many times on our show, which have released a series of scientific AI products, right? Applying AI to science to kind of like further it, create new cures and stuff. One of the main ones being this product called AlphaFold, kind of does a similar thing to what GPT-4B-Micro has done, but with a specific focus on curing certain diseases and creating kind of like solutions for, you know, specific diseases versus kind of reversing aging. So we've got kind of like the preventative method happening with GPT-4B-Micro, which is like, let's cure aging and therefore we'll save everyone from any kind of disease. And then you've got AlphaFold over here from Google, which is taking the approach Ejaaz: [18:59] of let's be more proactive and cure people from the disease that they're experiencing right now. So we're seeing this kind of convergence from these two major AI model producers in applying AI to sciences. And I think we're going to see this as a growing trend for a lot of other companies going forwards. Josh: [19:16] This is cool. I think as we talk a lot about AI, we get caught up in the benchmarks and the very surface level applications like, oh, can it solve this PhD problem that I'll never actually need to know? And can it write this amazing code but the reality is is like when we think about ai it really is all-encompassing it's not just about like writing code and creating better software it's it's about really unlocking a lot of the the problems that we have as a species and biology is a huge subcategory of that that's generally been so complex because the numbers get so large and we talked about a number greater than the number of atoms in the universe it's just astronomical it's it's so difficult to understand but ai is is able to help decipher that and make it easier to understand so it's cool to see google approaching the preventative side then we have open ai on the reversal side and i mean the labs are competing but they're also competing towards something really exciting that benefits all of us so i think at the end of the day this this has me super optimistic because there are really smart people working on really cutting-edge tech that are applying it to things that will actually improve the lives of everyone around us and that to me is like that's that's pretty freaking awesome yep Ejaaz: [20:24] I am excited to get access to the fountain of youth the elixir of life for the cheap price of a thousand dollars and 99 cents a month i hope yeah i'm looking forward to the uh little. Josh: [20:39] The little vial that shows up in my mailbox that I can just like drop some blood into, send it to OpenAI, and it'll deliver me an injection that reverses my age by 10 years. That's going to be pretty cool. I'm like very excited for that. Josh: [20:51] So I will keep my subscription valid in order to hopefully take advantage of that one day. So how long does it take to get to this world that we're describing? I don't know. I mean, these trials generally take a long time. The assumption is probably somewhere around 7 to 12 years, so say a decade. This is in the United States. Of course, this technology always happens faster in other countries. So I'm sure we'll see the results of this happening soon, DM, unknown. But I mean, directionally, we're going in the right direction. Things are only going to be moving quicker. It seems as if we do actually have a solution. So you just maybe in seven to 10 years, this reality does actually become a reality. And we are actually able to. Listen, I'm not getting any younger. Ejaaz: [21:33] I'm not getting any younger. I'm trying my best. Dude, I saw. Josh: [21:36] A few white Ejaaz: [21:37] Hairs this morning. I plucked him out. I'm in denial. Please, Sam. Please save me. Josh: [21:43] We are begging you. But yeah, that's basically what happened here. OpenAI is in the news for science and biology, which is just, Josh: [21:50] it's a really cool thing. It's really cool research. It's really exciting that they're working on things like this. It's exciting to see Google working on things like this. We have some exciting content in the pipeline for future biology and sciences. A few weeks ago, or even last week, we had on Logan Kibach from Google. He said he would intro us to some interesting people in the science division at Google, the head of the science division even. So there's going to be someone much smarter than us about biology, talking about biology and really getting into the weeds from the perspective of someone who is literally building it. So that's going to be a really exciting episode, but that's all we have for you today on our little biology science corner of the AI roll-up. So thank you for watching. I hope you enjoyed. As always, if you did, please share with your friends. Don't forget to like, subscribe all the good things thank you for for watching and we'll see you Music: [22:36] Music